Selecting text in a readonly code block (comment/issue markdown) lost the
selection within seconds, making copy impossible, whenever the surrounding
view re-rendered — most reliably while a sibling agent task streamed over
WebSocket (a re-render roughly every ~100ms).
Root cause: the `code` renderer emits highlighted HTML via
`dangerouslySetInnerHTML={{ __html }}`, a fresh prop object every render. Each
unrelated parent re-render re-ran react-markdown, and React rewrote the
`<code>` innerHTML even though the HTML string was byte-identical, tearing down
and rebuilding all 161 hljs `<span>` nodes. The native selection is anchored to
those nodes, so it collapsed.
Fix: memoize the entire `<ReactMarkdown>` subtree on its only real inputs
(`processed` + `components`). A stable element reference lets React bail out of
the subtree on unrelated re-renders, so the code-block DOM is never rebuilt
while content is unchanged. Confirmed via an instrumentation probe: zero
`<code>` DOM mutations during streaming after the fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The run-confirm interception box gated its handoff-note field on the
preview round-trip's `handoff_supported`, so every open showed a
"checking…" wait before the note box could even be used — to learn
something the client already holds. For a concrete agent assignee the
target runtime is exactly that agent's, and its CLI version is already
warm in the prefetched agent + runtime caches, so the box can settle
synchronously, the same way the quick-create version gate does.
Add a frontend `handoffSupported` mirror of the server's
MinHandoffCLIVersion gate, resolve the agent → runtime → cli_version
locally, and drive the note box from that verdict without waiting on
loading. Squad / batch-status / unresolved-agent paths — whose resolved
trigger set is only known server-side — keep falling back to the
preview's `handoff_supported`, unchanged.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(editor): accept highlighted composer suggestion on Tab
Plain Tab now accepts the highlighted mention / slash-command suggestion,
matching Enter, across every composer built on the shared TipTap editor
(chat, issue description, comment/reply). A single shared isPickerAcceptKey
predicate centralizes the accept-key policy so the two picker lists stay in
sync instead of each re-deciding what counts as accept.
Shift+Tab and Ctrl/Cmd/Alt+Tab are intentionally NOT accept keys, so reverse
focus navigation and OS window switching are preserved. When no picker is
open, Tab keeps its existing behavior (list indent / focus traversal).
Adds unit coverage for both picker lists plus a plugin-order guard that fires
Tab through real ProseMirror dispatch with the caret inside a list item,
proving the suggestion layer outranks PatchedListItem's Tab -> sinkListItem.
Scope: web/desktop shared composer only; mobile and the generic combobox are
untouched.
Refs: MUL-3685
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(editor): make Tab/list-item priority guard actually exercise sinkListItem
The guard placed the caret in the first (only) bullet item, where
sinkListItem is a no-op (no preceding sibling to nest under), so the test
passed regardless of whether the suggestion layer intercepted Tab. Rebuild
the fixture as a two-item list with the caret in the SECOND item, where
Tab -> sinkListItem can fire, and add a sanity control proving a bare Tab
(no picker) does sink that item — so the accept-wins assertion is meaningful.
Production code unchanged.
Refs: MUL-3685
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
The per-project issue list rides the filtered myAll cache. Changing an
issue's project is a membership change, but the surgical patch
(patchIssueInBuckets) is filter-blind and never removes a card that no
longer matches the list's project filter — so a moved issue stayed visible
in the old project's list until a manual refetch (#4548 / MUL-3669).
Root cause: project_id was the only membership-affecting field with no
server *_changed flag. The WS handler fell back to diffing project_id
against its own cache, which breaks once onMutate has optimistically
overwritten the cached value on a local move.
- server: stamp project_changed on issue:updated (UpdateIssue + Batch),
alongside status_changed / assignee_changed.
- events.ts: surface project_changed (optional, additive — old clients ignore).
- ws-updaters: prefer the server flag, fall back to the cache diff only when
absent (older backend) so a new frontend on an old backend does not regress.
- mutations: onSettled invalidates myAll when project_id changed — a local
safety net that never depends on the WS echo (update + batch).
Tests: WS flag wins over a matching cache (local-move repro), explicit false
suppresses the legacy diff, the cache-diff fallback still fires, and both
mutations invalidate myAll on a project change.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Board column headers read byStatus[status].total, which #4415 began
maintaining purely client-side via patchIssueInBuckets ±1. That patch
adjusts the totals only when it can locate the issue in a loaded page,
so a status change on an issue outside the loaded window (very common
when an agent flips the status of something the viewer never scrolled
to) was a silent no-op. #4415 also removed the list invalidation that
used to reconcile counts, so the totals drifted until a full reload.
Thread the server's status_changed flag through onIssueUpdated and, when
a status-changed patch cannot be applied surgically (no-op because the
issue is off-screen), refetch just that one list to reconcile its
counts. The loaded/echoed fast path is untouched, so #4415's no-flicker
drag behavior is preserved.
Closes#4554
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Closes#4204
Add a daemon signal guard in resolveToken so daemon-managed subprocesses fail closed instead of falling back to the user-global config token when agent/task env markers are missing. Also adds boundary tests for MULTICA_DAEMON_PORT, MULTICA_SERVER_URL, explicit MULTICA_TOKEN priority, and normal config fallback.
Point Lark binding and issue-created links at the web app URL (MULTICA_APP_URL)
instead of the backend public URL, so users on split-domain / self-host
deployments get working links. appURLFromEnv falls back to FRONTEND_ORIGIN,
matching the backend's existing app-URL resolution.
When codex emits a single stdout line larger than the daemon's 10 MB
bufio.Scanner cap (typical trigger: thread/resume on a long-history
session), the reader goroutine returns scanner.Err()="token too long",
markProcessExited fails the in-flight RPC, and the lifecycle goroutine
enters its failure path. That path calls drainAndWait() — stdin.Close()
+ cmd.Wait() — before sending the failed Result. But cmd.Wait() never
returns: codex is alive and blocked writing the rest of the oversized
line into a stdout pipe nobody is reading, so it never reaches its
stdin read syscall and never sees the EOF. The lifecycle goroutine
therefore never sends Result{failed} to its caller, the outer daemon
blocks on the result channel, and the existing PriorSessionID-with-
empty-SessionID fallback never fires — the task is permanently
stalled and codex (Node wrapper + native Rust app-server) leaks until
the OS reaps them.
The cancel() that would have unblocked things via cmd.WaitDelay's
SIGKILL was registered as a defer AFTER drainAndWait, so LIFO defer
order put cancel last — drainAndWait blocks first, cancel never runs.
Fix:
1. drainAndWait now runs the existing graceful-then-cancel pattern
itself, in two bounded phases. Phase 1 waits for readerDone (capped
by codexGracefulShutdownTimeout, so we still give codex its OTEL
flush window on clean exits); on timeout it cancels the runCtx so
cmd.Cancel kills the tree and the reader unblocks. Phase 2 bounds
cmd.Wait() the same way for the scanner-overflow case, where
readerDone closed early but the process is still alive on a full
stdout pipe. The success-path cleanup that previously duplicated the
graceful-cancel pattern around readerDone collapses to a single
drainAndWait() call.
2. cmd.Cancel is set to send SIGKILL to the whole codex process group
(Setpgid via configureProcessGroup, signalProcessGroup on cancel)
instead of just the leader. This addresses YOMXXX's
orphaned-Codex-child concern: the Node wrapper and the native
app-server it spawns now both die when cleanup forces the kill,
rather than the native binary leaking as an orphan reparented to
init. configureProcessGroup is a no-op on Windows.
3. codexGracefulShutdownTimeoutNanos atomic.Int64 mirrors
opencodeTerminateGraceNanos so the regression test can shrink the
grace window from 10 s to 500 ms. Production code is unchanged
(default 10 s).
Outer daemon (daemon.go) already retries with a fresh session when
result.Status == "failed" && PriorSessionID != "" && result.SessionID
== ""; the failed Result now actually reaches it, so the recovery
fires on its own without any daemon-side change.
Tests:
- New regression TestCodexExecuteCleansUpWhenScannerOverflowsOnResume
spawns a fake codex that emits an 11 MB single-line thread/resume
response (trips the scanner cap) and then sleeps without re-reading
stdin. With the original drainAndWait body it blocks at the 10 s
executeFakeCodex deadline ("timeout waiting for result") — verified
by temporarily reverting just the helper body — and with the fix it
completes in ~1.3 s with Result.Status="failed",
Result.SessionID="" so the outer fallback can fire.
- Full codex test suite, full agent package, daemon + execenv +
repocache packages, go build ./..., and go vet on agent/daemon all
pass.
Out of scope (deferred to follow-up per YOMXXX): bumping the 10 MB
bufio.Scanner cap on codex / claude / copilot / cursor / hermes /
kimi / kiro / codebuddy / antigravity / qoder / openclaw / opencode
(pi already sits at 32 MB), and the shared bounded JSON-RPC line
reader that would eliminate the single-line-overflow risk class
entirely. Buffer size alone is not the fix — recovery behaviour is.
Refs: GH#4520
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* feat(slack): Socket Mode channel.Channel adapter (MUL-3516)
First slice of the Slack adapter: implements channel.Channel (Type/Connect/Disconnect/Send/Capabilities) over Slack Socket Mode, normalizes inbound events to channel.InboundMessage (DM, channel @mention, thread reply; bot-loop + edit/delete guards), decodes the per-installation config/secret blob, and registers the Factory under TypeSlack. No engine, core, or channel_* schema change. Unit-tested (translation, capabilities, config decode, chunking, Send via httptest). Resolvers + engine wiring + Block Kit binding replier follow.
Co-authored-by: multica-agent <github@multica.ai>
* fix(slack): address adapter review (MUL-3516)
- Propagate InboundHandler errors through dispatchEventsAPI/handleSocketEvent to Connect so an infra failure tears down the connection for Supervisor reconnect/backoff instead of being silently swallowed (ACK still happens first).
- Capabilities: declare only CapText | CapThreadReply; drop CapRichCard/CapAttachment/CapMessageEdit until those Send paths are wired.
- slackChatType: map mpim (multi-party DM) to group, not p2p, so the 'must address bot' filter applies; only 1:1 im is p2p.
- Document the group-addressing decision: explicit @bot mention required in groups; mention-free thread continuation deferred to the session-aware layer.
- Tests: handler-error propagation, slackChatType table, mpim-requires-mention, capabilities negative assertions.
Co-authored-by: multica-agent <github@multica.ai>
* refactor(channel): shared channel-agnostic ChatSession service (MUL-3516)
Extract the session/append//issue machinery — currently locked inside the Feishu-pinned lark.chatSessionService — into a shared engine.ChatSession parameterized by channel_type + session titles, so every IM adapter reuses it instead of re-implementing it. Logic is verbatim (find-or-create session+binding with unique-violation race re-read; append+touch+reply-target+in-tx dedup Mark; /issue parse with bare-command previous-message fallback) but channel-neutral: command-parse source is supplied by the adapter (enrichment is platform-specific). Backed by a narrow SessionQueries interface so it is unit-tested with an in-memory fake (no DB). /issue parser moved to engine.ParseIssueCommand. Next: migrate Feishu onto it and wire Slack's ResolverSet, removing the lark duplicate.
Co-authored-by: multica-agent <github@multica.ai>
* fix(channel): decouple session binding key from outbound target (MUL-3516)
Addresses Elon's round-2 review. engine.ChatSession.EnsureSession previously keyed the binding on a raw chat id (EnsureSessionInput.ChatID), so a resolver wiring Slack straight through would collapse every @bot thread in one channel into a single chat_session and overwrite last_thread_id. Make the API un-misusable:
- EnsureSessionInput.ChatID -> BindingKey: the explicit session-isolation key (Feishu: chat id; Slack DM: channel id; Slack channel: channel id + thread root), documented so a raw threaded-platform chat id is never passed straight through.
- Add EnsureSessionInput.BindingConfig (opaque) persisted on the binding's config column, so the real outbound channel/thread is preserved when BindingKey is composite — outbound routing stays separate from the isolation key.
- channel.sql CreateChannelChatSessionBinding now writes config (additive, uses the existing NOT NULL column; lark caller passes '{}', no schema change, no Feishu regression).
- Tests: TestEnsureSession_ThreadRootIsolation (two thread roots in one channel -> two sessions; same root reuses) and TestEnsureSession_StoresBindingConfig.
No production wiring change yet (per review, the not-yet-wired shared service is an accepted preparatory state); this makes the API correct before Feishu/Slack are migrated onto it.
Co-authored-by: multica-agent <github@multica.ai>
* feat(slack): Slack ResolverSet with thread-root session isolation (MUL-3516)
Wires Slack into the channel-agnostic engine.Router via a ResolverSet built on the generic channel_* queries (installation route by team_id, identity + workspace-membership recheck, two-phase dedup, audit) plus the shared engine.ChatSession. No new query, no schema change.
slackSessionRouting is the per-message isolation rule (Elon round-2 / Niko round-3): a DM is one session per channel; a channel/group message is isolated by thread root (key = channel:threadRoot, root = inbound thread_ts or the message ts for a top-level @mention), so two @bot threads in one channel are two sessions. The real channel id rides in BindingConfig for outbound; the reply thread is returned separately. Tests cover DM/channel/thread routing, config, and that distinct thread roots isolate while a same-thread follow-up reuses its key.
Not yet wired into router.go (still a preparatory commit, per review); Feishu migration onto the shared service, router/config wiring, and the Slack outbound path follow.
Co-authored-by: multica-agent <github@multica.ai>
* feat(slack): Markdown->mrkdwn outbound formatting (MUL-3516)
Slack renders mrkdwn, not Markdown, so an unconverted agent reply shows literal ** , ## and [text](url). Add formatMrkdwn — a faithful Go port of Hermes Agent's slack format_message (MIT) — and apply it in slackChannel.Send before chunking/posting. Protects fenced+inline code, converted links, and existing Slack entities behind placeholders; converts headers/bold/italic/strike/links; escapes control chars. Unit tests cover each construct plus fenced-code protection and a link nested in bold.
Co-authored-by: multica-agent <github@multica.ai>
* docs(slack): preserve Hermes MIT notice for ported mrkdwn converter (MUL-3516)
Addresses Niko's review. formatMrkdwn is a substantial port of Hermes Agent's slack format_message; MIT requires preserving the copyright + permission notice. Add the full Hermes MIT copyright/permission notice + source URL as a header on mrkdwn.go (no repo-level third-party notice file exists, and the header cannot get separated from the ported code). Also add the suggested Send-layer regression test (TestSend_AppliesMrkdwn) that pins the wiring: slackChannel.Send converts Markdown to mrkdwn before posting.
Co-authored-by: multica-agent <github@multica.ai>
* refactor(lark): migrate Feishu onto shared engine.ChatSession, drop duplicate (MUL-3516)
Completes 'every IM reuses one shared session service' and removes the dual-path the reviewers flagged as temporary. Feishu's ResolverSet now drives the channel-agnostic engine.ChatSession (channel_type=feishu, Lark session titles preserved) instead of the Feishu-specific lark.chatSessionService, which is deleted. Behavior is unchanged: engine.ChatSession is the verbatim port of the old logic and is unit-tested; the new Feishu binder param-mapping (BindingKey=chat id, CommandText=un-enriched CommandBody from Raw) is covered by feishu_resolvers_test.go.
- Delete chat_service.go (chatSessionService + helpers) and issue_command.go/_test.go (parser now engine.ParseIssueCommand). Relocate the shared TxStarter interface to tx.go (still used by binding-token + registration services).
- chat.go keeps only the AuditLogger seam; remove the now-dead ChatSessionService / EnsureChatSessionParams / AppendUserMessageParams / AppendResult / IssueCommand types.
- router.go constructs engine.NewChatSession for Feishu; inbound_enricher_test + doc.go updated.
make-test parity: go build ./..., go vet, gofmt, and go test ./internal/integrations/{lark,channel/...,slack} all pass (full Feishu suite green).
Co-authored-by: multica-agent <github@multica.ai>
* feat(slack): wire Slack adapter + ResolverSet + outbound into router (MUL-3516)
Activates the full Slack pipeline, gated by MULTICA_SLACK_SECRET_KEY (the bot/app-token decryption key). When unset the block is skipped, so existing deployments are unaffected and Feishu is untouched.
- router.go registers slack.RegisterSlack (Socket Mode connect/send Factory) + channelRouter.Register(TypeSlack, NewSlackResolverSet) (inbound pipeline) + slack.NewOutbound(...).Register(bus) (outbound).
- New slack/outbound.go: an EventChatDone subscriber mirroring the Feishu Patcher. It finds the Slack chat binding for the finished session, recovers the real channel from the binding config (the channel_chat_id may be a composite thread-isolation key) + the reply thread from last_thread_id, and posts via slackChannel.Send (reusing formatMrkdwn / chunking / threading). Sessions with no Slack binding are ignored, so it coexists with the Feishu Patcher on the shared bus.
- Tests: posts to the bound channel/thread with the real channel id; ignores non-Slack sessions, empty completions, revoked installations, and non-chat events.
Slack now shares engine.ChatSession, channel_* tables, IssueService and TaskService with Feishu. Remaining: config-driven installation provisioning (an operator currently creates the channel_type='slack' row; the config block shape — which workspace/agent — is a product decision) and a live end-to-end smoke. go build ./..., go vet, gofmt, and go test ./internal/integrations/{slack,channel/...,lark} all pass.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
`useFileUpload` exposed a single `uploading: boolean` shared across all
concurrent upload calls. When the user drag-dropped N images into the
Quick Create modal, the first upload's `finally` flipped the flag back
to false while N-1 uploads were still in flight. The submit gate (which
only checked `uploading`) re-enabled, and on submit:
- `getMarkdown()` ran `stripBlobUrls()` and erased every still-pending
`` placeholder from the prompt.
- `pendingAttachments` only contained the first-completed upload, so
`activeAttachmentIds` shipped a single ID.
- The remaining attachment rows never got linked to the new issue —
their `issue_id` stayed NULL and the UI showed "Attachment doesn't
exist".
Fix:
1. Replace the boolean with an in-flight counter in `useFileUpload`.
`uploading = inFlight > 0` so the flag stays true until ALL concurrent
uploads resolve (or reject — the `finally` decrements either way).
The public return shape is unchanged; every existing call site keeps
working.
2. Belt-and-suspenders: add `editorRef.current?.hasActiveUploads()` to
the quick-create submit gate. The editor already tracks per-node
`uploading` attrs (the source of truth for "is there a blob preview
still resolving"); checking it on submit guarantees the strip step
can never run against an in-flight image even if some future code
path mis-clears `uploading`.
3. Regression coverage in `use-file-upload.test.ts`:
- Fire two concurrent uploads, resolve them out of order, assert
`uploading` stays true until BOTH resolve. Confirmed to fail
against the pre-fix code.
- Reject one of the concurrent uploads, assert the counter still
decrements correctly (the `finally` runs on rejection).
The mock ContentEditor in `quick-create-issue.test.tsx` gains a
`hasActiveUploads: () => false` stub so the new defense call doesn't
explode in existing tests.
Verified: `pnpm test` on @multica/core (663 tests) and @multica/views
(1471 tests) both green; typecheck + lint clean on both packages.
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* ci(frontend): path-filter the frontend job to skip irrelevant PRs (MUL-3667)
The frontend job (~6min) is the CI bottleneck and runs in full on every
PR, including pure backend-only / docs-only ones that change no frontend
code.
Gate it on a paths filter: a 'changes' job (dorny/paths-filter) decides
whether anything the frontend job validates changed; the frontend job
always runs but its steps are individually gated, so on an irrelevant PR
all steps skip and the job reports a genuine green — the required
'frontend' check stays satisfied with no branch-protection change, and no
top-level 'paths' that would also gate the shared backend/installer jobs.
Push to main always runs the full job.
Also fix a stale comment: mobile-verify filters packages/core/**, not
packages/core/types/**.
An earlier revision of this PR also cached apps/web/.next/cache. Two
back-to-back CI runs (cold vs warm) showed it cut the web build compile
4.3min -> 2.0min but did NOT move the job wall (6m13s -> 6m14s): the
floor is a cluster of typecheck/test tasks (web:typecheck ~2m13s,
views:test, desktop:typecheck) co-critical with web:build and bound by
the 4-vCPU runner, not the web build alone. Dropped the cache since it is
a no-op on its own; the real wall-clock levers (turbo remote cache /
larger runner) are tracked separately.
Co-authored-by: multica-agent <github@multica.ai>
* ci(frontend): include .npmrc in the frontend path filter (MUL-3667)
Address review: root .npmrc (shamefully-hoist=true) affects the pnpm
install layout, so a .npmrc-only PR must still run the frontend job. It
was missing from the filter's install-graph group, which would have made
such a PR a silent skip — exactly what the filter must avoid.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Honor an already-configured self-host server_url/app_url when re-running `multica setup self-host` without flags, instead of silently resetting to http://localhost:8080. The existing config is added as a fallback step in the URL resolution chain (flag → env → existing config → localhost), and the overwrite confirmation now shows URL changes as `old -> new`.
Closes#4536
MUL-3660
The sliding-window Lua script used the nanosecond timestamp as both the
ZSET score and member. Two requests landing in the same nanosecond
collided on an identical member, so ZADD updated in place instead of
inserting and the window under-counted — letting requests through past
the limit. This surfaced as a flaky CI failure in
TestRedisWebhookIPRateLimiter_HasSeparateBudgetFromTokenLimiter.
Keep the timestamp as the score (so ZREMRANGEBYSCORE trimming is
unchanged) and use a per-request UUID as the member so each admitted
request is counted exactly once.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
On cancellation/timeout the opencode backend closed the stdout read end
immediately, leaving the child writing into a closed pipe. Every write then
returns EPIPE and, per anomalyco/opencode#33653, can spin an orphaned process
at 100% CPU — surfacing as high idle CPU after a cancelled task or daemon
restart (MUL-3655).
Cleanup now runs opencode in its own process group and, on cancel, drives a
graceful group-wide SIGTERM → grace → SIGKILL, closing the stdout pipe only as
a last-resort unblock once the tree has been signalled (SIGKILL is uncatchable,
so no member can write again — no EPIPE window). The group signal also reaps
tool subprocesses opencode spawned instead of orphaning them. WaitDelay remains
the hard backstop.
Adds unix tests covering the graceful path and the SIGTERM-ignored → SIGKILL
escalation, asserting the whole process group is reaped and the run never
deadlocks on the scanner. Windows behaviour is unchanged (no process groups).
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Add a dedicated checkbox/task-list toggle button to the editor bubble menu, between the List dropdown and Quote. It turns the current line(s) into a `- [ ]` task item or back to a paragraph in one tap, reusing the existing toggleTaskList() command and the same Toggle/Tooltip pattern as the neighboring controls. Adds bubble_menu.task_list locale keys for en/zh-Hans/ja/ko.
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* fix(daemon): resolve skill bundles per-skill with size-scaled timeout (MUL-3650, #4505)
Cold-start skill resolution downloaded the agent's entire bundle in one
atomic request bounded by the shared 30s control-plane http.Client timeout.
On a slow/jittery link a large bundle (15+ skills) could not finish the body
read in 30s, and because the cache was only written after the whole batch
succeeded, nothing was persisted on failure — so every dispatch re-downloaded
the full bundle and timed out again, never converging.
Resolve each missing bundle in its own request and cache it the moment it
arrives:
- daemon: per-skill resolve with a deadline scaled to the bundle's declared
size (floor 30s, cap 5m, ~50KB/s floor throughput) instead of the fixed
control-plane timeout; each success is persisted independently, so a
dispatch that fails on one skill still caches the rest and the next dispatch
only re-fetches what is missing.
- client: dedicated bundleClient with no fixed Timeout (deadline comes from
ctx), a singular ResolveSkillBundle, and a short transient-retry schedule.
Tests cover the size-scaled timeout and the cross-dispatch incremental
caching / convergence (a failed skill does not discard its siblings, and
cached skills are not re-fetched).
Co-authored-by: multica-agent <github@multica.ai>
* fix(daemon): accept server-side skill updates in per-skill resolve (MUL-3650)
Address review on #4530: resolveSkillBundle validated the returned bundle
against the claim-time ref, which pinned it to the requested hash. The resolve
endpoint intentionally serves the agent's current bundle and hash when the
requested hash is stale (the skill can be edited between claim and prepare), so
a legitimate updated bundle was rejected as invalid and the task failed.
Confirm only that the server returned the requested skill (source/id), then
validate self-consistency against a ref derived from the returned bundle and
cache it under its own hash — matching the documented endpoint contract. Adds a
regression test covering a stale-hash request answered with an updated bundle.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The lark_*->channel_* cutover (MUL-3515) is deployed to prod, and the
MULTICA_LARK_HUB_DISABLED park-switch was a one-time scaffold for that
rollout — the end state intentionally does not use it (prod never set the
env). Remove the env-gated branch from cmd/server/main.go so the channel
supervisor always starts when built; its existing nil-guard and shutdown
join are unchanged. Trim migration 124's now-obsolete switch runbook to a
short historical note (comment-only; 124 is already applied, so this does
not re-run).
Refs MUL-3515
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* docs: document MULTICA_<PROVIDER>_ARGS default agent argument env vars
The backend default-args env-var layer (MULTICA_CLAUDE_ARGS / MULTICA_CODEX_ARGS /
MULTICA_CODEBUDDY_ARGS) shipped in #1807 but was never added to the docs site
environment-variables page. Document the variable, its precedence relative to
per-agent custom_args, POSIX shell-word parsing, and the shared blocked-flags
filter. Closes the docs follow-up requested in #1467.
* docs: refine MULTICA_<PROVIDER>_ARGS wording and sync zh/ja/ko translations
Reword the English section so daemon-wide default args read as a default
baseline rather than a hard ceiling (per-agent custom_args are appended
afterward and can override), and drop the uncertain --max-budget-usd example.
Sync the new env var row and section into the zh/ja/ko docs pages.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Adapts OpenClaw execenv prep to the 2026.6.x agents schema (agents.list config path removed; agents live in a sqlite registry). Case-insensitive key-missing guard + registry fallback on read, version-aware emission on write so per-task workspace pinning keeps working.
Closes#3028
MUL-3643
* docs: add 2026-06-24 changelog entry
Co-authored-by: multica-agent <github@multica.ai>
* docs(changelog): refine 2026-06-24 entry wording and terms
- Surface the flagship features in the title (Feishu collaboration channel
upgrade + feature rollout) instead of leading with an improvement and a
vague "runtime rollout" phrase
- Fix glossary term: Autopilot -> 自动化 (was 自动任务) in zh
- Make Feishu naming consistent within the entry (was mixing 飞书/Lark)
- Reconcile cross-language mismatch (Gemini CLI removal + Qoder/CodeBuddy/
Antigravity guidance now stated the same in all locales)
- Replace internal/jargon phrasing with product language across en/zh/ja/ko
MUL-3640
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: J <j@multica.ai>
The race detector caught a flaky failure on main (passes on retry):
Supervisor.startSupervisor does s.wg.Add(1) under s.mu, while Supervisor.Wait
calls s.wg.Wait() with no lock. Calling WaitGroup.Add concurrently with
WaitGroup.Wait is a data race and undefined per the WaitGroup contract — so it
only trips occasionally (it passed locally and in PR CI).
Wait now blocks on stopChan (closed by Run's defer when Run returns) before
calling wg.Wait(). Run is the sole caller of startSupervisor, so once Run has
returned no further Add can happen and wg.Wait is race-free. WaitWithTimeout
inherits the fix (it calls Wait), and its timer still bounds shutdown.
This latent race existed in the original lark.Hub.Wait too; fixed properly in
the generalized Supervisor.
Verified: go test -race -count=300 on the flagged test and -count=8 on the
whole engine package, all clean; no deadlock from the stopChan gate (every
caller pairs Wait with a started Run + cancelled ctx).
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
A run could post running progress/plan narration as issue comments, and a
review run surfaced its in-progress narration as the result instead of a
conclusion (MUL-3605).
Add one rule to the Output section's issue-task branch, in both the
legacy and slim briefs: post exactly one comment per run — the final
result, before the turn exits — and keep plans/progress in the agent's
own reasoning. The pre-existing "Final results MUST be delivered … a task
that finishes without a result comment is invisible" line already makes
the comment mandatory, and "state the outcome, not the process" already
rules out progress dumps, so no second rule is added.
Chat / quick-create / autopilot keep their own delivery channels. Adds a
regression test across both brief paths.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(channel): add channel-agnostic engine Supervisor (MUL-3620)
Stage-1 (MUL-3515) shipped the channel abstraction but nothing drove it.
Add the generic engine that does:
- channel.InboundHandler + Config.Handler: the single shared inbound entry
the engine injects into every adapter (Hermes set_message_handler model).
- channel.Channel.Connect now blocks for the connection lifetime (doc), so
the supervisor can tie lease renewal to connection liveness.
- new package channel/engine: Supervisor, generalized out of lark.Hub. It
enumerates active installations across ALL channel types (no hard-coded
feishu), fences each behind the WS lease CAS, builds the platform Channel
via channel.Registry, drives Connect/Disconnect with backoff+jitter, and
restarts on credential rotation. Knows nothing about any platform.
channel.Channel is now driven by an engine; integrations/channel has an
external consumer. Feishu adapter + boot cutover follow next.
Tests: supervisor_test.go covers lease CAS, reclaim, reap-on-revoke,
rotation restart + token fencing, backoff on build error, lease-loss
teardown, bounded release, shutdown timeout. Race-clean.
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): drive Feishu through the channel engine; remove lark.Hub (MUL-3620)
Refactor Feishu into the first channel.Channel and cut boot over to the
channel-agnostic engine.Supervisor, removing the Feishu-only Hub.
- feishuChannel implements channel.Channel: Connect runs the existing
WS long-conn connector for one installation; Send posts a text reply
via the Lark IM API; Capabilities declares Feishu's feature set.
RegisterFeishu wires it to channel.TypeFeishu — adding a platform is
now 'register a Factory', no engine edit.
- FeishuRuntime extracts the former Hub.handleEvent / scheduleReply:
runs the Dispatcher and drives the detached typing indicator +
OutcomeReplier off the connector ACK path. main.go drains it on
shutdown after the supervisor stops delivering events.
- channelInstallationStore (engine.InstallationStore) enumerates active
installations across ALL channel types via the new de-hardcoded query
ListAllActiveChannelInstallations; the Supervisor routes each row to
its registered Factory by channel_type. Generic per-row fingerprint
replaces the feishu-specific one.
- boot: engine.Supervisor replaces lark.Hub.Run; MULTICA_LARK_HUB_DISABLED
keeps its name for runbook compatibility.
- delete hub.go / hub_pgx.go / hub_test.go; relocate the connector
contract (EventConnector/EventEmitter), uuidString, and the reply-path
tests (-> feishu_runtime_test.go) so coverage is preserved.
No channel_* schema change. Feishu behaviour unchanged; lark + channel +
engine tests green under -race; go build/vet ./... clean.
Remaining (follow-up): lift the Dispatcher pipeline into a channel-
agnostic engine.Router over channel.InboundMessage + resolver interfaces,
so the inbound core stops being Lark-shaped and adding a channel needs
zero core edits (validated by Slack, MUL-3516).
Co-authored-by: multica-agent <github@multica.ai>
* feat(channel): add channel-agnostic engine.Router (inbound pipeline) (MUL-3620)
Generalize lark.Dispatcher's inbound pipeline into engine.Router: the single
shared channel.InboundHandler the Supervisor injects into every Channel. It
routes by ChannelType to a registered ResolverSet and runs the same ordered
pipeline for every platform (install route -> two-phase dedup -> group @bot
filter -> identity+membership -> ensure session -> append+mark -> /issue ->
debounced run), then drives the detached OutboundReplier + typing indicator.
Platform specifics live behind resolver interfaces (InstallationResolver,
IdentityResolver, Deduper, SessionBinder, Auditor, OutboundReplier,
TypingNotifier) + shared services (IssueCreator/TaskEnqueuer/SessionReader).
Adding a platform is 'register a ResolverSet', not 'edit the Router'. Outcome
/ DropReason values match the legacy lark ones 1:1.
Additive: lark.Dispatcher untouched and still wired; the feishu ResolverSet,
the cutover, and the old-path removal land next. channel.InboundMessage gains
ForceFresh (the normalized /fresh affordance). Batcher moved into engine.
router_test.go covers the pipeline invariants (routing, dedup finalize
states, group filter, identity, membership, ensure/append, /issue, debounce,
flush offline, force-fresh, drain) with generic fakes; race-clean.
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): cut Feishu over to engine.Router; remove lark.Dispatcher; core no longer Lark-shaped (MUL-3620)
Wire the channel-agnostic engine.Router (added in the prior commit) as the
shared inbound handler and refactor Feishu into a ResolverSet, completing the
generic-engine cutover. The inbound core (engine.Router) now contains zero
platform specifics.
- Feishu ResolverSet (feishu_resolvers.go): InstallationResolver,
IdentityResolver, Deduper, SessionBinder, Auditor, OutboundReplier,
TypingNotifier — each backed by the existing ChannelStore / ChatSessionService
/ OutcomeReplier / typing indicator, translating at the channel.InboundMessage
boundary (platform fields read from Raw). origin_type stays 'lark_chat'.
- feishuChannel now produces a normalized channel.InboundMessage and hands it to
the engine handler via channel.Config.Handler; the old Raw round-trip through
lark.Dispatcher is gone.
- Remove lark.Dispatcher, FeishuRuntime, and lark's pending_batcher (the engine
owns the pipeline + batcher now); their behavioural coverage moved to
engine.Router tests. Surviving native types (InboundMessage / Outcome /
DispatchResult) relocated to feishu_types.go.
elon review nits addressed:
- The channel engine (Registry + Router + Supervisor) is now built
UNCONDITIONALLY, outside the MULTICA_LARK_SECRET_KEY gate, so a non-Lark
deployment runs it; Feishu registers its Factory + ResolverSet only when its
key is present.
- channel.Config.Raw is now genuinely the platform config JSONB
(channel_installation.config): the feishu factory builds a credentials-only
Installation from it, and the workspace/agent identity is resolved per message
by the Router — no full-db-row marshaling.
- feishuChannel gains direct unit tests: factory config decode, Send text +
reply-target mapping, Capabilities, inbound normalization + Raw round-trip,
msg-type + result mapping.
No channel_* schema change. go build/vet ./... clean; channel + engine + lark
green under -race. Feishu behaviour preserved (pipeline logic lifted verbatim,
only generalized).
Co-authored-by: multica-agent <github@multica.ai>
* docs(channel): fix stale comments on the channel engine boot (MUL-3620)
Address Elon's review nit: three comments still described the pre-cutover
behavior.
- handler.go: ChannelSupervisor is built UNCONDITIONALLY now, not nil when
MULTICA_LARK_SECRET_KEY is unset.
- main.go: same — the supervisor always exists; only MULTICA_LARK_HUB_DISABLED
parks it.
- router.go: with no platform registered the store still lists active rows;
Registry.Build returns ErrUnknownType and the supervisor backs off (it does
not 'find no installations').
Comment-only; no behavior change.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
ListChildIssues and ListChildrenByParents ordered by
`position ASC, created_at DESC`. position is assigned by
NextTopPosition as MIN(position)-1 scoped to (workspace, status),
not relative to siblings, so a parent's children interleave
unpredictably across creation batches and statuses.
Order by `number` (a per-workspace monotonic counter) instead.
ASC keeps sub-issues in stable creation order (oldest first), so a
parent's plan reads top-to-bottom in the order tasks were added.
Adds ordering tests covering both queries with scrambled positions
and mixed statuses.
Closes#4232
MUL-3362
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Add ProxyURL field to GorillaDialer so deployments behind a corporate
proxy can route Lark WebSocket connections through an HTTP CONNECT proxy.
- GorillaDialer.ProxyURL: optional proxy URL parsed and applied to the
underlying gorilla/websocket dialer before each DialContext call.
Empty value preserves the default ProxyFromEnvironment behaviour.
- Router reads MULTICA_LARK_WS_PROXY_URL env var and sets it on the
production dialer.
- Three new unit tests cover invalid URL, proxy-applied, and empty-URL
default paths.
Closes#4032
Co-authored-by: multica-agent <github@multica.ai>
* docs: tidy agent runtime provider pages, add per-runtime FAQ (MUL-3617)
- Remove the Gemini CLI provider from install-agent-runtime and providers
across all four languages (Google folded the standalone CLI into
Antigravity). Update tool counts 12 -> 11 and the dependent
session-resumption, MCP, and skill-path sections.
- Add the Hermes profile custom_args workaround as a per-runtime FAQ note
under providers#hermes (supersedes #4497, which placed it in agents-create).
- Fix stale Japanese install copy that claimed only Claude Code reads
mcp_config and linked to a non-existent anchor.
Co-authored-by: multica-agent <github@multica.ai>
* docs: add Qoder and CodeBuddy runtimes to provider pages (MUL-3617)
Document the two newly added runtimes on install-agent-runtime and
providers across all four languages:
- Qoder (Alibaba): ACP-over-stdio CLI `qodercli`, shares the transport
with Hermes/Kimi/Kiro; session/resume, ACP mcpServers, dynamic model
discovery, native skills at .qoder/skills/.
- CodeBuddy (Tencent): Claude Code-compatible CLI `codebuddy`, driven via
stream-json; --resume, --mcp-config, dynamic models, .claude/skills/.
Update tool counts 11 -> 13 and the MCP section (now ten of thirteen
consume mcp_config; the other three still ignore it).
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The MUL-3560 slim runtime brief — kind-driven dispatcher, per-section
gating, prose compression for ~7k chars saved on the typical
comment-triggered task — now ships behind the `runtime_brief_slim`
feature flag wired via the framework-level service from MUL-3615.
Default: OFF in every environment (production stays on the legacy
brief that has shipped for ~2 years). Staging opts in via the YAML
rule set; ops can override per-process with `FF_RUNTIME_BRIEF_SLIM=true`.
Production is held back until staging has burned in long enough that
we are confident the slim brief does not regress agent behaviour.
Architecture (one toggle point, two code paths, both fully tested):
buildMetaSkillContent (runtime_config.go)
│
└─ useSlimBrief() → false (default)
│ → fall through to the legacy verbose body that ships on
│ main today — byte-for-byte unchanged, no migration risk
│
└─ useSlimBrief() → true
→ buildMetaSkillContentSlim (runtime_config_sections.go)
→ classifyTask → 5-way kind switch → per-section writers
BuildCommentReplyInstructions takes the same gate, so the per-turn
comment prompt and the runtime brief stay in sync on which template
they emit.
What's in this PR:
- runtime_config_flag.go (new): package-scope `runtimeFlags` atomic
pointer + `SetFeatureFlags` setter + `useSlimBrief` toggle point.
Nil-safe: a daemon that forgets to wire the service falls back to
legacy, no panic.
- runtime_config_kind.go (new): `taskKind` enum + `classifyTask` +
`hasIssueContext` predicate. Used only by the slim path.
- runtime_config_sections.go (new): the slim brief itself —
`buildMetaSkillContentSlim` + per-section `writeXxx` helpers
+ `writeAvailableCommandsQuickCreate` minimal variant +
`writeBackgroundTaskSafetySlim` compressed safety section. The
Section × Kind matrix is documented inline on
`buildMetaSkillContentSlim` and the test below checks the
dispatcher does not diverge from the spec.
- reply_instructions.go: `BuildCommentReplyInstructions` gains a
short slim-or-legacy prelude; new `buildCommentReplyInstructionsSlim`
is the compressed cookbook (defers the shell-hazard rationale to
`## Comment Formatting`).
- runtime_config.go: `buildMetaSkillContent` gains a 2-line
dispatcher at the top; the legacy body is otherwise untouched.
- runtime_config_kind_test.go (new): canaries for both paths.
- TestClassifyTask: 5 kinds + 3 tiebreak cases.
- TestTaskKindHasIssueContext: predicate semantics.
- TestSlimFlagOffUsesLegacy: nil flag service → legacy path
(renders "Get full issue details.", a legacy-only substring).
- TestSlimFlagOnUsesSlim: flag on → slim path (renders "full
issue.", a slim-only one-liner) AND must NOT render legacy
"Get full issue details.".
- TestBuildMetaSkillContentSlimKindMatrix: locks the per-kind
section set; heading match is line-anchored so inline references
don't trip absence assertions.
- TestSlimQuickCreateAvailableCommands: locks the minimal-variant
content for quick-create (issue create present, every other
Core command absent).
- TestSlimBriefIsSubstantiallyShorter: ≥ 30% reduction guard so
a future change can't accidentally re-bloat the slim path back
to legacy levels.
- cmd/server/main.go: now calls `execenv.SetFeatureFlags(flags)`
immediately after constructing the feature flag service.
Measured impact (slim vs legacy, claude provider, realistic fixture
with 2 repos + 2 skills + member initiator):
legacy = 19567 chars
slim = 11868 chars Δ = -7699 (-39.3%)
Verification:
- go vet ./internal/daemon/... ./cmd/server/... ok
- go test ./internal/daemon/... ok
- go test ./pkg/featureflag/... ok
- TestSlimBriefIsSubstantiallyShorter logs the 39.3% ratio
- TestSlimFlagOffUsesLegacy + TestSlimFlagOnUsesSlim pass both
directions, so the dispatcher is locked in code.
The pre-existing `internal/handler` test failures
(TestLeaveWorkspace_RevokesOwnRuntimes,
TestDeleteMember_CancelsTasksFromAgentReassignment,
TestDeleteMember_NoRuntimes_DeletesMember) reproduce on plain
`origin/main` with the same `relation "channel_user_binding" does
not exist` SQL error — they are a missing-migration bug from the
recent channels foundation PR (ce28d0aa0), not anything this PR
touched.
Rollout plan:
1. Merge this PR. Production daemons keep emitting the legacy brief
(flag default false).
2. Add a YAML rule to staging's
`MULTICA_FEATURE_FLAGS_FILE`:
runtime_brief_slim:
default: true
Staging daemons start emitting the slim brief on next restart.
3. Watch `agent prompt prepared` logs + agent behaviour for 7 days.
4. If staging is clean, flip the prod YAML to `default: true`.
Legacy code path stays in the binary as a kill-switch
(`FF_RUNTIME_BRIEF_SLIM=false` to revert without a deploy).
5. After ~30 days clean in prod, follow up with a PR that deletes
the legacy body and the flag — same pattern as docs/feature-flags.md
recommends ("plan the death of the flag at birth").
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* feat(featureflag): framework-level feature flag system (MUL-3615)
Introduces a reusable feature flag framework so future features can adopt
flags without writing infrastructure code.
Backend: server/pkg/featureflag (Go)
- Service / Provider / Decision separation per Martin Fowler's Toggle
Point / Toggle Router / Toggle Configuration pattern.
- Providers: StaticProvider (rules in source control), EnvProvider
(FF_<KEY> overrides for ops kill switches), ChainProvider
(first-hit-wins composition).
- EvalContext carried through context.Context with WithEvalContext /
EvalContextFrom; supports user_id, workspace_id, free-form attributes.
- PercentRollout via deterministic FNV-1a bucketing; same user always
lands in the same bucket so experiments do not flap between requests.
- Nil-safe Service: a nil *Service or missing flag returns the caller's
default so business code never panics on a missing flag.
- 100% unit-test coverage with -race; go vet clean.
Frontend: packages/core/feature-flags (TypeScript)
- Same vocabulary as the Go side (Decision, EvalContext, Rule,
PercentRollout). FNV-1a parity ensures cross-tier bucket agreement.
- FeatureFlagService + StaticProvider + ChainProvider in pure TS.
- React glue: FeatureFlagsProvider, useFlag(key, default),
useVariant(key, default). Hooks fall back to the default when no
provider is mounted so Storybook / unit tests stay simple.
- Vitest tests for service, providers, hash, and React hooks.
Docs: docs/feature-flags.md — wiring, EvalContext, toggle points,
backend-protection note, and the standard best-practice checklist.
The framework intentionally has no third-party Go deps and no API
surface beyond what real callers will need. New providers (DB, remote
config, LaunchDarkly) plug in by implementing Provider; no existing
caller has to change.
Co-authored-by: multica-agent <github@multica.ai>
* fix(featureflag): cross-tier hash parity + variant only when enabled (MUL-3615)
Two must-fix issues from the PR review on #4496:
1. TS hash had a trailing zero separator that Go did not emit, so the
same (key, identifier) bucketed differently on the two tiers. The
"user lands in the same bucket on server and client" promise was
broken. For example billing_new_invoice/user-42 was bucket 97 in Go
and bucket 11 in TS.
Fix: TS fnv1a now emits the zero separator BETWEEN parts only, never
after the last one, matching Go's hash.Write byte stream exactly.
Verified by parallel golden tests on both sides that pin five
(key, identifier) -> bucket triples; if either side drifts both tests
fail and one must be brought back in sync.
2. StaticProvider returned `Rule.Variant` regardless of whether the rule
evaluated to enabled=true. A 0%-rollout user, a deny-listed user, or
a default-off user would see variant="experiment-v2", so callers
branching on Variant() would route control users into the experiment
arm.
Fix: Rule.Variant is now the ON-variant only. When the rule evaluates
to enabled=false the Decision's variant is the canonical "off",
regardless of what Rule.Variant says. Documented as a behavior
contract in the Rule godoc / JSDoc and covered by regression tests
on both sides.
Tests: - go test -race ./pkg/featureflag/... : all green (1.58s).
- pnpm --filter @multica/core test : 661/661 (3 new).
- pnpm --filter @multica/core typecheck: clean.
Co-authored-by: multica-agent <github@multica.ai>
* fix(featureflag): hash UTF-8 bytes on the TS side for cross-tier parity (MUL-3615)
Follow-up review on PR #4496 caught that the previous hash fix was only
correct for ASCII input. The TS side used `charCodeAt`, which returns
UTF-16 code units, while the Go side hashes the UTF-8 byte
representation. Any non-ASCII flag key or identifier — Chinese flag
names, accented user IDs, emoji — would bucket differently on backend
vs frontend, silently breaking the "same user, same bucket" promise the
PR description makes.
Concretely:
flag/é Go 53 vs TS-old 68
flag/🦄 Go 82 vs TS-old 75
实验/user-1 Go 90 vs TS-old 4
flag/用户-1 Go 95 vs TS-old 2
Fix: replace per-char charCodeAt with a module-level `TextEncoder`
('utf-8') and hash each encoded byte. After the fix all four cases above
match Go exactly, and the existing ASCII cases continue to match.
The cross-language golden tables on both sides now include the 5 new
non-ASCII cases alongside the 5 ASCII cases, so any future regression
that swaps UTF-8 for charCodeAt (or vice versa) will fail loudly on
both Go and TS simultaneously.
TextEncoder is part of WHATWG Encoding and is available in every
evergreen browser, in Node 11+, and in Hermes (React Native) >= 0.74,
which covers every runtime that imports @multica/core/feature-flags.
Tests: - go test -race ./pkg/featureflag/... : all green.
- pnpm --filter @multica/core test : 661/661.
- pnpm --filter @multica/core typecheck : clean.
Co-authored-by: multica-agent <github@multica.ai>
* feat(featureflag): wire into main app config — YAML file + env override (MUL-3615)
Follow-up requested by Yushen on PR #4496: make the feature flag
framework configurable through the existing main-program config system
instead of requiring Go code edits. multica's main app is purely env-var
driven (see .env.example) with optional MULTICA_*_FILE knobs for richer
config; feature flags now follow the same pattern.
server/pkg/featureflag/config.go
- LoadRulesFromYAMLFile(path) parses a YAML rule set into runtime
Rule structs. Empty files are a valid "no flags yet" state; missing
or malformed files surface a hard error so operators see misconfig
the same way DATABASE_URL parse errors do.
