fix(executor): bypass summary-incomplete refusal for PREMISE STALE: summaries

Address the code-review finding on the prior commit: a natural premise-stale
summary like "PREMISE STALE: the task has no remaining work — implementation
is already done on HEAD" matches /\b(incomplete|not implemented|not done|
not finished)\b/i with 'the task' inside the 40-char first-person window,
refusing fn_task_done with summary-claims-incomplete and deadlocking the
escape hatch.

When summary starts (case-insensitive) with PREMISE STALE:, skip the
dissent-pattern and scoped-incomplete summary checks. Pending-code-review
and bulk-step-completion guards still apply unchanged.

Add executor-task-done-premise-stale.test.ts covering: the deadlock case
now passes; dissent phrasing in a sentinel summary is allowed; case-
insensitive sentinel; sentinel must be at the start (mid-summary doesn't
bypass); REVISE verdict still blocks even with the sentinel.

packages/engine/src/__tests__/executor-task-done-premise-stale.test.ts

@@ -0,0 +1,85 @@
+import { describe, expect, it } from "vitest";
+import { evaluateTaskDoneRefusal } from "../executor.js";
+
+function createTask(stepStatuses: Array<"done" | "skipped" | "pending" | "in-progress">) {
+  return {
+    id: "FN-PREMISE-STALE",
+    title: "Premise stale",
+    description: "",
+    column: "in-progress",
+    dependencies: [],
+    steps: stepStatuses.map((status, index) => ({ name: `Step ${index + 1}`, status })),
+    currentStep: 0,
+    createdAt: new Date().toISOString(),
+    updatedAt: new Date().toISOString(),
+  } as any;
+}
+
+describe("preflight PREMISE STALE: escape hatch", () => {
+  it("allows fn_task_done when summary starts with PREMISE STALE: even if it contains 'done' near 'the task'", () => {
+    // Without the bypass, the scoped-incomplete regex matches 'done' and the
+    // 40-char window contains 'the task' → would refuse with
+    // summary-claims-incomplete. The bypass must let this through.
+    const task = createTask(["done", "skipped", "skipped", "skipped", "skipped"]);
+    const result = evaluateTaskDoneRefusal(
+      task,
+      { summary: "PREMISE STALE: the task has no remaining work — implementation is already done on HEAD" },
+      new Map(),
+    );
+    expect(result).toEqual({ ok: true });
+  });
+
+  it("allows fn_task_done when summary starts with PREMISE STALE: and contains 'I'm blocked' style dissent phrasing", () => {
+    // Natural premise-stale phrasing may accidentally include a dissent-pattern
+    // word ("blocked from", "to unblock", "requires follow-up"). The bypass
+    // must not refuse on the dissent regex when the sentinel is present.
+    const task = createTask(["done", "skipped", "skipped"]);
+    const result = evaluateTaskDoneRefusal(
+      task,
+      { summary: "PREMISE STALE: targeted reproduction passes on HEAD; nothing to unblock and no further work required" },
+      new Map(),
+    );
+    expect(result).toEqual({ ok: true });
+  });
+
+  it("is case-insensitive on the sentinel", () => {
+    const task = createTask(["done", "skipped"]);
+    const result = evaluateTaskDoneRefusal(
+      task,
+      { summary: "premise stale: this task is not done because main already shipped it" },
+      new Map(),
+    );
+    expect(result).toEqual({ ok: true });
+  });
+
+  it("does NOT bypass when sentinel appears later in the summary (must be at start)", () => {
+    // Defends against agents tacking the sentinel into the middle to dodge a
+    // genuine incomplete-work refusal.
+    const task = createTask(["done", "pending"]);
+    const result = evaluateTaskDoneRefusal(
+      task,
+      { summary: "The task is not done yet, but PREMISE STALE: I think it's stale anyway" },
+      new Map(),
+    );
+    expect(result.ok).toBe(false);
+    if (!result.ok) {
+      expect(result.refusalClass).toBe("summary-claims-incomplete");
+    }
+  });
+
+  it("still enforces pending-code-review-revise even with the sentinel", () => {
+    // The bypass only relaxes the summary-text checks. A genuine REVISE verdict
+    // on an in-progress step must still block fn_task_done.
+    const task = createTask(["done", "in-progress"]);
+    const verdicts = new Map<number, "REVISE">([[1, "REVISE"]]);
+    const result = evaluateTaskDoneRefusal(
+      task,
+      { summary: "PREMISE STALE: already done on HEAD" },
+      verdicts as any,
+    );
+    expect(result.ok).toBe(false);
+    if (!result.ok) {
+      expect(result.refusalClass).toBe("pending-code-review-revise");
+    }
+  });
+});

packages/engine/src/executor.ts

@@ -310,7 +310,15 @@ export function evaluateTaskDoneRefusal(
   }
 
   const summary = params.summary?.trim();
-  if (summary) {
+  // Preflight escape hatch: when the agent's preflight finds PROMPT.md is out
+  // of sync with HEAD (work already done on the base), it marks remaining
+  // steps `skipped` and calls fn_task_done with a `PREMISE STALE:` summary.
+  // Skip the summary-text refusals (dissent + scoped-incomplete) for this
+  // sentinel so a natural premise-stale explanation like "...the work is
+  // already done on HEAD" cannot deadlock the executor. The pending-review
+  // and bulk-step-completion guards above/below still apply.
+  const isPremiseStale = !!summary && /^premise stale:/i.test(summary);
+  if (summary && !isPremiseStale) {
     const dissentMatch = DISSENT_PATTERNS.find((pattern) => pattern.test(summary));
     if (dissentMatch) {
       const matchText = summary.match(dissentMatch)?.[0] ?? dissentMatch.source;