pi-ui-extend 1.0.24 → 1.0.30

package/dist/app/types.d.ts

@@ -284,6 +284,7 @@ export type StatusLineLayout = {
     compactToolsWidget?: StatusCompactToolsWidgetLayout;
     terminalBellSoundWidget?: StatusTerminalBellSoundWidgetLayout;
     promptEnhancerWidget?: StatusPromptEnhancerWidgetLayout;
+    agentPauseWidget?: StatusAgentPauseWidgetLayout;
     voiceWidget?: StatusVoiceWidgetLayout;
 };
 export type StatusQuickScrollWidgetLayout = {
@@ -318,6 +319,10 @@ export type StatusPromptEnhancerWidgetLayout = {
     startColumn: number;
     endColumn: number;
 };
+export type StatusAgentPauseWidgetLayout = {
+    startColumn: number;
+    endColumn: number;
+};
 export type StatusTerminalBellSoundTarget = {
     row: number;
     startColumn: number;
@@ -589,6 +594,11 @@ export type StatusPromptEnhancerTarget = {
     startColumn: number;
     endColumn: number;
 };
+export type StatusAgentPauseTarget = {
+    row: number;
+    startColumn: number;
+    endColumn: number;
+};
 export type StatusUserJumpTarget = {
     row: number;
     startColumn: number;

package/dist/config.js

@@ -55,7 +55,7 @@ const DEFAULT_OUTPUT_FILTERS = {
 };
 const THINKING_LEVELS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"];
 const DEFAULT_PROMPT_ENHANCER = {
-    modelRef: "zai/glm-5-turbo",
+    modelRef: "openai-codex/gpt-5.6-luna",
 };
 const DEFAULT_AUTOCOMPLETE = {
     modelRef: "zai/glm-5-turbo",

package/dist/default-pix-config.js

@@ -52,9 +52,9 @@ export const DEFAULT_PIX_CONFIG_JSONC = String.raw `{
     "antigravity/antigravity-claude-*": "error"
   },
 
-  "promptEnhancer": { "modelRef": "zai/glm-5-turbo" },
+  "promptEnhancer": { "modelRef": "openai-codex/gpt-5.6-luna" },
   "autocomplete": { "modelRef": "zai/glm-5-turbo", "debounceMs": 350, "timeoutMs": 3000, "maxTokens": 48, "maxPromptTokens": 1200, "includeRecentMessages": 0 },
-  "sessionTitle": { "modelRef": "zai/glm-5-turbo", "fallbackModels": ["openai-codex/gpt-5.3-codex-spark"] },
+  "sessionTitle": { "modelRef": "openai-codex/gpt-5.6-luna", "fallbackModels": ["zai/glm-5-turbo"] },
   "dictation": {
     "language": "en",
     "languages": {

package/external/pi-tools-suite/README.md

@@ -247,7 +247,7 @@ For an oh-my-openagent-style workflow, run `/ultrawork` or `/ulw` to ask the par
 
 ### Private browser QA and project auth
 
-The built-in `browser-qa` role runs on `openai-codex/gpt-5.4-mini`, with
+The built-in `browser-qa` role runs on `openai-codex/gpt-5.6-luna`, with
 `antigravity/gemini-3-flash-preview` and then `zai/glm-5.3` as fallbacks. Its browser
 workflow is an explicit private skill under `src/async-subagents/private-skills/`,
 outside normal Pi skill discovery. The role's first-class `isolatedSkills` setting launches the child with

package/external/pi-tools-suite/docs/browser-qa-subagent.md

@@ -4,7 +4,7 @@
 
 Provide a cheap, fast `browser-qa` async-subagent that reproduces browser bugs
 and proves fixes with deterministic assertions plus screenshot, video, and trace
-evidence. The role uses `openai-codex/gpt-5.4-mini`, falling back to
+evidence. The role uses `openai-codex/gpt-5.6-luna`, falling back to
 `antigravity/gemini-3-flash-preview` and then `zai/glm-5.3`.
 
 ## Private skill isolation
@@ -62,6 +62,9 @@ evidence. The role uses `openai-codex/gpt-5.4-mini`, falling back to
   trusted runner implements a bounded set of navigation, interaction,
   assertion, screenshot, and auth-rejection actions and never gives the flow a
   Playwright context or credential values.
+- Target discovery is a bounded preflight, not an open-ended research task. The
+  sub-agent invokes the runner within 45 seconds or returns `BLOCKED`; it does
+  not spend the full launcher budget reading source or probing prerequisites.
 - The launcher injects `PI_SUBAGENT_AGENT_DIR`, pre-creates a private
   `browser-qa/flows/` workspace, and clears stale browser QA files when an agent
   id is reused. The runner validates the directory's project/type metadata and
@@ -79,6 +82,32 @@ evidence. The role uses `openai-codex/gpt-5.4-mini`, falling back to
 - Auth rejection discovered by a QA flow is reported through the
   `authRejectedIf` action so the parent gets an update-required status.
 