- NewServiceFromEnv composes the standard provider chain:
1. EnvProvider("FF_") (runtime kill-switch path)
2. StaticProvider from YAML file (declarative rule set)
When MULTICA_FEATURE_FLAGS_FILE is unset, only the env layer is
active and every IsEnabled call falls through to the caller's
default, so the server can boot before any flag is authored.
server/cmd/server/main.go
- Construct the Service once at startup right after env-var warnings,
fail loudly on malformed YAML, log the loaded rule count via the
Service logger. The Service is held in a local `flags` variable
ready to be threaded into handler.Handler / service constructors
when the first flag user lands. Threading is deferred to the PR
that adds the first business consumer so this PR stays a pure
framework + config layer.
.env.example
- New "Feature flags" section documents MULTICA_FEATURE_FLAGS_FILE and
the FF_<KEY> override convention, with a minimal YAML schema example
inline.
docs/feature-flags.md
- Replace the "build a provider manually" example with the
NewServiceFromEnv pattern that now matches what main.go actually
does. Show the YAML schema in one place. Note the on-variant /
off semantics from the previous review round.
server/pkg/featureflag/doc.go
- Update package doc to mention the gopkg.in/yaml.v3 dependency
(already a server-level dep) instead of the now-inaccurate
"no third-party dependencies" claim.
Tests: - go test -race -count=1 ./pkg/featureflag/... all green; new
config_test.go covers: simple YAML, full-shape YAML, empty file,
missing file, malformed YAML, no env var, file-only, env-beats-file,
bad file surfaces error.
- go test -race -count=1 -run TestHealth ./cmd/server/... sanity
check that the main.go boot path with the new wiring still passes.
- go vet ./... clean.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
The header live chip derived its active-task state from the workspace-wide
agent-task-snapshot, while the right-panel Execution log read the per-issue
task list. Two queries, two endpoints, two independent refetches: the heavier
workspace snapshot lands later than the per-issue list, so the log could show
a running task while the header chip had not started yet.
Point the chip at the same `issueKeys.tasks(issueId)` cache the Execution log
uses (identical query options). Both surfaces now observe one cache entry and
update atomically. Drop the now-redundant workspace-id lookup and client-side
issue_id filter, since the endpoint is already issue-scoped.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Post-deploy of the new scheduled-dispatch scheduler (PR #4444), an
autopilot configured for "weekdays 17:10 Asia/Shanghai" fired at
~12:30 Beijing the day after deploy — ~4h 38m before the next
scheduled time the UI showed. Traced to a cold-start regression in
the planner hook:
Old behaviour
-------------
On the first tick after migration the hook found no
`sys_cron_executions` row for the trigger
(`latestPlan(...).Found == false`) and anchored on the trigger's
`created_at`, then applied the 24h replay cap:
after := cfg.CreatedAt
if oldest := now.Add(-replayWindow); after.Before(oldest) {
after = oldest // now - 24h
}
For a trigger created days/weeks earlier and last fired by the
legacy goroutine at Mon 17:10 Beijing (= Mon 09:10 UTC), this set
`after = Tue 04:13 UTC - 24h ≈ Mon 04:13 UTC`. The half-open
enumeration `(Mon 04:13 UTC, Tue 04:13 UTC]` STILL contained Mon
09:10 UTC — the occurrence the legacy code had already handled —
so the new scheduler dispatched it again the moment it took over.
The result: a SCHEDULED-source autopilot_run with planned_at = Mon
17:10 Beijing but a wall-clock dispatch at Tue ~12:30 Beijing.
Timezone math was correct; the bug was purely the cold-start
anchor not respecting prior-fire history.
Fix
Co-authored-by: multica-agent <github@multica.ai>
---
The `autopilot_trigger.last_fired_at` column is maintained by both
the legacy goroutine and the new scheduler (via
TouchAutopilotTriggerFiredAt), so it is the authoritative
"most-recent successful fire" cursor across the migration boundary.
The planner hook now anchors cold-start enumeration on it:
case latest.Found: after = latest.PlanTime
case lastFiredAt != zero: after = lastFiredAt
default: after = cfg.CreatedAt
For the regressed case, `after = Mon 17:10 Beijing`, the next
enumeration window is `(Mon 17:10, Tue 12:30]`, and Tue 17:10 is
in the future — the hook returns nothing and the trigger waits
quietly for Tue 17:10 as the UI promised. For brand-new triggers
(last_fired_at NULL), the original `created_at` path still
applies. For long-dormant triggers the `replayWindow` cap remains.
Changes
-------
* `ListSchedulableAutopilotTriggers` SQL now returns
`last_fired_at`.
* `autopilotTriggerConfig.LastFiredAt` is populated by the scope
provider on every tick.
* `autopilotPlansForScope` cold-start branch uses the new anchor.
Tests
-----
* TestAutopilotScheduleJobColdStartHonorsLastFiredAt — seeds the
exact dev-environment shape (created 3 days ago, last_fired_at
5 hours ago, no sys_cron_executions row), runs a tick, asserts
zero exec rows AND zero autopilot_run rows. Without the fix this
test produces one of each at a historical plan_time.
* TestAutopilotScheduleJobColdStartBrandNewTriggerStillFires —
asserts a brand-new trigger (last_fired_at NULL) still fires its
first due occurrence on cold start.
All existing `TestAutopilotScheduleJob*` tests still pass.
Refs MUL-3551
Co-authored-by: Eve <eve@multica-ai.local>
* feat(integrations): add platform-agnostic channel foundation
Introduce server/internal/integrations/channel — the contract every
inbound IM integration implements, so the core never learns a platform's
event JSON. Four pieces:
- Channel interface (Type/Connect/Disconnect/Send/Capabilities) + Factory
+ Config (channel_type + opaque JSON blob, maps to channel_installation).
- Normalized InboundMessage/OutboundMessage envelopes + Source/MediaRef/
ReplyCtx/MsgType/ChatType. Envelope holds only cross-platform-true
fields; platform specifics live in Raw, read only by the adapter.
- Capability bitmask: declaration only, no degrade logic in core.
- Registry: Type->Factory map, last-writer-wins, concurrency-safe.
Pure package (no DB/network/platform deps). Foundation for MUL-3515; the
lark cutover + lark_*->channel_* generalization land in follow-up PRs.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* feat(channel): generalize lark_* tables into channel_* (DB layer)
Migration 123 creates channel_installation / channel_user_binding /
channel_chat_session_binding / channel_inbound_message_dedup /
channel_inbound_audit / channel_outbound_card_message /
channel_binding_token. Each carries a channel_type discriminator and a
JSONB config for platform-specific identifiers/credentials; cross-platform
columns stay flat. Existing Feishu rows are backfilled (channel_type=
'feishu', app_secret_encrypted via base64). NO foreign keys / cascades
(MUL-3515 §4) — integrity moves to the app layer in the cutover.
queries/channel.sql ports the lark query surface to channel_*, JSONB-aware,
plus DeleteChannelUserBindingsByWorkspaceMember /
DeleteChannelChatSessionBindingBySession for the app-layer cleanup that
replaces the removed cascades.
lark_* tables/queries are left in place here and removed once the Go
cutover lands, so this commit ships green on its own.
Verified: sqlc generate, go build ./..., full migrate chain (1..123) on
Postgres 17, and a real-data backfill spot-check (base64 round-trip,
NULL-strip, functional unique index on (channel_type, app_id)).
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* fix(channel): name app_id query param + multi-IM install key + null-safe binding merge
Addresses review on MUL-3515 (PR #4412):
- GetChannelInstallationByAppID: explicitly name params and cast app_id to
::text so sqlc emits AppID string. A bare $2 next to `config ->> 'app_id'`
was mis-attributed to the JSONB config column, generating Config []byte.
- channel_installation uniqueness -> (workspace_id, agent_id, channel_type),
with the UpsertChannelInstallation conflict key matched. Lets one agent
hold one installation per IM (feishu + slack + ...) instead of a later
install clobbering an earlier one. Behaviorally identical in the current
feishu-only world; "one agent, at most one IM overall" stays an app-layer
rule per MUL-3515 §4, not a DB constraint.
- CreateChannelUserBinding merges jsonb_strip_nulls(EXCLUDED.config) so a
re-bind carrying {"union_id": null} no longer erases an already-captured
union_id, restoring the old COALESCE(EXCLUDED.union_id, ...) semantics.
Regenerated with sqlc v1.31.1. Verified on PG17: re-install replaces in
place, feishu+slack coexist, null re-bind keeps union_id, real union_id wins.
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): channel-backed Feishu store + fix base64 backfill wrapping
Cutover step 1 of switching the lark Go code from lark_* onto the channel_*
tables (MUL-3515). Introduces the JSONB config boundary the rest of the
cutover sits on, and fixes a latent backfill bug surfaced while building it.
- migration 123: strip newlines from the app_secret_encrypted base64 backfill.
PostgreSQL encode(...,'base64') MIME-wraps at 76 chars, and a secretbox-
sealed ~72-byte secret exceeds that. Go's encoding/json decodes a JSON
string into []byte with base64.StdEncoding, which rejects embedded newlines,
so without the strip every migrated installation would fail to decrypt its
app secret once reads move to channel_installation.config.
- store.go: flat domain types (Installation / UserBinding / ChatSessionBinding)
with field parity to the retired db.Lark* rows, plus the feishu config codec.
Row->domain mappers decode the JSONB config; the secret decoder is
whitespace-tolerant so legacy MIME-wrapped data still round-trips, while the
encoder emits unwrapped base64. Binding config encodes an absent union_id as
"{}" so the upsert's jsonb_strip_nulls merge never clobbers a stored union_id.
- store_test.go: 72-byte secret round-trip, MIME-wrapped tolerance, optional
null-strip, and flat-column preservation. Verified on PG17.
Field parity keeps the upcoming ~190 db.LarkInstallation call sites a
mechanical rename. No call sites switched yet; behavior unchanged.
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): route inbound integration onto channel_* + explicit membership checks
Cutover step 2 (MUL-3515): switch the Feishu Go code from the lark_* queries to
channel_* via a ChannelStore adapter, and replace the removed member foreign key
with explicit application-layer membership checks. No user-visible behavior change.
- channel_store.go: ChannelStore embeds *db.Queries and SHADOWS the ~24 lark
query methods with channel_*-backed equivalents, keeping the db.Lark*
signatures so the dispatcher/hub/services and their ~20k lines of tests stay
untouched; the feishu JSONB config is (de)coded by store.go. Adds
IsWorkspaceMember and a tx-aware WithTx. Only production wiring swaps
*db.Queries for *ChannelStore.
- Membership re-check (§4 removed the lark_user_binding -> member FK, so a
binding row no longer proves current membership):
* the dispatcher inbound identity step verifies membership after the binding
lookup; a former member's stale binding is dropped as non_workspace_member
+ audited and never reaches chat_session (§4.3 safety property).
* RedeemAndBind and BindInstallerTx replace the now-dead FK (23503) branch
with an explicit IsWorkspaceMember gate, preserving the existing
ErrBindingNotWorkspaceMember outcome without burning the token.
- router wires the ChannelStore into the patcher, typing indicator, dispatcher,
hub, and the union_id/region backfills; constructor-based services wrap
*db.Queries internally so their signatures and nil-check tests are unchanged.
Verified: go build ./... ; go vet ; gofmt ; go test -race ./internal/integrations/...
(full lark suite green unchanged + new membership drop/error tests). Adapter
field mappings (secret base64, union_id RMW, chat-id/open-id remaps, dedup,
token, card) checked end-to-end against a PG17 channel_* schema.
lark_* tables and queries remain (unused at runtime) until the S3 cleanup-hooks
and S4 drop-tables/rename commits.
Co-authored-by: multica-agent <github@multica.ai>
* fix(channel): renumber generalization migration 123 -> 124
main merged 123_issue_stage after this branch forked, so the branch's 123_channel_generalization now collides on the migration number. The runner keys schema_migrations by full version string and would still apply both, but a duplicate number is a merge hazard and convention violation, so move the channel migration to the next free slot (124).
issue_stage (ALTER issue ADD COLUMN stage) and the channel generalization touch disjoint tables; verified on PG17 that 123_issue_stage applies cleanly on a DB already carrying 124_channel_generalization, so the two are order-independent. sqlc regenerated (v1.31.1): only the migration-number comment changed.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* feat(channel): prune channel bindings on member removal + chat session delete
MUL-3515 §4 dropped every channel_* foreign key, so the old ON DELETE CASCADE that cleared a user's channel_user_binding when they left a workspace, and a chat's channel_chat_session_binding when its chat_session was deleted, no longer fires. Re-establish that integrity in the application layer, inside the existing transactions: revokeAndRemoveMember -> DeleteChannelUserBindingsByWorkspaceMember, DeleteChatSession -> DeleteChannelChatSessionBindingBySession.
Adds real-DB tests for both paths, including a scoping check that a remaining member's binding survives the prune. Verified on PG17: both new tests plus the existing revocation tests and the full handler package pass.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* fix(channel): scope Lark/Feishu store reads to channel_type='feishu'
The S2 cutover routed the Feishu integration onto channel_*, but the Lark-facing ChannelStore wrappers read installation / chat-session-binding / outbound-card rows across ALL channel_type values. Once a second IM exists, that would let the Lark hub supervise a non-Feishu installation, the Lark install list show it, /lark/installations/{id} revoke another channel's row, and the outbound patcher / typing indicator act on a non-Feishu chat binding or card.
Add a channel_type predicate to the six read/list channel queries and pass channelTypeFeishu from every wrapper: GetChannelInstallation, GetChannelInstallationInWorkspace, ListChannelInstallationsByWorkspace, ListActiveChannelInstallations, GetChannelChatSessionBindingBySession, GetChannelOutboundCardByTask.
The S3 cleanup deletes (DeleteChannelUserBindingsByWorkspaceMember / DeleteChannelChatSessionBindingBySession) stay all-channel on purpose: a member leaving or a chat_session being deleted should clear every IM's binding. Adds a real-DB test that seeds a Slack installation/binding/card next to the Feishu ones and asserts the Lark wrappers never return them.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* refactor(channel): replace db.Lark* translation layer with lark domain types
S2 introduced ChannelStore as a translation layer that read/wrote channel_* but kept the retired db.Lark* struct/param shapes so the dispatcher/hub/services and their ~20k lines of tests did not have to change. This collapses that layer: the store now takes and returns the package's flat domain types (Installation, UserBinding, ChatSessionBinding, InboundMessageDedup, BindingTokenRow, OutboundCardMessage) and the *Params types in params.go, with channel-neutral field names (ChannelUserID / ChannelChatID / ...). All call sites, fakes, and tests move to the domain types.
No behavior change: only channel_* is read/written (as before); db.Lark* is now unused, and the lark_* tables + queries/lark.sql are removed in the next commit. Verified on PG17: go build / vet / gofmt clean, go test -race ./internal/integrations/... green (the ~20k-line fake suite), and the lark + handler suites pass.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* refactor(channel): drop lark_* tables and queries (remove old path)
The Go cutover (previous commit) moved the lark package entirely onto channel_* and the domain types, leaving the lark_* tables, queries/lark.sql, and the generated db.Lark* models unused. Remove them per the design (§5: replace, do not keep both): migration 125 drops the seven lark_* tables (data already lives in channel_* since migration 124), and queries/lark.sql is deleted + sqlc regenerated, removing the db.Lark* models and lark query methods.
The 125 down recreates the authoritative pre-drop schema (bot_union_id, region, per-installation dedup PK, thread-reply columns). Verified on PG17: fresh migrate up ends with lark_* gone + channel_* present; isolated 125 down/up round-trips correctly; go build / vet / gofmt clean; go test -race ./internal/integrations/... and the handler suite pass.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* fix(migrations): remove trailing blank line at EOF of 125 down migration
git diff --check flagged a blank line at EOF of 125_drop_lark_tables.down.sql (a pg_dump-generation artifact). Whitespace only; the recreate SQL is unchanged.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* refactor(channel): defer lark_* table drop to a follow-up migration
Preflight deploy review: dropping lark_* in the same release that cuts over (old migration 125) is not rollback/rolling-safe — the v0.3.27 release still reads lark_*, so a rolling deploy or a post-deploy code rollback would hit "relation does not exist". Remove the drop and keep the old tables for one release (standard expand/contract): migration 124 already backfilled lark_* -> channel_*, the new code reads/writes only channel_*, and the physical drop moves to a separate cleanup migration once this ships and is observed.
The lark_* tables remain in the schema, so sqlc regenerates the (now unused) db.Lark* models; queries/lark.sql stays deleted (the new code uses channel_*). No code path reads lark_* — only the destructive drop is deferred, keeping the design's no-compat-layer / no-dual-write rule while being deploy-safe.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* fix(channel): skip orphaned installations in hub-boot active scan
Preflight deploy review: channel_installation dropped the workspace/agent FK (MUL-3515 §4), so unlike lark_installation it does not cascade away when its workspace is deleted or its agent is hard-deleted (e.g. runtime teardown). The hub-boot query then keeps opening a WebSocket for a bot whose owner is gone.
JOIN ListActiveChannelInstallations to live workspace + agent so an orphaned installation is never connected, uniformly for every deletion path. The JOIN matches the old ON DELETE CASCADE semantics (row existence, not agent archival), so an archived-but-present agent's installation is still listed; the orphaned row's encrypted secret is thereby never decrypted/used.
Tests: a real-DB handler test asserts a deleted-workspace/agent installation and a non-Feishu one are both excluded; the lark scope test's active-list assertion moved there since the JOIN now needs real workspace/agent fixtures. (Physically deleting dormant orphaned channel rows on workspace/agent deletion is a separate app-layer-cleanup follow-up.)
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* docs(channel): document non-rolling cutover constraint for the lark->channel migration
Elon deploy review: keeping the lark_* tables (deferred drop) stops old v0.3.27 code from crashing, but is not full expand/contract. Migration 124 is a one-time backfill; afterwards new code runs on channel_* (lease + dedup on channel_*) while pre-cutover code runs on lark_* (lease + dedup on lark_*). If both run concurrently during a rolling deploy, each side claims the same Feishu bot's WS lease on its own table and double-processes inbound events.
This release therefore requires a NON-ROLLING cutover (stop the old hub before applying migration 124 + starting new code; rollback is not lossless once new code writes channel_*). Documented where deployers/reviewers see it: migration 124 header gains a ROLLOUT note; the channel_store.go header is corrected (lark_* tables are retained one release for rollback safety, not "gone"; the store still never touches them). Comment-only — no schema/codegen/behavior change.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): add MULTICA_LARK_HUB_DISABLED switch for the channel cutover
The lark_*->channel_* cutover needs a way to make the Feishu bot briefly unavailable WITHOUT taking down the whole multica-api process — the Lark hub is a goroutine inside it, not a separate Deployment. MULTICA_LARK_HUB_DISABLED=true parks the hub at startup: the API serves HTTP normally but never claims a WS lease or opens a Feishu connection.
Rollout (see migration 124 ROLLOUT note): ship the new release with the flag SET so new pods run API-only while old pods (hub on lark_*) drain during the rolling deploy — the two hubs never overlap. After the old pods are gone and migration 124 has run, flip the flag off; the new hub comes up on channel_*. The old backend does NOT need this switch — its hub stops when k8s terminates the old pods, not via a flag. Nil-ing LarkHub reuses the existing not-configured path so both the startup start and the shutdown join skip it.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* docs(channel): point migration 124 ROLLOUT note at the hub-disable switch
Refine the rollout note to use MULTICA_LARK_HUB_DISABLED for a bot-only cutover (new pods serve API with the hub parked while old pods drain; flip the switch off after the migration), instead of the earlier whole-API recreate. Comment-only.
MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
* docs(channel): fix migration 124 rollout order and document self-host cutover
The previous ROLLOUT note shipped the new (channel_*) build before
running migration 124, so the channel_*-backed HTTP paths (installation
list/install/revoke, chat-session delete, member revoke) would 500 in
the window between new-pod boot and the deferred migration. Restate the
runbook around two explicit invariants — channel_* must exist before the
new build serves those paths, and the old/new hubs must never overlap —
and order the steps so channel_* is created first (park old hub -> snapshot
-> deploy parked new build -> unpark). Document that default self-host
(entrypoint migrate + single-replica Recreate) satisfies both invariants
automatically and needs no manual steps; only prd / multi-replica rolling
self-host needs the switch procedure. Clarify in main.go that the
hub-park switch is generation-agnostic (parks whichever hub the build
carries), which is what enables the preparatory release.
Refs MUL-3515
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(autopilot): migrate scheduled dispatch to scheduler.Manager
PR 3 of 3 for the scheduled-Autopilot refactor on MUL-3551.
Replaces the legacy cmd/server/autopilot_scheduler.go goroutine
(30 s app-clock polling, app-time cron advancement, weak crash
recovery) with a JobSpec registered on the existing
scheduler.Manager. sys_cron_executions is now the lease + audit
table for scheduled Autopilot occurrences, and the unique key on
(job_name, scope_kind, scope_id, plan_time) is the primary
guarantee that the same planned fire time cannot produce two runs.
What changed
* server/internal/scheduler/jobs_autopilot.go
New AutopilotScheduleDispatchJob factory:
- scope_kind = "autopilot_trigger", scope_id = trigger.id
- PlansForScope hook (from PR 1) enumerates cron occurrences
in (lastPlan, dbNow] and collapses missed fires to the most
recent one (CatchUpLatestOnly — same policy the legacy
goroutine had, now provable via a one-row-per-tick audit).
- Handler re-loads trigger + autopilot inside the handler so a
between-tick state change (paused, disabled, deleted) takes
effect immediately and is recorded as a no-op SUCCESS row
with skipped_reason in the result JSON.
- Calls AutopilotService.DispatchAutopilotForPlan (from PR 2)
for the actual run creation; that path is itself idempotent
on (trigger_id, planned_at), so a stale-steal retry reuses
the run created by the prior attempt instead of duplicating.
- RunTimeout=2m, StaleTimeout=5m, HeartbeatInterval=30s,
AllowStaleReentry=true, MaxAttempts=3, RetryBackoff
[1m, 5m, 15m], MaxPlansPerTick=5 (safety cap).
* server/internal/scheduler/manager.go
Manager.runOnce promoted to RunOnce (exported) so external test
packages can drive deterministic ticks; existing call sites in
this package + cmd/server tests updated.
* server/internal/service/cron.go
NextOccurrenceAfterUTC and NextOccurrencesUTC: cron evaluators
that take an explicit "now" instant. Callers pass dbNow() so
schedule decisions stay consistent across app instances with
clock skew. Legacy ComputeNextRun is preserved (delegating to
NextOccurrenceAfterUTC with time.Now()) for the display-only
autopilot_trigger.next_run_at write path — scheduling decisions
no longer use it.
* server/pkg/db/queries/autopilot.sql
ListSchedulableAutopilotTriggers replaces the legacy
ClaimDueScheduleTriggers (the new path no longer mutates
autopilot_trigger.next_run_at on claim). RecoverLostTriggers
removed — sys_cron_executions lease theft now handles crash
recovery without an in-handler restart sweep.
* server/cmd/server/main.go
The "go runAutopilotScheduler(...)" line is gone. The new
JobSpec is registered alongside TaskUsageHourlyJob on the
existing schedulerMgr (still using sweepCtx for lifecycle).
* server/cmd/server/autopilot_scheduler.go DELETED.
Tests
* server/internal/service/cron_test.go — unit tests for the cron
helpers: timezone-aware enumeration, half-open (after, until]
window, plan_time-exclusive "after", invalid inputs surface
parse errors, and the "ignores wall clock" property the
scheduler relies on.
* server/cmd/server/autopilot_schedule_job_test.go — DB-backed
integration tests:
- DispatchesOnce: one tick → 1 SUCCESS exec row + 1
autopilot_run with planned_at set; a second tick does not
regress the count.
- MissedSchedulesCollapse: an hour of missed */5 fires
produce a single autopilot_run, not 12.
- CrashRecovery: simulated stale RUNNING lease at the same
plan_time → second tick reclaims it and DOES NOT duplicate
autopilot_run.
- TwoRunnersSingleWinner: two concurrent
scheduler.Manager instances on the same trigger →
per-plan_time uniqueness holds (sys_cron_executions never
has two RUNNING rows at the same plan_time, autopilot_run
count == exec row count).
- DisabledTriggerSkips: a trigger disabled between
scope-list and tick produces no exec row.
- PausedAutopilotSkipsAtHandler: an autopilot paused after
the first tick does not produce a new exec row.
- BadCronFailsLoudly: an invalid cron expression never fires
dispatch (parse error surfaces in the plan hook).
Existing autopilot listener / squad / dispatch tests still
pass.
* server/internal/scheduler/plans_for_scope_test.go from PR 1
still passes (RunOnce rename only).
Verification
* go build ./...
* go vet ./...
* go test ./internal/scheduler ./internal/service ./cmd/server
./internal/handler — all green.
Rollback
* Reverting this commit re-introduces the legacy goroutine.
Migration 124 (PR 2) and the scheduler hook (PR 1) stay in
place. Autopilot data on disk is forward- and backward-
compatible: planned_at columns are nullable, the legacy
goroutine never reads planned_at and the new job never reads
autopilot_trigger.next_run_at.
Refs MUL-3551
Co-authored-by: multica-agent <github@multica.ai>
* fix(autopilot): scheduler hook retries FAILED plans + tighten tests
Review fix for #4444 (MUL-3551).
Blocker: hook planner skipped the FAILED-with-retry plan_time
`autopilotPlansForScope` unconditionally set
`after = latest.PlanTime` when `latest.Found`, then enumerated cron
occurrences in the half-open interval `(after, dbNow]`. That
EXCLUDED the FAILED plan_time itself, so `tryClaim`'s
"FAILED-with-retry" branch — which only fires when the planner
returns the same plan_time — never ran. A claim + crash sequence
left the FAILED row stuck at attempt<max_attempts forever and the
scheduled occurrence was lost (MUL-3551 acceptance ③).
Fix: hook now branches on `latest.RetryEligible(now)` BEFORE
computing `after`. When the most recent stored row is FAILED with
attempts remaining and next_retry_at <= dbNow, the hook returns
`[latest.PlanTime]` unchanged. tryClaim's retry-from-FAILED path
fires, attempt increments, the run is retried, and the audit row
reaches SUCCESS at the same plan_time. Mirrors the cadence
planner's `info.RetryEligible(now)` branch in manager.plansForTick.
Tests
* TestAutopilotScheduleJobCrashRecovery rewritten to actually
pin the retry contract instead of just "no duplicate run":
- assert first attempt completes at attempt=1 with a real
task_id linkage (the "complete" snapshot the retry must
reuse);
- simulate a crash mid-dispatch (status=RUNNING, expired
stale_after, ghost lease_token);
- assert tick 2 transitions the SAME exec row (same plan_time)
to status=SUCCESS at attempt=2 (proving the planner did
NOT skip past the FAILED bucket);
- assert autopilot_run stays at exactly one row, reused from
the first attempt — proving DispatchAutopilotForPlan's
complete-run reuse path is what closes the loop.
* TestAutopilotScheduleJobPausedAutopilotSkipsAtHandler rewritten
to invoke `job.Handler` directly (the previous version drove
`mgr.RunOnce` which short-circuited at the scope-list SQL
filter and never reached the handler). The new test pauses the
autopilot AFTER setup, calls the handler with a fabricated
HandlerInput, and asserts the handler returns
skipped_reason=autopilot_inactive without creating an
autopilot_run.
* TestAutopilotScheduleJobBadCronFailsLoudly renamed to
TestAutopilotScheduleJobBadCronStaysSilent and updated to
match the real implementation: a parse error in the plan hook
surfaces as a manager-level warning log, NOT a
sys_cron_executions row (no plan_time was ever claimed). The
test now asserts zero exec rows AND zero autopilot_run rows,
documenting that bad cron is a permanent configuration error
(caught at HTTP create/update time first), not a transient
failure that belongs in the retry envelope.
Refs MUL-3551
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(scheduler): add PlansForScope hook for non-cadence jobs
The current Manager.plansForTick assumes a uniform Cadence grid:
plan_times are derived via FloorPlan(db_now, Cadence). That works for
rollup_task_usage_hourly but not for the upcoming Autopilot schedule
dispatch job, where each trigger has its own cron expression and the
plan_times do not snap to a single global grid.
This change adds an optional JobSpec.PlansForScope hook. When set:
* Manager loads the latest stored plan for (job, scope) and passes
a new LatestPlanInfo to the hook (exported from the previously
private latestPlanInfo). The hook returns the plan_times to attempt
this tick.
* Cadence, CatchUpMode and CatchUpWindow are bypassed; the hook is
in full control of plan_time selection.
* MaxPlansPerTick still acts as a safety cap on the hook's output.
* All other timing fields (RunTimeout / StaleTimeout /
HeartbeatInterval / MaxAttempts / RetryBackoff / AllowStaleReentry)
and the lease/heartbeat/terminal-write SQL primitives are reused
unchanged.
JobSpec.validate now allows Cadence=0 when PlansForScope is set, and
makes the every_plan MaxPlansPerTick > 0 invariant fire only on
Cadence-driven every_plan jobs. Existing rollup_task_usage_hourly
behaviour is unchanged — that JobSpec leaves PlansForScope nil.
Tests:
* TestJobSpecValidatePlansForScopeRelaxesCadence — validate() rules.
* TestManagerPlansForScopeHookDrivesPlans — end-to-end hook delegation
through the manager (DB-backed), proving that hook-returned
plan_times go through the same tryClaim path, MaxPlansPerTick
truncates without erroring, and LatestPlanInfo is populated on the
second tick.
* TestManagerPlansForScopeHookEmptyIsNoOp — empty hook output is a
valid no-op.
No behaviour change for callers that don't set PlansForScope.
Refs MUL-3551
Co-authored-by: multica-agent <github@multica.ai>
* refactor(autopilot): add planned_at + DispatchAutopilotForPlan for occurrence idempotency
PR 2 of 3 for the scheduled-Autopilot refactor on MUL-3551.
Adds dispatch-layer idempotency for scheduled triggers. This is the
second line of defence behind the primary uq_sys_cron_execution
guarantee in sys_cron_executions: if a runner crashes between
"create autopilot_run" and "write SUCCESS in sys_cron_executions",
the next stale-steal retry re-enters dispatch with the SAME
(trigger_id, planned_at). Without a row-level guard, that retry
would create a duplicate autopilot_run, issue, and task.
Changes:
* Migration 124: ALTER TABLE autopilot_run ADD COLUMN planned_at
TIMESTAMPTZ + partial unique index on (trigger_id, planned_at)
WHERE both are NOT NULL. Manual / webhook / api dispatch leaves
planned_at NULL so they keep the existing semantics unchanged.
* autopilot.sql: CreateAutopilotRun now takes planned_at;
GetAutopilotRunByTriggerAndPlanned is the fast-path lookup used
by DispatchAutopilotForPlan to detect a prior attempt's row
without burning an INSERT.
* service.DispatchAutopilotForPlan: new entry point for scheduled
triggers that already know the canonical UTC plan_time of the
occurrence they are firing. Looks up an existing run for
(trigger_id, planned_at) and reuses it on a stale-steal retry;
otherwise dispatches normally with planned_at stamped on the
new run.
* service.DispatchAutopilot keeps its current signature for
manual / webhook / api callers (planned_at stays NULL).
* recordSkippedRun also threads planned_at so the skip path
participates in the same partial-unique guarantee.
* sqlc v1.31.1 regenerated autopilot.sql.go + models.go.
Unrelated workspace.sql.go drift restored.
Tests (against local Postgres):
* TestDispatchAutopilotForPlanIsIdempotent — first call creates a
run; second call with same (trigger, planned_at) reuses it
(autopilot_run row count stays at 1); third call with a different
planned_at on the same trigger creates a second run (proves we
are not collapsing legitimate occurrences).
* TestDispatchAutopilotForPlanRejectsZeroArgs — invalid trigger_id
and zero planned_at both fail loudly so callers cannot silently
disable the idempotency guard.
* Existing autopilot listener / squad / dispatch tests all still
pass.
This PR has no scheduler / handler / UI behaviour change on its own:
the new entry point exists but is not yet wired into the schedule
goroutine. PR 3 will register the autopilot_schedule_dispatch
JobSpec that consumes it and remove the legacy
cmd/server/autopilot_scheduler.go path.
Refs MUL-3551
Co-authored-by: multica-agent <github@multica.ai>
* fix(autopilot): DispatchAutopilotForPlan recovers partial-state runs
Review fix for #4443 (MUL-3551).
Before this change, DispatchAutopilotForPlan returned ANY existing
autopilot_run for (trigger_id, planned_at), including the
half-written rows produced when a runner crashed between
"CreateAutopilotRun" and "create downstream issue/task". The
scheduler handler would then write SUCCESS in sys_cron_executions
even though no issue or agent task was ever created, silently
losing the scheduled occurrence.
Fix:
* New isAutopilotRunComplete helper classifies an existing run:
- terminal status (completed / failed / skipped) → reuse.
- issue_created with valid issue_id → reuse (the issue
listener owns task creation from here).
- running with valid task_id → reuse (the task is queued).
- anything else → partial; must NOT short-circuit.
* New SQL RecoverPartialAutopilotRun marks a partial row FAILED
with a recovery reason AND clears its planned_at. The cleared
planned_at releases the partial-unique slot in
uq_autopilot_run_trigger_planned, letting the fresh dispatch
INSERT a new row at the same (trigger_id, planned_at) without
conflict.
* DispatchAutopilotForPlan now branches on the lookup:
complete run → return; partial run → recover-then-fresh-
dispatch; not-found → fresh dispatch. The fresh dispatch path
still goes through dispatchAutopilot, so the new row carries
the real issue_id / task_id by the time the handler returns.
* Tests: TestDispatchAutopilotForPlanRecoversPartialRun seeds a
partial run (status='running', task_id=NULL for run_only;
status='issue_created', issue_id=NULL for create_issue) and
asserts the retry:
- returns a DIFFERENT run row (no false reuse);
- leaves the partial row in status='failed', planned_at=NULL,
with a non-empty failure_reason for ops;
- produces a fresh row with planned_at preserved AND the
appropriate downstream linkage (task_id for run_only,
issue_id for create_issue);
- exactly one live row at (trigger_id, planned_at) after
recovery, so the partial-unique constraint is honoured.
Existing TestDispatchAutopilotForPlanIsIdempotent and
TestDispatchAutopilotForPlanRejectsZeroArgs still pass — the
complete-reuse path is unchanged for the realistic SUCCESS-state
case.
Refs MUL-3551
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
The current Manager.plansForTick assumes a uniform Cadence grid:
plan_times are derived via FloorPlan(db_now, Cadence). That works for
rollup_task_usage_hourly but not for the upcoming Autopilot schedule
dispatch job, where each trigger has its own cron expression and the
plan_times do not snap to a single global grid.
This change adds an optional JobSpec.PlansForScope hook. When set:
* Manager loads the latest stored plan for (job, scope) and passes
a new LatestPlanInfo to the hook (exported from the previously
private latestPlanInfo). The hook returns the plan_times to attempt
this tick.
* Cadence, CatchUpMode and CatchUpWindow are bypassed; the hook is
in full control of plan_time selection.
* MaxPlansPerTick still acts as a safety cap on the hook's output.
* All other timing fields (RunTimeout / StaleTimeout /
HeartbeatInterval / MaxAttempts / RetryBackoff / AllowStaleReentry)
and the lease/heartbeat/terminal-write SQL primitives are reused
unchanged.
JobSpec.validate now allows Cadence=0 when PlansForScope is set, and
makes the every_plan MaxPlansPerTick > 0 invariant fire only on
Cadence-driven every_plan jobs. Existing rollup_task_usage_hourly
behaviour is unchanged — that JobSpec leaves PlansForScope nil.
Tests:
* TestJobSpecValidatePlansForScopeRelaxesCadence — validate() rules.
* TestManagerPlansForScopeHookDrivesPlans — end-to-end hook delegation
through the manager (DB-backed), proving that hook-returned
plan_times go through the same tryClaim path, MaxPlansPerTick
truncates without erroring, and LatestPlanInfo is populated on the
second tick.
* TestManagerPlansForScopeHookEmptyIsNoOp — empty hook output is a
valid no-op.
No behaviour change for callers that don't set PlansForScope.
Refs MUL-3551
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
The @mention popup tracked the highlighted row with a positional integer
(selectedIndex into displayItems), but rows are rendered in the re-bucketed
order produced by groupItems() (current → recent → search → users → issues).
An async server "search" result is appended to the END of displayItems yet
hoisted near the TOP on render, so the highlighted row and the committed
item pointed at different entries — you navigate to one target but mention
its neighbour. A separate useEffect also force-reset selectedIndex to 0 on
every displayItems change, snapping an active selection back to the first row
whenever async results landed.
Root cause: a slot index is not a stable target for a list whose order and
length change asynchronously. Track selection by item identity instead:
- Replace selectedIndex state with selectedKey (the item's `type:id`).
- Derive groups/orderedItems (the exact rendered order) and resolve the
numeric index from selectedKey against orderedItems; fall back to row 0
when the pinned item is gone or nothing is picked yet.
- Keyboard nav, Enter, clicks, highlight, and scroll all index orderedItems,
so the highlighted row always equals the committed item.
- Drop the force-reset effect; identity-based selection self-heals across
reorders and async arrival without resetting an active pick.
Adds a regression test asserting the highlighted row equals the committed
item when groupItems reorders the list, both initially and after ArrowDown.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The @mention popup navigated and committed by indexing the flat
`displayItems` array, but rendered rows in the re-bucketed order from
`groupItems()` (current → recent → search → users → issues). In chat
(context mode) the async server-search results are appended to the end
of `displayItems` yet tagged `group:"search"`, so `groupItems()` hoists
them near the top. The highlighted row and the committed item then point
at different entries — you select one target but mention its neighbour
(the reported "@bohan picks the next one" off-by-one).
Make the flattened grouped order (`orderedItems`) the single index space
for `selectedIndex`, arrow keys, Enter, and clicks, so the highlighted
row is always the committed item. Plain issue-comment mentions (default
mode) were already safe — no group tags means `groupItems()` is
order-preserving — and stay unchanged.
Adds a regression test asserting highlighted row == committed item when
the list is reordered by a hoisted search result.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(comments): resolve-aware fold for agent comment reads (MUL-3555)
Agents reading a long issue paid tokens for settled discussion. The human
timeline already folds resolved threads, but the agent read path
(`comment list`) ignored resolved_at entirely — humans saw the conclusion,
agents got the full raw discussion.
Add an opt-in `fold=true` projection to ListComments that collapses each
resolved thread to root + conclusion (reply-resolved) or root only
(root-resolved), reusing the human timeline's deriveThreadResolution
semantics. The resolved thread's root carries `thread_resolved` +
`folded_count`; `--full` brings the dropped comments back. Fold is rejected
on partial-thread reads (since/tail) and roots_only, where a resolution
comment could be unfetched and silently dropped.
CLI `comment list` folds by default on the complete-thread reads (default,
--recent, untailed --thread) with a `--full` escape hatch; the agent
prompts and runtime brief document the fold + escape. No new endpoint, no
human UI change, no SQL/migration change — in-memory projection, same
precedent as summary/roots_only.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(daemon): dedupe fold prompt restatements per review (MUL-3555)
Howard's PR review flagged DRY redundancy: the resolve-fold rule was
restated in full in the task prompt (prompt.go:41/:182) and the brief
workflow steps (runtime_config.go:673/:692, reply_instructions cold
hint) even though the canonical command catalog (runtime_config.go:477)
— always present in the brief — already documents it in full, and the
task prompt explicitly defers to it ("follow the rule in your runtime
workflow file").
Keep the catalog entry full (the canonical reference); shrink the five
inline restatements to a short "resolved threads come back folded —
`--full` to expand" pointer. No loss of signal (the agent always has the
full catalog in context), ~80-120 tokens/run saved on the worst-case
assignment / cold paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
Cancelling a chat task restored the message into the input via a
separate `editorRestore` component state that also held a private copy
of the content and forced an editor remount (bumping `editorKey`).
That copy was never cleared and `activeRestore` re-derived on every
return to the draft's session, so `defaultValue={activeRestore?.content
?? inputDraft}` kept re-injecting text the user had already deleted —
the draft (`inputDraft`) cleared, but the stale copy did not.
The editor already self-syncs external `defaultValue` changes into a
live instance (content-editor.tsx defaultValue-sync effect, used for WS
description updates), so the remount mechanism was redundant. Drop the
whole `editorRestore` state and let `inputDraft` be the single source of
truth: restore just writes the draft, the editor's own sync displays it.
Now cancel-restore behaves exactly like normal typing — delete the text
and it stays gone across navigation.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
agy's --print-timeout defaults to 5m when the flag is omitted, but the
daemon treated "omit the flag" as "no cap". In the default no-cap config
every Antigravity turn was therefore silently capped at 5 minutes: any
run whose build/tests outlived the budget had agy abort mid-turn, print
"Error: timed out waiting for response", and exit 0 — which the backend
recorded as a successful "completed" with truncated output (the reported
"Antigravity disconnects", MUL-3570 / #4453).
- Always pass --print-timeout: the configured cap when positive, else a
large value (24h) that defers to the daemon's idle/tool watchdogs.
- Detect agy's print-mode timeout marker in the run log and surface the
result as a timeout instead of a truncated success.
Verified by reproducing against agy 1.0.8 and with new unit + end-to-end
backend tests.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* docs: add 2026-06-23 changelog entry
Co-authored-by: multica-agent <github@multica.ai>
* docs(changelog): sharpen 0.3.28 title and handoff wording
- Headline the two flagship features in the title: staged Sub-Issues and Qoder runtime support
- Rewrite the vague agent-handoff line to spell out the pre-trigger confirmation (whether/which agent will start, apply without running) and the handoff note as next-run context
- Apply across en/ja/ko/zh
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: J <j@multica.ai>
The transcript dialog opened from a running task's row showed only a
one-shot snapshot taken at open time: TranscriptButton fetched once via
api.listTaskMessages and cached it locally, never subscribing to the
shared ["task-messages", taskId] cache that the WS task:message stream
already seeds. New tool calls / thinking / text never appeared until the
task finished or the page reloaded.
Add a live-cache mode to the shared TranscriptButton: when isLive and the
parent provides no items and the task id is a persisted UUID, render from
the shared task-messages cache so the open dialog grows in real time. On
open (and again on the running→terminal transition) force a backfill via
api.listTaskMessages and merge it into the cache by seq — taskMessagesOptions
is staleTime:Infinity, so a plain subscription never heals a WS reconnect
gap. The cache observer is read-only (enabled:false) so React Query never
blind-replaces the cache; only the WS handler and the seq-merged backfill
write it. The subscription mounts only while the dialog is open, so closed
live rows add no baseline requests; terminal tasks keep the lazy one-shot
fetch.
Covers issue execution-log and agent activity. Autopilot issue-less
run_only live log is out of scope: the backend doesn't broadcast
task:message for tasks with no issue/chat session, so there's nothing to
subscribe to — backend broadcast unchanged.
Extract mergeTaskMessagesBySeq into packages/core/chat/queries.ts and route
both the realtime task:message handler and the new backfill through it, so
there is one seq-merge semantics for that cache instead of two.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): stop issue-trigger preview flicker
The pre-trigger preview re-rendered/refetched on every workspace task
event: WS task lifecycle invalidated issueTriggerPreviewAll (staleTime 0),
forcing a background refetch whose isFetching was surfaced as isLoading,
collapsing and reopening CreateRunHint's reveal band.
The assign source (create / assignee change) cancels existing tasks before
enqueuing, so its verdict can't shift from a task event at all; the status
source's pending dedup could, but the preview is advisory and the write
path re-evaluates authoritatively, so a rare stale label is harmless. Drop
the WS invalidation so the preview refetches only on input (signature)
change. Keep the comment-trigger invalidation — its verdict genuinely
changes mid-compose and its chips drive an immediate, unconfirmed send.
Align the hook's data handling with the comment-trigger preview:
keepPreviousData so an input switch swaps in place instead of collapsing,
and treat only the first load (no prior data) as loading.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(issues): skip run-confirm modal for backlog assign
Assigning a Backlog issue to an agent/squad never starts a run (the
parking lot — server/internal/service/issue_trigger.go), so the
pre-trigger confirm modal only rendered an empty "won't start" box with
a single Apply button. Apply directly instead: the single path checks
issue.status, the batch path skips only when every selected issue is
Backlog (mixed selections still confirm — the non-backlog ones trigger).
Mirrors the existing backlog short-circuit in handleBatchStatus.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(modals): run-confirm loading state + submit spinner
The dialog grew in height after open: it rendered the short "won't
start" variant while POST /api/issues/preview-trigger was in flight, then
the note box appeared when the predicate landed. Keep the note box
mounted (disabled) during loading so assign mode opens at its resolved
height, and show a Spinner + 'checking' headline while loading.
Submit had no feedback — buttons only disabled, which read as frozen for
note assigns (the request starts an agent server-side). Track which
footer action is in flight and show a Spinner on the clicked button.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(issues): show handoff note in execution-log trigger text
An assignment-triggered run that carried a handoff note showed the
generic "Initial run" label. Surface the note inline (truncated, like
comment triggers show their text) so the row reads as the handoff.
taskToResponse now populates handoff_note for all callers (dropping the
now-redundant explicit set in ClaimTaskByRuntime); the field is added to
the AgentTask type + zod schema (optional, additive — old clients ignore
it via the loose schema, new clients fall back to "Initial run").
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(landing): show live GitHub star count on the header GitHub button
Add a small client hook (useGithubStars) that fetches stargazers_count
from the GitHub API and a formatStarCount helper that renders it in
GitHub's compact repo-header style (e.g. "37.6k"). The landing header's
GitHub button now appends a star badge (faint divider + filled star +
count) on both the desktop and mobile menu entries.
Fetched client-side on purpose: LandingHeader is shared across every
marketing page, so one client fetch covers them all without threading a
server value through each render site, and each visitor calls the API
from their own IP, sidestepping the shared-outbound-IP rate limit the
server-side github-release fetcher works around with a PAT. The result
is memoized at module scope (plus in-flight dedupe); a failed fetch
caches null and the button degrades to the plain "GitHub" label.
* fix(landing): drop the star glyph from the GitHub star badge
In the GitHub button context the number already reads as the star count,
so the icon is redundant. Keep the divider + count only.
`GET /api/issues/:id/usage` endpoint (handler `GetIssueUsage`). Returns the
aggregated token usage for an issue, summed across all of its task runs.
The usage is already captured server-side and shown in the issue detail view,
but was not reachable from the CLI, so it couldn't feed billing/cost scripts.
This closes that gap with no backend changes.
$ multica issue usage MUL-139
INPUT_TOKENS OUTPUT_TOKENS CACHE_READ CACHE_WRITE RUNS
5625 83880 9190806 154078 1
$ multica issue usage MUL-139 --output json
{ "total_input_tokens": 5625, ... "task_count": 1 }
Accepts an issue key (`MUL-139`) or UUID, mirroring `issue runs`.
- Table cells use the existing `formatMetadataValue` helper so large cache-token
counts render as plain integers (not scientific notation).
- Scope is the per-issue aggregate the endpoint already returns (`task_count` =
run count); a per-run token breakdown is out of scope.
- `go test ./cmd/multica/ -run TestRunIssueUsage` (added) ✅
- `go vet ./cmd/multica/` ✅
- Verified against a live self-hosted server; numbers match the issue UI.
