Preserve completed Pi sub-agent answers

Do not reopen completed turns with time-pressure steering. Scope child reporting to delegated task results while retaining task-required safe outputs and inherited controls. Make smoke answers explicitly plain text and tool-free, with regressions and documentation. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
pelikhan · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026
commit fe36b03733b53daf92827e7516b32611badf3b31
diff --git a/.changeset/fix-pi-smoke-completion.md b/.changeset/fix-pi-smoke-completion.md
diff --git a/.github/workflows/smoke-pi-sub-agents.lock.yml b/.github/workflows/smoke-pi-sub-agents.lock.yml
diff --git a/.github/workflows/smoke-pi-sub-agents.md b/.github/workflows/smoke-pi-sub-agents.md
@@ -62,30 +62,33 @@ Provider-qualified model IDs pin the exact models rather than resolving aliases.
 description: Returns the Haiku model identity for smoke testing
 model: copilot/claude-haiku-4.5
 ---
-When asked `who am i?`, reply with exactly:
+When asked `who am i?`, reply with this single plain-text line, without Markdown
+or backticks:
 
-`claude-haiku-4.5`
+claude-haiku-4.5
 
-No extra words, punctuation, or formatting.
+Do not call tools. No extra words, punctuation, or formatting.
 
 ## agent: `mini-whoami`
 ---
 description: Returns the GPT-5 mini model identity for smoke testing
 model: copilot/gpt-5-mini
 ---
-When asked `who am i?`, reply with exactly:
+When asked `who am i?`, reply with this single plain-text line, without Markdown
+or backticks:
 
-`gpt-5-mini`
+gpt-5-mini
 
-No extra words, punctuation, or formatting.
+Do not call tools. No extra words, punctuation, or formatting.
 
 ## agent: `compact-whoami`
 ---
 description: Returns the GPT-4o mini model identity for smoke testing
 model: copilot/gpt-4o-mini
 ---
-When asked `who am i?`, reply with exactly:
+When asked `who am i?`, reply with this single plain-text line, without Markdown
+or backticks:
 
-`gpt-4o-mini`
+gpt-4o-mini
 
-No extra words, punctuation, or formatting.
+Do not call tools. No extra words, punctuation, or formatting.
diff --git a/actions/setup/js/pi_steering_extension.cjs b/actions/setup/js/pi_steering_extension.cjs
@@ -60,8 +60,8 @@ function piSteeringExtension(pi) {
     process.stderr.write(`[gh-aw/steering] Session started. timeout=${config.timeoutMinutes}min, warn<${config.timeWarningMinutes}min, critical<${config.timeCriticalMinutes}min\n`);
   });
 
-  pi.on("turn_end", async (/** @type {any} */ _event, /** @type {any} */ ctx) => {
-    if (startTime === undefined) {
+  pi.on("turn_end", async (/** @type {any} */ event, /** @type {any} */ ctx) => {
+    if (startTime === undefined || event.message?.stopReason === "stop") {
       return;
     }
 