+## Reliability and shutdown contract
+
+- The built-in `browser-qa` profile has a 120-second wall-clock budget unless
+  the caller explicitly supplies a task or spawn timeout. This bounds model
+  stalls as well as browser work.
+- The trusted runner has its own bounded lifecycle. Browser launch, context
+  setup, auth, flow execution, evidence finalization, and browser shutdown must
+  not wait forever; a timeout reports the last started stage without exposing
+  flow contents or credentials.
+- Trace sanitization runs in a memory-limited worker that can be terminated at
+  the cleanup deadline; synchronous archive work cannot defeat the watchdog.
+- The launcher always writes a small sanitized `progress.jsonl` journal in the
+  agent directory. It records lifecycle/RPC event types and tool names, but not
+  prompts, tool arguments, tool results, model text, or secrets. The browser
+  runner writes similarly sanitized stage entries under its private workspace.
+- On POSIX, newly launched agents own a process group. Settled, timed-out, and
+  explicitly stopped agents signal that group rather than only the Pi process;
+  timeout/settled shutdown escalates to `SIGKILL` after its grace period. On
+  Windows the existing recursive `taskkill /T /F` behavior remains in force.
+- Process-tree cleanup is scoped to a launcher-created process-group marker so
+  an old or externally-created PID is never treated as an owned process group.
+  User browser sessions outside that group must not be signalled.
+- Playwright can launch Chromium in its own POSIX process group. On runner
+  failure the runner snapshots and kills only its own descendants before it
+  exits, covering that detached browser tree without touching a user's browser.
+
 ## Acceptance criteria
 
 1. `browser-qa` resolves to the intended model/fallback and its self-contained
@@ -96,7 +125,10 @@ evidence. The role uses `openai-codex/gpt-5.4-mini`, falling back to
    deleting the run removes them while persistent auth config/state remains.
 6. Completed test runs report clickable screenshot, video, and trace links
    whenever those artifacts exist.
-7. Suite tests/typecheck, host checks, and suite sync pass.
+7. Timeout tests identify the last browser stage, launcher progress remains
+   available when full RPC logging is disabled, and process-tree tests prove a
+   descendant is terminated without signalling unrelated processes.
+8. Suite tests/typecheck, host checks, and suite sync pass.
 