- `server/cmd/multica/cmd_issue.go` — command + handler
- `server/cmd/multica/cmd_issue_test.go` — unit test
- `CLI_AND_DAEMON.md` — docs
* feat(issues): unify run-enqueue decision behind WillEnqueueRun + preview endpoint
Collapse the issue update/batch enqueue copies into one service predicate
service.IssueService.WillEnqueueRun, shared verbatim with a new dry-run
endpoint POST /api/issues/preview-trigger so the four entry points stop
drifting (squad/self-loop/batch omissions, MUL-3375). The private-agent gate
stays at the HTTP boundary: write paths inject allow-all, preview injects the
real gate so it never leaks a private agent's readiness.
Add suppress_run to issue update/batch: the change applies but no run starts.
Remove the now-dead handler mirrors shouldEnqueueSquadLeaderOnAssign /
isSquadLeaderReady. service.Create and the comment trigger chain are untouched.
Tests: preview behavior, preview<->write-path match, batch aggregation,
member no-trigger, suppress_run skip, malformed-body 400.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): inject handoff note into assigned runs via first-class task field
Add an optional handoff_note carried by issue assign/promote into the run's
opening prompt and issue_context.md, via a dedicated agent_task_queue column
(migration 122) and a daemon assignment-handoff render branch — never a
fabricated comment, never trigger_comment_id (MUL-3375 §6.1).
Thread the note through enqueueIssueTask/enqueueMentionTask + WithHandoff
public variants and dispatchIssueRun; suppress_run or a parked write drops it
(no run = nothing to inject). Soft version gate: MinHandoffCLIVersion +
HandoffSupported, surfaced per-trigger as handoff_supported in the preview so
the UI can gray the note box on old daemons; the assignment never hard-fails.
Tests: daemon prompt + issue_context render via the assignment branch (not
quick-create/comment), version helper matrix, note persists on the task,
suppressed assign enqueues nothing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): leave a display-only handoff record on the timeline
When an assign/promote with a handoff note starts a run, write one
type='handoff' timeline record via TaskService.RecordHandoff — a direct
Queries.CreateComment + timeline event that bypasses Handler.CreateComment, so
it never reaches triggerTasksForComment and cannot start a second run
(MUL-3375 §6.2, the must-not-retrigger invariant). Author is the actor who
handed off; body is the note. Migration 123 admits the 'handoff' comment type.
Recorded only on a real run start: suppress_run or a parked write writes
nothing. enqueueSquadLeaderTask now reports whether it enqueued so the trace
is gated on an actual dispatch.
Test: exactly one handoff record on assign-with-note, exactly one task (no
re-trigger), and no record when suppressed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): frontend plumbing for issue-trigger preview + handoff (core)
Add api.previewIssueTrigger + IssueTriggerPreviewSchema (zod parseWithFallback),
the use-issue-trigger-preview hook, issueKeys.issueTriggerPreview(+All) with WS
queue-state invalidation, suppress_run/handoff_note on UpdateIssueRequest, the
'handoff' CommentType, and stripping of the control fields from optimistic
update/batch cache patches (MUL-3375 §9).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): exclude handoff records from new-comment counting
type='handoff' is a display-only timeline record, not conversation. Exclude it
from CountNewCommentsSince so a handoff note never inflates the count of
"new comments to catch up on" fed to a claiming agent (MUL-3375 §12). Analytics
already excludes it (RecordHandoff is a direct write that emits no analytics
event), and the comment-trigger path is already bypassed.
Test: a handoff record does not bump the new-comment count; a real comment does.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): pre-trigger preview UI, handoff note, timeline card (web/desktop)
Wire the §9 frontend onto the preview endpoint + handoff fields:
- Delete the backlog blocking dialog (backlog-agent-hint*) and its modal type;
the over-eager nag is gone. Backlog awareness is now a passive label.
- RunConfirmModal: single assign + batch assign/status route here. Shows the
backend predicate's verdict ("将启动 @X" / "将启动 N 个" / parked), an optional
handoff note (assign only, soft-gated by handoff_supported), and 暂不启动 —
then applies via update/batch. No frontend guessing.
- create modal: passive CreateRunHint ("将启动 @X" / backlog parked).
- single status change stays a direct apply (unchanged).
- timeline: render type='handoff' as a distinct, non-interactive handoff card.
- i18n run_confirm + handoff_card across en/ja/ko/zh-Hans; drop backlog action
keys; locale parity green.
Tests: use-issue-actions (assign → run-confirm modal, member → direct),
create-issue + comment-card suites updated/green; views typecheck + lint clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(issues): use a valid anchor in the handoff count-exclusion test
CountNewCommentsSince filters id <> @anchor_id; SQL id <> NULL is NULL and
excludes every row, so an empty anchor made the control assertion read 0. The
production caller always passes a real anchor — mirror that with a non-matching
sentinel uuid.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(issues): RunConfirmModal apply logic (start/suppress/note-gate/batch)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(core): preview schema malformed/missing/null fallback coverage
Cover IssueTriggerPreviewSchema via parseWithFallback (MUL-3375): well-formed
parse, top-level + item default fills (empty/older backend), and fallback to
{ triggers: [], total_count: 0 } for malformed shapes, a dropped required
issue_id, a wrong-typed total_count, and null/non-object bodies — so the four
entry points degrade to "nothing will start" instead of throwing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(issues): remove display-only handoff timeline record (留痕)
The handoff "留痕" timeline record (type='handoff' comment written on run
start) was judged superfluous and dropped per product call. This removes
only the display-only trace; the handoff NOTE injection into the run's
opening prompt + issue_context.md is untouched.
- backend: drop RecordHandoff + its call in dispatchIssueRun
- db: drop the `type <> 'handoff'` exclusion in CountNewCommentsSince and
migration 123 (comment_type_check reverts to the 4-type set from 001);
no production data exists for this unreleased feature
- frontend: drop the "handoff" CommentType, HandoffCard, and handoff_card
i18n (all locales)
- tests: drop handoff_count_test.go and the record-write assertions in
issue_trigger_preview_test.go (note-injection tests retained)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): dismissable run-confirm modal + team-handoff copy
Two fixes to the pre-trigger confirm modal (MUL-3375).
1. Dismissable: switch RunConfirmModal from AlertDialog to the standard
shadcn Dialog so it has the close (X) button + Esc + click-outside.
Previously the only choices were "start" / "don't start now" with no
way to abort the action entirely; dismissing now cancels with no write.
2. Copy: rework the action-surface wording away from the backend term
"run" toward team-handoff voice — 指派 / 开始 / 交接 (run stays only on
record surfaces). Unifies the note's three names to "交接说明", and
parallels the rewrite across en/ja/ko.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* chore(agent): bump handoff note min CLI version to 0.3.28
The daemon release that renders handoff notes ships in 0.3.28 (0.3.27
was the prior tag), so move the soft-gate threshold up. Below this the
note is silently dropped and the frontend grays the note box — assignment
is never blocked.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(issues): skip run-confirm when batch-moving issues to backlog
A move into backlog never starts a run (service/issue_trigger.go), so the
pre-trigger confirm modal degenerated to an empty "won't start" box with a
single Apply button — pure friction. Apply directly instead, matching the
single-issue status path. Other target statuses still route through the
modal.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(issues): refine pre-trigger preview hint and copy
- Move the create-issue run hint to a reveal band (grid 0fr→1fr) above the
property toolbar. It was sharing the footer button row and, lacking a
width constraint, reflowed the submit buttons whenever it appeared.
Restyle to a borderless, comment-style avatar+caption that is purely a
caption (non-interactive avatar).
- Distinguish squad from agent in the pre-trigger copy: a squad's leader
evaluates and delegates rather than "starting work" itself. Add
will_start_named_squad / will_start_squad / create_will_start_squad across
en/zh/ja/ko (reusing the squad_leader_* evaluate→arrange vocabulary) and
branch run-confirm + the create hint on squad assignees.
- Bold the assignee name in the run-confirm headline via a language-safe
sentinel split (no per-language prefix/suffix keys).
- Align zh "开始处理" → "开始工作" on the single-assign copy.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(issues): stub ActorAvatar in create-issue suite
CreateRunHint now renders an ActorAvatar for agent/squad assignees, which
pulls in getActorInitials/getActorAvatarUrl + the workspace/presence/navigation
hook tree. This form-focused suite only stubbed getActorName, so the
squad-forwarding test crashed with "getActorInitials is not a function". Stub
the avatar inert — its own behavior is covered elsewhere.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Walt <walt@multica.ai>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): stop kanban card snapping back on drag
A cross-column drag on a non-position-sorted board left the card in its
origin column for the whole request, then jumped to the target only when
the mutation settled — the "snaps back, then moves" glitch. Root cause was
three coupled choices in the optimistic path:
- board-view never updated local columns on drop for sortBy != "position"
(onDragOver is a no-op there), so the card relied on the settle refetch
to move across.
- useUpdateIssue invalidated the whole list on settle, replacing the column
and re-landing the card even on success.
- patchIssueInBuckets appended a moved card to the column tail instead of
its position slot, so any later cache refresh teleported it to the end.
Fixes:
- board-view: optimistically move the card into the target column on drop
for the non-position path (insertIdByPosition), and reconcile local
columns from the cache on settle for both paths (revert on error now that
the list is no longer refetched).
- mutations: reconcile via onSuccess surgical patch of the returned entity;
drop the list/detail invalidation from onSettled (aggregates still flush).
- cache-helpers: patchIssueInBuckets inserts the moved/reordered card at its
position slot; a plain field update still keeps its slot.
Adds cache-helpers and drag-utils unit tests.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): patch My-Issues / Project board caches on move too
The drag fix made the board reconcile local columns from its feeding cache
on settle. The workspace board rides issueKeys.list (patched by onMutate),
but the My-Issues and Project boards ride the myList cache, which the
mutation did not patch — so a successful move snapped back on those boards.
useUpdateIssue now patches/snapshots/rolls back every bucketed list cache
(workspace list + myList), selected by the ListIssuesCache `byStatus` shape
so grouped (assignee) and flat (gantt) caches are skipped. Adds renderHook
regression tests covering both-cache optimistic move, both-cache rollback,
and no-list-invalidation-on-settle.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): drop redundant WS position->list invalidate
onIssueUpdated already surgically patches the non-filtered workspace board
via patchIssueInBuckets (cross-status move + same-column reorder). The extra
`if (position) invalidateQueries(list)` re-pulled the whole board on top of
that, re-introducing drag flicker through the echoed-back WS event. Removed.
Filtered myAll lists still invalidate (membership can change there) — the
client-side membership reconciliation for those is a separate follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): batch update patches myList + stops list refetch on settle
- onMutate now patches both issueKeys.list and the filtered issueKeys.myAll
bucketed caches, so a batch edit on a My-Issues / Project board is
optimistic too. Previously only the workspace board was patched, so batch
edits on those boards relied entirely on the settle refetch.
- onSettled no longer invalidates issueKeys.list: the optimistic patch is a
complete reconcile for these bucketed boards (batch changes status /
priority / project, never a server-computed value), so a full-board
refetch only re-introduced the flicker the single-issue path removed.
Aggregate / grouped caches still refresh.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): list view optimistically moves row on non-position drag
The sortBy != "position" branch called onMoveIssue without moving the row in
local columns, so the row sat in its origin group for the whole request and
only jumped across on settle -- the same snap-back the board view had before
its fix. Now mirrors board-view: setColumns(insertIdByPosition) on drop so
the settle rebuild is a visual no-op.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): keep My-Issues/Project boards in place on non-membership WS change
onIssueUpdated now surgically patches the filtered myList (myAll) caches and
only invalidates them when the change can actually move an issue in/out of the
filter: an assignee change (covers My-Issues direct-assignee + the involves leg
+ actor panels) or a project change (Project board). A pure status / position /
priority / label change reconciles in place -- no refetch -- removing the last
drag flicker on filtered boards.
Uses the assignee_changed flag the server already sends on issue:updated
(surfaced on IssueUpdatedPayload + forwarded by the realtime dispatch); project
change is diffed client-side against the cached value. No predicate replication,
no backend change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): add settle-lock to swimlane drag (no clobber mid-flight)
The swimlane drag had no settle window: the resync useEffect (and the issueMap
freeze) guarded only isDraggingRef, so a cache change landing after drop but
before the move settled could rebuild localCells out from under the optimistic
move. Adds isSettlingRef + settleVersion (mirroring board-view / list-view): the
lock is held from drop until onMoveIssue settles, then released, forcing a
single resync from the reconciled cache.
onMoveIssue now accepts the same optional onSettled callback board/list already
use; the parent handleMoveIssue supplies it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(issues): extract shared useDragSettle hook for board + list
board-view and list-view carried byte-identical drag/settle scaffolding (the
local columns mirror, the dragging/settling locks, the post-move animation-frame
throttle, and the settle callback). That duplication is exactly what let
list-view silently drift earlier (it had lost the optimistic-move half of the
fix, and its position-branch settle callback omitted the settleVersion bump).
Extract the primitive into useDragSettle so both surfaces share one
implementation and can't drift again.
Behavior-preserving for board-view. For list-view the one intended alignment:
its position-branch failed move now reverts, gaining the settleVersion bump
board-view already had.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
Agents creating sub-issues only saw the runtime brief's Sub-issue
Creation section, which taught the manual todo/backlog serial chain and
never mentioned stages — the `--stage` flow was documented only in the
multica-working-on-issues skill, which an agent reads only if it opens
it. So agents defaulted to hand-managed backlog chains and rarely reached
for stages.
- Add an "Ordering with stages" paragraph to the brief's Sub-issue
Creation section nudging agents to group ordered/waiting sub-issues
with --stage instead of hand-promoting a backlog chain.
- List --stage on the brief's issue create / update command lines and
add multica issue children to the Core command list for discoverability.
- Extend the brief test with the new stage assertions.
The Sub-issue Creation section stays gated to issue-bound runs (skipped
for chat/quick-create/autopilot), unconditional on parent_issue_id, and
free of parent-notification guidance — all existing canaries still pass.
MUL-3508
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Replace the # (Hash) icon for the Stage property with the Milestone icon
across the picker trigger, dropdown option rows, and the Add-property menu.
Shrink the Stage dropdown option font to text-xs (scoped to the Stage
picker only; the shared PickerItem keeps text-sm so other property
dropdowns are unaffected).
* feat(issues): stage sub-issues so the parent wakes per stage, not per child
Sub-issues under a parent can be grouped into ordered stages (issue.stage).
The child-done -> parent notification + assignee wake now fire only when a
stage barrier closes: every sub-issue in the lowest unfinished stage has
reached a terminal status (done/cancelled). An unstaged sibling set is one
implicit stage, so the parent is woken once when the last sub-issue finishes
instead of on every child — the default fix for the fire-on-every-child
cascade reported in discussion #4320 / MUL-3508.
Stage advancement stays agent-driven: the server only detects the closed
barrier and wakes the parent assignee, who decides whether to promote the
next stage.
- DB: nullable issue.stage (CHECK >= 1) + sqlc regen
- API: stage on issue create/update/response and batch update
- CLI: `issue create`/`issue update` --stage; new `issue children` command
that lists sub-issues grouped by stage (table + json)
- stageBarrierClosed / stageProgressSummary in issue_child_done.go, with the
wake comment now stage-aware, plus unit tests
- skill docs (multica-working-on-issues SKILL.md + source map)
Web UI (create-form stage picker, sidebar edit, group-by-stage display) is a
follow-up; the API already returns stage for it to consume.
MUL-3508
Co-authored-by: multica-agent <github@multica.ai>
* fix(issues): address review on stage barrier (cancel, batch, unstaged)
Resolves the three blockers from the PR review:
1. Cancel can close a stage. The child-done barrier now fires on any
non-terminal -> terminal transition (done OR cancelled), not just done.
isTerminalChildStatus already treats cancelled as terminal (a cancelled
sibling never finishes, so it must not hold a stage open), so a cancelled
last-open child now closes its stage and wakes the parent. Keying on the
transition also makes a later cancelled -> done edit a no-op, avoiding a
lagging duplicate wake.
2. Batch update of stage no longer no-ops. `hasMutation` now includes
"stage", so `{"updates":{"stage":N}}` persists instead of returning
{"updated": 0}.
3. Unstaged children no longer participate in the staged frontier. In a
staged sibling set, NULL-stage children neither hold a stage open nor fire
on their own completion, and the wake comment no longer renders "Stage 0".
This matches migration 123 ("NULL does not participate in staged
grouping") and the CLI's separate unstaged group, removing the footgun
where an unstaged backlog child silently blocked Stage 1.
Tests: cancellation closes a stage (staged + unstaged), unstaged ignored in a
staged set, stage summary skips unstaged, and a stage-only batch update
persists.
MUL-3508
Co-authored-by: multica-agent <github@multica.ai>
* feat(web): stage UI — create picker, sidebar edit, group sub-issues by stage
Frontend for the sub-issue stage feature (web + desktop, shared via packages):
- core: `stage` on the Issue type + create/update request types; zod
IssueSchema parses it (defaults to null for older backends) with schema
tests for the numeric and omitted cases.
- StagePicker component (mirrors the other property pickers): "No stage" +
Stage 1..N, offering one beyond the current/sibling max.
- Create-issue modal: a Stage pill, shown only when a parent is selected,
threaded into the create payload.
- Issue detail sidebar: an editable Stage row + "add property" entry, gated to
sub-issues (issues with a parent).
- Sub-issue list grouped by stage with per-stage headers (flat when unstaged).
- i18n: stage keys across en / zh-Hans / ja / ko (parity test passes).
Verified: full typecheck (6/6), core (591) + views (1433) vitest suites,
lint clean (no new findings). Backend/CLI shipped earlier in this PR.
MUL-3508
Co-authored-by: multica-agent <github@multica.ai>
* test(issues): add stage to Issue fixtures merged from main
The merge brought in new Issue fixtures that predate the required `stage`
field: core issues/batch.test.ts, views batch-action-toolbar.test.tsx, and
the mobile EMPTY_ISSUE_FALLBACK sentinel. Add `stage: null` so they satisfy
the Issue type (mobile reuses core's IssueSchema for parsing, so only the
sentinel needs it).
MUL-3508
Co-authored-by: multica-agent <github@multica.ai>
* fix(web): feed StagePicker the sibling max stage so higher stages stay selectable
The StagePicker accepts maxStage to extend its option list beyond the
floored Stage 1-3, but neither call site passed it, so a parent with an
existing Stage 4/5 child could not pick that stage when creating a new
sub-issue or editing one in the sidebar.
- Compute the sibling max stage at both call sites: the create modal now
loads the parent's children (childIssuesOptions) and the detail sidebar
reuses the already-loaded parentChildIssues.
- Extract maxSiblingStage + stageOptions as pure helpers on stage-picker
and unit-test them (the regression: a Stage 5 sibling keeps Stage 5
selectable and offers Stage 6).
MUL-3508
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Fix GitHub pull_request and check_suite webhook routing so events are attributed to the workspace that registered the repository, with fallback to the installation workspace. Includes host-qualified repo matching, account-gated registry routing, deterministic matching, and regression coverage.
* feat(daemon): inject project description into the agent brief
Issues bound to a project only surfaced the project title in the runtime
brief; the project description (durable, project-wide context the owner
sets) was loaded but dropped. Carry it end-to-end:
- claim handler reads proj.Description onto the response (issue-bound and
quick-create paths)
- new ProjectDescription field on AgentTaskResponse, daemon Task, and
TaskContextForEnv
- rendered in the brief's `## Project Context` section and written to
.multica/project/resources.json as project_description
Empty descriptions render nothing (no extra heading). Updated the
projects-and-resources built-in skill docs in the same change.
MUL-3465
Co-authored-by: multica-agent <github@multica.ai>
* feat(projects): clarify project description is injected as agent context
The project description is now durable context injected into every task's
brief, but the UI still presented it as a plain "Description" field, so
existing descriptions could silently become agent input. Add a hint under
the description editor on the project detail page and in the create-project
modal, in all four locales, stating it is shared with agents as context for
every task in the project. No data-semantics change.
Addresses review feedback on PR #4395. MUL-3465
Co-authored-by: multica-agent <github@multica.ai>
* test(handler): assert project description flows through task claim
The execenv tests cover brief rendering, but nothing pinned the claim
handler boundary where proj.Description is read onto the response. Add
two tests — issue-bound and quick-create paths — so a regression in that
assignment fails loudly instead of silently dropping the description.
Addresses review feedback on PR #4395. MUL-3465
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(agent): Qoder ACP runtime, chat reconnect recovery, and task linkage
- Add Qoder CLI backend (ACP transport, model discovery, blocked-args policy)
- Wire daemon/runtime config, docs, and UI provider assets
- Retry terminal task reports; add backoff unit tests
- Chat: SQL attach user message to task; handler + optimistic cache reconcile
- Invalidate chat/task-messages caches on WS reconnect; extract helper + tests
Co-authored-by: Orca <help@stably.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore: drop non-Qoder changes (chat reconnect, task link, terminal report retries)
Keep only Qoder runtime, docs, daemon config/execenv, and UI provider assets.
Co-authored-by: Orca <help@stably.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(agent): harden Qoder ACP drain and wire project skills path
- Stop streaming to msgCh after reader wait so grace timeout cannot race close
- Resolve injected skills to .qoder/skills per Qoder CLI discovery
- Update AGENTS.md skill copy and add execenv tests
Co-authored-by: Orca <help@stably.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(qoder): add provider logo and wire MCP config into ACP sessions
- Add inline SVG QoderLogo component to provider-logo.tsx, replacing
the generic Monitor icon placeholder
- Add convertMcpConfigForACP helper to convert Claude-style MCP server
config (object map) into ACP array format for session/new and
session/resume
- Add unit tests for convertMcpConfigForACP covering stdio, SSE,
empty/nil, and multi-server cases
Co-authored-by: Orca <help@stably.ai>
* fix(test): capture both return values from InjectRuntimeConfig in Qoder test
Co-authored-by: Orca <help@stably.ai>
* fix(qoder): preserve remote MCP headers and promote provider errors
Addresses review feedback on #2461 (Bohan-J): two runtime-correctness
issues in the Qoder ACP backend.
1. Remote MCP headers were dropped. The bespoke convertMcpConfigForACP
only forwarded url/type, so an authenticated remote MCP server looked
configured in Multica but failed inside the Qoder session. Replace it
with the shared buildACPMcpServers helper (same path Hermes/Kimi/Kiro
use), which preserves headers as [{name, value}], sorts for
deterministic output, and handles remote transport aliases. Fail
closed on malformed mcp_config instead of silently dropping servers.
2. Provider failures could report as completed tasks. stderr was wired
via io.MultiWriter and the result was only promoted to failed when
output was empty, so a terminal upstream error (HTTP 429 / expired
token) racing a stopReason=end_turn with text still became
"completed". Switch to StderrPipe + an explicit copier, drain it
(bounded by the existing grace window, since qodercli can leave a
child holding the inherited fds) before the decision, and run the
shared promoteACPResultOnProviderError.
Tests: replace the convertMcpConfigForACP unit tests with two
end-to-end Qoder tests — one asserts the Authorization header reaches
the session/new payload as {name, value}, the other asserts a terminal
stderr error with non-empty output reports failed.
Co-authored-by: Orca <help@stably.ai>
* fix(qoder): align ACP session handling
Co-authored-by: Orca <help@stably.ai>
* fix(agent): guard qoder late output after drain
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Orca <help@stably.ai>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The batch action toolbar hardcoded status="todo", priority="none", and a
null assignee, so the status/priority/assignee pickers always checked a
fixed row regardless of the selected issues. The batch write itself worked,
but the picker mis-reported the current value, surfacing as "status always
defaults to todo" (MUL-3510). The same defect applied to priority and
assignee, across all five toolbar mount points.
Derive the shared status/priority/assignee of the selected issues via a new
commonIssueFields helper and feed it to the pickers; when the selection is
mixed, pass an empty value so no row is checked. Pickers now accept a
nullable current value, and AssigneePicker gains a `mixed` flag to
distinguish an all-unassigned selection (check "No assignee") from a mixed
one (check nothing). Each call site passes its issue universe, mirroring the
skill list's selected-rows approach.
Adds unit tests for commonIssueFields and a toolbar picker-wiring test.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
Extend the clean target beyond server binaries/temp files to also
remove Next.js/source/Expo/Electron build outputs, Turbo caches, and
tsbuildinfo files across apps and packages.
The header 'agent is working' chip previously required a click to reveal the
activity card. Open it on hover instead so the live signal reads as a
glanceable status surface. The hover config lives on Base UI's Popover.Trigger
(openOnHover + delay/closeDelay), and the trigger stays a real button so
click/keyboard access is retained for touch and a11y.
Add a regression test asserting openOnHover is wired on the trigger so a
click-only implementation can no longer pass.
MUL-3507
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Add a Join Discord promo card pinned to the bottom of the left sidebar
(above the help launcher). Dismiss state persists per-user in
localStorage so it stays hidden once closed.
Extract the shared DiscordIcon + invite URL into layout/discord.tsx so
the help launcher and the card reuse one source. i18n copy added for
en / zh-Hans / ja / ko.
MUL-3505
Co-authored-by: multica-agent <github@multica.ai>
Add a Discord invite (https://discord.gg/W8gYBn226t) in three places:
- Website footer: social icon + link in the Resources group (en/zh/ja/ko)
- In-app help menu: Discord item in the help launcher (all 4 locales)
- GitHub repo README: badge + link (README.md and README.zh-CN.md)
MUL-3492
Co-authored-by: multica-agent <github@multica.ai>
When a sub-issue transitions to done, notifyParentOfChildDone posts a
system comment on the parent and wakes the parent's assignee. A parent
deliberately parked in backlog should stay inert: waking it lets the
assignee promote sibling backlog sub-issues into todo, which is the
unwanted auto-activation reported in GitHub Discussion #4320 (MUL-3497).
Add a backlog guard alongside the existing done/cancelled guard so the
whole notification (comment + mention + trigger) is skipped until the
user explicitly moves the parent out of backlog.
MUL-3497
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The daemon auto-grants Codex item/permissions/requestApproval requests by
echoing back the network / fileSystem profile scoped to the current turn.
Previously a malformed params payload and any permission key outside
network / fileSystem were dropped silently, so a future app-server
protocol that adds a new permission shape would be narrowed away with no
trace in daemon logs.
Log both cases (parse failure and dropped keys) without changing the
granted response. Addresses review nits on #4346 / MUL-3451.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(lark): reply inside the originating thread (话题) instead of the group
When a user @-mentions the bot inside a Lark topic/thread, the bot now
replies back into that thread rather than posting a fresh message at the
chat level. Behavior is automatic and scoped: only triggers that were
themselves inside a thread get a threaded reply, so normal group/p2p
chats are unchanged.
The outbound path is event-driven and decoupled from the inbound
message, so the trigger message_id + thread_id are persisted on
lark_chat_session_binding (migration 122) at ingest time. The patcher
then routes the agent reply (text / markdown card / error card) and the
OutcomeReplier notices (/issue confirmation, offline/archived) through
Lark's reply endpoint with reply_in_thread=true when a thread is present,
falling back to a chat-level send if the threaded reply fails.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(lark): classify thread-reply failures before chat-level fallback
Only retry a threaded reply at the chat level when Lark returns an
explicit "this message/topic cannot receive a threaded reply" error
(recalled trigger, topic gone, topics disabled, aggregated message,
etc.). Transport errors, 5xx, timeouts, rate limits, and ambiguous
failures are now logged and returned as failures instead of being
retried, so we never duplicate a reply or leak a thread-only reply
into the main group chat.
The three reply-capable send methods now return a structured *APIError
carrying the Lark business code, and isThreadReplyUnsupported drives
the fallback via an allowlist. sendWithThreadFallback is promoted to a
package-level function so the immediate OutcomeReplier sends (/issue
confirmation, offline/archived notices) share the same classified
fallback path instead of silently swallowing thread-reply failures.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: kun <kuen@micous.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Skill import builds raw.githubusercontent.com URLs by hand and fetches them
via fetchRawFile, which sent no Authorization header. GitHub API calls were
authenticated by #2215 (doGitHubAPIGet/addGitHubAuthHeader) but the raw
content download path was missed, so importing a skill from a private/internal
GitHub repo listed the directory fine and then 404'd on the actual file
download, surfacing as a generic 502.
Attach the existing GITHUB_TOKEN bearer header in fetchRawFile, but only when
the URL host is raw.githubusercontent.com. fetchRawFile is shared with
clawhub.ai / skills.sh downloads, so the token must not leak to those hosts.
The host gate is extracted into newRawFileRequest so it is unit-testable
without a live round-trip.
MUL-3496
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
When an issue is assigned to a squad, only the leader is triggered. The
leader briefing's Squad Roster listed each member's name, type, role, and
mention link — but not the member's assigned skills, so the leader had to
infer capability from the free-text role label when deciding who to
delegate to.
renderMemberRow now loads each agent member's assigned skills via
ListAgentSkillSummaries and formatRosterRow renders them as
"skills: a, b" (or "no skills assigned" when the agent has none). Builtin
multica-* skills are excluded (they live outside agent_skill); human
members carry no skills segment; a skill-lookup error degrades to the
prior name+role row rather than asserting a misleading "no skills".
Operating-protocol step 1 now tells the leader to match the task to each
member's listed skills.
Updates the multica-squads builtin skill and its source map to document
the new roster content, and adds
TestBuildSquadLeaderBriefing_MemberSkillsInRoster.
Co-authored-by: hal9000botagent <hal9000botagent@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(auth): autofocus OTP input on verification step
The email-verification step renders the OTP input without focus, so
users must click the field before typing the code. This is friction on
every login, especially when switching accounts.
Add `autoFocus` to the InputOTP so the cursor lands in the field as
soon as the step mounts. Mirrors the existing email-step input and the
mobile OTP component, both of which already autofocus.
* test(web): polyfill document.elementFromPoint for input-otp in jsdom
Autofocusing the OTP input makes input-otp run its focus-time DOM
measurement, which calls document.elementFromPoint. jsdom doesn't
implement it, so the web login test threw an unhandled error.
packages/views/test/setup.ts already stubs this for the same reason;
mirror the stub in the web test setup (which already stubs
ResizeObserver for input-otp).
* test(auth): assert OTP input autofocuses on verification step
Guards the autofocus behavior: the test fails if the autoFocus prop is
removed from the verification-step InputOTP. Lives in packages/views
since it covers shared component behavior, mocking @multica/core.
* feat(web): add opt-in react-grab dev element inspector
Loads the react-grab overlay (hold ⌘C / Ctrl+C + click to copy an
element's source path + component stack) only when REACT_GRAB is set in
a local, gitignored apps/web/.env.local. Both the NODE_ENV and REACT_GRAB
guards are evaluated server-side in the root layout, so the <Script> tag
is omitted from the HTML for anyone who hasn't opted in — no effect on
other developers or production.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(desktop): add opt-in react-grab dev element inspector
Mirrors the web wiring for the Electron renderer: injects the react-grab
overlay (hold ⌘C / Ctrl+C + click to copy an element's source path +
component stack) only when VITE_REACT_GRAB is set in a local, gitignored
apps/desktop/.env.development.local. Guarded by import.meta.env.DEV so the
branch is tree-shaken out of production builds; never activates for other
developers. No CSP/sandbox blocks the unpkg script (webSecurity is off).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(web): unify react-grab opt-in var to VITE_REACT_GRAB
Use the same env var name as the desktop renderer so one variable name
controls both apps. The desktop renderer is bundled by Vite, which only
exposes VITE_-prefixed vars to client code, so the shared name must carry
the VITE_ prefix; web reads it server-side where the name is unconstrained.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
## Summary
- expose `--callback-host` on `multica setup` and `multica setup cloud`, reusing the existing login callback override
- print an SSH tunnel hint when browser login runs in an SSH session with a loopback callback
- show local flags in parent-command help so `multica setup --help` documents the callback option
Fixes#4357
## Tests
- `go test ./cmd/multica -run 'TestCallbackHostFlagValueReadsParentSetupFlag|TestSetupHelpShowsCallbackHostFlag|TestSetupCallbackHostFlagWiring|TestBrowserLoginInstructionsSSHRemoteHint'`
- `go run ./cmd/multica setup --help`
- `go run ./cmd/multica setup cloud --help`
- `git diff --check`
- `go test ./...`
* fix(projects): require admin for project deletion
* test(projects): clean up orphaned member rows in delete-permission helper
The schema uses no foreign keys or cascades, so deleting the test user
left its member row behind in the shared test workspace, polluting later
tests in the package. Delete the member row before the user in both the
pre-seed cleanup and t.Cleanup.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* fix(realtime): drop WS frames without a string type (MUL-3418)
The onmessage handler dispatched every parsed frame to onAny and the
ws.on subscribers without validating its shape. A frame whose parsed JSON
lacked a string `type` (an out-of-protocol frame injected by a proxy /
extension, or a bare JSON primitive) reached the realtime-sync onAny
dispatcher, where `msg.type.split(":")` threw an uncaught TypeError out of
onmessage. The browser reported it via window.onerror, flooding PostHog
`$exception` (~12k events). Non-fatal (no crash, connection stays up), but
pure noise.
Validate once at this trust boundary: a frame must be an object carrying a
string `type`, otherwise drop it as a no-op. The bad-frame log is
rate-limited to one entry per connection — a misbehaving source can repeat
the frame hundreds of times per session.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(analytics): drop benign ResizeObserver exceptions from telemetry
ResizeObserver "loop ..." errors are the dominant `$exception` bucket
(~31k events / ~1.5k users). They are a benign, self-recovering browser
quirk with no actionable signal and otherwise drown real failures and burn
the event budget. Drop them entirely in before_send (ahead of redaction
and the dedupe fuse, which only caps repeats). The match is narrow — only
the benign "loop" phrasing — so a genuine ResizeObserver bug still reports.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test: enable -race detector in Go test pipeline (WOR-61)
Add the -race flag to all three Go test invocation sites so the existing
concurrency regression harness (workdir_race_test.go for #3999,
runtime_gone_test.go, runtime_profile_drift_test.go) actually exercises
the race detector. The daemon package alone has 28+ goroutine launch
points with no automated race coverage before this change.
Sites updated:
- Makefile:299 (make test, local)
- .github/workflows/ci.yml:101 (CI backend job)
- .github/workflows/release.yml:55 (release verify job)
go test already runs a vet subset by default, so no separate -vet flag
is added. No production code touched.
Co-authored-by: multica-agent <github@multica.ai>
* test(execenv): serialize runtimeGOOS-mutating test (WOR-61)
TestInjectRuntimeConfigIssueMetadataCodexFormattingUnchanged called
t.Parallel() while mutating the package-level runtimeGOOS to drive the
windows/linux branches, racing with the other parallel tests that read
runtimeGOOS in buildMetaSkillContent. The -race flag enabled in the
prior commit surfaced it as 3 WARNING: DATA RACE reports and 11
"race detected" failures in CI (only the execenv package failed).
Drop t.Parallel() and add the "// Not parallel: mutates the package-level
runtimeGOOS." comment already used by the six sibling writer tests across
execenv_test.go and reply_instructions_test.go. This is test-isolation
only; no production code, no mutex/atomic, no signature change.
Verified locally:
go test -race -count=1 ./internal/daemon/execenv/ -> ok 2.276s
go test -race -count=1 ./internal/daemon/... -> all 3 pkgs ok
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: hzz <331380069@qq.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(daemon): discover local skills from ~/.agents/skills (MUL-3333)
Upgrade local skill discovery and import from a single provider root to an
ordered multi-root scan: the runtime's own skill directory (e.g.
~/.claude/skills) first, then the cross-tool universal root ~/.agents/skills.
- Rename localSkillRootForProvider -> localSkillRootsForProvider, returning
ordered roots [provider, universal] with a kind classifier.
- listRuntimeLocalSkills iterates the roots, gives each root its OWN visited
set (so a cross-root symlink alias is not collapsed), dedupes strictly by
Key with the provider root winning, and sorts once after the merge.
- loadRuntimeLocalSkillBundle walks the same priority order and only falls
through to the next root on os.IsNotExist; any other stat error is returned
so import never silently resolves a different same-key skill.
- Add a Root ("provider" | "universal") field to the local skill summary
(daemon + handler structs and the TS RuntimeLocalSkillSummary type) so the
UI can label a skill's origin without a future schema break.
Backward compatible: every skill visible today keeps its Key, SourcePath and
FileCount; the universal root only surfaces additional, non-conflicting skills.
Out of scope (follow-up issues): execution-time injection of ~/.agents/skills
into runtimes (e.g. Codex seedUserCodexSkills) and workspace-relative
.agents/skills discovery.
Tests cover universal-root discovery + import, provider-wins conflict
priority, both-roots merge, missing/both-missing roots, nested layouts,
IsNotExist fallback, the no-fallthrough-on-read-error guarantee, and the
per-root visited cross-root symlink alias. Docs updated in en/zh/ja/ko.
Co-authored-by: multica-agent <github@multica.ai>
* fix(daemon): fall through to next root when a same-key dir has no SKILL.md
loadRuntimeLocalSkillBundle previously only fell through to the next root on
os.IsNotExist for the skill DIRECTORY. A provider-root directory that shares a
skill's key but contains no SKILL.md (so listRuntimeLocalSkills descends past
it and surfaces the universal-root skill instead) made load stop on the
invalid provider dir and error — list and load disagreed, and the import the
user picked from the list could not be fetched.
Make the validity predicate match list: a root "has" the skill at a key only
when it is a directory containing a SKILL.md. A missing entry, a non-directory,
or a directory without a SKILL.md all mean "this root doesn't have it" and we
continue to the next root. Only a genuine non-IsNotExist stat error or an
unreadable existing SKILL.md (permission/IO) is returned, so we still never
silently substitute a different-content same-key skill from a lower-priority
root (Eve review #1, preserved by the existing read-error guard test).
Adds regression tests for the provider-dir-without-SKILL.md and provider-non-dir
fall-through cases.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
In context mode (chat) a query merges context items (current page / recently
viewed) with search results into one popup. The old contextLayout made only the
"Recent" group scrollable (`min-h-0`) while every other group was `shrink-0`, and
the Recent `<section>` did not clip its own overflow. When the search groups
(Users/Issues) filled the height, the flex algorithm squeezed Recent toward zero
and its un-clipped rows painted on top of the groups below — the overlap users saw.
Collapse the two render branches into a single `overflow-y-auto` flex column so
every group just stacks and the whole popup scrolls; no group can collapse onto
another. The context variant only differs in width / max-height / chrome.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(desktop): coerce commit-hash versions to valid semver
normalizeGitVersion checked only the first character (/^\d/) to tell a real
version from a bare commit hash. A hash beginning with a digit (e.g.
'2f24057b') passed that check and was stamped as the app version, but bare
'2f24057b' is not valid semver, so electron-updater threw on launch.
Require a full major.minor.patch prefix; anything else (including a
digit-leading hash) falls back to 0.0.0-<hash> as before.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(desktop): prefix bare-hash fallback with `g` for valid semver
normalizeGitVersion coerced an untagged build's bare commit hash into
`0.0.0-<hash>`, but that is not guaranteed valid semver: an all-digit
short hash with a leading zero (e.g. `0123456`) produced `0.0.0-0123456`,
and a numeric semver pre-release identifier must not have a leading zero.
electron-updater then threw on launch for exactly the untagged builds
this fallback exists to protect.
Prefix the hash with `g` (mirroring `git describe`'s own `g<hash>`
shorthand) so the pre-release is always a single alphanumeric identifier.
Add a regression test for the all-digit leading-zero case.
Addresses PR #4183 review.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
#4288 swapped the chip cap from a fixed `max-w-72` to `max-w-[min(18rem,100%)]`.
A percentage max-width on a flex item is dropped while its flex-container wrapper
(`<a class="inline-flex">`) computes its own max-content size, so the wrapper
ballooned to the untruncated title width while the chip truncated to the cap —
leaving an empty, clickable strip after the visible chip.
Fix it as standard atomic-inline behavior instead of a fixed magic cap:
- IssueChip caps at `max-w-full` with the title truncating to an ellipsis, so it
wraps to the next line as a unit and only truncates once a whole line can't
hold it.
- Drop `inline-flex` from the editor NodeView `<a>` (issue + project) and the
markdown AppLink so the chip's only wrapper is a plain inline box with a
definite (line-based) percentage basis — no second flex container to balloon.
ProjectChip uses a definite `max-w-72`, so it never hit the gap; left as-is.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds a dry-run-first operator command for historical Codex usage cache correction, packages it in the backend image, documents the operator-job flow, and covers execution with DB-backed tests.
Bundles the MUL-3404 disk-usage feature with the two Preflight BLOCK fixes.
- feat(daemon): `disk-usage --all-profiles` aggregates across every workspace
root (default + each ~/.multica/profiles/* root, incl. the Desktop app's),
with a per-root breakdown and combined grand total; the cross-root hint now
also fires when the current root is non-empty.
- fix(db): drop DB-level foreign keys/cascades from the new autopilot_subscriber
and comment.source_task_id migrations (resolved in the app layer — autopilot
delete now removes subscribers in a transaction); the autopilot_subscriber
down-migration relabels reason='autopilot' to 'manual' instead of deleting.
- fix(server): readiness verifies every required migration is applied, not just
the lexically-last one, so an out-of-order migration can't be masked.
MUL-3404.
* fix(daemon): reclaim autopilot_run workdir on terminal status (MUL-3403)
Autopilot run workdirs are never reused — there is no PriorWorkDir path
that hands a later run the same directory, so every run gets a fresh one.
Yet GC waited the full GCTTL (default 24h) before reclaiming a terminal
run's dir. Combined with one fresh dir per run, high-frequency autopilots
piled up hundreds of stale dirs (508 dirs / 22GB in the field report).
Drop the TTL gate so a terminal run (completed/failed/skipped/
issue_created) is reclaimed immediately, mirroring gcDecisionQuickCreate.
Existing safety constraints are untouched: active-env-root short-circuit,
404 -> orphanByMTime, non-404 error -> skip, and the local_directory
override all still apply.
Co-authored-by: multica-agent <github@multica.ai>
* docs(daemon): fix GetAutopilotRunGCCheck comment — completed_at is not a TTL anchor
The endpoint comment still claimed the daemon uses completed_at as the TTL
anchor for terminal runs. GC now decides purely on terminal status (the
workdir is never reused, so a terminal run is reclaimed on sight);
completed_at is returned for the API contract / diagnostics only. Addresses
the review nit on #4287.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The create-workspace and onboarding UI hardcoded `multica.ai/` as the
workspace slug URL prefix, so self-hosted deployments showed the wrong
domain. Add a `workspaceUrlHost` helper that derives the host from the
deployment's app URL (`daemon_app_url` from `/api/config`, via the config
store) and falls back to the brand host when none is configured, then use
it in both views. Fixes#4263.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Disambiguate client-side model pricing by provider: generic ids (e.g. `auto`) resolve ${provider}/${model} first, so they only price under their real provider instead of borrowing Cursor's rate. Provider is LOWER()-normalized on read and write so mixed-case historical rows merge.
Closes#4199. MUL-3346
Extract three module-local helpers in content-editor.tsx and route the
duplicated call sites through them — no behavior change:
- normalizeMarkdown(md) / normalizeEditorMarkdown(editor): the single
definition of the "strip blob URLs + trimEnd" canonical form, replacing
the five editor-markdown sites and the one incoming-string site.
- hasUploadingNode(editor): replaces the two byte-identical document scans
(content-sync Guard 0 and hasActiveUploads).
The imperative getMarkdown() is deliberately left untrimmed; a safety-net
test pins its exact current return value.
Scope: WOR-59 upload-readiness review. file-upload.ts untouched; follow-up
items B1/F4/F5 are out of scope.
Co-authored-by: hzz <331380069@qq.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
The `codebuddy` case in ProviderLogo was aliased to ClaudeLogo when the
backend was integrated (#3186), so CodeBuddy runtimes rendered the Claude
mark and were visually indistinguishable from Claude in the runtime list,
runtime detail, agent runtime picker, and onboarding.
Add a dedicated CodeBuddyLogo using the official mark shipped in Tencent's
own @tencent-ai/codebuddy-code CLI package (dist/web-ui/logo.svg) — the
same CLI this runtime spawns. The artwork overflows the 24×24 box and is
cropped by a clipPath whose id is per-instance (useId), mirroring KiroLogo,
so multiple logos on one page don't collide on a shared id.
Presentation-only: no provider string, backend/API, DB, daemon/runtime,
CLI, mobile, or copy changes. Shared views package, so web + desktop both
pick it up.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(chat): workspace-scoped attachment binding + fire-and-forget send
Uploads are now workspace-scoped: the chat session is created and
attachments are bound to the message at send time, so a paste/drop no
longer creates an empty session the user never sends.
- LinkAttachmentsToChatMessage returns the ids it actually bound; the
client diffs requested-vs-bound and warns on partial bind, replacing
an extra listChatMessagesPage fetch.
- Cancelling an empty chat task detaches attachments before deleting the
user message (attachment FK is ON DELETE CASCADE) and returns them via
cancelled_chat_message.attachments, so a restored draft can re-bind.
- SendChatMessageResponse.attachment_ids has no omitempty: "requested but
bound zero" serializes [] so the client can tell it apart from an older
server and still warn.
- Send is fire-and-forget: it no longer steals focus when the user has
navigated to another session (guarded on the live store + new-chat agent
id); the reply surfaces via the unread dot. commitInput gets clearEditor
so a navigated-away commit doesn't wipe the editor now showing another
session, while still clearing the sent draft's data.
- Draft restore is session-aware so a failed fire-and-forget send restores
into the session it was sent from, never the one the user moved to.
- Removed the now-unreferenced migrateInputDraft store action.
Verified: core/views typecheck, chat-input (15) / store (3) / api client
(24) unit tests, go build + vet, handler SendChatMessage + CancelTaskByUser
DB tests. Full make check / E2E left to CI.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(chat): guard attachment survival on empty-chat cancel
Cancelling an empty chat task deletes the user message, and
attachment.chat_message_id is ON DELETE CASCADE (migration 083), so the
detach-before-delete in finalizeCancelledChatMessage is the only thing
keeping the user's attachment from being silently destroyed. Nothing
covered it.
Add a DB regression test that binds an attachment to the cancelled user
message and asserts: the row survives the cascade (chat_message_id NULL,
chat_session_id retained), the cancel response returns it via
cancelled_chat_message.attachments, and a resend re-binds it to the new
message. Verified red when the detach step is removed.
Related issue: MUL-3364
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(comment): pessimistic submit for comment/reply composers
The comment and reply composers cleared the editor after `await onSubmit`
returned, with no in-flight lock. On a slow send the WS `comment:created`
event already dropped the real comment into the timeline while the box
still held the same text + spinner, so it read as two comments. And
because `submitComment`/`submitReply` swallow errors (toast, no rethrow),
a failed send still reached `clearContent` and silently discarded the
user's draft.
Recover the comment/reply portion of the closed#4236: make the submit
callback resolve a success boolean (true on success, false on the caught
failure), lock the editor while in flight (pointer-events-none + dimmed
wrapper + aria-busy, since ContentEditor can't toggle Tiptap `editable`
post-mount), keep the button spinning, and clear only on success — a
failed send keeps the draft. Chat composer is out of scope (already
reworked on this branch); attachment binding is untouched.
Adds two view tests (in-flight lock then clear-on-success; failed send
keeps the draft); both verified red against the un-fixed code.
Related issue: MUL-3364
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(autopilot): default subscriber template (MUL-2533) — server
Add per-autopilot member subscriber template that fans out to every
issue the autopilot spawns. New autopilot_subscriber table; extend
issue_subscriber.reason with 'autopilot' so the dispatch-time fanout
is distinguishable from manual subscriptions.
API: POST/PATCH /api/autopilots accept a `subscribers` array (member
user_type only for the first version); PATCH semantics are full-replace.
