fix(daemon): prevent Quick Create from inventing requirements beyond user input

The description rule in buildQuickCreatePrompt() instructed the agent to
"always provide a rich, self-contained description" and "spell out what
needs to be done", which caused the agent to fabricate detailed product
specs, implementation phases, and design decisions from a one-line input.

Replace with a faithfulness-first rule: enrich with factual context
(fetched PR details, linked resources) but never invent requirements,
design decisions, or constraints the user did not express.

Fixes MUL-1605

packages/views/autopilots/components/autopilot-detail-page.tsx

@@ -44,7 +44,9 @@ import {
 } from "./trigger-config";
 import type { TriggerConfig } from "./trigger-config";
 import type { AutopilotExecutionMode, AutopilotRun, AutopilotTrigger } from "@multica/core/types";
+import type { AgentTask } from "@multica/core/types/agent";
 import { ReadonlyContent } from "../../editor";
+import { TranscriptButton } from "../../common/task-transcript";
 import { AutopilotDialog } from "./autopilot-dialog";
 
 function formatDate(date: string): string {
@@ -63,11 +65,34 @@ const RUN_STATUS_CONFIG: Record<string, { label: string; color: string; icon: ty
   failed: { label: "Failed", color: "text-destructive", icon: XCircle },
 };
 
-function RunRow({ run }: { run: AutopilotRun }) {
+function RunRow({ run, agentId, agentName }: { run: AutopilotRun; agentId: string; agentName: string }) {
   const wsPaths = useWorkspacePaths();
   const cfg = (RUN_STATUS_CONFIG[run.status] ?? RUN_STATUS_CONFIG["issue_created"])!;
   const StatusIcon = cfg.icon;
 
+  // For runs with a task_id (run_only mode), build a minimal AgentTask so
+  // TranscriptButton can lazy-load the execution transcript.
+  const syntheticTask: AgentTask | null = run.task_id
+    ? {
+        id: run.task_id,
+        agent_id: agentId,
+        runtime_id: "",
+        issue_id: "",
+        status:
+          run.status === "running" ? "running" :
+          run.status === "completed" ? "completed" :
+          run.status === "failed" ? "failed" :
+          "queued",
+        priority: 0,
+        dispatched_at: null,
+        started_at: run.triggered_at || null,
+        completed_at: run.completed_at || null,
+        result: null,
+        error: run.failure_reason || null,
+        created_at: run.created_at,
+      }
+    : null;
+
   const content = (
     <>
       <StatusIcon className={cn("h-4 w-4 shrink-0", cfg.color, cfg.spin && "animate-spin")} />
@@ -83,6 +108,14 @@ function RunRow({ run }: { run: AutopilotRun }) {
       <span className="w-32 shrink-0 text-right text-xs text-muted-foreground tabular-nums">
         {formatDate(run.triggered_at || run.created_at)}
       </span>
+      {syntheticTask && !run.issue_id && (
+        <TranscriptButton
+          task={syntheticTask}
+          agentName={agentName}
+          isLive={run.status === "running"}
+          title="View execution log"
+        />
+      )}
     </>
   );
 
@@ -438,7 +471,7 @@ export function AutopilotDetailPage({ autopilotId }: { autopilotId: string }) {
             ) : (
               <div className="rounded-md border overflow-hidden">
                 {runs.map((run) => (
-                  <RunRow key={run.id} run={run} />
+                  <RunRow key={run.id} run={run} agentId={autopilot.assignee_id} agentName={getActorName("agent", autopilot.assignee_id)} />
                 ))}
               </div>
             )}

server/internal/daemon/prompt.go

@@ -44,7 +44,7 @@ func buildQuickCreatePrompt(task Task) string {
 	fmt.Fprintf(&b, "User input:\n> %s\n\n", task.QuickCreatePrompt)
 	b.WriteString("Field rules:\n")
 	b.WriteString("- title: required. A concise but semantically rich summary that lets a reader understand what the issue is about at a glance. If the user input references external resources (PRs, issues, URLs, etc.), use your judgment to decide whether fetching the resource would produce a meaningfully better title — if so, fetch it and incorporate the relevant context. For example, \"review PR #123\" is much less useful than \"Review PR #123: Refactor auth module to OAuth2\". Strip filler words but preserve key semantic information.\n")
-	b.WriteString("- description: always provide a rich, self-contained description. The created issue will be picked up and executed by an agent — the more accurate context the description contains, the better the agent will understand and execute the task. Use your judgment to gather context: if the user input contains URLs or references, fetch them and summarize the relevant parts. Spell out what needs to be done, what the background is, and any constraints or details that would help the executing agent avoid misunderstandings. Never echo the title here.\n")
+	b.WriteString("- description: stay faithful to the user's original input — do NOT invent requirements, design decisions, implementation plans, or constraints that the user did not express. The description should enrich the user's input with factual context only: if the input contains URLs or references (PRs, issues, docs), fetch them and summarize the relevant parts. Restate the user's intent clearly so the executing agent understands the task, but do not expand scope or add made-up details. Keep it concise. Never echo the title here.\n")
 	b.WriteString("- priority: one of `urgent`, `high`, `medium`, `low`, or omit. Map P0/P1 → urgent/high; \"asap\" → urgent. If unspecified, omit.\n")
 	b.WriteString("- assignee:\n")
 	b.WriteString("    - When the user names someone (\"assign to X\" / \"@X\"), call `multica workspace members --output json` and find the matching member by display name (case-insensitive substring match is fine). On a clean match, pass `--assignee <name>`. On no match or ambiguous match, do NOT pass `--assignee` — instead append a final line to the description: `Unrecognized assignee: X`.\n")