Files
multica/server/pkg/agent/agent.go
Bohan Jiang 60bae62622 feat(codex): add per-exec_command watchdog to escape dropped function_call_output (MUL-2337) (#2779)
* feat(codex): add per-exec_command watchdog to escape dropped function_call_output (MUL-2337)

Codex app-server can drop the second function_call_output when two
exec_command calls fan out in the same turn and both async-yield through
the yield_time_ms boundary (observed 2026-05-18, MUL-2334 — Trump Agent
wedged for 6+ min with no semantic activity events to drive any existing
timer). The model then waits forever for the missing output; only the
10-minute semantic inactivity timeout would eventually rescue the run.

Add a per-call watchdog in the codex client that tracks open
exec_command / commandExecution items by call_id and fails the turn
quickly (default 2 min, configurable via ExecOptions.ExecCommandStuckTimeout)
when one stays open without progress. outputDelta events reset the
per-call progress timestamp so long-running streaming commands aren't
flagged.

This is a daemon-side mitigation only — codex itself still has the
upstream race, but the daemon no longer burns the full inactivity budget
before the run is marked failed and a new run can recover.

Co-authored-by: multica-agent <github@multica.ai>

* feat(codex): track legacy exec_command_output_delta in watchdog (MUL-2337)

Mirrors the raw v2 item/commandExecution/outputDelta refresh on the legacy
codex/event protocol so a long-running streaming exec doesn't get falsely
flagged as stuck after begin + 2 min.