GET returns subscribers on the detail endpoint; the list endpoint omits
them to avoid an N+1.
Dispatch: dispatchCreateIssue lists the template inside the same tx as
the issue insert and writes the rows with reason='autopilot' before
EventIssueCreated fires, so notification listeners see the full
subscriber set on the first event.
Co-authored-by: multica-agent <github@multica.ai>
* feat(autopilot): default subscriber template (MUL-2533) — frontend
New SubscriberMultiSelect picker (members-only search + chips) wired
into the create / edit AutopilotDialog. The detail page renders the
saved template as read-only chips; edits flow through the dialog.
TS types expose the new `subscribers` field on Autopilot, plus an
AutopilotSubscriberInput shape for the create/update wire payloads.
Co-authored-by: multica-agent <github@multica.ai>
* fix(autopilot): notify template subscribers on issue creation (MUL-2533)
The autopilot create-issue path fans out template subscribers into
issue_subscriber inside the same tx as the issue insert, but the
issue:created notification listener only matches handler.IssueResponse
payloads and only direct-notifies the assignee + @mentions. The autopilot
publishes a map[string]any payload, so the listener falls through and the
template subscribers never receive an inbox item for the creation event —
breaking OQ3 ("reason='autopilot' subscribers receive all subscription
events, consistent with reason='manual'").
Fix it where the divergence lives: in dispatchCreateIssue, right after
EventIssueCreated fires, write an inbox_item (type='issue_subscribed',
severity='info') for each member subscriber and publish EventInboxNew so
the recipient's inbox WS feed updates in real time. The write is after
the tx commit so an inbox hiccup can't roll back the issue; failures are
logged, not propagated. The manual path is unchanged — manual subscribers
don't exist at creation time, so there is nothing to notify there.
Adds a new InboxItemType 'issue_subscribed' (en/zh labels) and two
covering tests in autopilot_subscriber_test.go: one asserts the inbox
row lands for a template subscriber on dispatch, the other asserts the
no-subscriber autopilot stays silent.
Co-authored-by: multica-agent <github@multica.ai>
* fix(autopilot): align subscriber PR with current main
Co-authored-by: multica-agent <github@multica.ai>
* fix autopilot subscriber template transaction
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The runtimes page header has three outline action buttons — "Add runtime",
"Cloud Runtime", and "Add a computer" — that previously rendered with
inconsistent dimensions and responsive behavior:
- "Add a computer" used `h-8 w-8 ... md:w-auto md:px-2.5` (icon-only
below md, expands to icon+label on md+), with an `aria-label` for the
icon-only state and the label wrapped in `<span class="hidden md:inline">`.
- "Add runtime" had no responsive className, no aria-label, and the label
stayed visible at every width.
- "Cloud Runtime" had the same gaps as "Add runtime", plus the Cloud
icon was sized `h-3 w-3` instead of `h-3.5 w-3.5` like the others.
Visually they read as three different button styles crammed into one
header. Aligning all three to the "Add a computer" pattern gives the
header a single, consistent action group at all breakpoints, fixes the
mobile layout (header no longer wraps onto multiple lines for users with
both Cloud Runtime enabled and admin role), and makes every icon-only
state screen-reader accessible.
Changes:
- Apply `h-8 w-8 gap-1 px-0 md:w-auto md:px-2.5` to all three buttons.
- Add `aria-label` to "Add runtime" and "Cloud Runtime".
- Wrap the label of "Add runtime" and "Cloud Runtime" in
`<span className="hidden md:inline">` so they collapse to icon-only
below md, matching "Add a computer".
- Normalize the Cloud icon size from `h-3 w-3` to `h-3.5 w-3.5`.
Verification: typecheck + lint + 1395 unit tests across @multica/views
all pass.
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
The Custom runtimes dialog repeated the same idea in three places: the
header tagline, the empty-state body, and the right-panel default state.
Per MUL-3367, tighten the prose so each surface adds new information:
- dialog_description: short tagline that points at the protocol families
list directly below.
- empty_description: drops the "such as Claude or Codex" example (already
enumerated in the protocols list right under it) and the "your daemon
should run" filler.
- detail.default_title / default_description: replace the
"Manage custom runtimes" + paragraph (which restated the header) with
a quiet "Select a runtime" placeholder.
- detail.default_builtin_hint: dropped entirely. The supported-protocols
section is already visible in the left column and each row carries a
Reference badge — pointing at it from the right panel is noise.
Updated en, zh-Hans, ja and ko locales in lockstep, plus the dialog test
assertions.
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* fix(markdown): don't auto-link bare filenames as external URLs
Agent comments that mention a project file like `plan.md` were turned into
clickable links to https://plan.md (dead external site). linkify-it fuzzy
detection matches `plan.md` as a domain because its extension is also a valid
TLD (md = Moldova; likewise io, sh, rs, py).
Suppress schemeless (fuzzy) linkify matches whose token is a bare filename
(single segment ending in a known source/config extension). Explicit schemes
(`https://plan.md`) and real domains (`example.com`) are unaffected. The file
extension list is now shared between the file-path and bare-filename detectors
so they can't drift.
Fixes#4222
Co-authored-by: multica-agent <github@multica.ai>
* docs(markdown): drop inaccurate .io example from bare-filename comment
io is not in the FILE_EXTENSIONS list, so .io domains are never suppressed.
Listing it as an example was misleading.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The runtime detail page disabled the Delete button — and the list-page
kebab dropped its only action — whenever the runtime was an online
local daemon (isSelfHealingRuntime). The intent was to spare users a
delete that the daemon would silently undo via re-register, but the
side effect was an Owner staring at an action they had every
permission for, with no obvious explanation.
The fix moves the warning into the confirmation dialog instead of
hiding the action:
- runtime-detail.tsx: drop the disabled+tooltip branch around the
Diagnostics card Delete button. canDelete (workspace owner/admin OR
runtime owner) now fully governs visibility, and the button is
always clickable when present.
- runtime-list.tsx: the row kebab no longer hides itself for
self-healing runtimes. showActions follows along.
- delete-runtime-dialog.tsx: drop the defensive self-healing guard in
handleConfirm (it returned early with a toast on the very click the
user just confirmed). Add a SelfHealNotice banner that renders in
both LightBody and CascadeBody when isSelfHealingRuntime is true,
so the user sees the trade-off before pressing the destructive
button rather than after.
- locales: drop delete_disabled_tooltip and the cascade
self_healing_blocked_toast across en/zh-Hans/ja/ko; add a single
detail.delete_dialog.self_heal_notice key that powers the banner.
- tests: flip the row-menu self-healing assertion (kebab is now
visible), add a banner-present / banner-absent / confirm-proceeds
triplet to delete-runtime-dialog.test.tsx, and pin the
enabled-Delete-for-owner case on runtime-detail-visibility.test.tsx.
Verified with @multica/views typecheck (clean) and a targeted vitest
run across the four affected suites — 73/73 tests pass. Lint stays at
0 errors on the package.
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* fix(github): surface in-flight CI on PRs and recover out-of-order check_suite events
Two bugs caused PR cards to render "checks not reported yet" while CI
was actually running (MUL-2392):
1. handleCheckSuiteEvent dropped every action except `completed`, so
`requested`/`rerequested` events (status queued/in_progress) never
landed in the suite table. Aggregated `checks_pending` stayed at 0
until the first suite finished, and the frontend fell through to the
unknown bucket. Persist all actions; the ListPullRequestsByIssue
aggregation already counts status<>completed as pending.
2. A check_suite for an unmirrored PR was logged and dropped, with no
replay path. Add a `github_pending_check_suite` stash keyed by
(workspace, repo, pr_number, suite_id); the pull_request webhook
drains it after the PR upsert and replays each entry through the
normal check_suite upsert. One-shot drain via DELETE … RETURNING
keeps it idempotent and free of retry storms.
Follow-ups for fork PRs (empty `pull_requests[]`) and a more specific
frontend placeholder ship in separate issues.
Co-authored-by: multica-agent <github@multica.ai>
* fix(github): guard pending check_suite stash against out-of-order events
UpsertPendingCheckSuite previously overwrote unconditionally on
conflict, so an older `requested/in_progress` event arriving after a
newer `completed/success` for the same suite would roll the stash
back to pending. The subsequent PR upsert then drained the stale
state and the PR card stuck on "pending" until the next suite. Mirror
the suite_updated_at guard from UpsertPullRequestCheckSuite and add a
regression test covering the PR-missing path.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): unify trigger chip copy to will-start phrasing
Make the comment trigger chip's on/off states symmetric around the verb
'start' instead of mixing natural language with the 'trigger' jargon:
- on: Will start when sent
- skipped: Won't start this time
- all skipped: No agents will start
- multi on: N agents will start when sent
Updates all four locales (en/zh-Hans/ja/ko); CJK on-state copy already
reads as future-conditional so only the skip states are realigned to the
'start' verb. Updates the component test expectations to match.
Refs MUL-3211
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): align restore hint to will-start phrasing
Carry the trigger-chip copy unification into the suppressed-agent restore
hint (trigger_click_to_restore), the last surface still mixing 'trigger'
with the chip's 'start' wording:
- en: Won't start this time. Click to restore.
- CJK: skip-state term realigned to the 'start' verb, rest unchanged.
Refs MUL-3211
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): trim trigger preview popover copy and drop redundant reason lines
The active hover popover stacked header + reason + presence, repeating the
same fact across lines and again against the chip. Tighten it:
- Drop the reason line for assignee / @mention: the header (name · source)
already conveys why they fire. Keep reason only for squad-leader (the link
is non-obvious) and the unknown fallback, both trimmed of the duplicated
name.
- Shorten presence (Starts right away. / Offline now — starts once online.)
and de-jargon the skip/manage hints (no more 'trigger').
- Align the popover title to the chip wording (Will start when sent).
All four locales updated; removes the two now-unused reason keys.
Refs MUL-3211
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The agent record already carries a top-level `thinking_level` field —
exposed by `agent get --output json`, settable from the web inspector,
and accepted/validated by `PUT /api/agents/{id}` — but the CLI had no
flag to write it. Scripted or version-controlled agent management could
set `--model` but not the thinking depth, forcing a drop to raw HTTP.
Add `--thinking-level` to `agent create` and `agent update`, mirroring
`--model`: a thin pass-through to the top-level `thinking_level` field.
On update an empty string clears it back to the runtime default (the
server reads it as a tri-state pointer: omitted = no change, "" = clear,
value = set). The CLI deliberately does not enumerate valid levels —
they are runtime/model-specific and the server already owns the catalog
(`agent.IsKnownThinkingValue`, `server/pkg/agent/thinking.go`), returning
a 400 for an unknown value or a runtime with no thinking concept, which
the CLI surfaces verbatim.
- server/cmd/multica/cmd_agent.go: register the flag on both commands,
Changed-gate it into the request body, add it to the no-fields error.
- server/cmd/multica/cmd_agent_test.go: cover create/update send,
unset-omission, empty-clears, the flag-exposed guard, and that a
server-side rejection surfaces to the user.
- multica-creating-agents builtin skill + source map: document the new
CLI write surface and re-derive shifted cmd_agent.go line numbers.
Closes#4170🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Matt Voska <voska@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
* feat(analytics): session-level $exception dedupe in before_send
A runaway client error (a render loop, a polling fetch that keeps
throwing) emits 100+ identical $exception events per session, which
showed up as a top PostHog cost/noise source after exception
autocapture landed (MUL-3331 / MUL-3330).
Add a per-tab-session fuse in before_send, after redaction: fingerprint
the already-redacted exception (type + redacted value + one deterministic
stack frame incl. colno), keep the first 3 per (session, fingerprint),
drop the rest. State lives in sessionStorage as a hash->count blob, so no
PII is persisted. Every storage failure fails open (keep the event);
before_send never throws.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(diagnostics): global 60s cooldown for client_unresponsive
A single sustained freeze is delivered as several long-task entries, so
emitting per entry made client_unresponsive volume grow without bound
with the freeze length (MUL-3331). Cap it with a module-level (page-
lifetime) global cooldown: at most one event per 60s window. No route
bucketing — a global window is the most direct cap on volume.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(analytics): add exception-dedupe safety matrix
This file was written alongside the dedupe implementation but missed the
original commit, so the $exception fail-open / cap / fingerprint matrix
never landed on the branch. No implementation change — the tests pass as
written against the existing exception-dedupe.ts.
Covers: first-3-then-drop, fingerprint independence, colno discrimination,
hash-only storage (no PII), degraded/missing frames, undefined / throwing
/ corrupt-JSON sessionStorage fail-open, setItem-failure under-counts, and
the distinct-fingerprint cap (51st new fingerprint kept).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
`multica version` now outputs two lines (version + go info). The old
`awk '{print $2}'` captured both lines, causing the version comparison
to always mismatch and triggering unnecessary upgrades.
Fixes https://github.com/multica-ai/multica/issues/4226
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* MUL-3332: daemon picks up new custom runtime profiles without restart
The workspaceSyncLoop's already-tracked branch refreshed only settings and
repos via refreshWorkspaceRepos and never re-fetched runtime profiles, so
a custom runtime profile created via the web UI / CLI did not become a
registered runtime row until the daemon restarted (or a runtimeGone
recovery happened to fire).
Detect server-side profile drift each sync tick by hashing the workspace's
profile list with profileSetSignature(), caching the digest on
workspaceState.profileSetSig, and triggering reregisterWorkspaceAfterRuntimeGone
when the live signature differs from the cached one. Steady-state syncs cost
exactly one extra GetRuntimeProfiles round trip; only real drift fans out to
a Register call.
The fetch is best-effort: a 404 / network blip preserves the cached signature
so a transient failure cannot loop the daemon into spurious re-registrations.
Tests in runtime_profile_drift_test.go cover digest stability under reorder,
field-by-field drift detection (add / enable-flip / command_name /
protocol_family / fixed_args / visibility), the no-drift hot path (no
re-register), the new-profile drift path (single re-register + index update +
sig converges), and best-effort fetch error handling.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3332: split orphan recovery from profile drift; converge to zero
Addresses two blocking review concerns on #4225 (raised by GPT-Boy):
1. Profile drift must not kill running tasks on existing runtimes.
The first cut reused reregisterWorkspaceAfterRuntimeGone, which after
re-register calls /recover-orphans for every returned runtime ID. The
server's RecoverOrphanedTasksForRuntime hard-fails every
dispatched/running/waiting_local_directory row on that runtime — the
correct response when a runtime row was actually deleted server-side,
but a catastrophic false positive on profile drift: a built-in runtime
still actively executing the user's tasks would have its work killed
just because the user added an unrelated sibling custom profile.
Fix: extract applyRegisterResponseInPlace as the shared in-place state
converger between the two paths, and stop calling /recover-orphans from
the drift path. reregisterWorkspaceAfterRuntimeGone keeps the
/recover-orphans call because in that path the rows really were gone.
2. Disabling the only profile on a custom-only daemon must converge.
The first cut hit registerRuntimesForWorkspace's len(runtimes)==0 guard
and bailed out, so the disabled profile's runtime stayed alive in
local tracking and on the server (still polling, still heartbeating,
still online for the full 150 s stale-heartbeat window).
Fix: introduce ErrNoRuntimesToRegister as a sentinel, have
registerRuntimesForWorkspace return profileSig even on the empty case
(so the drift path can cache the converged-empty signature), and have
the drift refresh's error handler take a convergeWorkspaceRuntimesToZero
branch that clears local runtimeIDs / runtimeIndex entries and
Deregisters the orphaned IDs so the server marks them offline
immediately. The same Deregister step also runs on partial drift (a
built-in survives, the disabled profile's runtime drops) so the user
sees the dropped runtime go offline within the next sync tick instead
of after the 150 s sweep.
Tests:
- TestRefreshWorkspaceRuntimeProfiles_DriftWithRunningRuntimeSkipsOrphanRecovery
(mixed built-in + custom, add another profile, asserts zero
/recover-orphans calls).
- TestRefreshWorkspaceRuntimeProfiles_DisableConvergesCustomOnlyDaemon
(custom-only daemon, disable only profile, asserts local state
cleared, signature converges to empty digest, Deregister called with
the orphaned ID, no recover-orphans, follow-up tick is no-op).
- TestRefreshWorkspaceRuntimeProfiles_DisableOneOfManyDeregistersDroppedID
(partial drift: only the dropped ID is Deregistered, surviving
built-in is left alone and not orphan-recovered).
- TestRefreshWorkspaceRuntimeProfiles_NewProfileTriggersReregister
extended to also assert no /recover-orphans calls.
- TestRegisterRuntimes_SkipsProfileNotOnPath strengthened to assert the
ErrNoRuntimesToRegister sentinel and that profileSig is still returned
on the empty path.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR3 (CLI): multica runtime profile subcommands + local path override
- cmd_runtime_profile.go: `multica runtime profile` group — list / create /
update / delete against /api/workspaces/{id}/runtime-profiles, plus set-path
/ unset-path for a per-machine command override. protocol-family validated
client-side via agent.IsSupportedType / agent.SupportedTypes; visibility
validated; update only sends changed flags (protocol_family immutable);
delete surfaces the server 409 body when agents are still bound.
- internal/cli/config.go: ProfileCommandOverrides map[string]string on
CLIConfig (omitempty), through the existing marshal/unmarshal so set/unset
round-trips without dropping other fields.
- internal/daemon: Config.ProfileCommandOverrides, loaded from CLIConfig;
appendProfileRuntimes now prefers an override path when set AND executable,
else falls back to exec.LookPath(command_name), else skips+logs as before.
- Tests: cmd_runtime_profile_test.go (registration, create/update/delete incl.
bad-family + missing-flag + 409 surfacing, set/unset path round-trip,
relative-path rejection, config preservation); cli/config round-trip;
daemon prefers-override / falls-back-when-not-executable.
Verified: go build ./..., go vet, go test ./cmd/multica/... ./internal/daemon/...
./internal/cli/... all pass.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR3 (Web): custom runtime profiles in the Runtime page
Single-list integration — no new page, no tabs/grouping. Built-in protocol
families and custom profiles render mixed in one catalog, each row badged
built-in vs custom (progressive disclosure).
- packages/core: RUNTIME_PROFILE_PROTOCOL_FAMILIES (single-source 13-family
whitelist, matches server agent.SupportedTypes + migration 120 CHECK) and
RuntimeProtocolFamily / RuntimeProfile types; api client
list/get/create/update/deleteRuntimeProfile against
/api/workspaces/{id}/runtime-profiles; runtimes/profiles.ts query +
mutation hooks and a 409 "agents still bound" conflict parser.
- packages/views/runtimes: runtime-profile-catalog (mixed built-in+custom
rows), runtime-profiles-dialog (header "+ Add runtime" → step 1 pick
protocol family → step 2 display_name/command_name/description; edit form
for custom; admin-gated), delete-runtime-profile-dialog (confirm + graceful
409), runtimes-page / runtime-list integration.
- i18n: new strings added to all four locales (en, zh-Hans, ja, ko).
- a11y: dialogs are focus-trapped, Esc-closable, labelled; full
create/edit/delete flow is keyboard + screen-reader operable.
Iron rule honored: no generic per-agent args UI here (those stay on Agent
config). fixed_args is not surfaced as a general args field.
Verified: turbo typecheck + lint + test pass for @multica/core, @multica/views,
@multica/web; the @multica/web production build succeeds.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR3: hide fixed_args from Web + CLI (not yet wired to launch)
Review fix. fixed_args was surfaced as a working feature, but the daemon does
not splice it into the agent launch command — exposing it promised admins a
no-op. Per the call, remove it from every user-facing surface while keeping the
underlying column/struct "carried but not exposed".
- Web (runtime-profiles-dialog.tsx + runtime-profile-catalog.ts): drop the
detail row, the create body field, the update patch field, and the form
textarea; remove the parseFixedArgs/fixedArgsToText helpers and the
fixedArgs form value. Left a NOTE pointing at the daemon TODO.
- i18n: removed the fixed_args strings from all four locales (en/zh-Hans/ja/ko).
- CLI (cmd_runtime_profile.go): removed the `--fixed-arg` flag from create and
update and stopped sending `fixed_args`; updated the "no fields" message.
Test now asserts the CLI never sends fixed_args.
Untouched (the carried-but-not-exposed layer): the runtime_profile.fixed_args
column, the server handler's accept/return, and the daemon's RuntimeProfile
field — all keep the existing TODO(MUL-3284) to wire it into the launch path
(with a test proving args reach the backend) before any UI/CLI re-exposes it.
Verified: turbo typecheck+lint+test pass for @multica/core and @multica/views;
go build/vet/test pass for ./cmd/multica/.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR3: stop exposing profile visibility=private (server forces workspace)
Double-review (Eve) caught a fixed_args-shaped hole: visibility=private was a
user-facing toggle (Web form + detail + CLI), but the three server read paths
(ListRuntimeProfiles, daemon ListEnabledRuntimeProfilesForWorkspace,
DaemonRegister) never enforce it — so a "private" profile's name/command would
leak to other members and could be registered by other machines' daemons
(lateral data leak). Same "don't paint a pie" fix as fixed_args: hide the
control everywhere and force the stored value.
- Server (runtime_profile.go): drop `visibility` from the create + update
request structs; CreateRuntimeProfile always stores 'workspace'
(runtimeProfileDefaultVisibility); UpdateRuntimeProfile no longer accepts it;
removed validRuntimeProfileVisibility. The column + response field stay
(always 'workspace') as the carried-but-not-exposed layer.
- Web (runtime-profiles-dialog.tsx): removed the visibility form fieldset,
the VisibilityOption component, the detail row, the visibility state, and the
create/update submit fields.
- i18n: removed the profile visibility strings from all four locales
(profiles.detail.visibility, profiles.visibility.*, profiles.form.visibility_*).
Top-level runtime/agent visibility strings are untouched.
- CLI (cmd_runtime_profile.go): removed `--visibility` from create/update and
the VISIBILITY list column; removed validateVisibility; stopped sending the
field.
- Tests: new TestCreateRuntimeProfile_ForcesWorkspaceVisibility (POST
visibility:"private" -> response and DB row are 'workspace'); CLI create test
now asserts visibility is never sent.
Follow-up MUL-3308 tracks implementing real creator-visibility (and wiring
fixed_args to the launch path); TODOs left in server/Web/CLI point to it.
Verified: turbo typecheck+lint+test pass (@multica/core, @multica/views);
go build/vet pass; go test ./cmd/multica/... and the full ./internal/handler/
suite pass against a migrated Postgres 17.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284: add runtime_profile schema (custom runtime PR1)
Schema-only foundation for custom runtimes. Additive migration 120:
- New workspace-level `runtime_profile` table: the shared, team-visible
definition of a custom runtime (e.g. an in-house Codex wrapper).
protocol_family is CHECK-constrained to the exact backend list in
agent.New() (server/pkg/agent/agent.go). The only args column is
`fixed_args` (args every agent on the runtime must inherit); there is
deliberately no generic per-agent args field — those stay on
agent.custom_args.
- `agent_runtime.profile_id` (nullable, FK -> runtime_profile ON DELETE
CASCADE): NULL = built-in runtime, non-NULL = a registered instance of
a custom profile.
- Partial unique index agent_runtime_workspace_daemon_profile_key on
(workspace_id, daemon_id, profile_id) WHERE profile_id IS NOT NULL.
The legacy UNIQUE (workspace_id, daemon_id, provider) constraint is left
INTACT so the existing registration upsert
(ON CONFLICT (workspace_id, daemon_id, provider) in runtime.sql) keeps
resolving its arbiter and the server stays green. Converting that key to
a partial (WHERE profile_id IS NULL) index and making the upsert
profile-aware is PR2's registration work, not this migration.
Verified up + down against Postgres 17: full `migrate up` applies 120;
schema shows the table, column, partial index and intact legacy
constraint; functional checks pass (partial index blocks dup
(ws,daemon,profile), allows same profile on another daemon; CHECK and
display_name uniqueness reject bad input; legacy ON CONFLICT still
resolves; profile delete cascades to instances); down/up round-trip is
clean.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284: drop DB FKs/cascade from runtime_profile migration (review fix)
Per review (house rule: no new database foreign keys / cascades; relational
integrity lives in the application layer):
- runtime_profile.workspace_id: drop REFERENCES workspace ON DELETE CASCADE
-> plain UUID NOT NULL.
- runtime_profile.created_by: drop REFERENCES "user" ON DELETE SET NULL
-> plain UUID.
- agent_runtime.profile_id: drop REFERENCES runtime_profile ON DELETE CASCADE
-> plain UUID.
CHECK constraints, UNIQUE (workspace_id, display_name), the workspace index,
and the partial unique index agent_runtime_workspace_daemon_profile_key are
unchanged. The legacy UNIQUE (workspace_id, daemon_id, provider) constraint
remains untouched.
Behavioral consequence: the database no longer auto-removes a profile's
agent_runtime instance rows on profile delete. That cleanup moves into PR2's
profile-delete path. Up-migration comments document this; down-migration
comment no longer references FKs/cascade.
Re-verified on Postgres 17: migrate up applies 120; no FK constraints exist on
the new columns; partial index still blocks dup (ws,daemon,profile_id); CHECK
and display_name uniqueness still reject bad input; deleting a profile now
leaves the runtime row orphaned (proving cascade is gone); down/up round-trip
clean with the legacy constraint intact.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR2 (server): runtime_profile CRUD + profile-aware registration
Server/DB half of the custom-runtime feature.
- Migration 121: convert the legacy UNIQUE (workspace_id, daemon_id, provider)
constraint on agent_runtime into a partial unique index scoped to built-in
rows (WHERE profile_id IS NULL). With 120's partial index on profile_id this
lets one daemon host the built-in provider AND custom profiles of the same
protocol family without collision.
- Queries: runtime_profile CRUD; ListEnabledRuntimeProfilesForWorkspace
(daemon-facing); CountAgentsByProfile + DeleteAgentRuntimesByProfile for the
app-layer cascade; profile-aware UpsertAgentRuntimeWithProfile; the built-in
UpsertAgentRuntime ON CONFLICT now spells out WHERE profile_id IS NULL so it
targets the right partial index. sqlc regenerated.
- agent.SupportedTypes / IsSupportedType: single-source protocol_family
whitelist, in lockstep with agent.New and the migration 120 CHECK.
- Handlers + routes: runtime_profile CRUD (member-read, admin-write) with
protocol_family whitelist validation, display_name uniqueness (409), and
fixed_args validation (no generic per-agent args — iron rule); a
daemon-token endpoint GET /api/daemon/workspaces/{id}/runtime-profiles;
DeleteRuntimeProfile does the app-layer cascade (delete instance rows then
profile, in one tx) and refuses (409) while active agents are bound.
- DaemonRegister accepts an optional per-runtime profile_id: validates the
profile belongs to the workspace and is enabled, registers via the
profile-aware upsert, and skips legacy hostname merge for custom rows.
AgentRuntimeResponse now carries profile_id.
Verified on Postgres 17: migrate up through 121; built-in + custom codex
coexist on one daemon; both upsert arbiters are idempotent; delete-by-profile
cascade removes only the custom instance; migrate down reverses 121 then 120
and replays clean. go build ./... and go vet pass; handler test package
compiles.
Daemon-side wiring (fetch profiles, PATH-resolve command_name, register with
profile_id, exec uses command_name) lands in a follow-up commit on this branch.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR2 (daemon): pull profiles, PATH-resolve, register, exec command
Daemon-side half of custom runtime profiles, against the server contract on
this branch.
- client.go: GetRuntimeProfiles(workspaceID) -> GET
/api/daemon/workspaces/{id}/runtime-profiles (mirrors GetWorkspaceRepos);
RuntimeProfile / RuntimeProfilesResponse types.
- types.go: Runtime gains profile_id (parsed from the register response so
runtimeIndex carries it).
- daemon.go:
* appendProfileRuntimes — called inside registerRuntimesForWorkspace before
the empty-runtimes guard. Best-effort fetch (older server 404s are logged
and swallowed; never fails registration). Per enabled profile: resolve
command_name via PATH (exec.LookPath, behind a `lookPath` test hook),
skip+log when absent, best-effort version probe, record the resolved
absolute path keyed by profile_id, and append a registration entry
{name, type=protocol_family, version, status:online, profile_id}. A
custom-only host (no built-in agents) still registers.
* profileCommandPaths map (guarded by d.mu) + recordProfileCommandPath /
customCommandPathForRuntime helpers.
* runTask: looks up the claimed task's RuntimeID -> profile command path and
overrides the executable path, synthesizing an AgentEntry so a custom
runtime runs even when the host has no built-in agent of the same
provider. provider (=protocol_family) is unchanged so agent.New still
selects the right backend.
- Tests: GetRuntimeProfiles request shape; profile runtime appended + path
recorded (custom-only host); profile skipped when command not on PATH;
profiles-fetch-404 is best-effort; customCommandPathForRuntime bookkeeping.
- agent: lockstep test pinning SupportedTypes to agent.New and the migration
120 protocol_family CHECK.
Iron rule honored: profile carries no generic per-agent args. fixed_args are
parsed and carried but intentionally NOT wired into the launch command yet
(optional/best-effort; explicit TODO(MUL-3284) in appendProfileRuntimes).
Verified: go build ./... clean; go vet ./internal/daemon/... clean;
go test ./internal/daemon/... pass (existing + 5 new); full
go test ./internal/handler/ suite passes against a migrated Postgres 17;
agent lockstep test passes.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284 PR2: profile delete runs full archived-agent cascade (fix 500)
Review fix. DeleteRuntimeProfile previously guarded only on ACTIVE agents, but
agent.runtime_id is ON DELETE RESTRICT — a profile whose runtimes had only
ARCHIVED agents passed the guard, then DeleteAgentRuntimesByProfile hit the FK
and the handler 500'd.
Now it mirrors the mature runtime-delete cascade (DeleteAgentRuntime): in one
transaction it enumerates the profile's runtime rows, refuses (409) any with
active agents or active squads led by archived agents, then for each runtime
pauses autopilots pinned to its archived agents, drops archived squads led by
them, and hard-deletes the archived agents before removing the runtime rows
and the profile. No code path can now fall through to a raw FK error.
- queries: ListAgentRuntimeIDsByProfile (sqlc regen). Reuses the existing
per-runtime teardown queries (CountActiveSquadsWithArchivedLeadersByRuntime,
ListArchivedAgentIDsByRuntime, PauseAutopilotsByAgentAssignees,
DeleteSquadsByArchivedAgentsOnRuntime, DeleteArchivedAgentsByRuntime).
- tests: TestDeleteRuntimeProfile_ArchivedAgentCascade (archived-only profile
deletes cleanly: 204, runtime + archived agent + profile gone) and
TestDeleteRuntimeProfile_ActiveAgentBlocks (active agent → 409, survives).
Verified against Postgres 17: both new tests pass; full handler suite, daemon
tests, and agent lockstep test pass; go vet clean.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284: add runtime_profile schema (custom runtime PR1)
Schema-only foundation for custom runtimes. Additive migration 120:
- New workspace-level `runtime_profile` table: the shared, team-visible
definition of a custom runtime (e.g. an in-house Codex wrapper).
protocol_family is CHECK-constrained to the exact backend list in
agent.New() (server/pkg/agent/agent.go). The only args column is
`fixed_args` (args every agent on the runtime must inherit); there is
deliberately no generic per-agent args field — those stay on
agent.custom_args.
- `agent_runtime.profile_id` (nullable, FK -> runtime_profile ON DELETE
CASCADE): NULL = built-in runtime, non-NULL = a registered instance of
a custom profile.
- Partial unique index agent_runtime_workspace_daemon_profile_key on
(workspace_id, daemon_id, profile_id) WHERE profile_id IS NOT NULL.
The legacy UNIQUE (workspace_id, daemon_id, provider) constraint is left
INTACT so the existing registration upsert
(ON CONFLICT (workspace_id, daemon_id, provider) in runtime.sql) keeps
resolving its arbiter and the server stays green. Converting that key to
a partial (WHERE profile_id IS NULL) index and making the upsert
profile-aware is PR2's registration work, not this migration.
Verified up + down against Postgres 17: full `migrate up` applies 120;
schema shows the table, column, partial index and intact legacy
constraint; functional checks pass (partial index blocks dup
(ws,daemon,profile), allows same profile on another daemon; CHECK and
display_name uniqueness reject bad input; legacy ON CONFLICT still
resolves; profile delete cascades to instances); down/up round-trip is
clean.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3284: drop DB FKs/cascade from runtime_profile migration (review fix)
Per review (house rule: no new database foreign keys / cascades; relational
integrity lives in the application layer):
- runtime_profile.workspace_id: drop REFERENCES workspace ON DELETE CASCADE
-> plain UUID NOT NULL.
- runtime_profile.created_by: drop REFERENCES "user" ON DELETE SET NULL
-> plain UUID.
- agent_runtime.profile_id: drop REFERENCES runtime_profile ON DELETE CASCADE
-> plain UUID.
CHECK constraints, UNIQUE (workspace_id, display_name), the workspace index,
and the partial unique index agent_runtime_workspace_daemon_profile_key are
unchanged. The legacy UNIQUE (workspace_id, daemon_id, provider) constraint
remains untouched.
Behavioral consequence: the database no longer auto-removes a profile's
agent_runtime instance rows on profile delete. That cleanup moves into PR2's
profile-delete path. Up-migration comments document this; down-migration
comment no longer references FKs/cascade.
Re-verified on Postgres 17: migrate up applies 120; no FK constraints exist on
the new columns; partial index still blocks dup (ws,daemon,profile_id); CHECK
and display_name uniqueness still reject bad input; deleting a profile now
leaves the runtime row orphaned (proving cascade is gone); down/up round-trip
clean with the legacy constraint intact.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
* fix(editor): don't wipe in-flight uploads on external content sync
When a brand-new chat's first file upload triggers lazy session creation,
`setActiveSession(null → uuid)` flips ChatInput's draft key mid-upload, which
changes `defaultValue` to the new (empty) session draft. ContentEditor's
"sync external defaultValue" effect then ran `setContent` over a document that
still held the `uploading` image/fileCard node, wiping it — so the upload's
finalize could no longer find the node. The file vanished and the draft was
left with an empty `!file[name]()`.
The editor was never remounted (instance stays alive); the node was removed by
the content-sync effect. An uploading node is local state an external sync must
not overwrite, exactly like the existing dirty/focused guards. Add a guard that
bails the sync while any `uploading` node is present.
Pure frontend; affects only the first upload in a new chat (subsequent uploads
hit an existing session, so no draft-key flip).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test(editor): cover the in-flight-upload content-sync guard
The content-sync effect now reads `editor.state.doc.descendants` on every run
to detect uploading nodes; the mocked editor didn't implement it, crashing all
ContentEditor tests. Add `descendants` (driven by `editorState.uploadingNodes`)
to the mock and a regression test asserting an external `defaultValue` change
does not setContent while an upload is in flight, and resumes once it settles.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(chat): migrate new-chat draft onto the session id on lazy create
The first file upload in a brand-new chat lazily creates the session, flipping
ChatInput's draft key from `__new__:agent` to the session id mid-upload. The
in-progress (empty-href) file-card markdown the editor had already written into
the `__new__:agent` draft was neither migrated nor cleared, so it stayed
stranded under that key — and resurfaced as a stale `!file[name]()` the next
time a new chat opened for the same agent (the send only cleared the
session-keyed draft).
Migrate the `__new__:agent` draft onto the new session id the moment the
session is created (upload path only — text send already clears the pre-flip
key via `keyAtSend`). Add a shared `newSessionDraftKey` helper so ChatInput and
ensureSession agree on the slot name, and a `migrateInputDraft` store action.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(execenv): switch agent prompt to --content-file to prevent heredoc flag swallowing (#4182)
The Linux/macOS reply template recommended --content-stdin with a quoted
HEREDOC. That pattern is safe for the trivial single-flag comment-add case
that BuildCommentReplyInstructions emits, but as soon as a model wraps
extra flags around the heredoc on multica issue create / update — assignee,
project — the bash heredoc/flag boundary is fragile in two ways the model
cannot see:
- A 'BODY \\' terminator with a trailing token is not recognised as the
heredoc end, so flag lines after it are swallowed into the description
(OXY-78: residual flag text leaked into the description, command exit 0).
- A clean terminator turns the trailing '--assignee ...' line into a
separate failing shell statement, while the create itself already exited
0 with no assignee (OXY-76: assignee silently dropped, no residual text).
In both cases the CLI never receives the swallowed flags, the API request
omits the fields, and the daemon has no visibility. The created issue lands
with assignee_id: null / project_id: null.
This commit:
* Switches the Linux/macOS branch of BuildCommentReplyInstructions to
--content-file with a 3-step recipe (write file, post, rm) so the body
never reaches the shell and all flags live on one shell-token line.
There is no heredoc boundary for flags to leak across.
* Adds a parallel cleanup step (Remove-Item) to the Windows branch so the
cross-platform template is one shape.
* Rewrites the runtime_config.go ## Comment Formatting non-Windows section
to mandate --content-file and explicitly ban --content-stdin HEREDOC for
agent-authored comments, citing #4182.
* Reorders the Available Commands menu lines for issue create / update /
comment add to put --content-file / --description-file ahead of the
stdin variant and add a per-line note pointing at #4182.
* Updates and renames the affected tests
(TestBuildCommentReplyInstructionsCodexLinux,
TestBuildCommentReplyInstructionsNonCodexLinux,
TestInjectRuntimeConfigLinuxCommentFormattingEmphasizesFile,
TestInjectRuntimeConfigIssueMetadataCodexFormattingUnchanged) so the
new file-first contract is pinned and the old HEREDOC mandate is in the
banned-strings lists.
This converges Linux/macOS with the long-standing Windows file-only path,
so the cross-platform guidance is now one shape. It also strictly improves
on the previous MUL-2904 guardrail by eliminating shell exposure of the
body entirely (no body ever reaches the shell, so backtick / $() / $VAR
substitution cannot corrupt it).
Closes GitHub multica-ai/multica#4182.
No CLI or backend changes — --content-file / --description-file already
exist.
Co-authored-by: multica-agent <github@multica.ai>
* docs(prompt): correct stale BuildPrompt comment to file-first (#4182)
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: CC-Girl <cc-girl@multica.ai>
* feat(lark): add proxy support for WebSocket connections
- Add Proxy field to GorillaDialer (func(*http.Request) (*url.URL, error))
- Default to http.ProxyFromEnvironment when Proxy is nil, so standard
HTTPS_PROXY/HTTP_PROXY/NO_PROXY env vars are respected automatically
- Allow explicit override via GorillaDialer.Proxy for custom proxy auth
or fixed proxy URLs
- Add unit tests for proxy defaults and error forwarding
Closes#4032
Co-authored-by: multica-agent <github@multica.ai>
* fix(lark): add missing net/url import in ws_connector_test.go
TestGorillaDialerProxyDefaults and TestGorillaDialerProxyForwardsError
use *url.URL in their Proxy func signatures but net/url was not imported.
Co-authored-by: multica-agent <github@multica.ai>
* fix(lark): preserve configured websocket proxy
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: J <j@multica.ai>
* fix(web): preserve CLI callback params across Google OAuth redirect
When 'multica login' runs in a headless/WSL2 environment, the CLI generates
a login URL with cli_callback and cli_state query parameters. These params
were being lost during the Google OAuth redirect because:
1. The login page did not encode cli_callback/cli_state into the Google
OAuth state parameter (only platform and nextUrl were included).
2. The callback page had no code path to redirect the JWT back to the
CLI's local HTTP listener after Google OAuth completed.
Fix:
- Login page: encode cli_callback and cli_state into the Google OAuth
state parameter alongside existing platform/nextUrl values.
- Callback page: parse cli_callback/cli_state from the returned state,
validate the callback URL, and redirect the JWT token to the CLI's
local HTTP listener after successful Google login.
Closes#3049
Co-authored-by: multica-agent <github@multica.ai>
* refactor(auth): reuse redirectToCliCallback helper in OAuth callback
Export the existing redirectToCliCallback helper from @multica/views/auth
and reuse it in the Google OAuth callback page instead of duplicating the
token+state redirect string inline, so the CLI callback URL contract lives
in one place.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: J <j@multica.ai>
* feat(analytics): capture JS exceptions to PostHog
Turn on posthog-js exception autocapture (window.onerror + unhandled
rejections, with stack) and add a buffered captureException() wrapper for
boundary-caught React errors those handlers can't see. Wire the web
route-level global-error boundary to report through it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(diagnostics): add shared freeze watchdog
Long-task observer (>=2s) emits client_unresponsive via captureEvent;
client_type super-property tags desktop vs web for free. Installed once in
CoreProvider so web and desktop share one in-thread, SSR-safe detector for
recoverable freezes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(desktop): report true hangs and crashes via breadcrumb
A real hang or crashed renderer can't report itself. The main process now
persists a breadcrumb on unresponsive / render-process-gone, and the next
renderer boot flushes it to PostHog (client_unresponsive / client_crash).
A recovered hang clears its breadcrumb so it isn't double-counted by the
in-thread watchdog.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(analytics): scrub PII from $exception before send
Error messages can interpolate user input (typed values, URLs with tokens).
Add a before_send hook that redacts emails, URL query strings, and long
opaque tokens from the exception message and $exception_list values, keeping
type + stack frames (code locations, not user data). Addresses the privacy
gap from leaving capture_exceptions on with no sanitizer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* test: cover breadcrumb state machine and freeze watchdog
The breadcrumb persist/clear orchestration is the correctness-critical part
and was untested. Cover: hang->write, recover->clear (no double-count),
recover-before-delay->no-op, force-quit->retained, crash->write-and-never-
clear, clean-exit->no-write. Add watchdog tests (threshold, idempotent,
SSR/PerformanceObserver no-op) via a fake observer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(desktop): breadcrumb field precedence + document limits
Spread the persisted context FIRST so explicit event fields (source,
recovered) always win over a future colliding context key. Document why
preload-error skips the breadcrumb and the single-slot last-write-wins
undercount limitation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two related cleanups to the issue comment/reply/edit composer:
- Drop the trigger-preview "context" copy added in #4147 (chip prefix
`trigger_context_*` and per-context popover titles `trigger_preview_title_*`).
The actual "align context" fix in #4147 was the backend/hook work; the copy
was redundant decoration. Removes the `context` prop, the dead i18n keys
across en/zh/ja/ko, and the corresponding test assertions; the popover title
falls back to the original single `trigger_preview_title`.
- Edit-comment footer: lay the trigger chip on a single row with the action
cluster (📎 Cancel Save) on the right, attachments on their own full-width
row above. The 📎 now sits with the action buttons, matching the new-comment
and reply composers.
- Unify composer buttons on shadcn `Button`: `FileUploadButton` renders a
ghost icon button instead of a hand-rolled circle, and the reply submit
button uses `Button` (icon-xs, ghost-when-empty / primary-when-typed) instead
of a hand-rolled element. Sizes: 📎 and reply submit are both icon-xs (24px).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Clicking a row's ⋯ kebab (or any in-row control) full-page reloaded the
app. The row was a whole-row <AppLink>, so a child's stopPropagation
stopped the event before AppLink's onClick (which calls preventDefault to
cancel native anchor navigation and do an SPA push) could run — leaving
the browser to perform the native <a> navigation, i.e. a full reload. It
was also invalid HTML: interactive content (button/menu) nested in an <a>.
Rework all five ListGrid row surfaces (agents, runtimes, skills,
autopilots, squads) to a plain <div> row whose whole-row navigation is a
mouse onClick (new useRowLink hook): left-click pushes, cmd/ctrl/middle
opens a background tab. Interactive cells (checkbox, kebab) stopPropagation
so they never trigger row nav — and with no <a> ancestor there is no native
navigation to cancel, so the reload class of bug is gone. Names are plain
text since the row itself is the click target. projects is unchanged — its
inline-editable cells make it a deliberate name-link exception.
Also fixes two adjacent defects found in the same menus:
- agents/runtimes kebab triggers reused the shared <Button>, which lacks
the data-popup-open styling the other surfaces have, so the trigger
vanished and lost its background while its menu was open. Switch them to
the bare-button trigger with data-popup-open: visible + highlighted.
- agents archive menu items used className="text-destructive" instead of
variant="destructive", so the base focus style overrode the red on hover.
Switch to variant (list row + detail page).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(issues): render thread replies in chronological order (#3691)
collectThreadReplies walked the parent_id tree depth-first, so an agent
reply forced to nest under its trigger comment rendered before earlier
sibling replies (A-D-B-C instead of A-B-C-D) whenever the agent returned
late. Sort the collected subtree by created_at (id tie-break) so the
thread reads in arrival order — the same order the server already feeds
agents via `comment list --thread` (ListThreadCommentsForIssue).
All other consumers of the array (resolution derivation, fold bars,
counts, deep-link) are order-independent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): rebuild skills list on shared Linear-style list grid
- new ListGrid primitives (subgrid: single source of truth for column tracks)
- skills list: sortable columns, used-by avatar stack, source/creator columns,
row kebab + batch toolbar with add-to-agent and delete
- skill view store in core; addAgentSkills client method; HoverCheck extracted
to views/common (issues header now imports the shared copy)
- locale keys for list actions/filters and the reworked detail page
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): rework detail page into overview/files tabs
- tabs directly under the breadcrumb header: overview (default) and files
- overview: identity block + rendered SKILL.md as the main column, right
rail with metadata card (source/creator/updated, inline name+description
edit toggle) and used-by panel with bind/unbind
- files: file tree + viewer/editor unchanged; SKILL.md "edit" jumps here
- header kebab menu (copy skill ID, delete); page-level save bar shared by
both tabs; tab state persisted in ?tab=
- file tree: ARIA tree roles + roving-tabindex keyboard navigation
- drop the old right sidebar (metadata dl, permissions paragraph)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* revert(skills): restore detail page to main, keep branch list-only
Drop the overview/files tabs rework from this branch so the PR scope is
the list rebuild only. skill-detail-page.tsx and file-tree.tsx are back
to the main versions; the locale detail/file_tree sections are restored
to match. The detail rework is preserved on stash/skills-detail-tabs
for a follow-up PR.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): drop description column from skills list
Description is agent-facing routing metadata, not a scannable list
property — Linear's display options expose no description column for
the same reason. Removes the cell, column key, display toggle, lg grid
track, skeleton cells, and the now-dead table.description /
table.no_description locale keys.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): drive list column hiding by container width, drop by priority
Replace viewport sm:/lg: breakpoints with Tailwind v4 container query
variants (@2xl/@4xl) on the list wrapper, so an open sidebar or split
pane narrows the column set instead of squashing tracks. Remove the
min-w-fit + overflow-x-auto horizontal-scroll fallback: when space runs
out, low-priority columns (created/source/creator, then updated) drop
and return as the container widens; name and usedBy never drop. ListGrid
conventions comment updated — this is the template for all list pages.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): virtualize list rows with @tanstack/react-virtual
Linear-style headless virtualization: the virtualizer computes the
visible index range and offsets; offsets land as padding on the
scrolling ListGridBody so mounted rows stay direct subgrid children and
column alignment is untouched. Fixed 48px rows skip per-row measurement.
Hideable column tracks move from max-content to deterministic widths
(CSS vars) — with only the visible slice mounted, content-driven tracks
would resize during scroll. A user-hidden column zeroes its var so the
track still collapses; per-cell max-w caps move into the tracks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(skills): list tiers must fit their container trigger width
The @4xl tier's track sum (~1080px with gaps) exceeded its 896px
trigger; with the horizontal-scroll fallback gone, the right-side
columns were clipped unreachably between 896-1080px. Move tier 3 to
@5xl (1024px), trim usedBy/source/creator tracks, and document the
fit invariant with its arithmetic next to the template and in the
ListGrid conventions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): show description as subtext under the skill name
Lives in the name track as a second truncated line (max-w 36rem,
title attr for the full text) — no track, no header, no slot in the
responsive arithmetic. Both lines fit the fixed 48px row, so the
virtualizer contract is untouched; rows without a description center
the name.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Revert "feat(skills): show description as subtext under the skill name"
This reverts commit f39721301b.
