Stop arming the critique evidence gate for greybeard review intent

Author: TheGreatAxios

src/subagent/index.test.ts

@@ -34,6 +34,7 @@ import {
   preferCompletedSubAgentReply,
   resolveSubAgentCatchOutcome,
   resolveSubAgentDeadlineMs,
+  shouldRequireEvidence,
   subAgentToolName,
   SUBAGENT_DEADLINE_MARGIN_MS,
   SUBAGENT_PLUGIN_SPAWN_TEARDOWN_LIMITS,
@@ -45,6 +46,8 @@ import {
 } from "./index.js";
 
 import { type } from "arktype";
+import { formatDirectorSystemPrompt } from "../agent/directors/identity.js";
+import { DIRECTOR_REGISTRY } from "../agent/directors/registry.js";
 import type {
   ReactorAction,
   ReactorCapabilities,
@@ -277,6 +280,28 @@ describe("sub-agent stop helpers", () => {
     ).toBe("complete");
   });
 
+  test("shouldRequireEvidence is armed for CritiqueDirector prompt", () => {
+    expect(
+      shouldRequireEvidence({
+        systemPromptRole: formatDirectorSystemPrompt(DIRECTOR_REGISTRY.critique),
+      }),
+    ).toBe(true);
+    expect(
+      shouldRequireEvidence({
+        systemPromptRole: DIRECTOR_REGISTRY.critique.systemPrompt,
+      }),
+    ).toBe(true);
+  });
+
+  test("shouldRequireEvidence is off for greybeard even with intent=review", () => {
+    expect(
+      shouldRequireEvidence({
+        intent: "review",
+        systemPromptRole: formatDirectorSystemPrompt(DIRECTOR_REGISTRY.greybeard),
+      }),
+    ).toBe(false);
+  });
+
   test("evaluateSubAgentStop does not complete a review/critique with empty readCounts even with a full envelope", () => {
     const thrashState = {
       totalToolCalls: 1,
@@ -295,7 +320,7 @@ describe("sub-agent stop helpers", () => {
         thrashState,
         requireEvidence: true,
       }),
-    ).not.toBe("complete");
+    ).toBe("incomplete-report");
   });
 
   test("evaluateSubAgentStop completes a review when readCounts has file evidence", () => {
@@ -319,6 +344,27 @@ describe("sub-agent stop helpers", () => {
     ).toBe("complete");
   });
 
+  test("evaluateSubAgentStop completes greybeard spawn-only envelope when requireEvidence is off", () => {
+    const thrashState = {
+      totalToolCalls: 1,
+      readCounts: new Map(),
+      editedPaths: new Set<string>(),
+    };
+    expect(
+      evaluateSubAgentStop({
+        hasToolCalls: false,
+        everHadToolCalls: true,
+        turnsCompleted: 2,
+        maxTurns: 10,
+        consecutiveIdentical: 0,
+        repeatLimit: 2,
+        lastAssistantText: FULL_REPORT_ENVELOPE,
+        thrashState,
+        requireEvidence: false,
+      }),
+    ).toBe("complete");
+  });
+
   test("evaluateSubAgentStop returns never-acted when the run never used tools", () => {
     expect(
       evaluateSubAgentStop({

src/subagent/index.ts

@@ -116,6 +116,7 @@ export {
   coreSubAgentWebTools,
   createSubAgentRunController,
   runSubAgent,
+  shouldRequireEvidence,
   type SubAgentRunController,
 } from "./run.js";
 

src/subagent/nudge-director.ts

@@ -93,7 +93,7 @@ export class SubAgentDirector extends DefaultDirector {
   private readonly repeatLimit: number;
   /** When true (intent=implement), tool-less finish without edits salvages as never-edited. */
   private readonly requireEdit: boolean;
-  /** When true (intent=review / critique), empty readCounts is not a successful complete. */
+  /** When true (CritiqueDirector), empty readCounts is not a successful complete. */
   private readonly requireEvidence: boolean;
   private turnsCompleted = 0;
   private everHadToolCalls = false;

src/subagent/run.ts

@@ -89,6 +89,7 @@ import {
 } from "./dispose.js";
 import { createTaskTool } from "./task-tool.js";
 import type { RunSubAgentParams, SubAgentProvider } from "./types.js";
+import type { TaskIntent } from "./report.js";
 import { runWithSubAgentIdentity } from "./identity-context.js";
 
 export type {
@@ -223,6 +224,21 @@ export function createSubAgentRunController(
   };
 }
 
+/**
+ * Arm requireEvidence only for CritiqueDirector. Greybeard is also
+ * intent=review and may spawn-only then envelope; that is not a fake
+ * review — do not pull it into the empty-readCounts gate.
+ */
+export function shouldRequireEvidence(input: {
+  intent?: TaskIntent;
+  systemPromptRole?: string;
+}): boolean {
+  return (
+    typeof input.systemPromptRole === "string" &&
+    input.systemPromptRole.includes("CritiqueDirector")
+  );
+}
+
 // Spin up an isolated, autonomous agent loop, hand it one task, and return
 // its final report. `params.cwd` is either the dispatcher's own cwd (shared
 // mode) or a worktree snapshotted from the dispatcher's last commit
@@ -417,9 +433,7 @@ export async function runSubAgent(params: RunSubAgentParams): Promise<string> {
         modelFamilyPolicy.subAgentStallTimeoutMs,
         Date.now,
         params.intent === "implement",
-        params.intent === "review" ||
-          (typeof params.systemPromptRole === "string" &&
-            params.systemPromptRole.includes("CritiqueDirector")),
+        shouldRequireEvidence(params),
       ),
   });
 

src/subagent/stop-policy.ts

@@ -275,7 +275,7 @@ export type SubAgentStopReason =
  * (Summary, Findings, Blockers, Paths). Omitting `lastAssistantText`
  * still completes (back-compat). Missing envelope nudges
  * once (`incomplete-report`) then salvages (`incomplete-report-stop`).
- * When `requireEvidence` is set (review/critique), an empty `readCounts`
+ * When `requireEvidence` is set (CritiqueDirector), an empty `readCounts`
  * is not complete even with all four headings — same incomplete-report
  * nudge then salvage, so a wrap-up envelope cannot fake a real review.
  */
@@ -297,7 +297,7 @@ export function evaluateSubAgentStop(input: {
    */
   requireEdit?: boolean;
   /**
-   * When true (intent=review / critique leaf), a tool-using run that never
+   * When true (CritiqueDirector leaf), a tool-using run that never
    * read or searched a file is not a successful complete — even a four-heading
    * envelope is incomplete-report so the parent does not treat a wrap-up
    * narration as a finished review.
@@ -317,7 +317,7 @@ export function evaluateSubAgentStop(input: {
   // read/searched (no edit_file/write_file/delete_file) is never-edited —
   // both hard-block identical re-dispatch. After those, a tool-less turn
   // following tools is complete only with a report envelope (or when
-  // lastAssistantText is omitted). Review/critique additionally requires
+  // lastAssistantText is omitted). CritiqueDirector additionally requires
   // at least one read/search in thrashState.readCounts.
   if (!input.hasToolCalls) {
     if (!input.everHadToolCalls) return "never-acted";