Co-authored-by: multica-agent <github@multica.ai>

---------

Co-authored-by: multica-agent <github@multica.ai>
2026-05-18 15:14:45 +08:00

174 lines
6.6 KiB
Go

// Package agent provides a unified interface for executing prompts via
// coding agents (Claude Code, Codex, Copilot, OpenCode, OpenClaw, Hermes,
// Gemini, Pi, Cursor, Kimi, Kiro). It mirrors the happy-cli AgentBackend
// pattern, translated to idiomatic Go.
package agent
import (
"context"
"encoding/json"
"fmt"
"log/slog"
"time"
)
// Backend is the unified interface for executing prompts via coding agents.
type Backend interface {
// Execute runs a prompt and returns a Session for streaming results.
// The caller should read from Session.Messages (optional) and wait on
// Session.Result for the final outcome.
Execute(ctx context.Context, prompt string, opts ExecOptions) (*Session, error)
}
// ExecOptions configures a single execution.
type ExecOptions struct {
Cwd string
Model string
// SystemPrompt is consumed only by providers that can pass or safely inline
// developer/system instructions. Hermes ACP intentionally ignores it and
// relies on cwd-scoped context files such as AGENTS.md instead.
SystemPrompt string
MaxTurns int
Timeout time.Duration
SemanticInactivityTimeout time.Duration
// ExecCommandStuckTimeout bounds how long a single exec_command (or v2
// commandExecution item) may stay open without a matching end / progress
// event before the turn is declared stuck. Currently honoured by the codex
// backend, which has a known bug where two parallel exec_command calls
// racing through the yield_time_ms boundary can drop the second
// function_call_output, leaving the model waiting forever. Zero means use
// the backend default.
ExecCommandStuckTimeout time.Duration
ResumeSessionID string // if non-empty, resume a previous agent session
ExtraArgs []string // daemon-wide default CLI arguments appended before CustomArgs; currently read by claude and codex backends only
CustomArgs []string // per-agent CLI arguments appended after ExtraArgs
McpConfig json.RawMessage // if non-nil, MCP server config to pass via --mcp-config
}
// Session represents a running agent execution.
type Session struct {
// Messages streams events as the agent works. The channel is closed
// when the agent finishes (before Result is sent).
Messages <-chan Message
// Result receives exactly one value — the final outcome — then closes.
Result <-chan Result
}
// MessageType identifies the kind of Message.
type MessageType string
const (
MessageText MessageType = "text"
MessageThinking MessageType = "thinking"
MessageToolUse MessageType = "tool-use"
MessageToolResult MessageType = "tool-result"
MessageStatus MessageType = "status"
MessageError MessageType = "error"
MessageLog MessageType = "log"
)
// Message is a unified event emitted by an agent during execution.
type Message struct {
Type MessageType
Content string // text content (Text, Error, Log)
Tool string // tool name (ToolUse, ToolResult)
CallID string // tool call ID (ToolUse, ToolResult)
Input map[string]any // tool input (ToolUse)
Output string // tool output (ToolResult)
Status string // agent status string (Status)
Level string // log level (Log)
SessionID string // backend session id (Status), for early resume-pointer pinning
}
// TokenUsage tracks token consumption for a single model.
type TokenUsage struct {
InputTokens int64
OutputTokens int64
CacheReadTokens int64
CacheWriteTokens int64
}
// Result is the final outcome after an agent session completes.
type Result struct {
Status string // "completed", "failed", "aborted", "timeout", "cancelled"
Output string // accumulated text output
Error string // error message if failed
DurationMs int64
SessionID string
Usage map[string]TokenUsage // keyed by model name
}
// Config configures a Backend instance.
type Config struct {
ExecutablePath string // path to CLI binary (claude, codex, copilot, opencode, openclaw, hermes, gemini, pi, cursor, kimi, kiro-cli)
Env map[string]string // extra environment variables
Logger *slog.Logger
}
// New creates a Backend for the given agent type.
// Supported types: "claude", "codex", "copilot", "opencode", "openclaw", "hermes", "gemini", "pi", "cursor", "kimi", "kiro".
func New(agentType string, cfg Config) (Backend, error) {
if cfg.Logger == nil {
cfg.Logger = slog.Default()
}
switch agentType {
case "claude":
return &claudeBackend{cfg: cfg}, nil
case "codex":
return &codexBackend{cfg: cfg}, nil
case "copilot":
return &copilotBackend{cfg: cfg}, nil
case "opencode":
return &opencodeBackend{cfg: cfg}, nil
case "openclaw":
return &openclawBackend{cfg: cfg}, nil
case "hermes":
return &hermesBackend{cfg: cfg}, nil
case "gemini":
return &geminiBackend{cfg: cfg}, nil
case "pi":
return &piBackend{cfg: cfg}, nil
case "cursor":
return &cursorBackend{cfg: cfg}, nil
case "kimi":
return &kimiBackend{cfg: cfg}, nil
case "kiro":
return &kiroBackend{cfg: cfg}, nil
default:
return nil, fmt.Errorf("unknown agent type: %q (supported: claude, codex, copilot, opencode, openclaw, hermes, gemini, pi, cursor, kimi, kiro)", agentType)
}
}
// DetectVersion runs the agent CLI with --version and returns the output.
func DetectVersion(ctx context.Context, executablePath string) (string, error) {
return detectCLIVersion(ctx, executablePath)
}
// launchHeaders maps each supported agent type to the user-visible skeleton
// that the daemon spawns before any custom_args are appended. This is
// intentionally minimal — only the command + subcommand (or a short mode
// label when there is no subcommand). Internal flags, transport values, and
// environment variables are deliberately omitted so the string is a hint
// about *what* users are extending, not a dump of the full command line.
var launchHeaders = map[string]string{
"claude": "claude (stream-json)",
"codex": "codex app-server",
"copilot": "copilot (json)",
"cursor": "cursor-agent (stream-json)",
"gemini": "gemini (stream-json)",
"hermes": "hermes acp",
"openclaw": "openclaw agent (json)",
"opencode": "opencode run (json)",
"pi": "pi (json mode)",
"kimi": "kimi acp",
"kiro": "kiro-cli acp",
}
// LaunchHeader returns the user-visible launch skeleton for agentType, or an
// empty string if the type is unknown. Callers render this as a preview so
// users understand which command their custom_args get appended to.
func LaunchHeader(agentType string) string {
return launchHeaders[agentType]
}