* fix(skills): anchor batch toolbar to the page, not the viewport
fixed bottom-6 left-1/2 centered the bar on the window; with the
sidebar open the list's visual center sits ~120px right of the window
center, so the bar looked off-center (worse with desktop split panes).
Page root becomes the positioning context (relative) and the bar uses
absolute — same rule applies to future list pages.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills): show matching count next to search while list is narrowed
"n / total" appears right of the search box only when search or
filters are active — idle state would duplicate the total already in
the page header.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(autopilots): derive trigger kinds, next run, last run status in list
The list endpoint only selected the autopilot table, so the list UI
could not answer "is this automation working" without N+1 detail
calls. Each list row now carries trigger_kinds + next_run_at (enabled
triggers only — the columns describe how it fires today) and
last_run_status (most recent run). Fields are omitempty and absent
from detail/create/update responses; clients must treat them as
optional per the API compatibility rules.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(autopilots): list schema, parsed client, and view store in core
- listAutopilots now runs through parseWithFallback with a zod schema
(this endpoint was a bare fetch — overdue per the API compatibility
rules); malformed bodies degrade to an empty list, old-server rows
without assignee_type or the new derived fields parse cleanly, and
enum drift passes through as plain strings
- Autopilot type gains the three optional list-only derived fields
- New autopilots view store (scope/sort/columns/filters, persisted per
workspace): status is the promoted scope dimension so it does NOT
appear in filters — one dimension lives in exactly one place
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(autopilots): rebuild list on shared ListGrid with scope buttons
Same skeleton as the skills list (container-query tiers, deterministic
var-width tracks with documented fit arithmetic, virtualized 48px rows,
sortable headers, filter + display toolbar, page-anchored batch
toolbar), plus the autopilots-specific pieces:
- Status is the promoted SCOPE dimension: 全部/运行中/已暂停/已归档
segmented buttons with full-set counts; "all" = active+paused
(archived gets its own visible home, Linear archive semantics);
status is therefore absent from the filter dropdown
- Columns: name (paused marker inline), assignee (agent/squad),
trigger kind badges, last run (outcome dot + time, enum-drift safe
default), next run; mode/creator/created opt-in hidden
- Filters: assignee, trigger kind, mode, creator (composite type:id
values for polymorphic actors); sort name/lastRun/nextRun/created
with lastRun desc default
- Row kebab (pause/resume/archive/unarchive/delete) and batch toolbar
share one delete dialog; status changes ride useUpdateAutopilot's
optimistic cache
- Fix noUncheckedIndexedAccess errors the branch had never typechecked
(skills virtual rows, UsedByCell, added_toast)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(autopilots): scope buttons follow the issues header pattern
Replace the bespoke segmented-pill control with the existing scope
button convention from the issues page: outline buttons with bg-accent
active state on md+, collapsing to a radio dropdown below md. Counts
stay (stage inventories from the full set).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(skills,autopilots): toolbar small-screen treatment follows issues header
Below md: the search box (and its result count) disappear entirely,
and the filter/display controls collapse to square icon-only buttons
(labels and the clear-X are md+), matching the issues header's
responsive pattern.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(skills,autopilots): two-zone columns — WYSIWYG with scroll escape valve
Static width tiers silently hid user-enabled columns (toggle on,
nothing appears — autopilots' mode/creator/created sat behind a 1280px
container gate no laptop reaches; skills' source/created behind
1024px). Tiers can't know how many columns are enabled, so the
mechanism is replaced, not retuned:
- ≥@2xl container: every enabled column renders; the grid carries
min-width = Σ(enabled tracks + gaps) (pure constants, no
measurement) and the wrapper scrolls horizontally only when the
enabled set outgrows the container
- <@2xl: static core set (skills: name+usedBy; autopilots:
name+assignee), no scroll, toggles don't apply
Per-tier templates and the hand-maintained fit arithmetic retire;
ListGrid conventions updated accordingly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(skills,autopilots): widen name column minimums (120px base, 200px wide)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(autopilots): drop the archived scope and the list search box
Archiving never existed as a UI flow (the DB status value is only
reachable via direct API; the detail page disables its switch when
archived), so the list stops inventing it: no archived scope, no
archive/unarchive row or batch actions. API-archived rows are excluded
everywhere; a persisted retired scope value falls back to "all".
The search box goes too — scope buttons already partition the small
set, search is redundant (product call). Skills keeps its search (no
scope there).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(skills,autopilots): quiet outline create buttons in page headers
Page-header chrome shouldn't carry the loudest element on the page:
the create button becomes outline with text on md+ and collapses to a
square plus icon below md (same responsive treatment as the toolbar
controls). Primary stays reserved for empty-state CTAs. Agents follows
when its list migrates to ListGrid.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(agents): rebuild list on shared ListGrid with identity rows
Same skeleton as the skills/autopilots lists (two-zone container
responsiveness, deterministic var tracks + min-width scroll escape
valve, virtualized fixed-height rows, issues-style scope buttons,
page-anchored batch toolbar, quiet outline create button), plus the
agents-specific decisions:
- Identity rows: the documented exception to the single-line rule —
avatar + name + description two-line cells, 64px rows (agents are
few, identity-rich entities); the italic "no description"
placeholder is gone, empty descriptions just center the name
- Scope: Mine (historical default) | All | Archived with full-set
counts; archived ignores the ownership lens; no search box
- The 7d sparkline column is replaced by a sortable "Last active"
column derived from the same 30-day activity buckets (zero API
change) — per-row-normalized mini bars can't be compared across
rows, and the default sort finally has a visible anchor; the
detailed histogram stays on the hover card / detail page
- Workload folds into the status cell ("Online · 2 tasks") — a 0-2
integer doesn't earn a column
- Columns: status, runtime, last active, runs (30d); model/created
opt-in hidden; filters: availability, runtime
- Operations unchanged: row kebab reuses AgentRowActions
(cancel-tasks/duplicate/archive/restore with permissions); batch
archive (confirmed) + restore; no delete — the API has none
- View store extended (scope incl. archived, sort, columns, filters);
agent-columns.tsx (DataTable columns) deleted
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): trim status track to its real worst case (160 -> 144px)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(runtimes): machine detail's runtime table on the shared ListGrid
The master-detail console keeps its shape (machines are few and
strongly categorized; left list, charts, update section untouched) —
only the right pane's runtimes table moves from TanStack DataTable to
the ListGrid family, taking the paradigm pieces that earn their keep
at 1-5 rows: subgrid template + var tracks, two-zone container
responsiveness (the pane is squeezed by the machine list, so the
core-set collapse below @2xl matters more here than on full-width
pages), min-width scroll escape valve, shared header/row/hover visual
language. Deliberately NOT taken: virtualization, sorting, filters,
column toggles, and batch selection — dead weight at this row count,
and batch-deleting runtimes (a cascade-confirm operation) is unsafe
by design.
Workload folds into the health cell ("Online · Working 2") like the
agents status cell; the owner column keeps its only-when-multiple-
owners rule via a zeroed track var. runtime-columns.tsx is deleted;
the row-menu/CLI tests render the exported cells directly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(runtimes): collapse the kebab track when no row has actions
On a healthy local machine every row's only action (delete) is hidden
by the self-healing rule, leaving a permanent ~64px dead zone after
the CLI column. The action track now follows the owner column's
conditional-var mechanism: zeroed unless at least one row will show
the menu.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(runtimes): drop doubled header border, align create button with convention
PageHeader already carries border-b; the content wrappers' border-t
stacked a second line right under it (the only list page doing this).
"Add a computer" follows the chrome-button convention: outline with
text on md+, square plus icon below md — primary stays reserved for
the empty state CTA.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(runtimes): health cell load suffix matches the agents status cell
"Healthy · 2 tasks" instead of the old workload vocabulary
("Working 2 +1q") — the count is unit-bearing and both surfaces now
speak one language. The queued-anomaly distinction the old words
hinted at belongs to the health layer if it ever earns surfacing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(lists): pin overflow-y-hidden on the horizontal-scroll wrappers
CSS coerces overflow-x:auto into overflow:auto on both axes, which
silently armed the list wrappers with a vertical scrollbar they were
never meant to have. Combined with the h-full grid's percentage
resolution across scrollbar-induced reflows, the wrapper's vertical
bar and horizontal bar fed each other in a non-converging layout loop
(visible as two stacked, flickering scrollbars on the agents list —
the same latent loop exists in all four wrappers; agents' wider
min-width and 64px rows just hit the trigger zone first). Vertical
scrolling belongs solely to ListGridBody; declare overflow-y-hidden
explicitly to break the loop.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): single scroll container for the list (trial before rollout)
Both scroll axes move to the outer wrapper; the grid drops h-full and
the rows wrapper drops its own overflow. Kills the percentage-height
bridge between the two scroll elements that fed the flickering double
scrollbars and clipped the last row under the horizontal scrollbar.
Sticky header pins inside the scroller; vertical scrollbar now spans
the full pane (Linear's structure). Skills/autopilots follow after
visual confirmation.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(lists): roll single scroll container out to skills/autopilots, add bottom clearance
ListGridBody retires its own scrolling entirely (the agents trial
confirmed the structure): both axes live on the single outer wrapper,
grids drop the h-full percentage bridge, virtualizers point at the
wrapper. The rows wrapper gains LIST_GRID_BOTTOM_CLEARANCE (64px)
appended to the virtualization padding so the last row scrolls clear
of the chat FAB (~48px at bottom-right) and the batch toolbar (~62px).
Runtimes' machine table is untouched: content-height at the top of a
tall pane, no bridge and no practical FAB overlap.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(squads): rebuild list on shared ListGrid (identity rows, minimal)
The last list joins the family. Squads are the fewest entity (1-5 rows),
so this is the agents identity-row shell on the runtime-list minimal
skeleton: ListGrid subgrid + var tracks + two-zone responsiveness +
single scroll container, but NO virtualization, checkbox, or batch.
- Identity two-line rows (squad avatar + name + description, 64px) like
agents; columns: name / leader / members (polymorphic ActorAvatar
stack from member_preview), creator + created opt-in hidden
- Scope Mine/All (creator-based, issues-header styling, <md dropdown);
no archived scope (list API hard-filters archived + no restore
endpoint), no search (scope-bearing), no filters (set too small)
- Sort name (default) / members / created
- Row kebab = Archive (= the delete endpoint, which archives + transfers
issues/autopilots to the leader); workspace owner/admin only, so the
kebab track collapses for non-admins. Reuses the existing
archive_dialog copy. No batch.
- View store extended (scope + sort + columns); zero API change — pure
frontend (member_preview/count already in the list payload)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(agents,squads): owner/created-by columns + owner filter
Surface ownership as a real column on both lists, named by what the
field actually means in each permission model:
- Agents: "Owner" — owner_id is the creator (set at creation, never
transferred) and carries management rights. Promoted to a default-
visible column (avatar + name); the half-baked inline owner avatar in
the name cell is removed ("You" badge stays).
- Squads: "Created by" (NOT Owner) — creator_id holds no rights
(archiving is workspace-admin only), so Owner would mislead. Now a
default-visible column with avatar + name.
Agents also gains an Owner filter, kept orthogonal to the Mine scope by
the single-axis rule: "Mine" is the clean no-filter personal view, so
applying any filter (owner or otherwise) leaves Mine for All, and
clicking Mine clears all filters. Owner and Mine therefore never
coexist — no "mine + owner=someone-else = empty" contradiction. Squads
keep the plain Mine/All toggle (too few rows for a creator filter).
Both lists keep a Created (date) column, opt-in hidden.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): backfill new filter dimensions on rehydrate (owners crash)
A view payload persisted before the owners filter existed overwrote the
default filters wholesale on rehydrate, dropping filters.owners to
undefined and crashing the list's filter predicate (.length on
undefined). The store merge now deep-merges filters over
EMPTY_AGENT_FILTERS so newly-added dimensions always get their default.
Regression test added.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(skills,autopilots): deep-merge filters on rehydrate too
Same latent crash the agents store just hit: the copied view-store
merge spread persisted.filters wholesale, so adding a new filter
dimension later would drop it to undefined for users with older
persisted state. Harden skills and autopilots the same way (merge over
their EMPTY_*_FILTERS) before that bug can ship.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(projects): rebuild table view on ListGrid + filters + pin/delete kebab
Projects is the dual-view list: the compact table moves onto the shared
ListGrid (subgrid tracks, two-zone responsiveness, single scroll
container, FAB bottom clearance) while the comfortable card grid stays
as the alternate view, toggled by a restyled view switch (Table/Cards
outline buttons, active = bg-accent). Inline editing is preserved —
rows are NOT whole-row links; the name navigates and status/priority/
lead stay click-to-edit (matching prior behaviour, no navigate-vs-edit
conflict).
- View store extended: viewMode + sort (name/priority/status/progress/
created) + hidden columns + filters (status/priority/lead); merge
deep-merges filters (migration-safe). No scope (lead optional/often
an agent; status is a 5-value lifecycle → filter, not scope).
- Toolbar: search (kept — scopeless list) + result count + Filter
(status/priority/lead) + Display (sort+columns, table view only).
- Row kebab: Pin/Unpin (any member, reuses the existing project pin
API — zero new endpoints) + Delete (workspace admin). Pin is the
flexible per-user favourite the list previously lacked.
- Zero API change; status/priority filtering is client-side like the
other lists.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(projects): GRID_COLS must be a literal string (Tailwind can't see interpolation)
The table view's grid-cols template interpolated ${STATUS_WIDTH}px, so
Tailwind never generated the arbitrary-value class — the grid collapsed
to one column and every cell stacked vertically. Inline the literal
116px. This is the documented ListGrid rule (keep the class literal so
Tailwind scans it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(projects): single view-toggle button, decouple Display from view mode
Two fixes from the same principle — view mode is pure presentation and
must not couple to anything:
- The view switch is now ONE button that flips table ⇄ cards (shows the
current view's icon+label, tooltip names the target), instead of two
side-by-side buttons.
- The Display (sort/columns) control no longer disappears when you
switch to cards — it was gated on isCompact, so flipping the view
made it vanish (the "filter gone after switching" weirdness). It's
always present now; only the columns *section* inside the popover is
table-only (cards have no columns). Sort applies to both views.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(projects,squads): projects multi-select + squads FAB clearance/toast
Cross-list consistency audit fixes:
- projects: add multi-select (checkbox column + select-all header +
page-anchored batch toolbar) — it's a dozens-scale full-page list
like skills/autopilots/agents but was the only one missing it. Batch
ops: Pin all (any member) + Delete (workspace admin). Table view
only (cards have no checkboxes). GRID template + min-width updated
for the checkbox track.
- squads: add the FAB bottom clearance the other full-page lists have
(last row/kebab was sliding under the chat FAB).
- squads: archive success toast was showing the dialog's question
title ("Archive this squad?"); use a proper "Squad archived" key.
Intentional and left as-is (documented): squads/runtimes have no
multi-select/virtualization (1-5 rows); projects table isn't
virtualized yet (dual-view + card grid; tracked as low-risk debt).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(agents,squads): close the filter/column consistency gaps
Apply the principle "every categorical column is filterable" where it
was missing:
- agents: add a Model filter (model was a categorical column with no
filter). Distinct non-empty models from the in-scope rows.
- squads: add filters entirely (it had leader/creator columns + a
column-toggle panel but no Filter button — the only such outlier).
Leader (agent) + Creator (member) filters, with the result count and
the same Filter dropdown shape as the other lists. Store gains
SquadListFilters + toggleFilter/clearFilters + migration-safe
filters deep-merge.
autopilots creator stays default-hidden per product call (not every
"who made it" must be visible). Filter stores' partialize tests
updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(autopilots): match list-page root to flex-1 convention
skills/agents/projects roots use `relative flex flex-1 min-h-0 flex-col`;
autopilots used `h-full`. Both anchor the batch toolbar correctly, but
align the flex sizing for consistency across the six list surfaces.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fixes Cursor agent token usage parsing for top-level camelCase, nested camelCase, and legacy nested snake_case result usage shapes. Includes tests for the locally verified nested camelCase stream-json output.
The two prior MUL-3254 fixes preserved draft/description state across a
modal close, but Desktop still could not RENDER the reopened image: in
CloudFront signed-URL mode every URL the renderer holds after reopen is
unloadable. The persisted record strips the expired signed download_url,
the raw CDN url is unsigned (403 on a signed distribution), and the
durable /api/attachments/<id>/download endpoint needs credentials that a
cross-site file:// <img> fetch cannot carry (web works via the same-site
session cookie, which is why the bug was desktop-only).
Two changes close the last mile:
- /api/config now reports cdn_signed when CloudFront signing is enabled,
and pickInlineMediaURL stops picking the raw (unsigned) CDN url in
that mode — it is a guaranteed 403.
- The Attachment renderer upgrades an auth-gated media URL to a freshly
signed one via authenticated GET /api/attachments/<id> (the same
re-sign the click-time download path already does), but only on
clients without a same-origin /api proxy (api.getBaseUrl() non-empty:
Desktop, mobile webview). Cached via TanStack Query with a 20-minute
staleTime, inside the server's 30-minute signed-URL TTL.
Old servers omit cdn_signed; the schema defaults it to false so behavior
is unchanged there. Non-CloudFront deployments return the API path again
from the metadata fetch and the renderer keeps the original URL.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(openclaw): support connecting to existing OpenClaw gateway (#3260)
When the daemon host is a lightweight dev machine or CI coordinator, the
heavy agent work (LLM inference, code execution, tool use) often belongs
on a more powerful remote server already running an OpenClaw gateway.
Multica historically hard-coded `openclaw agent --local`, forcing every
turn to execute in-process on the daemon host.
This change adds an opt-in gateway routing mode controlled per-agent via
`runtime_config`:
{
"mode": "gateway",
"gateway": { "host": "...", "port": 18789, "token": "...", "tls": false }
}
- Backend: ExecOptions gains OpenclawMode + OpenclawGateway; buildOpenclawArgs
drops `--local` when mode == "gateway". Per-task openclaw-config.json
wrapper pins gateway.{host,port,auth.{mode,token},tls} so users do not
need to edit the daemon host's `~/.openclaw/openclaw.json` to point at
a different endpoint.
- Daemon: AgentData carries the raw runtime_config; decoding is fail-soft
(malformed JSON falls back to local mode rather than blocking dispatch).
- API: gateway.token is masked to "***" on every GET; PATCH replays the
sentinel back, and the update handler restores the persisted token so
the round-trip never destroys the secret. Defense-in-depth masking on
WS broadcasts, plus String/MarshalJSON masking on the in-memory struct
to block stray `%+v` / json.Marshal leaks.
- UI: openclaw-only "Routing" tab on the agent detail page with mode
selector + structured endpoint form. Token uses a "saved — submit a
new value to rotate" UX and matching backend preserve hook.
Empty `runtime_config` keeps the historical embedded behaviour, so
existing agents are unaffected.
* fix(openclaw): address #3664 review — drop dead gateway field, gate pin on mode
Per Bohan-J's review:
- Remove the dead ExecOptions.OpenclawGateway field (+ its String/MarshalJSON and
the daemon.go construction block). It carried the plaintext bearer token but was
never read — buildOpenclawArgs only consumes OpenclawMode and the live gateway
path runs through execenv.OpenclawGatewayPin — so this narrows the secret's
footprint.
- Gate the gateway pin on mode=="gateway" in decodeOpenclawRuntimeConfig: a
{"mode":"local","gateway":{...,"token"}} payload no longer writes the token into
the 0o600 per-task wrapper that --local makes openclaw ignore.
- Warn on an unrecognized non-empty mode (e.g. "gatway") instead of silently
falling back to local.
- Run preserveMaskedGatewayToken in CreateAgent too, so a literal "***" at create
time can't persist as a real bearer token.
- Document the gateway host:port trust boundary (SSRF note for shared daemon hosts).
Adds regression tests for the local-mode pin drop and the unknown-mode warning.
* fix: flush issue description editor on close
Co-authored-by: multica-agent <github@multica.ai>
* fix: make unmount flush opt-in via flushPendingOnUnmount
The unconditional unmount flush re-emitted discarded content into
composers that clear their draft and then unmount (comment edit cancel,
create-issue / feedback submit), resurrecting the cleared draft.
- Add flushPendingOnUnmount prop (default false); only the issue-detail
description editor opts in.
- Cache the pending markdown in a ref at onUpdate time and emit that
cached copy on unmount, instead of reading the editor instance during
teardown.
- Regression tests: default drops the pending update on unmount, opt-in
flush emits the cached value even when the editor is already
destroyed, no double-emit after the debounce fired, and issue-detail
pins the opt-in wiring.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
remark-math defaults to singleDollarTextMath: true, so any paragraph
containing two dollar amounts (e.g. "costs $120/mo (~$85 net)") has
the text between them parsed as inline TeX and rendered by KaTeX in an
italic math font, with ~ treated as a non-breaking space. Disable
single-dollar parsing in both web render paths, matching GitHub's
behavior; explicit $$...$$ math still renders.
Co-authored-by: Matt Voska <voska@users.noreply.github.com>
* fix: keep issue draft attachment records
Co-authored-by: multica-agent <github@multica.ai>
* fix: avoid persisting signed draft attachment urls
Co-authored-by: multica-agent <github@multica.ai>
* fix: reuse resolved media url for draft previews
Co-authored-by: multica-agent <github@multica.ai>
* fix: address draft attachment review nits
- Backfill an empty caller download_url from the in-session upload on id
collision so a just-pasted image first-paints from the signed URL
instead of detouring through markdown_url.
- Prune draft attachments no longer referenced by the persisted
description when the create dialog reopens.
- Backfill EMPTY_DRAFT defaults on draft-store rehydrate so drafts
persisted before the attachments field existed get a stable shape.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The comment card exposed two identical add-reaction affordances: a
QuickEmojiPicker in the header's top-right actions and the add button
inside the bottom ReactionBar. Keep only the bottom one.
- Drop QuickEmojiPicker from the root header and reply-row headers
- Always show the ReactionBar add button (it is the only entry point
now), removing the isLongContent gating
- Remove the now-unused hideAddButton prop from ReactionBar
MUL-3262
Co-authored-by: Lambda <lambda@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Adds migration 119 creating idx_user_created_at on "user"(created_at)
using CREATE INDEX CONCURRENTLY, matching the repo convention for
index-only migrations (114/115).
Co-authored-by: multica-agent <github@multica.ai>
Fixes GitHub issue #3999 by moving the daemon StartTask transition behind workdir provisioning and extending the active env-root guard through completion metadata writes.
- suggest other profile workspace roots when disk-usage sees an empty selected root
- include the default profile in reverse suggestions and shell-quote profile arguments
- keep JSON output and explicit --workspaces-root behavior unchanged
MUL-3232
* feat(skills): structured conflict + overwrite path for local skill re-import
Local-skill re-import previously failed (or silently skipped) on a same-name
collision and, on delete+reimport, changed the skill UUID and dropped agent
bindings. This adds a structured conflict result and a creator-only overwrite
write path so a re-import can update the existing skill in place.
- New terminal import status `conflict` carrying { existing_skill_id,
existing_created_by, can_overwrite }; can_overwrite = requester is the
skill creator (canOverwriteSkillByLocalImport — intentionally narrower than
canManageSkill: admins edit in-app, not via re-import).
- Conflict is detected at daemon-report time (the effective name is only known
once the bundle arrives) via GetSkillByWorkspaceAndName, with the unique
constraint as a race backstop.
- Import requests carry action=overwrite + target_skill_id, persisted through
both the in-memory and Redis LocalSkillImportStore (the heartbeat → daemon
payload is unchanged; overwrite is resolved server-side).
- overwriteSkillWithFiles updates by target_skill_id in one tx: re-checks
existence (workspace-scoped) and creator permission, then replaces
description/content/config and fully replaces files (pruning files absent
from the new bundle). Preserves id, created_by, created_at, name, and
agent_skill bindings. Publishes skill:updated (not skill:created).
- Boundaries: target deleted or permission lost → failed (no fallback to
create-by-name); any mid-write error rolls back the tx, leaving the original
skill untouched. Retrying a terminal request is a no-op.
Tests cover: creator/non-creator conflict (can_overwrite), overwrite preserves
UUID + agent binding + prunes removed files, non-creator overwrite fails,
deleted target fails without create fallback, retry idempotency, and Redis
round-trip of the new fields.
Backend half of MUL-2701. Contract change: same-name local imports now return
status `conflict` instead of `failed` — the Desktop/core client must be updated
to consume it (sibling task).
MUL-2800
Co-authored-by: multica-agent <github@multica.ai>
* fix(skills): gate structured conflict behind client opt-in; guard overwrite target name
Addresses review feedback on PR #3498 (MUL-2800).
Backward compatibility: a same-name local import now returns the new `conflict`
status only when the initiating client opts in via `supports_conflict` (an
overwrite request implies it). Older clients — already-installed Desktop builds
whose poll loop only understands `failed`/`timeout` — keep the legacy `failed`
+ "a skill with this name already exists" behavior, so upgrading the backend
ahead of the client no longer regresses the import UX. This is the installed-app
API-compat boundary the repo's CLAUDE.md calls out.
Also: the overwrite write path now verifies the incoming effective name matches
the target skill's current name (errSkillOverwriteNameMismatch -> failed),
preventing a stale/wrong target_skill_id from writing one skill's content onto
another. Creator-only + workspace scoping already prevent privilege escalation;
this narrows the API so it can't be misused.
Refactored LocalSkillImportStore.Create to a LocalSkillImportRequestInput params
struct (the signature had grown to 8 positional args; the opt-in flag pushed it
over). supports_conflict is persisted in both the in-memory and Redis stores.
Tests: conflict tests now opt in; added a legacy-client test (no flag ->
failed + legacy message) and an overwrite name-mismatch test.
MUL-2800
Co-authored-by: multica-agent <github@multica.ai>
* feat(skills): resolve local import conflicts in desktop
Co-authored-by: multica-agent <github@multica.ai>
* fix(skills): preserve bulk flow after conflict resolution
Co-authored-by: multica-agent <github@multica.ai>
* fix(skills): show creator name instead of UUID in import conflict UI
When a local skill import hits a name conflict with a skill owned by
another user, the locked-creator message rendered the raw
existing_created_by UUID via the {{creator}} placeholder, which is
unreadable.
Resolve the UUID against the workspace member list and render the
display name instead. When the creator has left the workspace (or the
member list hasn't loaded), fall back to the unbranded conflict_locked
message rather than leak the UUID.
Adds two test cases covering both branches.
MUL-2701
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: Eve <eve@multica-ai.local>
Two bugs prevented the Lark binding flow from completing for already-logged-in users:
1. The useEffect ran before AuthInitializer's getMe() returned, setting state to
needs-auth; the guard then blocked re-entry once auth loaded.
2. The sign-in redirect used ?redirect= but the login page reads ?next=.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
collectThreadReplies walked the parent_id tree depth-first, so an agent
reply forced to nest under its trigger comment rendered before earlier
sibling replies (A-D-B-C instead of A-B-C-D) whenever the agent returned
late. Sort the collected subtree by created_at (id tie-break) so the
thread reads in arrival order — the same order the server already feeds
agents via `comment list --thread` (ListThreadCommentsForIssue).
All other consumers of the array (resolution derivation, fold bars,
counts, deep-link) are order-independent.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* fix(agent): clear stale session id when a resumed ACP session is gone
When an agent's stored ACP session no longer exists on the runtime side,
session/resume still succeeds — hermes echoes the requested sessionId
back — so the failure only surfaces when session/prompt returns JSON-RPC
-32603 "Session not found". The backend then reported Status=failed with
the stale SessionID still set, which kept the daemon's resume-failure
fallback (gated on SessionID == "") from ever firing. The failed task
never updates the stored session, so every future mention on the same
(agent, issue) dispatched against the same dead id, forever (#4010).
handleResponse now returns a structured acpRPCError instead of a flat
string (rendered text unchanged), and the hermes/kimi/kiro prompt-error
paths clear the session id when the error is session-not-found class on
a resumed session. The daemon's existing retry then re-executes with a
fresh session and stores the replacement id, healing the mapping.
* fix(agent): clear stale session id when set_model hits a dead resumed session
With a model override, session/set_model runs before session/prompt,
so a resumed session that is gone on the agent side surfaces there
instead of at the prompt — and the error branch returned the stale
SessionID, so the daemon's fresh-session retry (gated on
SessionID == "") never fired. Apply the same clear-the-id fix in the
set_model error branch of all three backends.
Also relax isACPSessionNotFound to accept -32602: kimi-cli raises
RequestError.invalid_params({"session_id": "Session not found"}) for
every unknown-session path (src/kimi_cli/acp/server.py), so pinning
-32603 made the fix dead code for kimi. The wording gate keeps
unrelated invalid_params errors (e.g. "model not available") on the
preserve-the-id path.
Regression tests for all three backends: resumed session + model
override + set_model failing with each runtime's observed
session-not-found shape must yield status=failed with an empty
SessionID.
CLI backends key their session stores to the cwd (Claude Code looks
sessions up under ~/.claude/projects/<encoded-cwd>/), so a prior session
id can only resolve when the task runs in the exact workdir the session
was recorded against. When the prior workdir no longer exists (GC'd
after the issue went done, daemon reinstall, manual cleanup),
execenv.Reuse falls back to a fresh Prepare but the stale session id was
still passed to the backend: claude exited within a second and the run
failed before doing any work — permanently, because the failed run
records no session_id and the next claim serves the same stale pointer
again.
Gate ResumeSessionID on the workdir actually being reused, and correct
PriorSessionResumed so the runtime brief uses the cold-path wording when
the session is dropped.
Fixesmultica-ai/multica#3854 (MUL-3221)
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* feat(skills): structured conflict + overwrite path for local skill re-import
Local-skill re-import previously failed (or silently skipped) on a same-name
collision and, on delete+reimport, changed the skill UUID and dropped agent
bindings. This adds a structured conflict result and a creator-only overwrite
write path so a re-import can update the existing skill in place.
- New terminal import status `conflict` carrying { existing_skill_id,
existing_created_by, can_overwrite }; can_overwrite = requester is the
skill creator (canOverwriteSkillByLocalImport — intentionally narrower than
canManageSkill: admins edit in-app, not via re-import).
- Conflict is detected at daemon-report time (the effective name is only known
once the bundle arrives) via GetSkillByWorkspaceAndName, with the unique
constraint as a race backstop.
- Import requests carry action=overwrite + target_skill_id, persisted through
both the in-memory and Redis LocalSkillImportStore (the heartbeat → daemon
payload is unchanged; overwrite is resolved server-side).
- overwriteSkillWithFiles updates by target_skill_id in one tx: re-checks
existence (workspace-scoped) and creator permission, then replaces
description/content/config and fully replaces files (pruning files absent
from the new bundle). Preserves id, created_by, created_at, name, and
agent_skill bindings. Publishes skill:updated (not skill:created).
- Boundaries: target deleted or permission lost → failed (no fallback to
create-by-name); any mid-write error rolls back the tx, leaving the original
skill untouched. Retrying a terminal request is a no-op.
Tests cover: creator/non-creator conflict (can_overwrite), overwrite preserves
UUID + agent binding + prunes removed files, non-creator overwrite fails,
deleted target fails without create fallback, retry idempotency, and Redis
round-trip of the new fields.
Backend half of MUL-2701. Contract change: same-name local imports now return
status `conflict` instead of `failed` — the Desktop/core client must be updated
to consume it (sibling task).
MUL-2800
Co-authored-by: multica-agent <github@multica.ai>
* fix(skills): gate structured conflict behind client opt-in; guard overwrite target name
Addresses review feedback on PR #3498 (MUL-2800).
Backward compatibility: a same-name local import now returns the new `conflict`
status only when the initiating client opts in via `supports_conflict` (an
overwrite request implies it). Older clients — already-installed Desktop builds
whose poll loop only understands `failed`/`timeout` — keep the legacy `failed`
+ "a skill with this name already exists" behavior, so upgrading the backend
ahead of the client no longer regresses the import UX. This is the installed-app
API-compat boundary the repo's CLAUDE.md calls out.
Also: the overwrite write path now verifies the incoming effective name matches
the target skill's current name (errSkillOverwriteNameMismatch -> failed),
preventing a stale/wrong target_skill_id from writing one skill's content onto
another. Creator-only + workspace scoping already prevent privilege escalation;
this narrows the API so it can't be misused.
Refactored LocalSkillImportStore.Create to a LocalSkillImportRequestInput params
struct (the signature had grown to 8 positional args; the opt-in flag pushed it
over). supports_conflict is persisted in both the in-memory and Redis stores.
Tests: conflict tests now opt in; added a legacy-client test (no flag ->
failed + legacy message) and an overwrite name-mismatch test.
MUL-2800
Co-authored-by: multica-agent <github@multica.ai>
* feat(skills): resolve local import conflicts in desktop
Co-authored-by: multica-agent <github@multica.ai>
* fix(skills): preserve bulk flow after conflict resolution
Co-authored-by: multica-agent <github@multica.ai>
* feat(cli): add skill import conflict strategies
Co-authored-by: multica-agent <github@multica.ai>
* fix(i18n): sync skill import locale keys
Co-authored-by: multica-agent <github@multica.ai>
* docs: explain skill import conflict handling
Co-authored-by: multica-agent <github@multica.ai>
* docs: refresh skill import source map anchors
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The preview answer depends on live queue state (pending-task dedup), not
just the mention set, so three staleness bugs showed up around it:
- staleTime: Infinity pinned a "nobody triggers" snapshot taken while
the mentioned agent was still queued — the chip never appeared even
though sending really did wake the agent (create recomputes).
-> staleTime: 0, cached signatures revalidate in the background.
- The in-flight gap on a signature change rendered as an empty agent
list, flickering the chips and wiping the composer's suppressed-id
set via the pruning effect. -> placeholderData: keepPreviousData.
- Nothing refreshed an open composer when an agent's task finished.
-> the WS task-lifecycle handler now also invalidates the
commentTriggerPreviewAll prefix, so chips appear mid-typing the
moment the agent becomes triggerable again.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Five chip states get distinct copy instead of sharing one sentence and a
vague "not this time":
- single, will trigger: Starts working when sent (unchanged)
- single, skipped: Won't be triggered
- several, k will fire: {{count}} agents start working when sent —
the count covers only non-suppressed agents; skipped ones read as the
dimmed heads in the stack next to the number
- several, all skipped: No agents will be triggered
- popover row state: Skipped
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Table.configure had renderWrapper unset (defaults to false), so tables
rendered as bare <table> elements with no .tableWrapper div. The
overflow-x: auto rule in prose.css targets .tableWrapper and never
matched, so a wide table pushed the horizontal scrollbar onto the
issue detail's page-level scroll container instead of scrolling
within the table itself.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
- Copy: one fixed sentence for single and stacked chips — the avatar(s)
carry who and how many, the text carries condition + outcome
("发送后开始工作" / "Starts working when sent"), killing the
"is it already running?" misread. Drops the per-name and count keys.
- Color: sidebar-style resting state — muted-foreground until hover so
the strip reads as metadata, not content.
- Motion: pure fade-in (no slide offset).
- Spacing: reply composer reserves pb-9 so the chip strip reads as a
footer instead of a second content line glued to the text.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
The issue description editor bound pending uploads with
`md.includes(a.url)`, but the editor persists the durable markdownLink
(`/api/attachments/<id>/download` / markdown_url), never the raw storage
`a.url`. The filter therefore never matched, so description uploads were
never linked via `attachment_ids`.
After reload the attachment was absent from `issueAttachments`, so the
renderer could not resolve it to a freshly-signed CDN `download_url` and
fell back to the persisted auth-gated download endpoint. That endpoint
loads on web (same-site cookie / proxy) but fails as a native <img> on
Desktop/Electron (cross-origin file:// renderer carries no auth), leaving
the image broken — while comments rendered fine because they already bind
via contentReferencesAttachment.
Switch the description binding to contentReferencesAttachment, matching
the comment/reply/chat composers, so description images resolve to the
signed CDN URL on every client. Add a regression test pinning the
absolute-host markdown_url shape.
* Add comment trigger preview suppression
Co-authored-by: multica-agent <github@multica.ai>
* Use TanStack Query for trigger preview
Co-authored-by: multica-agent <github@multica.ai>
* Test note comments skip create triggers
Co-authored-by: multica-agent <github@multica.ai>
* feat(issues): redesign comment trigger chips as avatar chips
Single agent renders as avatar + presence dot + full sentence; several
agents collapse to an overlapping stack + active count, mirroring the
header working chip. Per-agent skip moves into a click-opened popover
(hover layers stay read-only tooltips); suppression reads as brightness,
not a ban glyph. Loading and preview errors render nothing.
Also: share one tooltip body across chip and popover rows, invalidate
cached previews after a comment lands (the enqueued task changes the
dedup answer), move the preview query key into issueKeys, and drop the
now-unconsumed status field from useCommentTriggerPreview.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor(server): drop comment trigger wrappers kept only for tests
enqueueMentionedAgentTasks and shouldEnqueueSquadLeaderOnComment had no
production callers after the compute/enqueue split — the comment path
goes through computeCommentAgentTriggers. Tests now exercise the compute
functions directly via package-local helpers, so the legacy adapters
cannot drift from the real path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(skills): sync mentioning/squads source maps with shared trigger computation
The squads source map still pointed the comment-trigger contract at the
pre-refactor call chain (comment.go:940 -> shouldEnqueueSquadLeaderOnComment),
and the mentioning skill referenced the deleted wrapper. Re-anchor both
to computeCommentAgentTriggers / computeAssignedSquadLeaderCommentTrigger
/ computeMentionedAgentCommentTriggers with current line numbers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Comment / issue / chat images uploaded inside the Desktop app rendered
as the broken-image fallback. The editor was persisting a site-relative
`/api/attachments/<id>/download` URL into markdown — that path only
resolves when the document origin proxies /api to the API host (apps/web
via Next.js rewrite). On Electron's file:// origin it never resolved.
Per GPT-Boy's plan, move the durable-URL choice from the client to the
server so the persisted shape is correct regardless of which client
performed the upload.
Server:
- AttachmentResponse gains a markdown_url field, computed by
buildMarkdownURL from the deployment policy:
• storage URL is already absolute + unsigned (public CDN, S3 public
bucket, LocalStorage with MULTICA_LOCAL_UPLOAD_BASE_URL on https) →
use it verbatim;
• CloudFront-signed mode → never expose the raw S3 URL (private
bucket); return cfg.PublicURL + /api/attachments/<id>/download so
the server can re-sign on every request;
• LocalStorage relative + cfg.PublicURL set → same prefixed API
endpoint;
• cfg.PublicURL unset → fall back to site-relative path so web's
Next.js rewrite still works.
- isDurablePublicURL helper rejects URLs carrying CloudFront / S3
signature query params, so a freshly-signed download_url can never
leak into persistence — the original MUL-3130 bug stays closed.
Frontend:
- Attachment type + AttachmentResponseSchema (and apps/mobile mirror)
carry markdown_url. Schema lenient-defaults to '' so a backend old
enough to predate this field doesn't break clients.
- useFileUpload picks markdownLink with three-layer fallback:
(1) att.markdown_url (modern server),
(2) attachmentDownloadPath(att.id) — legacy site-relative shape,
retained for backends old enough to omit markdown_url,
(3) att.url — no-workspace avatar branch with no attachment-row id.
- attachment.tsx keeps the relative→absolute absolutize pass, but
reframed as the legacy-compat fallback for already-persisted
/api/attachments/<id>/download or /uploads/<key> URLs in old
bodies. New content writes absolute URLs and skips this path.
- ContentEditor still tracks freshly-uploaded records into
AttachmentDownloadProvider so Quick Create's editor can swap the URL
via the resolver during the same session even before the server-side
binding lands.
Tests:
- server/internal/handler/file_test.go: 5 new buildMarkdownURL matrix
tests (public CDN passthrough, CloudFront-signed swap, relative
prefixing, PublicURL unset fallback, trailing-slash strip) + 15
table-driven isDurablePublicURL cases.
- packages/core/hooks/use-file-upload.test.ts: new file, 4 cases
covering modern server / legacy server / no-id avatar / oversize.
- packages/views/editor/attachment.test.tsx + content-editor.test.tsx:
10 cases for the absolutize matrix and in-session attachment merge.
- 6 existing test fixtures updated to include markdown_url.
Verification: 1236 @multica/views tests pass; 514 @multica/core tests
pass (4 new); server handler package tests pass for the new matrix
plus all pre-existing TestAttachmentToResponse* and TestDownload*
cases. Typecheck green for views/core/web/desktop. Lint clean on
touched files.
Quick Create attachment_ids binding (orphaned attachment relationship
on the resulting issue) is a follow-up — it requires a new --attachment-id
CLI flag and daemon prompt-template work and is intentionally scoped
out of this PR.
Refs: MUL-3192
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
* fix(realtime): invalidate per-issue caches on WS reconnect (MUL-3189)
Per-issue caches (timeline, reactions, subscribers, usage, attachments,
tasks) are keyed without wsId, so the issueKeys.all(wsId) prefix in
invalidateWorkspaceScopedQueries never reached them. With the
staleTime: Infinity default they rely entirely on WS events for
freshness, so a comment:created event lost during a disconnect (e.g.
macOS sleep) left the timeline stale until a full view reload — the
inbox showed the agent's new comment while the issue's comment area
stayed empty.
Add *All prefix helpers for the per-issue key families and invalidate
them in the reconnect / WS-instance-change recovery path. Inactive
caches are only marked stale and refetch on next mount; the mounted
issue refetches immediately, matching its existing useWSReconnect
behavior, so this does not reintroduce the MUL-1941 memo thrash.
Fixes#3953
Co-authored-by: multica-agent <github@multica.ai>
* refactor(core): define issueKeys.tasks via tasksAll prefix helper
Review nit on #3992 — keep the per-issue key families consistently
defined in terms of their *All prefix helpers. No behavior change.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
A thread could hold multiple resolved comments at once: ResolveComment
was a plain per-row setter that never cleared the prior resolution, and
"replacing" one was a display-only illusion (deriveThreadResolution
picks the max resolved_at). The stale rows stayed resolved in the DB and
the optimistic update flashed the new resolution, then reverted.
Make single-resolution-per-thread a write invariant:
- ClearOtherThreadResolutions: thread-scoped clear via a RECURSIVE CTE
(root + descendants of the target, id <> target), returns each cleared
row.
- ResolveComment handler runs the clear + set in one tx so the replace
is atomic. It emits comment:unresolved per cleared sibling (granular
realtime consumers patch a single comment in place and would otherwise
keep showing the stale resolution). Target keeps its COALESCE
idempotency and the re-resolve event suppression.
- Frontend optimistic update mirrors the invariant: resolving clears
every other resolution in the same thread, so the cache never shows
two at once. Unresolve still only clears its own row.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(autopilots): show creator in autopilot detail properties (MUL-3139)
The autopilot creator was already persisted end-to-end (created_by_type /
created_by_id on the autopilot table, exposed via AutopilotResponse and the
frontend Autopilot type) but never rendered. Add a "Created by" field to the
detail page Properties section, mirroring the existing assignee field and the
issue-detail creator row, reusing ActorAvatar + getActorName.
Creator may be a member or an agent (the HTTP create path stamps member today,
but backend logic also writes created_by_type=agent), so the display resolves
both actor types and does not assume member. List rows are intentionally left
unchanged, matching the issue convention (creator lives in detail, not lists).
Adds the field_created_by label to all four locale bundles (en/zh-Hans/ja/ko);
locale parity test enforces full coverage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(autopilots): show creator in autopilots list (MUL-3139)
Add a Created by column to the autopilots list, mirroring the detail
page. Secondary columns (creator, mode, last run) are hidden below lg
so small screens keep only name, agent, and status.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(daemon): report OS in /health response
The desktop app reads daemon liveness over HTTP but starts/stops it via the
native CLI, which acts on the host process namespace. On Windows with the
daemon in WSL2, /health is reachable via localhost forwarding yet the daemon's
process is unreachable — so the app needs a signal to tell a daemon it manages
from one it merely sees. Expose runtime.GOOS as `os` so the desktop can
compare it against its own host OS. MUL-3154, #3916
Co-authored-by: multica-agent <github@multica.ai>
* fix(desktop): disable auto-start/stop for an unmanageable daemon
When the daemon runs in an environment the app can't drive — e.g. Linux in
WSL2 behind a Windows desktop, reachable only via localhost forwarding — the
Auto-start/Auto-stop toggles silently did nothing: the lifecycle CLI acts on
the host process namespace and never reaches the daemon's PID.
Detect it by comparing the daemon's reported OS (new /health `os` field)
against the host OS, and only when a daemon is actually running. When they
differ: disable both toggles with an explanatory note, skip the version-match
restart on auto-start, and skip the no-op stop on quit. Fails safe — a missing
`os` (older daemon) or a matching OS keeps the toggles live, so native
Mac/Windows/Linux daemons are unaffected.
MUL-3154, #3916
Co-authored-by: multica-agent <github@multica.ai>
* fix(desktop): centralize externally-managed guard at the lifecycle boundary
Review follow-up. The first cut only disabled the Settings toggles, but the
same unmanageable daemon (WSL2 etc.) could still be Stop/Restart-ed from the
Runtime card and from automatic lifecycle entries (logout, user switch,
reauth, first-workspace restart) — each of which would shell out to a native
CLI that can't reach the daemon's process.
Move the guard into the main-process lifecycle functions so every entry point
is covered by construction: stopDaemon() and restartDaemon() no-op for an
externally-managed daemon, and ensureRunningDaemonVersionMatches() treats it
as up-to-date (no misleading restart). The per-branch checks in the auto-start
handler and before-quit are removed — the boundary now covers them. The
Runtime card hides Stop/Restart and shows a 'Managed outside the app' hint,
mirroring the Settings tab. Adds a component test for the card's two states.
MUL-3154, #3916
Co-authored-by: multica-agent <github@multica.ai>
* fix(desktop): preflight the lifecycle guard against live /health
Review follow-up. The guard read a cached lastExternallyManaged, which only
fetchHealth() updates — but not every lifecycle entry polls before calling
stop/restart. syncToken()'s user-switch branch calls restartDaemon() directly
after its own fetchHealthAtPort(), without refreshing the cache; on a fresh
launch / account switch (no poll yet) the cache is still the initial false, so
restartDaemon() would shell out to the native CLI and hit the very WSL/native
PID-namespace problem this PR avoids.
Make stopDaemon()/restartDaemon() preflight against a live /health read each
call instead of trusting the poll cache. The decision is extracted to a pure
daemonLifecycleUnreachable(readDaemonOS, hostOS) so a unit test can prove the
*live* value (not a cache) drives it. lastExternallyManaged is removed — the UI
already reads the per-status externallyManaged field, so it had no other
consumer.
MUL-3154, #3916
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* refactor(transcript): reuse taskMessageToPayload in WS broadcast
The ReportTaskMessages WebSocket broadcast hand-built the payload and
duplicated the created_at formatting that taskMessageToPayload already
does. Reuse the helper with the just-inserted row, which carries the
same redacted values and the DB-assigned timestamp.
Co-authored-by: multica-agent <github@multica.ai>
* test(transcript): cover coalesce created_at behavior
Lock in that coalescing streaming fragments carries the latest
created_at, and falls back to the previous timestamp when the merged
fragment has none.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
The initiator_user_id column (added in 117) carried a foreign key to the
"user" table. Adding that FK also locks the hot "user" table at migration
time, which made the ALTER time out on a busy production deploy. The
column only feeds a best-effort name/email lookup at claim time (a stale
id just yields no `## Task Initiator` section), so referential integrity
is not load-bearing.