 ## Real-browser regression test
 

package/external/pi-tools-suite/src/async-subagents/async-subagents.sample.jsonc

@@ -71,7 +71,7 @@
     "model": "zai/glm-4.5-air",
     // Ordered router model fallbacks tried when the primary routing model is
     // unavailable or fails. The current parent model is always tried last.
-    "fallbackModels": ["openai-codex/gpt-5.3-codex-spark"],
+    "fallbackModels": ["openai-codex/gpt-5.6-luna"],
     "maxTaskChars": 1200,
     "maxTokens": 512,
     "maxRetries": 3,
@@ -116,7 +116,7 @@
         "research": { "model": "zai/glm-5-turbo", "thinking": "low" },
         "docs": { "model": "zai/glm-4.5-air", "thinking": "low" },
         "frontend": { "model": "antigravity/gemini-3-flash-preview", "fallbackModels": ["zai/glm-5.3"], "thinking": "medium" },
-        "browser-qa": { "model": "openai-codex/gpt-5.4-mini", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "medium" },
+        "browser-qa": { "model": "openai-codex/gpt-5.6-luna", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "low" },
         "tests": { "model": "zai/glm-5-turbo", "thinking": "medium" },
         "review": { "model": "zai/glm-5.3", "thinking": "high" },
         "implement": { "model": "zai/glm-5.3", "thinking": "high" },
@@ -132,7 +132,7 @@
         "research": { "model": "openai-codex/gpt-5.6-terra", "fallbackModels": ["zai/glm-5-turbo"], "thinking": "low" },
         "docs": { "model": "openai-codex/gpt-5.6-luna", "fallbackModels": ["zai/glm-4.5-air"], "thinking": "low" },
         "frontend": { "model": "openai-codex/gpt-5.6-terra", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "medium" },
-        "browser-qa": { "model": "openai-codex/gpt-5.4-mini", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "medium" },
+        "browser-qa": { "model": "openai-codex/gpt-5.6-luna", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "low" },
         "tests": { "model": "openai-codex/gpt-5.6-terra", "fallbackModels": ["zai/glm-5-turbo"], "thinking": "medium" },
         "review": { "model": "openai-codex/gpt-5.6-sol", "fallbackModels": ["zai/glm-5.3"], "thinking": "high" },
         "implement": { "model": "openai-codex/gpt-5.6-sol", "fallbackModels": ["zai/glm-5.3"], "thinking": "high" },
@@ -143,13 +143,13 @@
     "deep": {
       "description": "Use only enabled strong models, mixing GPT, Claude, and Gemini by role.",
       "types": {
-        "quick": { "model": "openai-codex/gpt-5.4-mini", "fallbackModels": ["zai/glm-4.5-air"], "thinking": "low" },
-        "scan": { "model": "antigravity/gemini-3-flash-preview", "fallbackModels": ["openai-codex/gpt-5.3-codex-spark", "zai/glm-4.5-air"], "thinking": "off" },
-        "research": { "model": "antigravity/gemini-3.1-pro-preview", "fallbackModels": ["openai-codex/gpt-5.4-mini", "zai/glm-5-turbo"], "thinking": "medium" },
-        "docs": { "model": "antigravity/gemini-2.5-flash", "fallbackModels": ["openai-codex/gpt-5.3-codex-spark", "zai/glm-4.5-air"], "thinking": "medium" },
-        "frontend": { "model": "antigravity/gemini-3.1-pro-preview-customtools", "fallbackModels": ["openai-codex/gpt-5.4-mini", "zai/glm-5.3"], "thinking": "low" },
-        "browser-qa": { "model": "openai-codex/gpt-5.4-mini", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "medium" },
-        "tests": { "model": "antigravity/antigravity-claude-sonnet-4-6", "fallbackModels": ["openai-codex/gpt-5.4-mini", "zai/glm-5-turbo"], "thinking": "high" },
+        "quick": { "model": "openai-codex/gpt-5.6-luna", "fallbackModels": ["zai/glm-4.5-air"], "thinking": "low" },
+        "scan": { "model": "antigravity/gemini-3-flash-preview", "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-4.5-air"], "thinking": "off" },
+        "research": { "model": "antigravity/gemini-3.1-pro-preview", "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-5-turbo"], "thinking": "medium" },
+        "docs": { "model": "antigravity/gemini-2.5-flash", "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-4.5-air"], "thinking": "medium" },
+        "frontend": { "model": "antigravity/gemini-3.1-pro-preview-customtools", "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-5.3"], "thinking": "low" },
+        "browser-qa": { "model": "openai-codex/gpt-5.6-luna", "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"], "thinking": "low" },
+        "tests": { "model": "antigravity/antigravity-claude-sonnet-4-6", "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-5-turbo"], "thinking": "high" },
         "review": { "model": "antigravity/antigravity-claude-sonnet-4-6", "fallbackModels": ["openai-codex/gpt-5.6-sol", "zai/glm-5.3"], "thinking": "high" },
         "implement": { "model": "openai-codex/gpt-5.6-sol", "fallbackModels": ["zai/glm-5.3"], "thinking": "high" },
         "deep": { "model": "antigravity/antigravity-claude-opus-4-6-thinking", "fallbackModels": ["openai-codex/gpt-5.6-sol", "zai/glm-5.3"], "thinking": "high" }
@@ -160,34 +160,38 @@
   "types": {
     "quick": {
       "description": "Use for tiny cheap tasks: answer a simple question, inspect one known file, or verify one fact. Not for broad repo search.",
-      "model": "zai/glm-5-turbo",
+      "model": "openai-codex/gpt-5.6-luna",
+      "fallbackModels": ["zai/glm-4.5-air"],
       "thinking": "off"
     },
 
     "scan": {
       "description": "Use for finding files, symbols, text, or inventory across a repo. Return paths/facts; do not judge code quality.",
-      "model": "zai/glm-5-turbo",
+      "model": "openai-codex/gpt-5.6-luna",
+      "fallbackModels": ["zai/glm-4.5-air"],
       "thinking": "off",
       "tools": ["read", "grep"]
     },
 
