Q113: a goal audit approves only with the decision alone on its last line

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/main/services/agent-goal-tools.ts

@@ -173,7 +173,7 @@ async function runAudit(
       "The executor's claim:",
       evidence,
       `Inspect the work before deciding: list_files and read_file for the workspace, list_documents and read_document for the library. An approval counts only after you have read some of it. ${canRun ? "run_command runs in a throwaway copy of the workspace: build and run the tests yourself instead of trusting claims that they pass; nothing you do there reaches the real workspace." : "You cannot run commands here, so a claim that something was run or passes is unverified unless the workspace holds its results."} Judge outcomes, not artifacts about outcomes: plans, specs, status reports, and skeletons do not make a goal achieved; working deliverables do. If the goal promises software, look for the actual implementation, not documents describing it.`,
-      "End your final message with exactly one marker: <approved/> if the goal is genuinely achieved, or <disapproved/> followed by one short sentence naming the decisive gap.",
+      "Keep your reasoning apart from your decision. End your final message with the decision alone on its last line: <approved/> by itself if the goal is genuinely achieved, or <disapproved/> followed by one short sentence naming the decisive gap. Use no marker anywhere else.",
     ].join("\n\n"),
   }];
   const instructions = "You are the independent completion auditor. You did not do the work and owe its author nothing; your only loyalty is to what actually exists. Be strict: an unverifiable claim is a false claim.";
@@ -225,10 +225,14 @@ async function runAuditTool(deps: AgentToolDeps, ctx: AgentToolContext, call: To
   }
 }
 
+/** Approved only when the last line is <approved/> alone and no other marker appears; quoted, indented or hedged markers do not count. */
 function parseVerdict(text: string): AuditVerdict {
-  const reason = text.replace(/<\/?(approved|disapproved)\s*\/?>/gi, " ").replace(/\s+/g, " ").trim();
-  if (/<approved\s*\/>/i.test(text)) return { approved: true, reason: reason || "no gaps found" };
-  return { approved: false, reason: reason || "the auditor did not state an explicit approval" };
+  const markers = text.match(/<\/?\s*(approved|disapproved)\s*\/?>/gi) ?? [];
+  const reason = text.replace(/<\/?\s*(approved|disapproved)\s*\/?>/gi, " ").replace(/\s+/g, " ").trim();
+  const lastLine = text.split("\n").map((line) => line.trimEnd()).filter(Boolean).at(-1);
+  if (markers.length === 1 && lastLine === "<approved/>") return { approved: true, reason: reason || "no gaps found" };
+  if (markers.some((marker) => /disapproved/i.test(marker))) return { approved: false, reason: reason || "the auditor disapproved without a reason" };
+  return { approved: false, reason: reason ? `${reason} (no <approved/> alone on the last line)` : "the auditor did not end with <approved/> alone on the last line" };
 }
 
 function actorIsCeoOrGoalOwner(deps: AgentToolDeps, ctx: AgentToolContext, ownerAgentId: string | null): boolean {

tests/unit/agent-goal-tools.test.ts

@@ -97,7 +97,7 @@ describe("complete_goal with independent audit", () => {
       const readResult = input.messages.at(-1);
       expect(readResult?.role).toBe("tool");
       expect(input.messages.at(-2)).toMatchObject({ role: "assistant", replay });
-      return turnOf("The implementation exists and runs. <approved/>");
+      return turnOf("The implementation exists and runs.\n<approved/>");
     };
 
     const readOnly = argsFor(h, h.ceoId, auditor, call({ evidence: "src/index.ts implements the app and runs." }));
@@ -128,7 +128,7 @@ describe("complete_goal with independent audit", () => {
         ranIn = results[1].output;
         return turnOf("Reading.", [{ id: "c3", name: "read_file", arguments: { path: "built.txt" } }]);
       }
-      return turnOf("Tests pass when run. <approved/>");
+      return turnOf("Tests pass when run.\n<approved/>");
     };
     const output = await completeGoalTool(argsFor(h, h.ceoId, auditor, call({ evidence: "All tests pass." })));
     expect(output).toContain("achieved");
@@ -199,7 +199,7 @@ describe("complete_goal with independent audit", () => {
     const auditor = async (input: AuditorInput): Promise<AuditorTurn> => {
       const last = input.messages.at(-1);
       prompts.push(last && "text" in last ? String(last.text) : "");
-      return turnOf("Looks done. <approved/>");
+      return turnOf("Looks done.\n<approved/>");
     };
     const output = await completeGoalTool(argsFor(h, h.ceoId, auditor, call({ evidence: "Trust me." })));
     expect(output).toContain("audit rejected");
@@ -221,7 +221,7 @@ describe("complete_goal with independent audit", () => {
       turnOf("History only.", [{ id: "c3", name: "read_document", arguments: { doc_id: docId, history: true } }]),
       turnOf("<approved/>"),
       turnOf("Reading the report.", [{ id: "c2", name: "read_document", arguments: { doc_id: docId } }]),
-      turnOf("The report shows customers. <approved/>"),
+      turnOf("The report shows customers.\n<approved/>"),
     ];
     const seen: string[][] = [];
     const histories: AuditorInput["messages"][] = [];
@@ -240,6 +240,30 @@ describe("complete_goal with independent audit", () => {
     h.connection.close();
   });
 
+  test("only a final message that ends in a single approval marker closes the goal", async () => {
+    const verdicts: Array<[string, string]> = [
+      ["<approved/> on style, but <disapproved/> the tests fail.", "active"],
+      ["I would write <approved/> if the tests passed. They do not.", "active"],
+      ["Everything checks out: `<approved/>`", "active"],
+      ["<approved/> <approved/>", "active"],
+      ["The example approval token is:\n> <approved/>", "active"],
+      ["Verified. <approved/>", "active"],
+      ["Example:\n    <approved/>", "active"],
+      ["The build and tests pass.\n<approved/>\n", "achieved"],
+    ];
+    for (const [verdict, expected] of verdicts) {
+      const h = createHarness();
+      let turns = 0;
+      const auditor = async (): Promise<AuditorTurn> => {
+        turns += 1;
+        return turns === 1 ? turnOf("Reading.", [{ id: "c1", name: "read_file", arguments: { path: "src/index.ts" } }]) : turnOf(verdict);
+      };
+      await completeGoalTool(argsFor(h, h.ceoId, auditor, call({ evidence: "src/index.ts works" })));
+      expect([verdict, goalStatus(h)]).toEqual([verdict, expected]);
+      h.connection.close();
+    }
+  });
+
   test("no explicit approval means not achieved", async () => {
     const h = createHarness();
     const auditor = async (): Promise<AuditorTurn> => turnOf("Looks plausible I guess.");