- Edit 117 to add a plain `UUID` column (no FK). The original timed-out
deploy never recorded 117, so its retry now runs the FK-free version.
- Add 118 to `DROP CONSTRAINT IF EXISTS` for environments that already
applied the constraint-bearing 117 (they skip the edited 117 by
version). All environments converge to a plain, FK-free column.
No code/codegen change: dropping the FK does not affect the Go column
type, so sqlc output is unchanged. Verified locally: 118 drops the FK and
keeps the column; sqlc regen produces no diff; build/vet/tests pass.
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
Threads the existing task_message.created_at column through the full stack (Go protocol -> REST/WS handlers -> TS types -> transcript dialog) so agent run transcripts show per-entry timestamps, helping users spot stalled runs. Additive, no migration.
Treats Cursor's stream-json terminal `result` event as the protocol completion boundary so a lingering Cursor worker process can no longer hold the daemon task open after the agent has produced its final result.
- Tighten `cmd.WaitDelay` to 500ms (set before `Start()`)
- Set `resultSeen` and `cancel()` on terminal `result`
- Preserve completed/failed status across the cancellation via two `!resultSeen` guards in the post-loop status decision
- Add unix fake-CLI coverage for success and `is_error` terminal results
* feat(issues): per-comment thread resolution with sticky collapse
Allow resolving any comment, not just roots. Resolving a root folds the
whole thread into one bar (existing); resolving a reply marks it as the
thread's resolution ("Resolve thread with comment") and folds the other
replies behind a "N comments" bar, with the resolution kept visible and
badged. Which comment is the resolution is a pure frontend derivation
(root wins, else latest resolved reply), so no write-side bookkeeping is
needed and any resolved_at combination renders one resolution.
- backend: drop the "only root comments can be resolved" guard
- views: deriveThreadResolution + reply-resolution rendering, sticky
collapse/fold bars (overflow-clip on the card so sticky resolves to the
timeline scroll parent), scroll the folded thread back into view on
collapse, ListChevronsDownUp icon, locales (en/ja/ko/zh-Hans)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(issues): sticky comment headers for long comments
Pin each comment's header (root + replies) to the timeline's scroll
parent while reading, so a long comment keeps its author + actions
visible instead of scrolling out of reach. Exactly one header is pinned
at a time:
- Reply headers stick within their own CommentRow box (release at the
reply's end).
- The root header is wrapped in a root-section container so its sticky
containing block spans only the header + root body — without it the
containing block is the whole thread and the root header stays stuck
behind every reply. Replies render outside the wrapper, gated on open.
- Skip the root header sticky whenever a resolution collapse bar already
owns the top-0 slot (root resolved+expanded, or reply-resolution
expanded) to avoid two bars stacking at the same offset.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Generalize SyncRunFromLinkedIssueTask beyond Codex no-progress: any
terminal create-issue task failure with no retry still in flight now fails
the linked autopilot run, so it can no longer hang in issue_created
(invisible to the failure-rate auto-pause monitor).
- fail the linked run for any terminal task failure, gated by the existing
HasActiveTaskForIssue wait-for-retry guard
- remove the isNoProgressTaskFailure classifier (subsumed; drops duplicated
pkg/agent marker literals)
- drop the redundant GetIssue/origin lookup; GetAutopilotRunByIssue leads
and short-circuits ordinary failures in one query
- tests: keep no-progress regression, add agent_error (non-retryable) and
retry-pending cases
Follow-up to #3927. VEN-661 / VEN-662 / MUL-3164
Opening an inbox comment notification on an issue with a running agent
shoved the whole desktop page — header included — off the top, and no
amount of scrolling brought it back; only toggling the right sidebar
(which reflows the panel group) restored it.
Root cause: the deep-link landing uses native
scrollIntoView({block:"center"}) on the target comment. Native
scrollIntoView is spec'd to scroll EVERY scrollable ancestor. On a cold
mount where the timeline is still streaming (is-working) and the sidebar
panel starts collapsed, the inner timeline scroller can't center the
target on its own, so the scroll propagates up and scrolls the desktop
shell's overflow:hidden wrapper (desktop-layout.tsx). That wrapper has no
scrollbar and doesn't auto-clamp, so the page stays shoved up until a
resize reflows it. Desktop-only: web sits in a document-scroll context
with a real scrollbar that self-corrects.
Fix: drive the timeline container's scrollTop directly and re-center
across rAF frames until async heights settle, instead of leaning on the
ancestor scroll. The scroll never touches an ancestor, so the header can
no longer be pushed off-screen.
Tests assert the user-facing contract (lands on + highlights the target
comment) rather than the scroll mechanism, which jsdom can't lay out.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Codex telemetry was never reaching the OTLP collector for tasks run by the
daemon. The per-task config (including the [otel] block) is copied into
CODEX_HOME correctly, but the lifecycle goroutine closed stdin and then
immediately cancelled the run context, which SIGKILLs the app-server. Codex's
OTEL batch exporters only force-flush on a graceful shutdown, so the buffered
spans/metrics/logs were dropped before they could be exported — short tasks
lost everything, long tasks lost the final batch.
Let codex exit on its own after stdin EOF (running its shutdown + flush path)
and only force-cancel after a bounded grace period if it doesn't, so the reader
goroutine still can't block forever. Also set cmd.WaitDelay, matching the other
long-lived backends (claude, copilot, cursor, …).
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* MUL-3130: persist a stable attachment download URL in comment markdown
Comment image attachments rendered as broken placeholders ~30 minutes
after upload because the editor was persisting a short-lived
HMAC-signed URL into the comment body. After PR #3903 (MUL-3132)
hardened /uploads/* with auth, `attachmentToResponse` started signing
`attachment.url` as `/uploads/<key>?exp=<unix>&sig=<HMAC>` for
LocalStorage so token-auth clients could keep loading inline images.
The signature has a 30-min TTL by design — but `useFileUpload` was
returning that signed value as `link` and the editor was writing
`` straight into the markdown, so the comment
permanently captured a URL that stopped working as soon as the
signature expired.
The fix is to persist a stable per-attachment URL that the server can
re-sign on every request:
* `useFileUpload` now returns `link = /api/attachments/<id>/download`
(avatar uploads without an id still fall back to `att.url` so the
pre-attachment-row code paths keep working).
* `DownloadAttachment` self-resolves the workspace from the attachment
row instead of reading X-Workspace-Slug / X-Workspace-ID headers,
and the route is registered under the auth-only group so a native
browser <img>/<video> resource load (which cannot attach those
headers) succeeds. Membership is checked inside the handler with
a 404 deny shape so the route does not act as an IDOR oracle.
* A new `GetAttachmentByIDOnly` SQL query supports the workspace-
derivation step.
* `AttachmentDownloadProvider` now extracts the attachment id from
the stable URL when matching markdown refs to attachment records,
with a fallback to the existing url-equality check for legacy
comments (and S3/CloudFront markdown that points straight at the
CDN).
* `contentReferencesAttachment` covers both URL shapes for the
composer / standalone-list dedup paths so an attachment uploaded
before the fix and one uploaded after both deduplicate cleanly.
Tests:
- New unit tests for the URL helpers (16 tests, packages/core).
- Backend regression test: bare `<img src>`-style request without
workspace headers now succeeds for a member (200) and 404s for a
non-member, replacing the previous "400 without workspace context"
contract.
- Existing TestDownload*, TestServeLocalUpload*, TestAttachmentTo
Response* and the 1220 frontend views tests all pass.
Refs: MUL-3130, GitHub issue #3891
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3130: address PR review — split markdown link from upload link, swap render src
Two follow-ups from GPT-Boy's review on PR #3937.
(1) Don't reroute every upload consumer through the workspace-gated
download endpoint.
The previous change made `useFileUpload`'s `link` field unconditionally
return `/api/attachments/<id>/download` whenever the upload had an id.
But `useFileUpload` is also used by avatar / logo pickers
(account-tab, workspace-tab, agents/avatar-picker, squads/squad-detail-page)
that persist `result.link` directly into `avatar_url`. Avatars are
referenced cross-workspace (mention chips, member lists, inbox
items), so binding their URL to a workspace-membership-gated
endpoint would silently break cross-workspace avatar visibility.
The fix splits the URL into two semantically distinct fields:
- `link` — same as `att.url` (legacy contract). Avatar /
logo callers continue to use this and remain
on whatever URL semantics the storage backend
dictates.
- `markdownLink` — the stable per-attachment URL
`/api/attachments/<id>/download`. Only the
editor's markdown-persisting flow consumes
this. Falls back to `link` for the
no-workspace upload branch (where there is
no attachment-row id to address).
`editor/extensions/file-upload.ts` switches `image.src` and
`fileCard.href` to `markdownLink ?? link` so comment markdown gets
the stable shape while avatar callers stay on `link` unchanged.
(2) Make the render-time img src loadable for token-mode clients.
Persisting the stable `/api/attachments/<id>/download` URL fixes the
expiry problem but the path itself sits behind `middleware.Auth`,
which expects either a `multica_auth` cookie or a Bearer token in
`Authorization`. Native `<img>`/`<video>` resource loads from
token-mode clients (Electron's default mode, the mobile app,
legacy-token web sessions) cannot attach the Authorization header,
so the bare URL would 401 immediately rather than 30 minutes later.
`Attachment.normalize` now runs the resolved record through a new
`pickInlineMediaURL` helper that returns:
- `record.download_url` when it's an absolute URL with a
recognised CDN signature query (CloudFront-signed
`Signature` / `Expires` / `Key-Pair-Id`, or
`X-Amz-Signature` for raw S3 presigns) — these load as
native resource src in any client.
- else `record.url`, which on the LocalStorage backend carries
a freshly-minted `/uploads/<key>?exp&sig` query whose
signature IS the auth (token-mode-loadable). On non-CF S3
backends this is the raw stored URL — same behaviour as
today.
- else the original input URL (legacy / unresolved markdown
keeps its existing path).
This gives the same effect for both `kind: "record"` and
`kind: "url"` attachment inputs: once a record is in hand, the
rendered media src is whichever URL the current backend exposes
a working signature on.
Tests:
- New `file-upload.test.ts` regression pinning that `markdownLink`
is what lands in the markdown body when the upload result returns
both a short-lived storage URL and a stable download path.
- Updated `attachment.test.tsx` to reflect the new render-time
swap (the rendered img src now follows the freshly signed URL,
not the raw storage URL) and added a record-mode regression
pinning the LocalStorage default — when `download_url` is the
bare /api/attachments/<id>/download path, the renderer must fall
through to the signed `record.url`.
- Updated `chat-input.test.tsx` makeUpload helper for the new
`markdownLink` UploadResult field.
- 1222 frontend views tests + 507 core tests + typecheck across
@multica/{core,ui,views} all pass.
Refs: MUL-3130, GitHub issue #3891. Builds on a740f7a35.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3130: chat upload map keys on persisted markdownLink, not the short-lived link
GPT-Boy's second-round review on PR #3937 caught a chat-only blocker
left over from the previous fix.
After the previous commit split `UploadResult.link` into `link`
(legacy avatar/logo URL) and `markdownLink` (stable per-attachment
URL persisted into markdown), the comment editor's image src + file
card href correctly switched to `markdownLink ?? link`. But chat
input still kept the upload-map key on the old `link`:
uploadMapRef.current.set(result.link, result.id)
…
if (content.includes(url)) activeIds.push(id)
In the LocalStorage backend `link` is the short-lived
`/uploads/<key>?exp=&sig=` URL. The editor persists the stable
`/api/attachments/<id>/download` URL into the message body, so
`content.includes(url)` never matches and the send call drops
`attachment_ids`. The attachment ends up bound only to the chat
session, not to the message — agents reading message-level metadata
see no attachments.
Fix: key the upload map on the same value the editor actually wrote
into the markdown body (`markdownLink || link`). The
`content.includes(url)` check then matches and the attachment id is
correctly forwarded on send.
Tests:
- Updated the chat-input mock editor to insert `markdownLink || link`
into its value, mirroring the real editor's persisted-URL choice
(uploadAndInsertFile in editor/extensions/file-upload.ts). Without
this the mock would silently paper over the bug.
- Added a regression test where the upload result returns a
short-lived `link = /uploads/...?exp&sig` and a stable
`markdownLink = /api/attachments/<id>/download`. Asserts (a) the
message body carries the stable URL and never the signed query,
and (b) the bound `attachment_ids` includes the attachment id.
All 1223 frontend views tests pass (was 1222, +1 new regression).
Typecheck and 507 core tests still green.
Refs: MUL-3130, PR #3937 review by GPT-Boy. Builds on f66a522d0.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
Fail create-issue autopilot runs that hang in issue_created after a Codex
no-progress / semantic-inactivity task failure, so they surface as failed
and count toward the failure-rate auto-pause monitor.
- route failed create-issue issue tasks (no direct autopilot_run_id) into linked run sync
- fail linked runs only for Codex no-progress / semantic-inactivity failures
- wait when an active retry task still exists for the issue
- add classifier coverage + a DB-backed listener regression
VEN-661 / VEN-662 / MUL-3164
Summary:
- Add CLI config schema for OpenClaw backend binary path and state dir overrides.
- Apply those overrides during daemon LoadConfig using the existing env-var based probe/spawn path.
- Cover backward compatibility, precedence, partial overrides, and fail-soft config loading.
Verification:
- go test ./internal/cli ./internal/daemon
- go vet ./internal/cli ./internal/daemon
- GitHub CI passed
* fix(cli): honor MULTICA_SERVER_URL in setup self-host
`multica setup self-host` resolved the backend URL only from the
--server-url flag, falling back to http://localhost:8080 when the flag
was absent. It never consulted MULTICA_SERVER_URL, even though that env
var is documented on the root --server-url flag and in `multica --help`,
and is honored by every other command via resolveServerURL. A self-host
user who set the env var instead of the flag still hit localhost and got
"Server at http://localhost:8080 is not reachable".
Route server-url and app-url through cli.FlagOrEnv so the documented env
vars (MULTICA_SERVER_URL / MULTICA_APP_URL) are honored when the matching
flag is not set, with the flag still taking precedence. userProvided now
reflects flag-or-env, so an env-sourced remote URL still triggers the
explicit app_url prompt. Not platform-specific despite the report.
Fixes GitHub #3912.
Co-authored-by: multica-agent <github@multica.ai>
* fix(cli): normalize MULTICA_SERVER_URL in setup self-host
MULTICA_SERVER_URL is documented as a ws:// daemon address
(ws://localhost:8080/ws) and every other command normalizes it via
NormalizeServerBaseURL before use. setup self-host consumed the resolved
value raw and probed <url>/health, so a self-hoster who set the
documented ws:// form would still fail the reachability check.
Run the flag/env value through normalizeAPIBaseURL (ws->http, wss->https,
strip /ws) so the documented form works and the stored server_url stays a
clean http(s) base. Add a normalization test case and a focused test for
the MULTICA_APP_URL env path (review nit).
Co-authored-by: multica-agent <github@multica.ai>
* docs(self-host): note setup self-host honors MULTICA_SERVER_URL / MULTICA_APP_URL
Document that `setup self-host` reads the env vars when the matching flag
is omitted (flag wins), and that MULTICA_SERVER_URL accepts the ws://…/ws
daemon form. Added to en/zh/ja/ko quickstart for parity.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* fix(projects): return 400 (not 500) for invalid project status/priority
CreateProject/UpdateProject passed an unvalidated status/priority straight to
the INSERT, so an unknown value (e.g. --status active) tripped the table's
CHECK constraint and surfaced as a blanket 500 'failed to create project'
with no server-side log to diagnose it (#3925).
Pre-validate both enums against the column CHECK lists and return a 400 with
the allowed values. Back it with isCheckViolation -> 400 for any other
constrained column, and log the underlying error on genuine 500s so transient
DB failures are diagnosable.
MUL-3153
Co-authored-by: multica-agent <github@multica.ai>
* fix(cli): validate project --status in create/update
project create and project update forwarded --status to the server without
checking it, while project status already validated. Share a single
validateProjectStatus helper across all three so a typo fails fast with the
valid list instead of a server round-trip.
MUL-3153
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: J <j@multica.ai>
Co-authored-by: multica-agent <github@multica.ai>
* docs(cli): add Error Messages conventions + refine sign-in copy (PR3)
Final pass of the CLI error-message work (MUL-3104).
- CLI_AND_DAEMON.md: new "Error Messages" section documenting the user-facing
contract — friendly single-line messages, server validation passthrough,
English default with automatic Chinese on a zh locale, the tiered exit codes
(0/1/2/3/4/5), --debug / MULTICA_DEBUG for the full chain, and
MULTICA_HTTP_TIMEOUT.
- cmd_auth.go: clarify three high-frequency sign-in errors so the message
states what failed and the next step — local login-callback server start
(hints at port/firewall), access-token creation, and token verification
(suggests retrying `multica login` and checking the token is valid/not
expired). All keep %w so exit-code tiering and --debug detail are preserved.
cmd_id_resolver.go is left as-is — its not-found / ambiguous-prefix messages
already point at `list --full-id` and need no change. The user-facing
FormatError layer is unchanged, so its existing PR1/PR2 test coverage still
applies; no test asserted the old verb strings.
Refs MUL-3104. PR3 of 3 (final).
Co-authored-by: multica-agent <github@multica.ai>
* fix(cli): make login failure guidance visible via typed user-message wrapper
Addresses 张大彪's PR3 review: the refined sign-in copy was wrapped with %w,
so FormatError returned the centralized *HTTPError/*NetworkError copy and the
new guidance only appeared under --debug.
- Add cli.UserMessageError + cli.WithUserMessage: a typed wrapper carrying a
user-facing message that FormatError surfaces by default, recognized before
the network/http branches. Unwrap() is preserved, so ExitCodeFor still
classifies by the underlying typed error and --debug still prints the full
original chain.
- cmd_auth.go: wrap the OAuth access-token-creation and PAT-verification
failures with WithUserMessage (OAuth copy no longer mentions a passed token,
since that flow has none), and move the token-specific 'valid / not expired'
hint to the real Enter your personal access token: verification site (was the generic
'invalid token: %w').
- Focused tests: under a wrapped *HTTPError(401) the default FormatError shows
the login hint, ExitCodeFor returns ExitAuth, and --debug retains the raw
chain; a wrapped *NetworkError still classifies as ExitNetwork.
- CLI_AND_DAEMON.md: narrow 'every error' to command errors returned to the
top-level handler, noting commands like setup's fast /health probe bypass it.
Refs MUL-3104, PR #3900.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3132: harden /uploads/* (auth, no listing, nosniff, tight CSP)
Closes the open hardening items from the SVG XSS disclosure
(security-findings-2026-06-02). The primary chain (PR #3023 / #3050)
is intact; this PR addresses every remaining recommendation from the
disclosure's hardening list except 'serve uploads from a separate
origin' (a structural change beyond this fix).
Changes:
- /uploads/* now requires authentication. The route is wrapped in
middleware.Auth so anonymous internet users can no longer fetch
workspace attachments by guessing the URL. A new ServeLocalUpload
handler then enforces the second layer:
- workspaces/{wsID}/* paths require membership in wsID (uses
MembershipCache for the hot path);
- users/{userID}/* paths allow any authenticated user (avatars
are referenced cross-workspace);
- any other prefix returns 404, so a future feature cannot drop
content under /uploads/<other-prefix>/ and inherit a relaxed
policy by accident.
Non-members see 404 (not 403) so the route does not act as an IDOR
oracle for workspace IDs.
- Directory listing on /uploads/* is rejected at the storage layer:
empty keys, trailing-slash keys, and any key that resolves to a
directory return 404 before http.ServeFile would render an HTML
index. UUID filenames were obscurity, but enumerating them
shouldn't be free.
- Every successful /uploads/* response carries
X-Content-Type-Options: nosniff and a tight per-response CSP
(default-src 'none'; sandbox; frame-ancestors 'none'), overriding
the application-wide CSP. This is belt-and-suspenders if a future
regression weakens the Content-Disposition: attachment path.
- UploadFile rejects HTML-family uploads at the edge (.html, .htm,
.xhtml, .shtml, .xht, .phtml, plus a content-type denylist for
text/html and application/xhtml+xml so renamed payloads cannot
bypass the extension check). SVG and JS remain allowed because
their existing serve-side defenses neutralize them and source-code
attachments preview as text/plain via /api/attachments/{id}/content.
Tests:
- storage: TestLocalStorage_ServeFile_RejectsDirectoryListing,
TestLocalStorage_ServeFile_HardeningHeaders.
- handler: TestIsUploadDenied (pure), TestUploadFile_RejectsHTMLByExtension,
TestUploadFile_RejectsHTMLByContentType, TestUploadFile_AllowsLegitimateImage,
and the full ServeLocalUpload matrix (RequiresAuth, MemberCanRead,
NonMemberDenied, RejectsDirectoryInPath, UnknownPrefixDenied,
UserPrefixAllowsAnyAuthedUser).
- Full server test suite passes.
Co-authored-by: multica-agent <github@multica.ai>
* MUL-3132: HMAC-signed query auth for /uploads/* (token-auth client compat)
Addresses J's Request Changes review on PR #3903.
Problem: PR #3903 wrapped /uploads/* in middleware.Auth, but native
<img>/<video>/<iframe> resource loads cannot attach Authorization
headers. Token-auth clients (Desktop default, legacy-token Web
sessions, mobile) were breaking on inline attachment rendering even
though the API itself authenticated fine.
Fix: implement HMAC-signed query parameters for /uploads/*, mirroring
S3 + CloudFront presigned URLs.
- storage.SignLocalUploadURL(rawURL, key, secret, expiry) appends
'?exp=<unix>&sig=<HMAC-SHA256(key|exp)>' query params; signature
is bound to one specific key, has a TTL matching CloudFront mode
(defaultAttachmentDownloadURLTTL = 30 min), constant-time compared
on verify.
- storage.VerifyLocalUploadSignature(key, exp, sig, secret, now)
rejects expired, tampered, wrong-secret, and key-mismatched
signatures.
- ServeLocalUpload now has two auth paths: signed-query (no Auth
middleware needed; signature itself is the authority) and
Bearer/cookie (membership-gated as before). Partial signed-query
fails closed.
- The route in router.go dispatches between the two: if both exp+sig
query params are present, route to inner handler unwrapped; else
wrap in middleware.Auth.
- attachmentToResponse appends signed query to URL when the storage
backend is *LocalStorage. CloudFront-signed download URLs and S3
paths are unchanged.
Tests:
- storage: TestSignAndVerifyLocalUploadURL_RoundTrip,
TestVerifyLocalUploadSignature_RejectsExpired, _RejectsTamperedSig,
_BoundToKey, _RejectsWrongSecret,
TestSignLocalUploadURL_PreservesExistingQuery,
TestLocalUploadSignatureFromQuery_EmptyOnAbsence (7 pure tests).
- handler: TestServeLocalUpload_{SignedQueryBypassesAuth,
SignedQueryRejectsExpired, SignedQueryRejectsTampered,
SignedQueryBoundToOneKey, PartialSignedQueryFailsClosed},
TestAttachmentToResponse_LocalStorageMintsSignedURL.
Full server test suite passes.
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Eve <eve@multica-ai.local>
Co-authored-by: multica-agent <github@multica.ai>
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
Guidance for Claude Code when working in this repository. Keep this file short and authoritative: rules here should be hard to infer from code or easy to get wrong.
## Conventions reference
## Conventions
The single source of truth for **code naming, the i18n translation glossary, and the Chinese voice guide** is the docs site:
The source of truth for code naming, i18n glossary, and Chinese product voice is:
Read it before editing translations in `packages/views/locales/`, naming routes/packages/files/DB columns/types, or writing Chinese UI/docs copy. Do not rely on `packages/views/locales/glossary.md`; it is only a redirect stub.
- Writing or editing translations (`packages/views/locales/`)
- Naming a new route, package, file, DB column, or TS type
- Writing Chinese product copy (UI strings, error messages, docs)
## Project Shape
The legacy `packages/views/locales/glossary.md` is now a stub redirecting to the docs page; do not rely on it.
Multica is an AI-native task management platform for small teams, with agents as first-class assignees that can own issues, comment, and change status.
## Project Context
-`server/`: Go backend, Chi router, sqlc, gorilla/websocket.
-`apps/web/`: Next.js App Router.
-`apps/desktop/`: Electron desktop app.
-`apps/mobile/`: Expo / React Native iOS app. Read `apps/mobile/CLAUDE.md` before touching it.
-`packages/core/`: headless business logic, API client, React Query hooks, Zustand stores.
-`packages/ui/`: atomic UI components only.
-`packages/views/`: shared business pages/components for web and desktop.
-`packages/tsconfig/`: shared TypeScript config.
Multica is an AI-native task management platform — like Linear, but with AI agents as first-class citizens.
Shared packages export raw `.ts` / `.tsx` and are compiled by consuming apps. Dependency direction is `views -> core + ui`; `core` and `ui` must stay independent.
- Agents can be assigned issues, create issues, comment, and change status
- Supports local (daemon) and cloud agent runtimes
- TanStack Query owns server state: issues, users, workspaces, inbox, agents, members, and anything fetched from the API.
- Zustand owns client state: selected workspace, filters, drafts, modals, tab layout, and navigation history.
- Shared Zustand stores live in `packages/core/`, never in `packages/views/` or app directories.
- React Context is for platform plumbing only, such as `WorkspaceIdProvider` and `NavigationProvider`.
- Only auth/workspace stores may call `api.*` directly. Other server interaction belongs in queries/mutations.
- Workspace-scoped query keys must include `wsId`.
- Mutations should be optimistic by default: patch locally, send request, roll back on failure, invalidate on settle.
- WebSocket events invalidate or patch Query cache; they never write directly to Zustand stores.
- Persist durable preferences/drafts/layout. Do not persist server data or ephemeral UI state.
- Zustand selectors must return stable references. Do not return freshly allocated objects/arrays from selectors without shallow comparison.
- Hooks that need workspace context should accept `wsId`; do not call `useWorkspaceId()` internally unless the hook is guaranteed to run under the provider.
-`server/` — Go backend (Chi router, sqlc for DB, gorilla/websocket for real-time)
-`apps/web/` — Next.js frontend (App Router)
-`apps/desktop/` — Electron desktop app (electron-vite)
-`apps/mobile/` — Expo / React Native iOS app. See `apps/mobile/CLAUDE.md`.
-`packages/core/` — Headless business logic (zero react-dom)
-`packages/ui/` — Atomic UI components (zero business logic)
-`packages/views/` — Shared business pages/components (zero next/* imports, zero react-router imports)
What lives where for sharing purposes is documented in *Sharing Principles* below — read it once.
These are hard constraints:
### Key Architectural Decisions
-`packages/core/`: no `react-dom`, `localStorage` (use `StorageAdapter`), `process.env`, or UI libraries.
-`packages/ui/`: no `@multica/core` imports and no business logic.
-`packages/views/`: no `next/*`, no `react-router-dom`, no stores. Use `NavigationAdapter`, `useNavigation()`, and `<AppLink>`.
-`apps/web/platform/`: only place for Next.js navigation/platform APIs.
-`apps/desktop/src/renderer/src/platform/`: only place for `react-router-dom` navigation wiring.
- Every workspace under `apps/` and `packages/` must declare directly imported external packages in its own `package.json`.
- Shared dependencies use `catalog:` from `pnpm-workspace.yaml`; `apps/mobile/` pins Expo/React Native related versions directly.
**Internal Packages pattern** — all shared packages export raw `.ts`/`.tsx` files (no pre-compilation). The consuming app's bundler compiles them directly. This gives zero-config HMR and instant go-to-definition.
## Sharing Rules
**Dependency direction:**`views/ → core/ + ui/`. Core and UI are independent of each other. No package imports from `next/*`, `react-router-dom`, or app-specific code.
Web and desktop share business logic, hooks, stores, components, and views through `packages/core/`, `packages/ui/`, and `packages/views/`.
**Platform bridge:**`packages/core/platform/` provides `CoreProvider` — initializes API client, auth/workspace stores, WS connection, and QueryClient. Each app wraps its root with `<CoreProvider>` and provides its own `NavigationAdapter` for routing.
If the same logic exists in both web and desktop, extract it unless it depends on platform APIs:
**pnpm catalog** — `pnpm-workspace.yaml` defines `catalog:` for version pinning. All shared deps use `catalog:` references to guarantee a single version across all packages. When adding new shared deps (including test deps), add to catalog first.
1. Next.js, Electron, or router APIs stay in the app/platform layer.
2. Headless logic belongs in `packages/core/`.
3. Shared UI or business views belong in `packages/views/`.
4. Shared primitives belong in `packages/ui/`.
### State Management
The architecture relies on a strict split between server state and client state. Mixing them is the most common way to break it.
- **TanStack Query owns all server state.** Issues, users, workspaces, inbox — anything fetched from the API lives in the Query cache. WS events keep it fresh via invalidation; no polling, no `staleTime` workarounds.
- **Zustand owns all client state.** UI selections, filters, drafts, modal state, navigation history. Stores live in `packages/core/` (never in `packages/views/`) so they're shared.
- **React Context** is reserved for cross-cutting platform plumbing — `WorkspaceIdProvider`, `NavigationProvider`. Don't reach for it for general state.
- **Auth and workspace stores are the only stores allowed to call `api.*` directly**, because they manage critical state that must exist before queries can run. They're created via factory + injected dependencies, registered by the platform layer.
**Hard rules — these are how the architecture stays coherent:**
- **Never duplicate server data into Zustand.** If it came from the API, it belongs in the Query cache. Copying it into a store creates two sources of truth and they will drift.
- **Workspace-scoped queries must key on `wsId`.** This is what makes workspace switching automatic — the cache key changes, the right data appears, no manual invalidation needed.
- **Mutations are optimistic by default.** Apply the change locally, send the request, roll back on failure, invalidate on settle. The user shouldn't wait for the server.
- **WS events invalidate queries — they never write to stores directly.** This keeps the cache as the single source of truth and avoids race conditions.
- **Persist what's worth preserving across restarts** (user preferences, drafts, tab layout). **Don't persist ephemeral UI state** (modal open/close, transient selections) or server data.
**Common Zustand footguns to avoid:**
- Selectors must return stable references. Returning a freshly built object or array on every call (e.g. `s => ({ a: s.a, b: s.b })` or `s => s.items.map(...)`) triggers infinite re-renders. Either select primitives separately or use shallow comparison.
- Hooks that need workspace context should accept `wsId` as a parameter, not call `useWorkspaceId()` internally — this lets them work outside the `WorkspaceIdProvider` (e.g. in a sidebar that renders before workspace is loaded).
## Sharing Principles
The monorepo splits into two share zones:
- **Web and desktop** share business logic, components, hooks, stores, and views through `packages/core/`, `packages/ui/`, and `packages/views/`. Existing model — keep using it.
- **Mobile (`apps/mobile/`) is independent.** It shares only **types and pure functions** from `@multica/core/`, with `import type` for types (zero runtime coupling). UI, state, hooks, providers, i18n, React version, build pipeline, release cadence — all mobile-owned.
Mobile is locked to the React version that Expo SDK / React Native ships (which lags React main by 6-12 months). Coupling mobile to the root `catalog:` React would block mobile from upgrading on its own schedule.
See `apps/mobile/CLAUDE.md` for the mobile rules and tech-stack baseline.
Mobile is independent. It may import types and pure functions from `@multica/core`, with `import type` for types, but owns its UI, state, hooks, providers, i18n, React version, build pipeline, and release cadence.
## Commands
Use the repo scripts as the source of truth. Common commands:
```bash
# One-command dev (auto-setup + start everything)
make dev # Auto-creates env, installs deps, starts DB, migrates, launches app
# Explicit setup & run (if you prefer separate steps)
pnpm ui:add badge # Adds component to packages/ui/components/ui/
# Infrastructure
make db-up # Start shared PostgreSQL (pgvector/pg17 image)
make db-down # Stop shared PostgreSQL
make db-reset # Drop + recreate current env's DB, then re-run migrations (local only; stop backend first)
make sqlc # regenerate sqlc code after SQL changes
pnpm install
pnpm dev:web
pnpm dev:desktop
pnpm build
pnpm typecheck
pnpm lint
pnpm test# TS/Vitest tests through Turborepo
pnpm exec playwright test
pnpm ui:add badge # shadcn/Base UI component into packages/ui
```
### CI Requirements
Worktrees share one PostgreSQL container and get isolated DB names/ports via `.env.worktree`. `make dev` auto-detects this. For manual setup use `make worktree-env`, `make setup-worktree`, and `make start-worktree`.
CI runs on Node 22 and Go 1.26.1 with a `pgvector/pgvector:pg17` PostgreSQL service. See `.github/workflows/ci.yml`.
### Worktree Support
All checkouts share one PostgreSQL container. Isolation is at the database level — each worktree gets its own DB name and unique ports via `.env.worktree`. Main checkouts use `.env`.
`make dev` auto-detects worktrees and handles everything. For explicit control:
```bash
make worktree-env # Generate .env.worktree with unique DB/ports
make setup-worktree # Setup using .env.worktree
make start-worktree # Start using .env.worktree
```
CI runs Node 22, Go 1.26.1, and a `pgvector/pgvector:pg17` PostgreSQL service.
## Coding Rules
- TypeScript strict mode is enabled; keep types explicit.
- Go code follows standard Go conventions (gofmt, go vet).
-Keep comments in code **English only**.
- Prefer existing patterns/components over introducing parallel abstractions.
- Unless the user explicitly asks for backwards compatibility, do **not** add compatibility layers, fallback paths, dual-write logic, legacy adapters, or temporary shims **for internal, non-boundary code** (a function calling another function in the same package, a component reading its own state, a store helper, etc.).
- This rule does **not** apply at API boundaries: the desktop app cannot assume the backend it talks to has the same shape as the one it was built against (older desktop installs will outlive any given server build). API response handling must follow the rules in **API Response Compatibility** below — that is a defensive boundary, not a legacy shim.
- If a flow or API is being replaced and the product is not yet live, prefer removing the old path instead of preserving both old and new behavior.
- Go follows standard conventions: `gofmt`, `go vet`, checked errors.
-Code comments must be English.
- Prefer existing patterns/components over new parallel abstractions.
- Avoid broad refactors unless required by the task.
-New global (pre-workspace) routes MUST use a single word (`/login`, `/inbox`) or a `/{noun}/{verb}` pair (`/workspaces/new`). NEVER add hyphenated word-group root routes (`/new-workspace`, `/create-team`) — they collide with common user workspace names and force endless reserved-slug audits. Reserving the noun (`workspaces`) automatically protects the entire `/workspaces/*` subtree.
-The reserved-slug list lives in **one** place: `server/internal/handler/reserved_slugs.json`. The Go side embeds the JSON; `packages/core/paths/reserved-slugs.ts` is generated from it by `pnpm generate:reserved-slugs`. Edit the JSON, run the generator, commit both. CI re-runs the generator and fails on any drift, so a stale TS file cannot land.
-When you change a CLI command or flag, an API request/response field, or product behavior that a built-in skill documents (`server/internal/service/builtin_skills/*`), update that skill's `SKILL.md`**and** its `references/*-source-map.md`in the same PR. The built-in skills are source-traced contracts shipped to agents — if the code moves and the skill doesn't, it silently teaches stale behavior.
-For internal, non-boundary code, do not add compatibility layers, fallback paths, dual writes, legacy adapters, or temporary shims unless explicitly requested.
-API boundaries are different: installed desktop clients can talk to newer backends, so response parsing must follow the API compatibility rules below.
-If a flow or API is being replaced and the product is not live, prefer removing the old path instead of preserving both.
- New global pre-workspace routes must be a single word (`/login`, `/inbox`) or `/{noun}/{verb}` (`/workspaces/new`). Do not add hyphenated root routes like `/new-workspace`.
- Reserved slugs live in `server/internal/handler/reserved_slugs.json`. Edit it, run `pnpm generate:reserved-slugs`, and commit the generated `packages/core/paths/reserved-slugs.ts`.
- When changing CLI commands/flags, API fields, or product behavior documented by built-in skills under `server/internal/service/builtin_skills/*`, update the relevant `SKILL.md` and `references/*-source-map.md` in the same PR.
### API Response Compatibility
## API Compatibility
The desktop app installed on a user's machine is older than any backend it talks to: a user on 0.2.26 will hit a server running 0.3.x, then 0.4.x, then beyond. Every response shape is a contract that **will** drift, and the frontend must survive drift without white-screening. Three concrete incidents already happened from violating this — #2143, #2147, #2192.
Frontend code must survive backend response drift, especially in installed desktop builds.
When writing code that consumes an API response, follow these rules:
- Parse API JSON with `parseWithFallback` in `packages/core/api/schema.ts` and a zod schema. Do not cast network JSON to `T`.
- Endpoint responses consumed by UI logic must pass through a schema before returning.
- Downstream UI should optional-chain and default fields defensively.
- Prefer explicit boolean checks (`=== true`) over truthy/falsy checks on server fields.
- Do not pin critical affordances to one backend boolean; combine signals when possible.
- Server-driven enum switches need a `default` branch.
- When adding or changing an endpoint, add/update the schema and include a malformed-response test.
- **Parse, don't cast.** Untyped JSON crossing the network is not `T`. Use `parseWithFallback` in `packages/core/api/schema.ts` with a `zod` schema and an explicit fallback. On validation failure it logs a warning and returns the fallback; it never throws into the UI.
- **No bare `as` casts on response bodies.** Every endpoint method whose response is consumed by UI logic must run through a schema before returning.
- **Optional-chain and default everywhere downstream.** Treat every field as possibly missing. Use explicit boolean checks (`=== true`) over truthy/falsy negation, which silently treats `undefined` and `null` as `false`.
- **Don't pin a UI affordance to a single backend field.** If a button or indicator depends on exactly one boolean from the server, a backend bug deletes it. Combine signals (cursor presence, page length, etc.) so the affordance stays available in the worst case.
- **Enum drift downgrades, not crashes.** A new server-side enum value should render a generic fallback. `switch` statements on server-driven strings must have a `default` branch.
- **When you add or change an endpoint:** add the schema in the same PR, and write at least one test that feeds a malformed response through it (missing field, wrong type, `null` array). The test fails closed if a future change breaks the contract.
## Backend UUID Rules
This is not premature defense — it is the *only* defense for an installed-app architecture. CSR-only browser apps can ship a fix in minutes; an Electron build sitting on a developer's laptop cannot.
In `server/internal/handler/`, always know where a UUID came from before using it in write queries.
### Backend Handler UUID Parsing Convention
- Resource path params that may be UUIDs or human-readable IDs must be resolved through loaders such as `loadIssueForUser`, `loadSkillForUser`, `loadAgentForUser`, or `requireDaemonRuntimeAccess`; subsequent writes use the resolved `entity.ID`.
- Pure UUID inputs from request boundaries use `parseUUIDOrBadRequest(w, s, fieldName)` and return immediately on `ok=false`.
- Trusted UUID round-trips from sqlc results or test fixtures use `parseUUID(s)`, which panics on invalid input.
- Outside handlers, `util.ParseUUID(s) (pgtype.UUID, error)` is the safe variant; always check the error.
Every Go handler in `server/internal/handler/` follows these rules. The convention exists because `util.ParseUUID` used to silently return a zero UUID on invalid input, which caused #1661 — a `DELETE` returning 204 success while the SQL `DELETE` matched zero rows.
## Web/Desktop Features
- **Resource path params that accept either a UUID or a human-readable identifier** (e.g. `chi.URLParam(r, "id")` for an issue, which accepts both `MUL-123` and a UUID) MUST be resolved through the dedicated loader (`loadIssueForUser` / `loadSkillForUser` / `loadAgentForUser` / `requireDaemonRuntimeAccess`). After resolution, all subsequent DB calls — especially `Queries.Delete*` / `Queries.Update*` — MUST use `entity.ID` from the resolved object. Never round-trip the raw URL string through `parseUUID` for a write query.
- **Pure-UUID inputs from request boundaries** (URL params that are always UUIDs, request body fields, query params, headers) MUST be validated with `parseUUIDOrBadRequest(w, s, fieldName)`. On invalid input it writes a 400 and returns `ok=false` — return immediately.
- **Trusted UUID round-trips** (sqlc-returned UUIDs being passed back into queries, test fixtures) use `parseUUID(s)` which calls `util.MustParseUUID` and panics on invalid input. A panic here means an unguarded user-input string slipped in — that is a real bug. `chi`'s `middleware.Recoverer` translates the panic into a 500 so the process keeps running.
- **`util.ParseUUID(s) (pgtype.UUID, error)`** is the only safe variant outside the handler package. Always check the error.
When adding a shared page or feature for web and desktop:
When adding a `Queries.Delete*` or `Queries.Update*` call, ask: "Where did this UUID come from?" If the answer is "raw user input that hasn't been validated," route it through `parseUUIDOrBadRequest` or a loader first.
1. Put the page/component in `packages/views/<domain>/`.
2. Add platform wiring in both `apps/web/app/` and the desktop router, unless the desktop flow is a transition overlay.
3. Use `useNavigation().push()` or `<AppLink>` in shared code.
4. Use shared guards/providers such as `DashboardGuard` from `packages/views/layout/`.
5. Keep platform-only UI in the app or inject it through props/slots.
6. Hooks that need workspace context should accept `wsId`.
### Dependency Declaration Rule
CSS for web/desktop is shared from `packages/ui/styles/`. Use semantic tokens such as `bg-background` and `text-muted-foreground`; avoid hardcoded Tailwind colors and duplicated base styles.
Every workspace (`apps/` and `packages/` directories) must explicitly declare all directly imported external packages in its own `package.json`. Relying on pnpm hoist to resolve undeclared imports (phantom deps) is prohibited — it causes production build failures when pnpm creates peer-dep variants.
## Desktop Rules
- Use `"pkg": "catalog:"` to reference the shared version from `pnpm-workspace.yaml`.
- CI enforces this via `eslint-plugin-import-x/no-extraneous-dependencies`.
- Exception: `apps/mobile/` uses pinned versions (not `catalog:`) for packages tied to its own React/Expo version.
Desktop routing has three categories:
### Package Boundary Rules
- Session routes: workspace-scoped tab destinations such as `/:slug/issues`.
- Transition flows: pre-workspace one-shot actions such as create workspace or accept invite. These are `WindowOverlay` state, not routes.
- Error/stale states: stale workspace tabs should auto-heal by dropping stale tab groups, not render desktop error pages.
These are hard constraints. Violating them breaks the cross-platform architecture:
More desktop constraints:
-`packages/core/` — zero react-dom, zero localStorage (use StorageAdapter), zero process.env, zero UI libraries. **Shared Zustand stores live here**, even view-related ones (filters, view modes) — stores are pure state, not UI.
-`packages/ui/` — zero `@multica/core` imports (pure UI, no business logic).
-`packages/views/` — zero `next/*` imports, zero `react-router-dom` imports, zero stores. Use `NavigationAdapter` for all routing.
-`apps/web/platform/` — the only place for Next.js APIs (`next/navigation`).
-`apps/desktop/src/renderer/src/platform/` — the only place for react-router-dom navigation wiring.
-New pre-workspace desktop flows register a `WindowOverlay` type in `stores/window-overlay-store.ts`; do not add them to `routes.tsx`.
-`setCurrentWorkspace(slug, uuid)` from`@multica/core/platform` is the active workspace source of truth.
-Code that leaves workspace context must call `setCurrentWorkspace(null, null)` explicitly.
-Leave/delete workspace flow order: read cached destination, clear current workspace, navigate, then run the mutation.
-Cross-workspace navigation must go through the navigation adapter so it can call `switchWorkspace(slug, targetPath)`.
- Full-window desktop views outside the dashboard shell must mount `<DragStrip />` from `@multica/views/platform` as the first flex child. Interactive controls in the top 48px need `WebkitAppRegion: "no-drag"`.
### The No-Duplication Rule (web + desktop)
## Mobile Rules
**If the same logic exists in both web and desktop, it must be extracted to a shared package.**
Read `apps/mobile/CLAUDE.md` before touching `apps/mobile/`. It contains the mandatory pre-flight process, import limits, parity rules, tech stack, UI rules, data helpers, realtime strategy, and mobile release flow.
This applies to everything between web and desktop: components, hooks, guards, providers, utility functions. The decision process:
Root-level reminders:
1. Does this code depend on Next.js or Electron APIs? → Keep in the respective app.
2. Does it depend on `react-router-dom` or `next/navigation`? → Keep in app's `platform/` layer.
3. Everything else → belongs in `packages/core/` (headless logic) or `packages/views/` (UI components).
- Mobile shares only `@multica/core` types and pure functions.
- Mobile must match web/desktop product semantics: counts, permissions, enums/transitions, and data identity.
- Mobile may differ in UI/interaction when the phone context requires it.
When the two apps need different behavior for the same concept (e.g., different loading UI), extract the shared logic into a component with props/slots for the differences. Don't duplicate the logic.
## UI Rules
### Cross-Platform Development Rules (web + desktop)
- Prefer shadcn/Base UI components over custom implementations. Add them with `pnpm ui:add <component>` from the repo root.
- Use design tokens and semantic classes; avoid hardcoded colors.
- Do not introduce extra local state unless the design requires it.
- Handle overflow, long text, scrolling, alignment, and spacing deliberately.
- If a component is identical between web and desktop, it belongs in a shared package.
When adding a new page or feature for web/desktop:
## Testing
1.**New page component** → add to `packages/views/<domain>/`. Never import from `next/*` or `react-router-dom`.
2.**Wire it in both apps** → add a route in `apps/web/app/` (Next.js page file) AND in the desktop router. **Exception**: pre-workspace transition flows (create workspace, accept invite) are NOT routes on desktop — they're `WindowOverlay` state. See *Desktop-specific Rules → Route categories*.
3.**Navigation** → use `useNavigation().push()` or `<AppLink>`. Never use framework-specific link/router APIs in shared code.
4.**Shared guards/providers** → use `DashboardGuard` from `packages/views/layout/`. Don't create separate guard logic per app.
5.**Platform-specific UI** → if a feature is web-only or desktop-only, keep it in the respective app. Use props slots (`extra`, `topSlot`) on shared layout components to inject platform-specific UI.
6.**New hooks that need workspace context** → accept `wsId` as parameter instead of reading from `useWorkspaceId()` Context, so they work both inside and outside `WorkspaceIdProvider`.
Tests follow the code:
### CSS Architecture (web + desktop)
| What is tested | Location |
| --- | --- |
| Shared business logic, stores, queries, hooks | `packages/core/*.test.ts` |
| Platform wiring such as cookies, redirects, search params | `apps/web/*.test.tsx` or `apps/desktop/` |
| End-to-end flows | `e2e/*.spec.ts` |
| Backend | `server/` Go tests |
Web and desktop share the same CSS foundation from `packages/ui/styles/`.
Rules:
-**Design tokens** → use semantic tokens (`bg-background`, `text-muted-foreground`). Never use hardcoded Tailwind colors (`text-red-500`, `bg-gray-100`).
-**Shared styles** → `packages/ui/styles/`. Never duplicate scrollbar styling, keyframes, or base layer rules in app CSS.
-**`@source` directives** → both apps scan shared packages so Tailwind sees all class names.
## Mobile-specific Rules
Rules for `apps/mobile/` live in `apps/mobile/CLAUDE.md`. Read it before touching anything in `apps/mobile/` — it covers what may be imported from `@multica/core/`, the React version policy, the build/release pipeline, and the locked tech-stack baseline.