diff --git a/actions/setup/js/pi_steering_extension.test.cjs b/actions/setup/js/pi_steering_extension.test.cjs
@@ -69,4 +69,15 @@ describe("Pi timeout steering", () => {
     expect(sendUserMessage).toHaveBeenCalledOnce();
     expect(sendUserMessage).toHaveBeenCalledWith(expect.stringContaining("CRITICAL"), { deliverAs: "steer" });
   });
+
+  it("does not reopen a completed answer, but still steers subsequent tool turns", async () => {
+    const { handlers, sendUserMessage } = harness();
+    await handlers.agent_start();
+    vi.advanceTimersByTime(9 * 60_000);
+    await handlers.turn_end({ message: { role: "assistant", stopReason: "stop" } }, {});
+    expect(sendUserMessage).not.toHaveBeenCalled();
+    await handlers.turn_end({ message: { role: "assistant", stopReason: "toolUse" } }, {});
+    expect(sendUserMessage).toHaveBeenCalledOnce();
+    expect(sendUserMessage).toHaveBeenCalledWith(expect.stringContaining("CRITICAL"), { deliverAs: "steer" });
+  });
 });
diff --git a/actions/setup/js/pi_subagent_extension.cjs b/actions/setup/js/pi_subagent_extension.cjs
@@ -38,7 +38,11 @@ async function runPiSubagent(agent, task, ctx, signal, launch = spawn) {
     const promptPath = path.join(dir, "system.txt");
     const systemPath = process.env.GH_AW_PI_SYSTEM_PROMPT;
     const system = systemPath ? fs.readFileSync(systemPath, "utf8") + "\n\n" : "";
-    fs.writeFileSync(promptPath, system + agent.prompt, { mode: 0o600 });
+    const delegationScope =
+      "You are a delegated sub-agent, not the workflow's main agent. Complete only the delegated task and return your answer to the parent, following the declared agent's output format exactly. " +
+      "The parent is responsible for finalizing the workflow and its required safe-output reporting. Do not emit noop safe outputs or workflow-completion reports for this child session. " +
+      "Safe-output actions required to perform the delegated task remain permitted.";
+    fs.writeFileSync(promptPath, system + agent.prompt + "\n\n" + delegationScope, { mode: 0o600 });
     process.stdout.write(JSON.stringify({ type: "gh_aw_subagent_dispatch", agent: agent.name, requested_model: agent.declaredModel || agent.modelId, resolved_model: agent.modelId }) + "\n");
     return await new Promise((resolve, reject) => {
       const child = launch(process.env.GH_AW_PI_COMMAND || "pi", subagentArgs(agent, promptPath), {

diff --git a/actions/setup/js/pi_subagent_extension.test.cjs b/actions/setup/js/pi_subagent_extension.test.cjs
@@ -93,6 +93,21 @@ describe("Pi managed delegation", () => {
     await expect(runPiSubagent(agent, "Read.", { cwd: dir }, undefined, launcher({ stopReason: "error", errorMessage: "Model unavailable" }))).rejects.toThrow("Model unavailable");
   });
 
+  it("inherits workflow controls without making the child finalize the workflow", async () => {
+    const systemPath = path.join(dir, "workflow-system.txt");
+    fs.writeFileSync(systemPath, "Workflow instructions and tool restrictions.");
+    vi.stubEnv("GH_AW_PI_SYSTEM_PROMPT", systemPath);
+    const launch = launcher();
+    await runPiSubagent(agent, "Read.", { cwd: dir }, undefined, (...args) => {
+      const prompt = fs.readFileSync(args[1].at(-1), "utf8");
+      expect(prompt).toContain("Workflow instructions and tool restrictions.\n\nRead only.");
+      expect(prompt).toContain("The parent is responsible for finalizing the workflow");
+      expect(prompt).toContain("Do not emit noop safe outputs");
+      expect(prompt).toContain("Safe-output actions required to perform the delegated task remain permitted");
+      return launch(...args);
+    });
+  });
+
   it("propagates nonzero child exits", async () => {
     await expect(runPiSubagent(agent, "Read.", { cwd: dir }, undefined, launcher({}, 1))).rejects.toThrow();
   });

diff --git a/docs/src/content/docs/engines/pi.md b/docs/src/content/docs/engines/pi.md
@@ -143,6 +143,8 @@ An omitted model inherits the parent's configured model. Bare model IDs use the
 
 Children retain the workflow's inference gateway, MCP configuration, steering, bash allowlist, and edit restrictions. Tool-call and denial budgets are shared with the parent. Children use compiler-managed execution arguments rather than inheriting `engine.args` overrides. Optional automatically discovered extensions are disabled in children; only the managed infrastructure extensions are loaded. Custom drivers do not support this delegation path.
 
+Children return task results to the parent instead of emitting workflow-completion reports or `noop` safe outputs. Task-required safe-output actions remain available. Time-pressure steering applies to ongoing tool turns, not completed answers, so it does not reopen a finished child session.
+
 Dispatch logs record both the requested alias and resolved model. `gh aw audit` reports `subagent_model_not_observed` when usage evidence exists but no agent inference matches a declared model. This means the agent may not have run or its model was not honored; model presence alone does not prove delegation. Detection requests are not treated as agent-model evidence.
 
 ## Additional providers