## Desktop-specific Rules
These rules apply to `apps/desktop/` only. Web has different constraints (URL bar, SSR, no tabs) and doesn't share these concerns. Every rule in this section was added after a concrete bug — treat them as enforced, not suggestions.
### Route categories
Every path in the desktop app falls into exactly one category. Choosing the wrong one reproduces bugs we've already fixed.
- **Session routes** — workspace-scoped pages (`/:slug/issues`, `/:slug/settings`). Rendered by the per-tab memory router under `WorkspaceRouteLayout`. These are legitimate tab destinations.
- **Transition flows** — pre-workspace / one-shot actions (create workspace, accept invite). **NOT routes.** They live as `WindowOverlay` state, dispatched when the navigation adapter sees `push('/workspaces/new')` or `push('/invite/<id>')`. The shared view (`NewWorkspacePage`, `InvitePage`) is the content; the overlay wrapper supplies platform chrome.
- **Error / stale states** — "workspace not available", tabs pointing at a revoked workspace. **NOT pages.**`WorkspaceRouteLayout` auto-heals by dropping the stale tab group from the store; the user never lands on an explicit error screen. Web keeps `NoAccessPage` (shareable URL makes the error state meaningful); desktop has no URL bar so stale = heal silently.
**Adding a new pre-workspace flow on desktop**: register a new `WindowOverlay` type in `stores/window-overlay-store.ts`. Do NOT add it to `routes.tsx`. If a shared view needs the flow on both platforms, add the route on web (`apps/web/app/(auth)/...`) AND the overlay type on desktop — the shared view component is identical.
### Workspace context
`setCurrentWorkspace(slug, uuid)` from `@multica/core/platform` is the single source of truth for the active workspace. `WorkspaceRouteLayout` sets it on mount; unmount does NOT clear it. Code that leaves workspace context (leave/delete workspace, force-navigate to overlay) must call `setCurrentWorkspace(null, null)` explicitly.
### Workspace destructive operations
Leave / Delete workspace flows must follow this order, otherwise concurrent refetches race and the renderer hard-reloads:
1. Read destination from cached workspace list.
2.`setCurrentWorkspace(null, null)`.
3.`navigation.push(destination)`.
4. THEN `await mutation.mutateAsync(workspaceId)`.
### Tab isolation
Tabs are grouped per workspace in `stores/tab-store.ts`. The TabBar shows only the active workspace's tabs; cross-workspace tab leakage is impossible by construction (no flat global tabs array).
Cross-workspace `push(path)` is detected by the navigation adapter (`platform/navigation.tsx`) and translated into `switchWorkspace(slug, targetPath)` — NOT a navigation within the current tab's router. Don't bypass the adapter; always go through `useNavigation()` from shared code.
### Drag region (macOS)
Every full-window desktop view (anything outside the dashboard shell) must mount `<DragStrip />` from `@multica/views/platform` as the first flex child of the page root, otherwise users can't drag the window. Interactive UI inside the top 48px needs `WebkitAppRegion: "no-drag"` to stay clickable.
## UI/UX Rules
- Prefer shadcn components over custom implementations. Install via `pnpm ui:add <component>` from project root — adds to `packages/ui/components/ui/`. All components use Base UI primitives (`@base-ui/react`), not Radix.
- Use shadcn design tokens for styling. Avoid hardcoded color values.
- Do not introduce extra state (useState, context, reducers) unless explicitly required by the design.
- Pay close attention to **overflow** (truncate long text, scrollable containers), **alignment**, and **spacing** consistency.
- **If a component is identical between web and desktop, it belongs in a shared package.** Do not copy-paste between apps.
## Testing Rules
### Where to write tests
Tests follow the code, not the app. This is the most important testing principle in this monorepo:
| What you're testing | Where the test lives | Why |
|---|---|---|
| Shared business logic (stores, queries, hooks) | `packages/core/*.test.ts` | No DOM needed, pure logic |
| End-to-end user flows | `e2e/*.spec.ts` | Real browser, real backend |
**Never test shared component behavior in an app's test file.** If a test requires mocking `next/navigation` or `react-router-dom` to test a component from `@multica/views`, the test is in the wrong place — move it to `packages/views/` and mock `@multica/core` instead.
### Test infrastructure
-`packages/core/` — Vitest, Node environment (no DOM)
All test deps are in the pnpm catalog for unified versioning.
### Mocking conventions
- Mock `@multica/core` stores with `vi.hoisted()` + `Object.assign(selectorFn, { getState })` pattern (Zustand stores are both callable and have `.getState()`).
-Never test shared component behavior in an app test file.
-`packages/views/` tests must not mock `next/*` or `react-router-dom`.
-Mock `@multica/core` stores with the Zustand callable-store shape (`selectorFn` plus `getState`).
- Mock `@multica/core/api` for API calls.
-In `packages/views/` tests: never mock `next/*` or `react-router-dom` — those don't exist here.
-In `apps/web/` tests: mock framework-specific APIs only for platform-specific behavior.
-E2E tests should use `TestApiClient`for setup/teardown.
-Prefer writing the failing test in the correct package before implementation when the change is behavioral.
### TDD workflow
## Verification
1. Write failing test in the **correct package** first.
2. Write implementation.
3. Run `pnpm test` (Turborepo discovers all packages).
4. Green → done.
For code changes, run the narrowest useful checks while iterating, then run broader verification when risk justifies it or when asked.
### Go tests
Standard `go test`. Tests should create their own fixture data in a test database.
### E2E tests
E2E tests should be self-contained. Use the `TestApiClient` fixture for data setup/teardown:
After writing or modifying code, always run the full verification pipeline:
Useful checks:
```bash
pnpm typecheck
pnpm test
make test
pnpm exec playwright test
make check
```
**Workflow:**
- Write code to satisfy the requirement
- Run `make check`
- If any step fails, read the error output, fix the code, and re-run
- Repeat until all checks pass
- Only then consider the task complete
Do not claim verification passed unless you ran it. If you skip checks because the change is docs-only or the user asked not to run them, say so.
**Quick iteration:** If you know only TypeScript or Go is affected, run individual checks first for faster feedback, then finish with a full `make check` before marking work complete.
## Commits and Releases
## CLI Release
- Commits should be atomic and use conventional prefixes: `feat(scope)`, `fix(scope)`, `refactor(scope)`, `docs`, `test(scope)`, `chore(scope)`.
- A production deployment requires a CLI release tag on `main`: create `v0.x.x`, push it, and let `release.yml` publish binaries and the Homebrew tap.
- Bump patch by default unless the user specifies a version.
**Prerequisite:** A CLI release must accompany every Production deployment.
## Domain Reminders
1. Create a tag on the `main` branch: `git tag v0.x.x`
2. Push the tag: `git push origin v0.x.x`
3. GitHub Actions automatically triggers `release.yml`: runs Go tests → GoReleaser builds multi-platform binaries → publishes to GitHub Releases + Homebrew tap
By default, bump the patch version each release (e.g. `v0.1.12` → `v0.1.13`), unless the user specifies a specific version.
## Multi-tenancy
All queries filter by `workspace_id`. Membership checks gate access. `X-Workspace-ID` header routes requests to the correct workspace.
## Agent Assignees
Assignees are polymorphic — can be a member or an agent. `assignee_type` + `assignee_id` on issues. Agents render with distinct styling (purple background, robot icon).
- All queries filter by `workspace_id`; membership gates access; `X-Workspace-ID` selects the workspace.
- Issue assignees are polymorphic: `assignee_type` plus`assignee_id` can reference a member or an agent.
You need at least one installed. The daemon registers each detected CLI as an available runtime.
@@ -220,6 +221,10 @@ Agent-specific overrides:
| `MULTICA_KIMI_MODEL` | Override the Kimi model used |
| `MULTICA_KIRO_PATH` | Custom path to the `kiro-cli` binary |
| `MULTICA_KIRO_MODEL` | Override the Kiro model used |
| `MULTICA_QODER_PATH` | Custom path to the `qodercli` binary |
| `MULTICA_QODER_MODEL` | Override the Qoder model used |
The daemon launches Qoder as `qodercli --yolo --acp`, matching Qoder’s ACP “bypass permissions” mode so tool runs do not block on interactive approval in headless runs.
`MULTICA_CLAUDE_ARGS` and `MULTICA_CODEX_ARGS` are parsed with POSIX shellword quoting, so values such as `--model "gpt-5.1 codex" --sandbox read-only` are split like a shell command line. Agent arguments are applied in this order: hardcoded Multica defaults, daemon-wide env defaults, then per-agent `custom_args` from the task.
# Aggregated token usage for an issue (sum across all its task runs)
multica issue usage <issue-id>
multica issue usage <issue-id> --output json
```
The `usage` command returns the aggregated token usage for an issue, summed across all of its task runs: input tokens, output tokens, cache read/write tokens, and the run count (`task_count`). It wraps `GET /api/issues/<id>/usage` — the same figures the issue detail view shows. Use `--output json` to feed billing/cost tooling.
The `runs` command shows all past and current executions for an issue, including running tasks. Table output uses short task UUID prefixes by default; pass `--full-id` to print canonical task UUIDs. The `run-messages` command accepts full task UUIDs directly; copied short task prefixes must be scoped with `--issue <issue-id>` so the CLI only checks that issue's runs. It shows the detailed message log (tool calls, thinking, text, errors) for a single run. Use `--since` for efficient polling of in-progress runs.
## Projects
@@ -648,14 +659,18 @@ multica autopilot create \
--title "Nightly bug triage"\
--description "Scan todo issues and prioritize."\
--agent "Lambda"\
--mode create_issue
--mode create_issue\
--subscriber "Alice"
multica autopilot update <id> --status paused
multica autopilot update <id> --description "New prompt"
`--mode` accepts `create_issue` (creates a new issue on each run and assigns it to the agent) or `run_only` (enqueues a direct agent task without creating an issue). `--agent` accepts either a name or UUID.
`--subscriber` accepts a workspace member name or user ID and may be repeated; on update it replaces the autopilot's subscriber template. Subscribers receive inbox notifications for issues created by a `create_issue` autopilot. Use `--clear-subscribers` to remove all autopilot subscribers.
### Manual Trigger
@@ -699,3 +714,79 @@ Most commands support `--output` with two formats:
multica issue list --output json
multica daemon status --output json
```
## Error Messages
The CLI funnels command errors returned to the top-level handler through a
single user-facing translation layer (`server/internal/cli/errors.go`) so that
what you see on the terminal is a short, actionable sentence rather than a raw
Go error, an HTTP status line, or an internal `resolve issue: ...` chain. (A
few commands print their own output or run deliberate fast probes — for example
`setup`'s short `/health` reachability check — and don't go through this
layer.) The underlying detail is still available on demand (see `--debug`).
### What you see
- **Friendly, single-line message.** Transport failures (timeout, DNS,
connection refused, TLS) and HTTP status failures (401/403/404/409/400·422/
429/5xx) are each rendered as one clear sentence with a next step — for
example a timeout suggests checking the network or raising
`MULTICA_HTTP_TIMEOUT`, and a 401 tells you to run `multica login`.
- **Server-provided validation messages are preserved.** For a 400/422 that
carries a message from the server, that message is shown verbatim
(`Invalid request: <server message>`); only when there is none do you get the
generic "check your values / run with --help" hint.
- **No leaked internals by default.** Raw URLs, status lines, JSON bodies, and
the internal verb chain are hidden unless you ask for them.
### Language
Messages default to **English**, matching the rest of the CLI's help output.
If a Chinese locale is detected in `LC_ALL`, `LC_MESSAGES`, or `LANG` (in that
precedence order), messages switch to **Chinese**. No flag is needed; set the
locale as usual:
```bash
LANG=zh_CN.UTF-8 multica issue get MUL-9999 # 错误信息显示为中文
```
### Exit codes
The process exit code is tiered so scripts can branch on the failure class:
@@ -30,7 +31,7 @@ Turn coding agents into real teammates — assign tasks, track progress, compoun
Multica turns coding agents into real teammates. Assign issues to an agent like you'd assign to a colleague — they'll pick up the work, write code, report blockers, and update statuses autonomously.
No more copy-pasting prompts. No more babysitting runs. Your agents show up on the board, participate in conversations, and compound reusable skills over time. Think of it as open-source infrastructure for managed agents — vendor-neutral, self-hosted, and designed for human + AI teams. Works with **Claude Code**, **Codex**, **GitHub Copilot CLI**, **OpenClaw**, **OpenCode**, **Hermes**, **Gemini**, **Pi**, **Cursor Agent**, **Kimi**, and **Kiro CLI**.
No more copy-pasting prompts. No more babysitting runs. Your agents show up on the board, participate in conversations, and compound reusable skills over time. Think of it as open-source infrastructure for managed agents — vendor-neutral, self-hosted, and designed for human + AI teams. Works with **Claude Code**, **Codex**, **GitHub Copilot CLI**, **OpenClaw**, **OpenCode**, **Hermes**, **Gemini**, **Pi**, **Cursor Agent**, **Kimi**,**Kiro CLI**, and **Qoder CLI**.
For larger teams, Squads add a stable routing layer: assign work to a group led by an agent, and the leader delegates to the right member.
@@ -114,7 +115,7 @@ multica setup # Connect to Multica Cloud, log in, start daemon
multica setup # Configure, authenticate, and start the daemon
```
The daemon runs in the background and auto-detects agent CLIs (`claude`, `codex`, `copilot`, `openclaw`, `opencode`, `hermes`, `gemini`, `pi`, `cursor-agent`, `kimi`, `kiro-cli`, `agy`) on your PATH.
The daemon runs in the background and auto-detects agent CLIs (`claude`, `codex`, `copilot`, `openclaw`, `opencode`, `hermes`, `gemini`, `pi`, `cursor-agent`, `kimi`, `kiro-cli`, `agy`, `qodercli`) on your PATH.
### 2. Verify your runtime
@@ -124,7 +125,7 @@ Open your workspace in the Multica web app. Navigate to **Settings → Runtimes*
### 3. Create an agent
Go to **Settings → Agents** and click **New Agent**. Pick the runtime you just connected and choose a provider (Claude Code, Codex, GitHub Copilot CLI, OpenClaw, OpenCode, Hermes, Gemini, Pi, Cursor Agent, Kimi, Kiro CLI, or Antigravity). Give your agent a name — this is how it will appear on the board, in comments, and in assignments.
Go to **Settings → Agents** and click **New Agent**. Pick the runtime you just connected and choose a provider (Claude Code, Codex, GitHub Copilot CLI, OpenClaw, OpenCode, Hermes, Gemini, Pi, Cursor Agent, Kimi, Kiro CLI, Antigravity, or Qoder CLI). Give your agent a name — this is how it will appear on the board, in comments, and in assignments.
### 4. Assign your first task
@@ -165,7 +166,7 @@ See the [CLI and Daemon Guide](CLI_AND_DAEMON.md) for the full command reference
`multica skill import --url <url>`의 기본값은 `--on-conflict fail`입니다. 같은 이름의 스킬이 이미 있으면 명령은 구조화된 `conflict` 결과로 종료되며 워크스페이스를 변경하지 않습니다.
기존 스킬을 만든 사용자이고, 스킬 ID와 에이전트 연결은 유지한 채 내용을 바꾸려면 `--on-conflict overwrite`를 사용하세요. 기존 스킬을 그대로 두고 복사본을 가져오려면 `--on-conflict rename`을 사용하면 `-2` 같은 접미사가 자동으로 붙습니다. 같은 이름의 항목을 그냥 건너뛰려면 `--on-conflict skip`을 사용하세요.
`multica autopilot create/update`는 `create_issue` 모드 autopilot이 만드는 이슈의 기본 구독자를 설정하는 `--subscriber`도 받습니다. `update`에서는 `--clear-subscribers`로 제거할 수 있습니다.
@@ -39,9 +39,9 @@ import { Callout } from "fumadocs-ui/components/callout";
## 이슈 참조하기
다른 이슈를 링크하려면 `MUL-123`처럼 이슈 키를 입력하세요. Multica는 댓글에서 실제 존재하는 이슈 키를 해석하여 내부적으로 `mention://issue/<uuid>` 링크로 저장합니다. 이슈 링크는 단순한 상호 참조일 뿐입니다. 사람에게 알림을 보내지 않으며 에이전트를 트리거하지도 않습니다.
다른 이슈를 링크하려면 댓글 mention 선택기에서 해당 이슈를 선택하세요. Multica는 이슈 링크를 명시적인 `[MUL-123](mention://issue/<uuid>)` mention 링크로 저장합니다. 이슈 링크는 단순한 상호 참조일 뿐입니다. 사람에게 알림을 보내지 않으며 에이전트를 트리거하지도 않습니다.
보통은 `[MUL-123](mention://issue/<uuid>)`을 직접 손으로 작성할 필요가 없습니다. 그 형식은 Multica가 키를 해석한 뒤에 사용하는 표준 내부 표현입니다.
`MUL-123` 같은 bare 이슈 키를 입력하면 일반 텍스트로 유지됩니다. 따라서 `feature/MUL-123` 같은 댓글 안의 브랜치 이름과 경로도 다시 작성되지 않습니다.
<Callout type="info">
Markdown 강조는 CommonMark 규칙을 따릅니다. 굵은 텍스트가 문장 부호나 닫는 따옴표로 끝나고 그 뒤에 한국어 조사가 바로 이어지면, 닫는 `**`가 인식되지 않을 수 있습니다.
@@ -39,9 +39,9 @@ Mentioning the same person multiple times in one comment still produces **only o
## Referencing issues
To link another issue, type its issue key, such as `MUL-123`. Multica resolves real issue keys in comments and stores them as an internal `mention://issue/<uuid>` link. Issue links are cross-references only: they do not notify people and they do not trigger agents.
To link another issue, choose it from the comment mention picker. Multica stores issue links as an explicit `[MUL-123](mention://issue/<uuid>)` mention link. Issue links are cross-references only: they do not notify people and they do not trigger agents.
You normally do not need to write `[MUL-123](mention://issue/<uuid>)` by hand. That format is the canonical internal representation after Multica has resolved the key.
Typing a bare issue key, such as `MUL-123`, keeps it as plain text. This also keeps branch names and paths, such as `feature/MUL-123`, from being rewritten inside comments.
<Callout type="info">
Markdown emphasis follows CommonMark rules. When bold text ends with punctuation or a closing quote and is immediately followed by a Korean particle, the closing `**` may not be recognized.
- **같은 데몬, 워크스페이스, 도구는 정확히 하나의 런타임을 만듭니다.** 데몬을 재시작해도 중복 레코드가 생기지 않습니다
- Multica UI의 **런타임** 페이지가 이 행들을 나열합니다
## 사용자 지정 런타임 프로필
기본 provider 감지는 일반적인 도구를 다룹니다. 팀에서 Multica가 지원하는 프로토콜 패밀리로 동작하지만 워크스페이스별 실행 명령이 필요한 AI CLI를 쓰는 경우에는 **custom runtime profile**을 정의할 수 있습니다. Runtimes UI 또는 CLI에서 관리합니다.
profile의 명령이나 인수 변경은 데몬이 다시 등록된 뒤 새로 가져오는 작업부터 적용됩니다. 이미 실행 중인 작업은 시작할 때의 인수를 유지합니다. 혼합 배포에서는 server를 먼저 업그레이드한 뒤 daemon을 순차적으로 업데이트하는 것을 권장합니다. `fixed_args` 입력은 server 쪽 Runtimes UI가 담당하고, `failed_profiles` 등록 보고도 server가 표시합니다. 오래된 구성요소는 알 수 없는 필드를 명확히 실패시키기보다 무시할 수 있으므로 server를 먼저 올리면 rollout을 더 잘 관찰할 수 있습니다.
<Callout type="info">
**클라우드 런타임이 곧 제공됩니다.** 현재는 대기자 명단 단계입니다. 제공이 시작되면 로컬 데몬을 실행하지 않고도 Multica Cloud에서 직접 에이전트 작업을 실행할 수 있습니다. [다운로드 페이지](https://multica.ai/download)에서 이메일로 등록하면 알림을 받을 수 있습니다.
Profile command or argument edits apply to newly claimed tasks after the daemon
re-registers. Running tasks keep the launch arguments they started with. For
mixed deployments, upgrade the server before rolling out newer daemons: the
server-side Runtimes UI stores `fixed_args`, and the server is what surfaces
`failed_profiles` registration reports. Older components may ignore fields they
do not understand instead of failing loudly, so treating the server upgrade as
the first step keeps the rollout observable.
<Callout type="info">
**Cloud runtimes are coming**, currently in a waitlist phase. Once available, you'll be able to execute agent tasks directly on Multica Cloud without running a local daemon. Sign up with your email on the [download page](https://multica.ai/download) to get notified.
@@ -181,9 +181,19 @@ S3 앞에 CloudFront를 두는 경우 세 가지 변수가 적용됩니다: `CLO
| `MULTICA_DAEMON_MAX_CONCURRENT_TASKS` | `20` | 최대 동시 작업 수 |
| `MULTICA_<PROVIDER>_PATH` | CLI 이름과 일치 | 각 AI 코딩 도구 실행 파일의 경로 (예: `MULTICA_CLAUDE_PATH`) |
| `MULTICA_<PROVIDER>_MODEL` | 비어 있음 | 각 AI 코딩 도구의 기본 모델 |
| `MULTICA_<PROVIDER>_ARGS` | 비어 있음 | 백엔드별 데몬 전역 기본 CLI 인자. 각 작업에 대해 각 에이전트 자체의 `custom_args`보다 먼저 적용됩니다. `MULTICA_CLAUDE_ARGS`, `MULTICA_CODEX_ARGS`, `MULTICA_CODEBUDDY_ARGS`를 지원 |
각 파라미터가 데몬 동작에 어떻게 영향을 미치는지에 대한 전체 설명은 [데몬과 런타임](/daemon-runtimes)을 참고하세요.
### 기본 에이전트 인자 (`MULTICA_<PROVIDER>_ARGS`)
백엔드에 대해 **플릿 전역 기본값** 계층의 CLI 플래그를 설정합니다. 각 에이전트의 `custom_args`를 일일이 수정하지 않고도 데몬의 모든 에이전트에 기본 비용·리소스 기준선(예: `--max-turns`)을 적용할 수 있는 편리한 방법입니다. 이는 넘을 수 없는 상한이 아니라 기본 계층입니다. 각 에이전트 자체의 `custom_args`가 뒤에 추가되어 이를 덮어쓸 수 있습니다(아래 **우선순위** 참고).
- **우선순위:** 기본 인자가 먼저 적용되고, 그다음 각 에이전트 자체의 `custom_args`가 추가됩니다. 값을 받는 플래그의 경우 다운스트림 CLI 자체의 인자 파서가 최종 적용 값을 결정합니다(대부분의 도구는 마지막 항목이 우선). 따라서 개별 에이전트는 데몬 기본값을 높일 수 있지만, 에이전트가 덮어쓰지 않은 부분에는 기본값이 계속 적용됩니다.
- **파싱:** 값은 POSIX 셸 단어 규칙으로 분할되므로 따옴표를 사용할 수 있습니다 — `MULTICA_CLAUDE_ARGS='--append-system-prompt "multi word"'`는 두 개의 토큰으로 파싱됩니다.
- **안전성:** 기본 인자 계층과 에이전트별 `custom_args` 계층 모두 동일한 blocked-flags 필터를 통과합니다. 따라서 프로토콜에 중요한 플래그(Claude의 `-p`, `--output-format`, `--input-format`, `--permission-mode`, `--mcp-config` 및 Codex의 `--listen` 등)는 어느 계층을 통해서도 주입할 수 없습니다.
@@ -181,9 +181,19 @@ The daemon runs on the user's local machine, and its config is read from local e
| `MULTICA_DAEMON_MAX_CONCURRENT_TASKS` | `20` | Max concurrent tasks |
| `MULTICA_<PROVIDER>_PATH` | matches the CLI name | Path to each AI coding tool's executable (for example `MULTICA_CLAUDE_PATH`) |
| `MULTICA_<PROVIDER>_MODEL` | empty | Default model for each AI coding tool |
| `MULTICA_<PROVIDER>_ARGS` | empty | Daemon-wide default CLI arguments for a backend, applied to every task before each agent's own `custom_args`. Supported for `MULTICA_CLAUDE_ARGS`, `MULTICA_CODEX_ARGS`, and `MULTICA_CODEBUDDY_ARGS` |
For a full explanation of how each parameter affects daemon behavior, see [Daemon and runtimes](/daemon-runtimes).
These set a **fleet-wide default** layer of CLI flags for a backend — a convenient way to apply a default cost or resource baseline (for example `--max-turns`) across every agent on a daemon without editing each agent's `custom_args` individually. This is a default layer, not a hard ceiling: per-agent `custom_args` are appended afterward and can override it (see **Precedence** below).
- **Precedence:** the default args are applied first, then each agent's own `custom_args` are appended after. For flags that take a value, the downstream CLI's own argument parser decides the winner (last occurrence wins for most tools), so an individual agent can raise a daemon default but the default still applies wherever the agent doesn't override it.
- **Parsing:** the value is split with POSIX shell-word rules, so quoting works — `MULTICA_CLAUDE_ARGS='--append-system-prompt "multi word"'` parses into two tokens.
- **Safety:** both the default-args and per-agent `custom_args` layers pass through the same blocked-flags filter, so protocol-critical flags (such as `-p`, `--output-format`, `--input-format`, `--permission-mode`, `--mcp-config` for Claude, and `--listen` for Codex) cannot be injected through either layer.
description: Multica는 사용자 기기에 설치된 AI 코딩 도구를 구동합니다. 이 페이지에서는 데몬이 도구를 감지할 수 있도록 지원되는 12종의 도구를 각각 설치하는 방법을 설명합니다.
description: Multica는 사용자 기기에 설치된 AI 코딩 도구를 구동합니다. 이 페이지에서는 데몬이 도구를 감지할 수 있도록 지원되는 13종의 도구를 각각 설치하는 방법을 설명합니다.
---
import { Callout } from "fumadocs-ui/components/callout";
Multica에서 **런타임**이란 사용자 기기의 데몬과, 데몬이 `PATH`에서 찾아낸 AI 코딩 도구 하나가 짝을 이룬 것입니다. 온보딩의 "런타임 연결" 단계에서 **지원되는 도구를 감지하지 못했습니다**라고 표시된다면, 데몬이 `PATH`를 스캔했지만 구동 방법을 아는 12종의 도구 중 어느 것도 찾지 못했다는 뜻입니다. 아래 도구 중 하나(또는 여러 개)를 설치한 다음 해당 단계로 돌아와 다시 스캔하세요. 몇 초 안에 런타임이 나타납니다.
Multica에서 **런타임**이란 사용자 기기의 데몬과, 데몬이 `PATH`에서 찾아낸 AI 코딩 도구 하나가 짝을 이룬 것입니다. 온보딩의 "런타임 연결" 단계에서 **지원되는 도구를 감지하지 못했습니다**라고 표시된다면, 데몬이 `PATH`를 스캔했지만 구동 방법을 아는 13종의 도구 중 어느 것도 찾지 못했다는 뜻입니다. 아래 도구 중 하나(또는 여러 개)를 설치한 다음 해당 단계로 돌아와 다시 스캔하세요. 몇 초 안에 런타임이 나타납니다.
이 페이지는 다음 문서의 설치 측면 동반 문서입니다.
@@ -31,9 +31,9 @@ multica daemon restart
또는 데스크톱 앱에서는 앱을 다시 실행하기만 하면 됩니다. 데몬은 시작될 때마다 `PATH`를 다시 스캔합니다.
## 지원되는 12종의 도구
## 지원되는 13종의 도구
대략 많이 쓰이는 순서대로 나열했습니다. 이미 자격 증명을 갖고 있는 것을 골라 사용하세요. 12종을 모두 설치할 필요는 없습니다.
대략 많이 쓰이는 순서대로 나열했습니다. 이미 자격 증명을 갖고 있는 것을 골라 사용하세요. 13종을 모두 설치할 필요는 없습니다.
### Claude Code (Anthropic)
@@ -48,7 +48,7 @@ multica daemon restart
### Codex (OpenAI)
더 세분화된 승인 게이트를 갖춘 JSON-RPC 2.0 전송 방식입니다. MCP 구성은 작업별 `$CODEX_HOME/config.toml`에 기록됩니다. **세션 재개 코드는 존재하지만 현재 도달할 수 없습니다** — 재개가 필요하다면 Claude Code 또는 ACP 계열 중 하나를 선택하세요.
더 세분화된 승인 게이트를 갖춘 JSON-RPC 2.0 전송 방식입니다. MCP 구성은 작업별 `$CODEX_HOME/config.toml`에 기록됩니다. **세션 재개가 동작합니다** — Multica는 Codex app-server의 `thread/resume`으로 재개하며, 오래되었거나 없는 thread는 새 thread로 폴백합니다.
| | |
|---|---|
@@ -58,7 +58,7 @@ multica daemon restart
### Cursor (Anysphere)
Cursor 에디터에 대응하는 CLI입니다. **세션 재개가 작동하지 않습니다** — Cursor의 CLI가 세션 id를 반환하지 않으므로 재개 시 전달하는 값은 항상 유효하지 않습니다.
Cursor 에디터에 대응하는 CLI입니다. **세션 재개가 동작합니다** — 현재 Cursor Agent는 stream-json 이벤트에서 `session_id`를 반환하고, Multica는 다음 실행 때 이를 `--resume <id>`로 전달합니다.
Gemini 2.5 및 3 시리즈를 지원합니다. 세션 재개와 MCP는 없습니다 — 단발성 작업에 적합합니다.
| | |
|---|---|
| 데몬이 찾는 이름 | `gemini` |
| 설치 | [github.com/google-gemini/gemini-cli](https://github.com/google-gemini/gemini-cli)의 공식 가이드를 따르세요. 일반적인 방법은 npm 패키지 `@google/gemini-cli`입니다. |
| 인증 | `gemini`를 실행하면 Google 계정 로그인을 요청하거나, `GEMINI_API_KEY`를 설정하세요. |
### OpenCode (SST)
오픈 소스 CLI 에이전트입니다. 자체 설정 파일에서 사용 가능한 모델을 동적으로 발견합니다 — 자신만의 모델 카탈로그를 직접 가져오려는 사용자에게 잘 맞습니다. `OPENCODE_CONFIG_CONTENT`를 통해 에이전트의 `mcp_config` 필드도 소비합니다.
@@ -147,6 +137,26 @@ ACP 프로토콜 에이전트입니다(Kimi와 전송 방식을 공유). 세션
| 설치 | Inflection의 CLI 문서 [pi.ai](https://pi.ai/)를 참고하세요. |
| 인증 | 공급사 문서에 따릅니다. |
### CodeBuddy (Tencent)
Claude Code 호환 CLI 에이전트입니다. Multica는 Claude Code와 동일한 stream-json 프로토콜로 구동합니다: 세션 재개는 `--resume`로 동작하고, MCP 구성은 `--mcp-config`로 전달되며, 스킬은 `.claude/skills/`에 배치됩니다. 모델은 동적으로 탐색됩니다.
| | |
|---|---|
| 데몬이 찾는 이름 | `codebuddy` |
| 설치 | 공식 CLI 문서 [codebuddy.ai/cli](https://www.codebuddy.ai/cli)를 참고하세요. |
| 인증 | 공급사 문서에 따릅니다. |
### Qoder (Alibaba)
stdio 위에서 ACP 프로토콜을 사용하는 에이전트형 코딩 CLI입니다(Hermes, Kimi, Kiro CLI와 전송 계층을 공유합니다). 세션 재개는 ACP `session/resume`를 통해 동작하고, MCP 구성은 ACP `mcpServers`로 전달되며, 모델 선택은 동적으로 탐색되고, 스킬은 `.qoder/skills/`로 복사됩니다.
| | |
|---|---|
| 데몬이 찾는 이름 | `qodercli` |
| 설치 | 공식 CLI 문서 [qoder.com/cli](https://qoder.com/cli)를 참고하세요. |
| 인증 | 공급사 문서에 따릅니다. |
### Antigravity (Google)
Google의 Antigravity CLI(`agy`)입니다. Google의 Antigravity 서비스와 짝을 이루며 Gemini 기반 모델을 실행합니다. 세션 재개는 `--conversation <id>`를 통해 작동하며, 데몬이 CLI 로그 파일에서 이를 캡처합니다. 모델 선택은 Antigravity CLI 자체 내부에서 관리됩니다 — Multica는 이 제공자에 대해 에이전트별 모델 선택기를 비활성화합니다. 스킬은 `.agents/skills/`에 기록됩니다(CLI가 Gemini CLI의 워크스페이스 스킬 레이아웃을 상속함 — [Antigravity 문서](https://antigravity.google/docs/gcli-migration) 참고).
description: Multica drives whichever AI coding tools you have on your machine. This page shows you how to install each of the 12 supported tools so the daemon can detect them.
description: Multica drives whichever AI coding tools you have on your machine. This page shows you how to install each of the 13 supported tools so the daemon can detect them.
---
import { Callout } from "fumadocs-ui/components/callout";
A **runtime** in Multica is the daemon on your machine paired with one AI coding tool the daemon found on your `PATH`. If the onboarding "Connect a runtime" step shows **No supported tools detected**, it means the daemon scanned `PATH` and didn't find any of the 12 tools it knows how to drive. Install one (or several) of the tools below, then come back to the step and re-scan — the runtime will show up within a few seconds.
A **runtime** in Multica is the daemon on your machine paired with one AI coding tool the daemon found on your `PATH`. If the onboarding "Connect a runtime" step shows **No supported tools detected**, it means the daemon scanned `PATH` and didn't find any of the 13 tools it knows how to drive. Install one (or several) of the tools below, then come back to the step and re-scan — the runtime will show up within a few seconds.
This page is the install-side companion to:
@@ -31,9 +31,9 @@ multica daemon restart
Or, in the desktop app, just relaunch the app. The daemon re-scans `PATH` on every start.
## The 12 supported tools
## The 13 supported tools
Listed roughly from most to least common. Pick whichever ones you already have credentials for — you don't need all 12.
Listed roughly from most to least common. Pick whichever ones you already have credentials for — you don't need all 13.
### Claude Code (Anthropic)
@@ -48,7 +48,7 @@ The most complete integration. Session resumption works, MCP works, and it consu
### Codex (OpenAI)
JSON-RPC 2.0 transport with finer-grained approval gates. MCP config is written into the per-task `$CODEX_HOME/config.toml`. **Session resumption code exists but is currently unreachable** — pick Claude Code or one of the ACP family if you need resume.
JSON-RPC 2.0 transport with finer-grained approval gates. MCP config is written into the per-task `$CODEX_HOME/config.toml`. **Session resumption works** through Codex app-server `thread/resume`; stale or missing threads fall back to a fresh thread.
| | |
|---|---|
@@ -58,7 +58,7 @@ JSON-RPC 2.0 transport with finer-grained approval gates. MCP config is written
### Cursor (Anysphere)
The CLI counterpart to the Cursor editor. **Session resumption is broken** — Cursor's CLI doesn't return a sessionid, so the value you pass on resume is always invalid.
The CLI counterpart to the Cursor editor. **Session resumption works** with current Cursor Agent releases: Multica reads `session_id` from the stream-json events and passes it back with `--resume <id>`.
| | |
|---|---|
@@ -77,16 +77,6 @@ Model routing goes through your GitHub account entitlement — the tool doesn't
| Authentication | Browser-based GitHub login through the CLI. |
| Notes | Requires an active GitHub Copilot subscription on the signed-in account. |
### Gemini (Google)
Supports the Gemini 2.5 and 3 series. No session resumption, no MCP — suitable for one-shot tasks.
| | |
|---|---|
| Daemon looks for | `gemini` |
| Install | Follow the official guide at [github.com/google-gemini/gemini-cli](https://github.com/google-gemini/gemini-cli). The standard route is the npm package `@google/gemini-cli`. |
| Authentication | `gemini` will prompt for a Google account login, or set `GEMINI_API_KEY`. |
### OpenCode (SST)
Open-source CLI agent. Dynamically discovers available models from its own configuration file — good fit for users who want to bring their own model catalog. Consumes the agent's `mcp_config` field through `OPENCODE_CONFIG_CONTENT`.
@@ -147,6 +137,26 @@ Minimalist. **Session resumption is unusual** — the resume id is the path to a
| Install | See Inflection's CLI docs at [pi.ai](https://pi.ai/). |
| Authentication | Per the vendor's docs. |
### CodeBuddy (Tencent)
A Claude Code–compatible CLI agent. Multica drives it with the same stream-json protocol as Claude Code: session resumption works via `--resume`, MCP config is passed through `--mcp-config`, and skills land in `.claude/skills/`. Models are discovered dynamically.
| | |
|---|---|
| Daemon looks for | `codebuddy` |
| Install | See the official CLI docs at [codebuddy.ai/cli](https://www.codebuddy.ai/cli). |
| Authentication | Per the vendor's docs. |
### Qoder (Alibaba)
Agentic coding CLI using the ACP protocol over stdio (shares the transport with Hermes, Kimi, and Kiro CLI). Session resumption works through ACP `session/resume`, MCP config is passed through ACP `mcpServers`, model selection is discovered dynamically, and skills are copied into `.qoder/skills/`.
| | |
|---|---|
| Daemon looks for | `qodercli` |
| Install | See the official CLI docs at [qoder.com/cli](https://qoder.com/cli). |
| Authentication | Per the vendor's docs. |
### Antigravity (Google)
Google's Antigravity CLI (`agy`). Pairs with Google's Antigravity service and runs Gemini-backed models. Session resumption works through `--conversation <id>`, captured by the daemon from the CLI log file. Model selection is managed inside the Antigravity CLI itself — Multica disables the per-agent model picker for this provider. Skills are written to `.agents/skills/` (the CLI inherits Gemini CLI's workspace skill layout — see [Antigravity docs](https://antigravity.google/docs/gcli-migration)).
@@ -28,12 +28,15 @@ The default resource type — checked out per task into an isolated worktree:
"resource_type": "github_repo",
"resource_ref": {
"url": "https://github.com/owner/repo",
"ref": "release/v2",
"default_branch_hint": "main"
}
}
```
`default_branch_hint` is optional — if present, the daemon surfaces it in the meta-skill so the agent knows which branch to base its work on.
`ref` is optional — if present, `multica repo checkout <url>` uses it as the default branch, tag, or commit for tasks in this project. An explicit `multica repo checkout <url> --ref <other-ref>` still wins for that one checkout.
`default_branch_hint` is optional prompt context. It is not used for checkout; use `ref` when the project should pin a branch, tag, or SHA.
# Generic escape hatch for any resource_type the server understands —
@@ -251,7 +255,7 @@ The repo list shown to the agent (`## Repositories` block in `CLAUDE.md` / `AGEN
This keeps the agent's working set tight: when a project is explicit about its repos, that's the authoritative answer. The structured resource list at `.multica/project/resources.json` always carries the full set, so a skill that wants to inspect everything still can.
The daemon mirrors this on the checkout side: when a task arrives with project-scoped `github_repo` URLs, those URLs are merged into the per-workspace allowlist *and* synced into the local repo cache before the agent spawns. So a project repo URL that isn't bound at the workspace level is still a valid argument to `multica repo checkout` — the daemon won't reject it as "not configured." The allowlist split is internal: workspace-bound URLs and task-scoped URLs are tracked separately, so a workspace-repos refresh doesn't accidentally revoke a project URL mid-run.
The daemon mirrors this on the checkout side: when a task arrives with project-scoped `github_repo` URLs, those URLs are merged into the per-workspace allowlist *and* synced into the local repo cache before the agent spawns. So a project repo URL that isn't bound at the workspace level is still a valid argument to `multica repo checkout` — the daemon won't reject it as "not configured." If the project resource includes `ref`, that ref becomes the default for `multica repo checkout <url>` during that task; passing `--ref` to the checkout command overrides it. The allowlist split is internal: workspace-bound URLs and task-scoped URLs are tracked separately, so a workspace-repos refresh doesn't accidentally revoke a project URL mid-run.
Nous Research が提供します。ACP プロトコルを使用します（Kimi とトランスポート層を共有します）。セッション再開が動作します。しかし**スキル注入パスは専用のものではなく汎用のフォールバック**（`.agent_context/skills/`）です — Hermes CLI 自体がこのパスを読み取らない場合、スキルが適用されないことがあります。テストで確認してください。
Nous Research が提供します。ACP プロトコルを使用します（Kimi とトランスポート層を共有します）。セッション再開が動作し、MCP 構成は ACP `mcpServers` として渡されます。しかし**スキル注入パスは専用のものではなく汎用のフォールバック**（`.agent_context/skills/`）です — Hermes CLI 自体がこのパスを読み取らない場合、スキルが適用されないことがあります。テストで確認してください。
Inflection AI が提供し、ミニマルです。**セッション再開の方式が独特です** — セッション ID が文字列 ID ではなく、ディスク上のファイルパス（`~/.pi/...`）です。他のツールでは再開 id は CLI が返す文字列ですが、Pi では再開 id はセッションファイルそのものです。
セッション再開のメカニズムは[タスク](/tasks#can-a-task-continue-from-the-previous-context)で扱います。**サポートされているすべてのツールがセッションを再開できます** — 再開 id を渡すと、タスクは以前のコンテキストから続行します。唯一の例外は Pi で、再開 id が文字列 ID ではなくディスク上のセッションファイルへのパスです（上記の [Pi](#pi) を参照）。
| 状態 | ツール | 意味 |
|---|---|---|
| ✅ 実際に動作 | Antigravity, Claude Code, Copilot, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw, Pi | 再開 id を渡すと以前のコンテキストから続行します |
description: Multica는 12개의 AI 코딩 도구를 지원합니다. 모두 동일한 인터페이스를 구현하지만, 기능 세부사항은 크게 다릅니다.
description: Multica는 13개의 AI 코딩 도구를 지원합니다. 모두 동일한 인터페이스를 구현하지만, 기능 세부사항은 크게 다릅니다.
---
import { Callout } from "fumadocs-ui/components/callout";
Multica는 **12개의 AI 코딩 도구**를 기본 지원합니다. 이들은 모두 동일한 인터페이스(대기열 적재, 디스패치, 실행, 결과 반환)를 구현하므로, 같은 Multica 보드에서 어느 것이든 구동할 수 있습니다. **하지만 기능 세부사항은 크게 다릅니다**: 세션 재개가 실제로 동작하는지, MCP를 지원하는지, 스킬 파일이 어디에 위치하는지, 모델을 어떻게 선택하는지. 이 페이지가 전체 대조표입니다.
Multica는 **13개의 AI 코딩 도구**를 기본 지원합니다. 이들은 모두 동일한 인터페이스(대기열 적재, 디스패치, 실행, 결과 반환)를 구현하므로, 같은 Multica 보드에서 어느 것이든 구동할 수 있습니다. **하지만 기능 세부사항은 크게 다릅니다**: 세션 재개가 실제로 동작하는지, MCP를 지원하는지, 스킬 파일이 어디에 위치하는지, 모델을 어떻게 선택하는지. 이 페이지가 전체 대조표입니다.
에이전트를 생성할 때 도구를 고르는 방법은 [에이전트 생성 및 구성](/agents-create)을 참고하세요.
@@ -15,16 +15,17 @@ Multica는 **12개의 AI 코딩 도구**를 기본 지원합니다. 이들은
|---|---|---|---|---|---|
| **Antigravity** | Google | ✅ (`--conversation <id>`) | ❌ | `.agents/skills/` | 동적 탐색(`agy models`) |
@@ -36,9 +37,13 @@ Google에서 제공합니다. CLI 바이너리 이름은 `agy`입니다. Google
Anthropic에서 제공합니다. **신규 사용자에게 첫 번째 선택지**이며, 가장 완전한 기능 세트를 갖추고 있습니다: 세션 재개가 실제로 동작하고, MCP 구성을 읽으며, `--max-turns`와 `--append-system-prompt` 같은 세부 조정 flag를 지원합니다. Anthropic API 키가 필요합니다.
### CodeBuddy
Tencent에서 제공합니다. Claude Code 호환 CLI 에이전트입니다 — Multica는 Claude Code와 동일한 stream-json 프로토콜로 구동하므로 세션 재개가 동작하고(`--resume` 경유), MCP 구성은 `--mcp-config`로 전달되며, 스킬은 Claude Code의 `.claude/skills/` 레이아웃을 사용합니다. 모델은 동적으로 탐색됩니다.
### Codex
OpenAI에서 제공합니다. JSON-RPC 2.0을 사용하고, 상태 유지 능력이 더 강하며, 더 세밀한 승인 메커니즘(`exec_command` 및 `patch_apply`에 대한 수동 승인)을 갖추고 있습니다. MCP 구성은 작업별 `$CODEX_HOME/config.toml`에 기록됩니다. **세션 재개 코드는 존재하지만 현재 도달할 수 없습니다** — 재개가 필요하다면 Claude Code나 ACP 계열 중 하나를 선택하세요.
OpenAI에서 제공합니다. JSON-RPC 2.0을 사용하고, 상태 유지 능력이 더 강하며, 더 세밀한 승인 메커니즘(`exec_command` 및 `patch_apply`에 대한 수동 승인)을 갖추고 있습니다. MCP 구성은 작업별 `$CODEX_HOME/config.toml`에 기록됩니다. **세션 재개가 동작합니다** — Multica는 Codex app-server의 `thread/resume`으로 재개합니다. 저장된 thread가 없거나 오래된 경우에는 새 thread로 폴백해 작업을 계속 실행합니다.
### Copilot
@@ -46,16 +51,20 @@ GitHub에서 제공합니다. 모델 라우팅은 GitHub 계정 권한을 거칩
### Cursor
Anysphere에서 제공하며, Cursor 에디터에 대응하는 CLI입니다. **세션 재개 코드는 존재하지만 실제로는 동작하지 않습니다** — Cursor CLI 이벤트 스트림이 세션 ID를 반환하지 않으므로, 전달하는 재개 값은 항상 무효입니다. 재개가 필요하다면 다른 것을 선택하세요.
### Gemini
Google에서 제공하며, Gemini 2.5 및 3 시리즈를 지원합니다. **세션 재개도 MCP도 지원하지 않습니다** — 긴 컨텍스트 기억이 필요 없는 일회성 작업에 적합합니다.
Anysphere에서 제공하며, Cursor 에디터에 대응하는 CLI입니다. **세션 재개가 동작합니다** — 현재 Cursor Agent의 stream-json 이벤트에는 `session_id`가 포함되며, Multica는 다음 실행 때 이를 `--resume <id>`로 다시 전달합니다. MCP 구성은 작업 워크스페이스의 `.cursor/mcp.json`에 기록되고, Cursor의 프로젝트 approval 파일은 작업별 `CURSOR_DATA_DIR` 아래에 기록되므로, 관리되는 MCP 서버는 사용자의 전역 Cursor approval에 의존하지 않습니다.
### Hermes
Nous Research에서 제공합니다. ACP 프로토콜을 사용합니다(Kimi와 전송 계층을 공유합니다). 세션 재개가 동작하고, MCP 구성은 ACP `mcpServers`로 전달됩니다. 하지만 **스킬 주입 경로는 전용 경로가 아니라 범용 fallback**(`.agent_context/skills/`)입니다 — Hermes CLI 자체가 이 경로를 읽지 않으면 스킬이 적용되지 않을 수 있습니다. 테스트로 확인하세요.
**Hermes profile 선택.** 특정 profile로 Hermes를 실행하려면 에이전트의 `custom_args`에 profile 플래그와 profile 이름을 두 개의 독립된 항목으로 설정하세요. 예를 들어 `research`라는 profile을 사용하려면:
```json
["-p", "research"]
```
`"-p research"`처럼 하나의 문자열로 합치지 마세요. Multica는 배열의 각 항목을 하나의 argv 항목으로 도구에 전달합니다. `custom_args`는 에이전트별로 설정합니다 — [에이전트 생성 및 구성](/agents-create)을 참고하세요.
### Kimi
Moonshot에서 제공하며, 중국 시장을 겨냥합니다. Hermes와 ACP 프로토콜을 공유하고 MCP 구성도 ACP `mcpServers`로 전달되지만, 스킬 경로 `.kimi/skills/`는 Kimi CLI의 기본 탐색 메커니즘으로 Hermes의 fallback과는 다릅니다.
@@ -76,23 +85,19 @@ SST에서 제공하는 오픈소스입니다. 사용 가능한 모델과 모델
Inflection AI에서 제공하며, 미니멀합니다. **세션 재개 방식이 특이합니다** — 세션 ID가 문자열 ID가 아니라 디스크상의 파일 경로(`~/.pi/...`)입니다. 다른 도구에서는 재개 id가 CLI가 반환하는 문자열이지만, Pi에서는 재개 id가 세션 파일 그 자체입니다.
### Qoder
Alibaba에서 제공합니다. 에이전트형 코딩 CLI입니다. stdio 위에서 ACP 프로토콜을 사용합니다(Hermes, Kimi, Kiro CLI와 전송 계층을 공유합니다). 세션 재개는 ACP `session/resume`를 통해 동작하고, MCP 구성은 ACP `mcpServers`로 전달되며, 모델 선택은 동적으로 탐색되고, 스킬은 네이티브 탐색을 위해 `.qoder/skills/`로 복사됩니다.
## 세션 재개: 실제로 지원하는 도구
세션 재개 메커니즘은 [작업](/tasks#can-a-task-continue-from-the-previous-context)에서 다룹니다. 다음은 도구별 **정확한 현재 상태**입니다:
| 상태 | 도구 | 의미 |
|---|---|---|
| ✅ 실제로 동작 | Antigravity, Claude Code, Copilot, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw, Pi | 재개 id를 전달하면 이전 컨텍스트에서 이어집니다 |
| ⚠️ 코드는 존재하지만 도달 불가 | Codex, Cursor | 코드에 재개 경로가 있지만 실제로는 도달하지 않습니다(Codex는 조용히 폴백하고, Cursor는 세션 id를 반환하지 않습니다) — **미지원으로 간주하세요** |
| ❌ 없음 | Gemini | CLI에 재개 메커니즘이 없습니다 |
**의사결정을 위해**: 워크플로에서 에이전트가 작업 간에 컨텍스트를 유지해야 한다면(실패 재시도, 수동 재실행, 대화형 반복), ✅ 행에 있는 도구만 선택하세요.
세션 재개 메커니즘은 [작업](/tasks#can-a-task-continue-from-the-previous-context)에서 다룹니다. **지원되는 모든 도구가 세션을 재개합니다** — 재개 id를 전달하면 작업이 이전 컨텍스트에서 이어집니다. 유일한 예외는 Pi로, 재개 id가 문자열 ID가 아니라 디스크상의 세션 파일 경로입니다(위의 [Pi](#pi) 참고).
## MCP 구성: 도구별 지원
**12개 도구 중 `mcp_config`를 실제로 소비하는 것은 7개입니다: Claude Code, Codex, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw**. 나머지 5개는 이 필드를 받아들이지만 **무시합니다** — 오류도, 경고도 없으며, 구성이 그저 효과를 내지 못합니다.
**13개 도구 중 `mcp_config`를 실제로 소비하는 것은 10개입니다: Claude Code, CodeBuddy, Codex, Cursor, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw, Qoder**. 나머지 3개는 이 필드를 받아들이지만 **무시합니다** — 오류도, 경고도 없으며, 구성이 그저 효과를 내지 못합니다.
각 도구의 연결 방식은 다릅니다: Claude Code는 `--mcp-config`와 `--strict-mcp-config`로 받고, Codex는 데몬이 관리하는 `mcp_servers` 블록을 작업별 `$CODEX_HOME/config.toml`에 기록하며, Hermes/Kimi/Kiro CLI는 ACP `mcpServers`로 받습니다. OpenCode는 `OPENCODE_CONFIG_CONTENT` 환경 변수로 인라인 구성을 받고, OpenClaw는 Multica의 작업별 config wrapper를 통해 `mcp.servers`를 받습니다. OpenCode 경로는 프로젝트의 `opencode.json`을 다시 쓰지 않습니다.
각 도구의 연결 방식은 다릅니다: Claude Code와 CodeBuddy는 `--mcp-config`와 `--strict-mcp-config`로 받고, Codex는 데몬이 관리하는 `mcp_servers` 블록을 작업별 `$CODEX_HOME/config.toml`에 기록하며, Cursor는 `.cursor/mcp.json`과 작업별 `CURSOR_DATA_DIR` 아래의 프로젝트 approval을 기록합니다. Hermes/Kimi/Kiro CLI/Qoder는 ACP `mcpServers`로 받습니다. OpenCode는 `OPENCODE_CONFIG_CONTENT` 환경 변수로 인라인 구성을 받고, OpenClaw는 Multica의 작업별 config wrapper를 통해 `mcp.servers`를 받습니다. OpenCode 경로는 프로젝트의 `opencode.json`을 다시 쓰지 않습니다.
<Callout type="warning">
에이전트 구성에서 `mcp_config`를 설정했더라도 MCP 열에 ✅가 없는 도구를 선택하면, MCP 서버가 해당 에이전트에 **아무런 효과**도 미치지 않습니다. MCP 연동은 도구별로 구현됩니다.
fallback 경로를 쓰는 도구가 실제로 이 디렉터리를 읽는지는 해당 도구 자체의 문서에 따라 달라지며 — 보장되지 않습니다. Gemini / Hermes / OpenClaw에서 스킬이 적용되지 않는다면, 먼저 이 점을 확인하세요.
fallback 경로를 쓰는 도구가 실제로 이 디렉터리를 읽는지는 해당 도구 자체의 문서에 따라 달라지며 — 보장되지 않습니다. Hermes / OpenClaw에서 스킬이 적용되지 않는다면, 먼저 이 점을 확인하세요.
기본 프로젝트 수준 경로에서는 저장소 범위 탐색이 의도된 동작입니다. 체크아웃된 저장소가 이미 해당 디렉터리를 포함하고 있으면, 기반 도구가 커밋된 스킬을 자체적으로 탐색할 수 있습니다. 해당 저장소에서 사용하기 위해 이러한 repo skills를 Multica로 먼저 가져올 필요는 없습니다. Multica는 이러한 저장소 파일을 그대로 둡니다. 워크스페이스 스킬의 자연 디렉터리 이름이 같으면 데몬은 `review-helper-multica` 같은 충돌 없는 형제 디렉터리에 워크스페이스 사본을 씁니다.
스킬의 생성과 사용은 [스킬](/skills)을 참고하세요.
@@ -126,4 +134,4 @@ fallback 경로를 쓰는 도구가 실제로 이 디렉터리를 읽는지는
- [에이전트 생성 및 구성](/agents-create) — 에이전트에 사용할 도구를 선택하세요
- [작업](/tasks) — 작업 생명주기와 세션 재개 메커니즘
- [데몬과 런타임](/daemon-runtimes) — 도구가 실행되는 곳과 Multica에 연결되는 방식
- [에이전트 런타임 설치](/install-agent-runtime) — 지원되는 12개 도구 각각의 설치 및 인증
- [에이전트 런타임 설치](/install-agent-runtime) — 지원되는 13개 도구 각각의 설치 및 인증
description: Multica supports 12 AI coding tools; they implement the same interface, but the capability details diverge significantly.
description: Multica supports 13 AI coding tools; they implement the same interface, but the capability details diverge significantly.
---
import { Callout } from "fumadocs-ui/components/callout";
Multica ships with built-in support for **12 AI coding tools**. They all implement the same interface — queue, dispatch, execute, return results — so you can drive any of them from the same Multica board. **But the capability details diverge significantly**: whether session resumption actually works, whether MCP is supported, where skill files live, how models are selected. This page is the full matrix.
Multica ships with built-in support for **13 AI coding tools**. They all implement the same interface — queue, dispatch, execute, return results — so you can drive any of them from the same Multica board. **But the capability details diverge significantly**: whether session resumption actually works, whether MCP is supported, where skill files live, how models are selected. This page is the full matrix.
For guidance on picking a tool when creating an agent, see [Creating and configuring agents](/agents-create).
@@ -15,16 +15,17 @@ For guidance on picking a tool when creating an agent, see [Creating and configu
@@ -36,9 +37,13 @@ From Google. CLI binary name is `agy`. Pairs with Google's Antigravity service a
From Anthropic. **First choice for new users** — the most complete feature set: session resumption actually works, it reads MCP configuration, and it supports fine-tuning flags like `--max-turns` and `--append-system-prompt`. Requires an Anthropic API key.
### CodeBuddy
From Tencent. A Claude Code–compatible CLI agent — Multica drives it with the same stream-json protocol as Claude Code, so session resumption works (via `--resume`), MCP config is passed through `--mcp-config`, and skills use Claude Code's `.claude/skills/` layout. Models are discovered dynamically.
### Codex
From OpenAI. Uses JSON-RPC 2.0, has stronger statefulness, and a finer-grained approve mechanism (manual approval for `exec_command` and `patch_apply`). MCP config is materialized into the per-task `$CODEX_HOME/config.toml`. **Session resumption code exists but is currently unreachable** — if you need resume, pick Claude Code or one of the ACP family.
From OpenAI. Uses JSON-RPC 2.0, has stronger statefulness, and a finer-grained approve mechanism (manual approval for `exec_command` and `patch_apply`). MCP config is materialized into the per-task `$CODEX_HOME/config.toml`. **Session resumption works** through Codex app-server `thread/resume`; if the saved thread is missing or stale, Multica falls back to a fresh thread so the task can still run.
### Copilot
@@ -46,16 +51,20 @@ From GitHub. Model routing goes through your GitHub account entitlement — the
### Cursor
From Anysphere, the CLI counterpart to the Cursor editor. **Session resumption code exists but doesn't actually work** — the Cursor CLI event stream doesn't return a session ID, so any resume value you pass is always invalid. If you need resume, pick something else.
### Gemini
From Google, supports the Gemini 2.5 and 3 series. **No session resumption and no MCP** — suitable for one-shot tasks that don't need long context memory.
From Anysphere, the CLI counterpart to the Cursor editor. **Session resumption works** with current Cursor Agent releases: the stream-json event includes a `session_id`, and Multica passes it back with `--resume <id>` on the next run. MCP config is materialized into the task workspace's `.cursor/mcp.json`, with Cursor's project approval file written under a per-task `CURSOR_DATA_DIR` so managed MCP servers do not depend on the user's global Cursor approvals.
### Hermes
From Nous Research. Uses the ACP protocol (shares a transport with Kimi). Session resumption works, and MCP config is passed through ACP `mcpServers`. But the **skill injection path is the generic fallback** (`.agent_context/skills/`), not a dedicated one — if the Hermes CLI itself doesn't read this path, skills may not take effect. Verify by testing.
**Selecting a Hermes profile.** To launch Hermes under a specific profile, set the agent's `custom_args` to the profile flag and the profile name as two separate entries — for example, for a profile named `research`:
```json
["-p", "research"]
```
Don't combine them into one string like `"-p research"`; Multica passes each array item as a separate argv entry. `custom_args` is configured per agent — see [Creating and configuring agents](/agents-create).
### Kimi
From Moonshot, aimed at the Chinese market. Shares the ACP protocol with Hermes, including MCP config through ACP `mcpServers`, but the skill path `.kimi/skills/` is Kimi CLI's native discovery mechanism — different from Hermes's fallback.
@@ -76,23 +85,19 @@ Open-source project, a CLI agent orchestrator. MCP config is materialized throug
From Inflection AI, minimalist. **Session resumption is unusual** — the session ID is a file path on disk (`~/.pi/...`) rather than a string ID. In other tools, the resume id is a string returned by the CLI; in Pi, the resume id is the session file itself.
### Qoder
From Alibaba. An agentic coding CLI. Uses the ACP protocol over stdio (shares a transport with Hermes, Kimi, and Kiro CLI). Session resumption works through ACP `session/resume`, MCP config is passed through ACP `mcpServers`, model selection is discovered dynamically, and skills are copied into `.qoder/skills/` for native discovery.
## Session resumption: who really supports it
The session resumption mechanism is covered in [Tasks](/tasks#can-a-task-continue-from-the-previous-context). Here's the **exact current state** per tool:
| Status | Tools | Meaning |
|---|---|---|
| ✅ Really works | Antigravity, Claude Code, Copilot, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw, Pi | Pass the resume id and it continues from the previous context |
| ⚠️ Code exists but unreachable | Codex, Cursor | Resume paths exist in the code but aren't actually reached (Codex silently falls back; Cursor doesn't return session id) — **treat as unsupported** |
| ❌ None | Gemini | The CLI has no resume mechanism |
**For your decision**: if your workflow needs agents to preserve context across tasks (failure retries, manual reruns, conversational iteration), pick only from the ✅ row.
The session resumption mechanism is covered in [Tasks](/tasks#can-a-task-continue-from-the-previous-context). **Every supported tool resumes sessions** — pass the resume id and the task continues from the previous context. The one quirk is Pi, whose resume id is a session file path on disk rather than a string id (see [Pi](#pi) above).
## MCP configuration: provider-specific support
**Of the 12 tools, seven consume `mcp_config`: Claude Code, Codex, Hermes, Kimi, Kiro CLI, OpenCode, and OpenClaw**. The other five accept the field but **ignore it** — no error, no warning, the config just has no effect.
**Of the 13 tools, ten consume `mcp_config`: Claude Code, CodeBuddy, Codex, Cursor, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw, and Qoder**. The other three accept the field but **ignore it** — no error, no warning, the config just has no effect.
The runtime paths are provider-specific: Claude Code receives it through `--mcp-config` paired with `--strict-mcp-config`; Codex writes a daemon-managed `mcp_servers` block into the per-task `$CODEX_HOME/config.toml`; Hermes, Kimi, and Kiro CLI receive ACP `mcpServers`; OpenCode receives inline config through `OPENCODE_CONFIG_CONTENT`; OpenClaw receives `mcp.servers` through Multica's per-task config wrapper. OpenCode's path does **not** rewrite the project's `opencode.json`.
The runtime paths are provider-specific: Claude Code and CodeBuddy receive it through `--mcp-config` paired with `--strict-mcp-config`; Codex writes a daemon-managed `mcp_servers` block into the per-task `$CODEX_HOME/config.toml`; Cursor writes `.cursor/mcp.json` plus per-task project approvals under `CURSOR_DATA_DIR`; Hermes, Kimi, Kiro CLI, and Qoder receive ACP `mcpServers`; OpenCode receives inline config through `OPENCODE_CONFIG_CONTENT`; OpenClaw receives `mcp.servers` through Multica's per-task config wrapper. OpenCode's path does **not** rewrite the project's `opencode.json`.
<Callout type="warning">
If you set `mcp_config` in an agent configuration but pick a tool not marked ✅ in the MCP column, your MCP servers have **no effect** on that agent. MCP integration is provider-specific.
@@ -105,6 +110,7 @@ Each tool uses **its own** skill discovery path. Before a task runs, the Multica
| Tool | Path | Native discovery? |
|---|---|---|
| Claude Code | `.claude/skills/` | ✅ Native |
| CodeBuddy | `.claude/skills/` | ✅ Native |
| Codex | `$CODEX_HOME/skills/` | ✅ Native |
| Copilot | `.github/skills/` | ✅ Native |
| Cursor | `.cursor/skills/` | ✅ Native |
@@ -112,12 +118,14 @@ Each tool uses **its own** skill discovery path. Before a task runs, the Multica
Whether a fallback-path tool actually reads this directory depends on the tool's own documentation — no guarantees. If your skills aren't taking effect for Gemini / Hermes / OpenClaw, check this first.
Whether a fallback-path tool actually reads this directory depends on the tool's own documentation — no guarantees. If your skills aren't taking effect for Hermes / OpenClaw, check this first.
For native project-level paths, repo-scoped discovery is expected: if the checked-out repository already contains a matching directory, the underlying tool can discover those committed skills on its own. You do not need to import those repo skills into Multica just to use them in that repo. Multica keeps the repo files intact. If a workspace skill has the same natural directory name, the daemon writes the workspace copy to a collision-free sibling such as `review-helper-multica`.
For creating and using skills, see [Skills](/skills).
@@ -126,4 +134,4 @@ For creating and using skills, see [Skills](/skills).
- [Creating and configuring agents](/agents-create) — pick a tool for your agent
- [Tasks](/tasks) — task lifecycle and session-resumption mechanics
- [Daemon and runtimes](/daemon-runtimes) — where the tools run and how they connect to Multica
- [Install an agent runtime](/install-agent-runtime) — installation and authentication for each of the 12 supported tools
- [Install an agent runtime](/install-agent-runtime) — installation and authentication for each of the 13 supported tools
플래그 대신 환경 변수를 선호한다면, 해당 플래그를 생략할 때 `setup self-host`가 `MULTICA_SERVER_URL`과 `MULTICA_APP_URL`을 읽습니다(둘 다 설정하면 플래그가 우선합니다). `MULTICA_SERVER_URL`은 [환경 변수](/environment-variables)에 나오는 `ws://…/ws` 데몬 형식도 허용하며 HTTP 기본 URL로 정규화합니다.
</Callout>
단일 호스트네임에서 프런트엔드와 백엔드를 모두 앞단에 두는(데몬과 웹 앱 모두에 필요한 WebSocket 지원 포함) 최소 Caddyfile은 다음과 같습니다.
Prefer environment variables over flags? `setup self-host` reads `MULTICA_SERVER_URL` and `MULTICA_APP_URL` when the matching flag is omitted — a flag still takes precedence over the env var. `MULTICA_SERVER_URL` also accepts the `ws://…/ws` daemon form from [Environment variables](/environment-variables) and normalizes it to the HTTP base.
</Callout>
A minimal Caddyfile that fronts both the frontend and the backend (with WebSocket support, which the daemon and the web app both need) on a single hostname:
- **ローカルスキル** — あなたのマシン上のディレクトリに存在します（各 AI コーディングツールには慣例的なデフォルトパスがあります。例: Claude Code の `~/.claude/skills/`）。あなたが要求すると、[デーモン](/daemon-runtimes)がマシンをスキャンし、どれをワークスペースに取り込むかを手動で選びます。
- **ローカルスキル** — あなたのマシン上のディレクトリに存在します。あなたが要求すると、[デーモン](/daemon-runtimes)がマシンをスキャンし、どれをワークスペースに取り込むかを手動で選びます。デーモンは**2 つのルートを優先順位順に**確認します。まずランタイム自身のスキルディレクトリ（各 AI コーディングツールには慣例的なデフォルトパスがあります。例: Claude Code の `~/.claude/skills/`）、次にツール横断の汎用ディレクトリ `~/.agents/skills/`（Codex や Gemini CLI などのエコシステムが共有する場所）です。同じスキル名が両方に存在する場合は**プロバイダー専用ディレクトリが優先**されるため、汎用ルートは*追加の*スキルを表示するだけで、既存スキルの解決結果を変えることはありません。
@@ -12,10 +12,16 @@ import { Callout } from "fumadocs-ui/components/callout";
Multica는 두 가지 스킬 소스를 지원합니다.
- **워크스페이스 스킬** — Multica 클라우드에 저장됩니다. 에이전트에 연결되면 작업 실행 시점에 여러분의 데몬으로 동기화됩니다. 이것이 **팀 전체에서 스킬을 공유하는 표준 방식**입니다.
- **로컬 스킬** — 여러분의 기기에 있는 디렉터리에 존재합니다(각 AI 코딩 도구마다 관례적인 기본 경로가 있습니다. 예: Claude Code의 `~/.claude/skills/`). 여러분이 요청하면 [데몬](/daemon-runtimes)이 기기를 스캔하고, 어떤 스킬을 워크스페이스로 가져올지 직접 고릅니다.
- **로컬 스킬** — 여러분의 기기에 있는 디렉터리에 존재합니다. 여러분이 요청하면 [데몬](/daemon-runtimes)이 기기를 스캔하고, 어떤 스킬을 워크스페이스로 가져올지 직접 고릅니다. 데몬은 **두 개의 루트를 우선순위 순서로** 확인합니다. 먼저 런타임 자체의 스킬 디렉터리(각 AI 코딩 도구마다 관례적인 기본 경로가 있습니다. 예: Claude Code의 `~/.claude/skills/`), 그다음 도구 간 공용 디렉터리 `~/.agents/skills/`(Codex, Gemini CLI 등 생태계가 공유하는 위치)입니다. 동일한 스킬 이름이 양쪽에 모두 있으면 **프로바이더 전용 디렉터리가 우선**하므로, 공용 루트는 *추가* 스킬만 노출할 뿐 기존 스킬의 해석 결과를 절대 바꾸지 않습니다.
대부분의 경우 **워크스페이스 스킬**을 원하게 됩니다. 한 번만 가져오면 모든 팀원의 에이전트가 사용할 수 있기 때문입니다. 로컬 스킬은 먼저 로컬에서 테스트하고 싶거나, 콘텐츠에 민감한 로컬 자료가 포함된 경우에 적합합니다.
## 저장소 범위 스킬
저장소 범위 스킬은 의도된 동작입니다. 일부 AI 코딩 도구는 `.claude/skills/`, `.cursor/skills/`, `.opencode/skills/`, `.agents/skills/`처럼 **저장소에 커밋된 프로젝트 수준 스킬**을 탐색합니다. Multica 작업이 해당 저장소를 체크아웃하면 이 파일들은 작업 디렉터리에 남아 있고, 기반 도구가 자체 기본 탐색 규칙에 따라 읽어 들일 수 있습니다. 해당 저장소에서 사용하기 위해 이러한 repo skills를 Multica로 먼저 가져올 필요는 없습니다. Multica도 이를 워크스페이스 스킬 레지스트리로 자동 가져오지 않습니다.
프로젝트 수준 탐색을 지원하는 도구에서는 워크스페이스 스킬도 같은 프로바이더 기본 위치로 동기화됩니다. 워크스페이스 스킬이 기존 저장소 스킬 디렉터리와 충돌하면, Multica는 저장소 파일을 덮어쓰지 않고 `review-helper-multica` 같은 형제 이름으로 워크스페이스 사본을 씁니다. 그러면 도구에는 조정된 이름을 가진 워크스페이스 사본을 포함해 두 스킬이 모두 보일 수 있습니다.
## 스킬 가져오기
워크스페이스 스킬은 네 가지 소스에서 가져옵니다.
@@ -54,7 +60,7 @@ GitHub나 ClawHub에서 가져온 스킬에는 스크립트와 실행 가능한
- **스킬** = 구조화된 **지식 팩**(정적 콘텐츠 + 지침). 에이전트는 스킬을 읽어 "문제 X를 만나면 이렇게 생각하고 이렇게 처리하라"를 학습합니다.
- **MCP**(Model Context Protocol) = **도구 채널**. 에이전트는 MCP를 사용해 외부 서비스(데이터베이스, 파일 시스템, 서드파티 API)에 연결하고 이를 **호출**합니다.
이 둘은 상호 보완적입니다. 현재 Multica에서 MCP 지원은 **도구별로 구현됩니다**: Claude Code, Codex, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw는 `mcp_config`를 사용하고, 다른 도구들은 이 필드를 받더라도 실제로 사용하지 않습니다. MCP 전용 섹션은 추후 릴리스에서 추가될 예정입니다.
이 둘은 상호 보완적입니다. 현재 Multica에서 MCP 지원은 **도구별로 구현됩니다**: Claude Code, Codex, Cursor, Hermes, Kimi, Kiro CLI, OpenCode, OpenClaw는 `mcp_config`를 사용하고, 다른 도구들은 이 필드를 받더라도 실제로 사용하지 않습니다. MCP 전용 섹션은 추후 릴리스에서 추가될 예정입니다.
@@ -12,10 +12,16 @@ A Skill is a **knowledge pack** for an [agent](/agents) — a `SKILL.md` plus op
Multica supports two skill sources:
- **Workspace skill** — stored in Multica's cloud. Once attached to an agent, it's synced down to your daemon at task execution time. This is the **standard way to share skills across a team**.
- **Local skill** — lives in a directory on your machine (each AI coding tool has a conventional default path, e.g. Claude Code's `~/.claude/skills/`). On your request, the [daemon](/daemon-runtimes) scans your machine, and you manually pick which ones to bring into the workspace.
- **Local skill** — lives in a directory on your machine. On your request, the [daemon](/daemon-runtimes) scans your machine, and you manually pick which ones to bring into the workspace. The daemon checks **two roots, in priority order**: first the runtime's own skill directory (each AI coding tool has a conventional default path, e.g. Claude Code's `~/.claude/skills/`), then the cross-tool universal directory `~/.agents/skills/` — a shared location used by ecosystems like Codex and Gemini CLI. When the same skill name exists in both, the **provider-specific directory wins**, so the universal root only ever surfaces *additional* skills and never changes what an existing skill resolves to.
Most of the time you want **workspace skills**: import once, every teammate's agent can use it. Local skills are a fit when you want to test locally first, or when the content involves sensitive local material.
## Repository-scoped skills
Repo-scoped skills are expected behavior. Some AI coding tools discover **project-level skills committed inside a repository**, such as `.claude/skills/`, `.cursor/skills/`, `.opencode/skills/`, or `.agents/skills/`. When a Multica task checks out that repository, those files remain in the workdir and the underlying tool can load them through its native discovery rules. You do **not** need to import those repo skills into Multica just to use them in that repo; Multica does **not** import them into the workspace skill registry automatically.
Workspace skills still sync into the same provider-native location for tools that support project-level discovery. If a workspace skill would collide with an existing repo skill directory, Multica writes the workspace copy to a sibling name such as `review-helper-multica` instead of overwriting the repo's files. The tool may then see both skills, with the workspace copy carrying the adjusted name.
## Importing a skill
Workspace skills come from four sources:
@@ -54,7 +60,7 @@ Both augment what an agent can do, but in different directions:
- **Skill** = a structured **knowledge pack** (static content + instructions). The agent reads a skill to learn "when I see problem X, here's how to think and what to do."
- **MCP** (Model Context Protocol) = a **tool channel**. The agent uses MCP to connect to external services (databases, filesystems, third-party APIs) and **invoke** them.
The two are complementary. In Multica today, MCP support is **provider-specific**: Claude Code, Codex, Hermes, Kimi, Kiro CLI, OpenCode, and OpenClaw consume `mcp_config`; other tools receive the field but don't actually use it. A dedicated MCP section will come in a later release.
The two are complementary. In Multica today, MCP support is **provider-specific**: Claude Code, Codex, Cursor, Hermes, Kimi, Kiro CLI, OpenCode, and OpenClaw consume `mcp_config`; other tools receive the field but don't actually use it. A dedicated MCP section will come in a later release.
@@ -292,6 +293,275 @@ export function createEnDict(allowSignup: boolean): LandingDict {
fixes:"Bug Fixes",
},
entries:[
{
version:"0.3.30",
date:"2026-06-25",
title:"Slack Channel Integration, a Smoother Editor, and Many Reliability Fixes",
changes:[],
features:[
"Slack conversations now run on the new unified collaboration channel, putting Slack on the same reliable footing as Feishu and Lark",
"The Issue composer now accepts the highlighted @mention or suggestion when you press Tab, so picking the right teammate or Issue is a single keypress",
"Task list items can be toggled from a one-click button in the editor's floating menu",
],
improvements:[
"Frontend continuous integration now skips automatically when a pull request does not touch frontend code, freeing up build time for the changes that actually need it",
"Command line subcommands have broader automated test coverage so everyday workflows stay stable across releases",
"Provider-specific default agent arguments now have explicit documentation, and a one-time Lark cutover flag was retired now that the unified channel adapter is fully in production",
],
fixes:[
"OpenClaw is more forgiving about config file mismatches and supports the newer 2026.6.x agents schema, keeping existing OpenClaw runtimes connected",
"Moving an Issue between projects now removes it from the old project list right away, and board column counts stay accurate when an Issue's status changes off-screen",
"Attachment previews open correctly even when files are served from a different origin",
"Command line agents wait for the daemon to be ready before falling back to a personal access token, and the self-host setup flow now respects existing configuration and surfaces server URL changes",
"Lark messages now link to the configured app URL instead of falling back to a generic web address",
"Codex runs clean up correctly even when their output overflows, Kiro runs preserve their goal completion state through close errors, and agent shutdown now terminates the entire opencode process group before closing",
"Quick-create reliably keeps every uploaded file attached when several uploads happen at the same time",
"Redis webhook rate limiting no longer throttles unrelated webhooks together, and daemon skill bundles load reliably even for large skill libraries",
"Issue label names now reject control characters so labels stay readable everywhere",
],
},
{
version:"0.3.29",
date:"2026-06-24",
title:"Feishu Channel Upgrade, Feature Rollout Controls, and More Reliable Autopilots",
changes:[],
features:[
"Feishu conversations now run on a new unified collaboration channel, making message handling more stable and consistent and laying the groundwork for more chat platforms",
"New feature rollout controls cover both the app and the daemon, so teams can open up risky changes gradually and to a limited audience",
"When agents read long Issue discussions, resolved threads now fold down to their key conclusion to keep the context focused",
"Feishu users can start a fresh conversation with the `/new` command, and Feishu WebSocket connections can use a configured proxy",
],
improvements:[
"Scheduled autopilots are more dependable: even with missed schedules, retries, or several runners working at once, they settle on the intended single run",
"Agent runtime briefings can switch to a slimmer version that drops redundant detail, with the full version still available as a fallback",
"Runtime provider docs now match the current provider list, with Qoder, CodeBuddy, and Antigravity guidance added and the outdated Gemini CLI runtime removed",
"The branch or version pinned in a project's repository settings now takes effect during local agent work, so agents no longer end up on the wrong branch",
],
fixes:[
"Sub-Issues now stay in stable creation order inside a parent Issue",
"Attachment previews now open correctly inside Issues",
"The @mention picker now selects the highlighted person or Issue even when search results reorder",
"Cancelled chat drafts stay deleted after you navigate away and come back",
"Autopilot cold starts, the agent status in the Issue header, and Antigravity provider errors now report more accurately",
],
},
{
version:"0.3.28",
date:"2026-06-23",
title:"Staged Sub-Issues and Qoder Runtime Support",
changes:[],
features:[
"Sub-Issues can now be organized into stages, so parallel work moves forward together and the parent Issue is updated only when a stage is complete",
"Assigning or batch-updating an Issue now confirms upfront whether it will start an agent — and which one — so you can apply the change without launching a run; when a run does start, you can attach a handoff note that the agent receives as context for that run",
"Qoder is now available as an agent provider, including model discovery and provider branding",
"Custom runtimes can include fixed launch arguments, with clearer feedback when a saved runtime cannot register",
],
improvements:[
"Project descriptions now travel with agent work, giving agents more durable context from the project they are working in",
"Command line workflows now cover comment resolve actions, Issue usage summaries, and autopilot subscriber management",
"Readonly code blocks include a copy button, and the marketing header now shows the live GitHub star count",
"Agent skill delivery is more efficient for newer daemons while keeping older daemons compatible",
],
fixes:[
"Issue batch edit menus now show the real shared status, priority, and assignee for the selected Issues",
"Dragging Issues across board and list views no longer snaps cards back before settling",
"GitHub PR links and check updates are routed to the workspace that owns the repository",
"Live task transcripts now keep updating while a run is still in progress",
"Custom runtime deletion now removes the saved profile instead of only removing a row that could return later",
],
},
{
version:"0.3.27",
date:"2026-06-22",
title:"Threaded Lark Replies and Smoother Team Workflows",
changes:[],
features:[
"Lark conversations now reply inside the original topic when a message starts from a topic, keeping team discussions easier to follow",
"Squad leaders can see member skills in the roster, making delegation more precise",
"Discord is now available from the website footer, help menu, README, and a dismissible in-app sidebar card",
],
improvements:[
"Agent activity in Issue headers opens on hover, so live work is easier to check at a glance",
"Desktop sidebars and pinned navigation feel smoother, clearer, and less noisy",
"Chat replies, assignment catch-up, and contributor guidance are tighter so agent work stays in the right place with less noise",
"Remote CLI setup and custom runtime deletion now give clearer guidance before users continue",
],
fixes:[
"Backlog parent Issues stay parked when child work finishes, avoiding unexpected follow-up automation",
"Project deletion now requires an owner or admin, and private GitHub skill imports work when a valid token is available",
"Login verification focuses the code field automatically, and detail sidebars no longer animate unexpectedly when pages open",
"Codex and daemon diagnostics are more reliable when permissions or slow task claims need investigation",
],
},
{
version:"0.3.25",
date:"2026-06-18",
title:"More Reliable Agent Work Across Skills, Autopilots, and Chat",
changes:[],
features:[
"Local skill libraries on a developer machine can now be picked up automatically for agent runs",
"Autopilots can include default subscribers so the right teammates are included when new Issues are created",
"Chat attachments now stay tied to the current workspace and messages can continue sending without blocking the conversation",
"Failed agent comments can be retried directly from the Issue timeline",
],
improvements:[
"Usage reporting is more accurate when the same model name is available from different providers",
"Older Codex usage records can be filled in for more complete usage history",
"Runtime storage reporting is more complete across multiple workspace locations",
"Background task guidance and release checks are stricter, helping catch risky changes earlier",
],
fixes:[
"Issue mention chips in chat and comments now fit their container and no longer overlap nearby text",
"Workspace links now use the correct deployment host more reliably",
"Autopilot run folders are cleaned up after terminal runs finish",
"Desktop builds now handle commit-based version names correctly",
"Tencent CodeBuddy shows the correct provider logo",
"Daemon claim responses are smaller and faster to transfer",
],
},
{
version:"0.3.24",
date:"2026-06-17",
title:"Custom Runtimes",
changes:[],
features:[
"Teams can create custom runtimes so agents use the right local tools and models",
"CLI agent create and update now supports thinking level",
],
improvements:[
"Runtime profiles sync faster and prefer the best match for the current environment",
"Client error and freeze reports now group duplicates",
"Issue trigger previews are easier to read",
],
fixes:[
"Office 365 email delivery is more reliable",
"GitHub installation context and pending CI display are more reliable",
"Codex runs fail quickly when the app server exits",
"Self-healing runtimes can be deleted again, and incompatible models are cleared on runtime switch",
"Unknown Issue icons and plain filenames are handled safely",
],
},
{
version:"0.3.23",
date:"2026-06-16",
title:"Issue Date Filters and More Stable Agent Runs",
changes:[],
features:[
"Issues can now be filtered by created or updated date, with quick ranges and custom date selections",
"Command line users can now delete runtimes with safer defaults and an explicit option for related data",
"Lark connections can now use network proxies, helping teams in restricted network environments connect reliably",
],
improvements:[
"Web and desktop failures are now easier to investigate with clearer reports for errors, freezes, and crashes",
"Project rows, comment previews, and comment composers are more consistent and easier to use",
],
fixes:[
"Reply and edit previews now show the right agents or squads before a comment is saved",
"Plain Issue IDs in comments now stay as text unless they are intentionally linked",
"Google sign-in from command line login now returns to the command line correctly after browser authentication",
"Chat file uploads wait until an active agent is ready, avoiding failed uploads during loading",
"Transcript actions remain visible on touch devices where hover is unavailable",
"Agent instructions for posting comments now avoid shell formatting problems that could drop assignees, projects, or other fields",
],
},
{
version:"0.3.22",
date:"2026-06-15",
title:"Faster Lists, Easier Runtime Setup, and Safer Issue Editing",
changes:[],
features:[
"Agents, autopilots, projects, runtimes, skills, and squads now use a faster, more consistent list experience with clearer rows, filters, selections, and actions",
"The command line can now manage workspace repositories, so local agents can pick up project repo context more easily",
"Cursor and OpenClaw are easier to set up: Cursor connection settings can be managed for you, and OpenClaw can connect through an existing gateway",
"When editing a comment, you can preview and control which agents or squads will run before saving",
],
improvements:[
"Desktop recovery prompts now include more page context, making stuck-window reports easier to understand",
"Long Issues and inbox views now keep their scroll position and comment anchors more reliably when you navigate away and return",
"Cursor usage and billing details are clearer for Composer, cached inputs, and newer Cursor agent output",
],
fixes:[
"Issue attachments, inline images, and file cards are more reliable across web, desktop, mobile, and shared token links",
"The editor and read-only Issue content now handle dollar amounts and email links more predictably",
"Desktop Cmd+W now closes the active tab first, then the window when no tab can be closed",
"Self-hosted Docker Compose uploads and default settings fail less often, with missing values caught earlier",
"Agent tasks now stop safely when their run credentials are invalid",
],
},
{
version:"0.3.21",
date:"2026-06-12",
title:"CodeBuddy Runtime",
changes:[],
features:[
"CodeBuddy can now run local Multica agents, with its available model and effort choices shown automatically",
"Quick-created Issues now keep uploaded files attached from the first draft through the final Issue",
],
improvements:[
"Skill import conflicts are clearer: locked skills show a person's name instead of an internal ID, and a single overwrite now completes in one click",
"Desktop recovery prompts now explain what happened first and give clearer details to include when reporting a stuck window",
"Views that sort or filter people by signup time can now load faster",
],
fixes:[
"Chat now keeps messages and drafts in sync when sending, stopping, or recovering from a failed send",
"Lark account binding now works reliably for users who are already signed in, and sign-in returns to the binding page",
"Local agent runs no longer announce that work has started before the task folder is ready",
],
},
{
version:"0.3.20",
date:"2026-06-11",
title:"Skill Imports, Cleaner Run History, and Resilient Agents",
changes:[],
features:[
"Skill imports now let you choose what happens when a skill already exists: stop, replace it, save a renamed copy, or skip it",
"Import results now clearly show which skills were added, updated, skipped, blocked by a conflict, or could not be imported",
],
improvements:[
"Execution logs now show the newest past runs first on web and mobile, so recent progress is easier to scan",
"Changelog content was cleaned up so the latest release notes stay grouped under the right release",
],
fixes:[
"Issue thread replies now stay in the order they arrived, even when a slower agent reply lands later",
"Agents can recover when a saved session disappears, starting fresh instead of failing again on every mention",
"Reviving an Issue from a new workspace folder now starts a fresh session instead of retrying one that only existed in the old folder",
],
},
{
version:"0.3.19",
date:"2026-06-10",
title:"Safer Comment Triggers, Reliable Agents, and Attachments",
changes:[],
features:[
"Comment boxes now show which agents or squads will start work before you send, with controls to avoid accidental runs",
"Run transcripts now include timestamps, making agent progress and handoffs easier to review",
"Autopilot detail pages now show who created each autopilot",
"Claude Fable 5 is now available in Multica's supported model and pricing list",
"Issue conversations can now resolve a specific reply, making long threads easier to close while keeping the final answer visible",
"Lark and Feishu conversations now show a typing reaction while Multica is preparing a reply, then clear it before the answer is sent",
"Agent runs now know who started each task, making handoffs, audit trails, and privacy-aware behavior more accurate",
"OpenClaw users can point Multica at a custom app location and data folder from their local configuration",
],
improvements:[
"Comment trigger indicators are quieter, clearer, and less likely to crowd long agent names",
"Desktop now disables daemon start and stop controls when the daemon is managed outside Multica, such as in WSL2",
"The active agent indicator in an Issue header is easier to read, with motion only while work is running and clearer queued wording otherwise",
"The CLI now gives clearer guidance around common errors, sign-in problems, and project setup values",
],
fixes:[
"Inline images and files in Issue descriptions now stay visible across web and desktop after reloads",
"Each Issue discussion thread now keeps only one resolved answer at a time, so replacing the conclusion is consistent for everyone",
"Issue pages refresh their data after realtime reconnects, avoiding stale timelines after a connection drop",
"Agent task initiator history now works more reliably for older task records",
"Sticky Issue comments keep a cleaner visual edge while scrolling",
"Newly posted attachments now use stable private download links, so images and files stay visible after temporary upload links expire",
"Autopilot runs started from newly created Issues now fail cleanly when the assigned task cannot complete, instead of staying stuck",
"Inbox deep links now scroll inside the Issue timeline without pushing the desktop window out of place",
"Cursor and Codex sessions now end more cleanly after terminal results, preserving completion state and final telemetry",
"Self-host setup now respects configured server URLs, and project creation returns clear validation errors instead of a generic failure",
"A previous upload hardening change was rolled back after it conflicted with attachment behavior",
Multica ships a framework-level feature flag implementation:
- **Backend**: `server/pkg/featureflag` — Go package.
- **Frontend**: `@multica/core/feature-flags` — TypeScript module with React hooks.
Both sides share the same vocabulary (`Decision`, `EvalContext`, `Rule`, `PercentRollout`) and the same FNV-1a percent bucketing, so a flag evaluated on the server and on the client lands in the same bucket for the same user.
The package is designed so new features can adopt feature flags without writing any infrastructure code — drop a rule into the static config, call `Service.IsEnabled` / `useFlag`, done.
- A **Toggle Point** is the single `if` in business code. It always calls the Service, never the provider directly.
- The **Service** (`Service` in Go, `FeatureFlagService` in TS) is the router. Business code never depends on which provider is behind it.
- A **Provider** is the configuration backend. Today we ship `StaticProvider` (in-memory rules), `EnvProvider` (Go only — env-var override), and `ChainProvider` (composition). A future DB or LaunchDarkly provider plugs in without changing any caller.
- A **Decision** is the structured result: `{ enabled, variant, reason, source }`. `IsEnabled` is the boolean projection, `Variant` is the raw string. Use `Decision` for diagnostic endpoints.
| **Experiment** | Hours–weeks | Product / Data | A/B test `checkout_algo` between `control` and `experiment-v2` |
| **Ops** | Short or evergreen | SRE | Kill switch `billing_disable_invoice_pdf` |
| **Permission** | Years | Product | `plan_gate_enterprise_dashboard` |
Manage them in the same provider but treat them differently: Release flags get deleted; Ops flags need fast override paths (`FF_<KEY>` env var); Permission flags use `Allow` lists; Experiment flags use `PercentRollout`.
---
## Backend (Go)
### Wiring at startup
The server constructs a `featureflag.Service` once in `cmd/server/main.go` via the standard helper:
slog.Error("feature flag configuration failed to load","error",err)
os.Exit(1)
}
```
`NewServiceFromEnv` reads two env vars — both follow the same `MULTICA_*_FILE` / `FF_*` conventions documented in `.env.example`:
| Env var | Role |
|---|---|
| `MULTICA_FEATURE_FLAGS_FILE` | Path to the YAML rule set (optional; absent = no static rules). |
| `FF_<FLAG_KEY>` | Per-flag runtime override. `FF_BILLING_NEW_INVOICE_EMAIL=false` / `25%` / `experiment-v2`. Beats the YAML, no redeploy. |
The provider chain is `EnvProvider → YAML StaticProvider`. The server can boot with zero flag config — every `IsEnabled` call falls back to the caller's default until someone authors a rule.
### Daemon-bound flags
Daemon-bound flags are evaluated by the server and delivered to local daemons
over the daemon heartbeat ack. This is for process-level daemon behavior where
operators need one rollout and kill-switch path across cloud runtimes, Desktop
embedded daemons, and user-run CLI daemons.
Only flags listed in `server/internal/featureflagdispatch/registry.go` are sent
to daemons. The registry is intentionally short:
```go
varDaemonBoundFlags=[]string{
"runtime_brief_slim",
}
```
On each HTTP or WebSocket heartbeat, the server evaluates every registered key
as a daemon/process-level decision. The snapshot EvalContext exposes
`daemon_id` only; workspace/runtime/task/user scoped rollout is intentionally
not part of this channel because the daemon stores one process-global snapshot.
The heartbeat ack carries a full snapshot:
```json
{
"feature_flags":{
"version":1,
"flags":{
"runtime_brief_slim":"on"
}
}
}
```
The daemon installs that snapshot into its process-level feature flag service.
The daemon provider order is:
1.`EnvProvider` (`FF_*`) for local emergency overrides.
2.`ServerSnapshotProvider` from the latest heartbeat ack.
3. local YAML `StaticProvider` as a fallback for old servers or self-hosted rescue.
4. the toggle point's caller-supplied default.
That means `FF_RUNTIME_BRIEF_SLIM=false` always suppresses a server snapshot
that enables `runtime_brief_slim`. New daemons talking to old servers receive no
`feature_flags` field and automatically fall back to local env/YAML behavior.
Old daemons talking to new servers ignore the unknown JSON field.
To add another daemon-bound process-level flag, add its key to the registry and
use the existing daemon feature flag service at the toggle point. Do not add
workspace percent rollout, task payload fields, or task-scoped readers for
daemon-bound flags unless a separate design explicitly introduces scoped daemon
flag evaluation.
### YAML schema
```yaml
# /etc/multica/feature-flags.yaml
billing_new_invoice_email:
default:true
checkout_algo:
default:false
variant:experiment-v2
percent:
percent:25
by:user_id
ops_disable_recommendations:
default:false
allow:["user-internal-1","user-internal-2"]
allow_by:user_id
```
Every field except `default` is optional. `variant` is the on-variant — see the multi-arm note below. An empty file is a valid "no flags yet" state. Malformed YAML fails startup the same way `DATABASE_URL` parse errors do, so misconfig surfaces loudly.
`Rule.Variant` is the **on-variant**: it is only returned when the rule evaluates to enabled=true (allow hit, percent hit, default-on). When the rule evaluates to disabled (deny hit, percent miss, default-off) the Service returns `"off"` so callers branching on `Variant()` cannot route control users into the experiment arm. This is exercised by `TestStaticProviderVariantOnlyWhenEnabled` and is the same on the TS side.
The Service is nil-safe and missing-key-safe: `(*Service)(nil).IsEnabled(ctx, "any", true)` returns `true`. Business code never needs to guard against a missing flag.
---
## Frontend (TypeScript / React)
### Mounting once at the root
```tsx
// apps/web/app/_providers.tsx (or the equivalent root)
When the backend pushes a fresh rule set (via an API response or WebSocket), call `service.setProvider(new StaticProvider(remoteRules))` and the whole tree re-evaluates.
Outside a `FeatureFlagsProvider` (Storybook, unit tests, error pages) `useFlag` / `useVariant` return the supplied default. You never have to mount the provider just to render a component in isolation.
### Security note: never rely on the frontend alone
A frontend feature flag controls what the user *sees*. It does NOT enforce access. Any API route exposing the same capability MUST evaluate the matching backend flag independently. The two flags can share a key but they live in two `Service` instances and the backend value is the source of truth.
---
## Best-practice checklist
Adopted from Martin Fowler, ConfigCat and Octopus.
- **Naming**: `{team}_{area}_{behavior}`, e.g. `billing_checkout_new_payment_flow`. No `enable_` / `disable_` prefixes (redundant).
- **One flag, one purpose**: never repurpose an old flag for a new feature. Add a new flag and delete the old one.
- **Plan the death of the flag at birth**: open a follow-up issue to remove the flag when the rollout completes. Release flags should live days, not quarters.
- **Convention**: `Off` is the legacy / safe state, `On` is the new behavior. Lets CI test "all-off (today)" and "all-on (tomorrow)".
- **Kill switch fast path**: ops-critical flags should be exposed via `EnvProvider` so SREs can flip them without a deploy.
- **Backend protection**: anything controlling access goes through the backend Service; the frontend flag is presentation only.
- **No secrets in flags**: variant values are not Secrets Manager / KMS. Use those for tokens, keys, and passwords.
See `docs/design.md` and `docs/timezone-architecture-rfc.md` for prior examples of how this pattern is used across the codebase.
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