     "research": {
       "description": "Use for multi-file codebase research: read several files and explain how something works. No edits.",
-      "model": "zai/glm-5-turbo",
+      "model": "openai-codex/gpt-5.6-terra",
+      "fallbackModels": ["zai/glm-5-turbo"],
       "thinking": "low",
       "tools": ["read", "grep"]
     },
 
     "docs": {
       "description": "Use for documentation work: README/API docs review, docs gaps, changelog, migration notes, examples.",
-      "model": "zai/glm-5-turbo",
+      "model": "openai-codex/gpt-5.6-luna",
+      "fallbackModels": ["zai/glm-4.5-air"],
       "thinking": "low"
     },
 
     "frontend": {
       "description": "Use for frontend UI/UX visual work: styling, layout, typography, animation, responsive states, component polish, accessibility. Avoid backend/business logic unless needed for UI behavior.",
       "model": "antigravity/gemini-3-flash-preview",
-      "fallbackModels": ["openai-codex/gpt-5.4-mini", "zai/glm-5.3"],
+      "fallbackModels": ["openai-codex/gpt-5.6-luna", "zai/glm-5.3"],
       "thinking": "medium",
       "promptAppend": [
         "Act as a frontend UI/UX engineer for visual and product-facing work.",
@@ -199,15 +203,17 @@
 
     "browser-qa": {
       "description": "Use for browser-based visual QA: reproduce UI bugs and verify fixes with deterministic assertions, screenshots, video, and traces.",
-      "model": "openai-codex/gpt-5.4-mini",
+      "model": "openai-codex/gpt-5.6-luna",
       "fallbackModels": ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"],
-      "thinking": "medium",
+      "thinking": "low",
+      "timeoutMs": 120000,
       "tools": ["read", "grep", "bash"]
     },
 
     "tests": {
       "description": "Use for tests: locate coverage, find gaps, run/check targeted test commands, diagnose failing tests.",
-      "model": "zai/glm-5-turbo",
+      "model": "openai-codex/gpt-5.6-terra",
+      "fallbackModels": ["zai/glm-5-turbo"],
       "thinking": "medium",
       "tools": ["read", "grep", "bash"]
     },

package/external/pi-tools-suite/src/async-subagents/core/config.ts

@@ -167,7 +167,7 @@ export const DEFAULT_MAX_CONCURRENT = 5;
 export const DEFAULT_ROUTING_CONFIG: ResolvedSubagentRoutingConfig = {
 	enabled: true,
 	model: "zai/glm-4.5-air",
-	fallbackModels: ["openai-codex/gpt-5.3-codex-spark"],
+	fallbackModels: ["openai-codex/gpt-5.6-luna"],
 	maxTaskChars: 1200,
 	maxTokens: 512,
 	maxRetries: 3,
@@ -181,18 +181,26 @@ const BUILTIN_CONFIG: SubagentConfig = {
 	types: {
 		quick: {
 			description: "Use for tiny cheap tasks: answer a simple question, inspect one known file, or verify one fact. Not for broad repo search.",
+			model: "openai-codex/gpt-5.6-luna",
+			fallbackModels: ["zai/glm-4.5-air"],
 			thinking: "off",
 		},
 		scan: {
 			description: "Use for finding files, symbols, text, or inventory across a repo. Return paths/facts; do not judge code quality.",
+			model: "openai-codex/gpt-5.6-luna",
+			fallbackModels: ["zai/glm-4.5-air"],
 			thinking: "off",
 		},
 		research: {
 			description: "Use for multi-file codebase research: read several files and explain how something works. No edits.",
+			model: "openai-codex/gpt-5.6-terra",
+			fallbackModels: ["zai/glm-5-turbo"],
 			thinking: "low",
 		},
 		docs: {
 			description: "Use for documentation work: README/API docs review, docs gaps, changelog, migration notes, examples.",
+			model: "openai-codex/gpt-5.6-luna",
+			fallbackModels: ["zai/glm-4.5-air"],
 			thinking: "low",
 		},
 		frontend: {
@@ -207,9 +215,10 @@ const BUILTIN_CONFIG: SubagentConfig = {
 		},
 		"browser-qa": {
 			description: "Use for browser-based visual QA: reproduce UI bugs and verify fixes with deterministic assertions, screenshots, video, and traces.",
-			model: "openai-codex/gpt-5.4-mini",
+			model: "openai-codex/gpt-5.6-luna",
 			fallbackModels: ["antigravity/gemini-3-flash-preview", "zai/glm-5.3"],
-			thinking: "medium",
+			thinking: "low",
+			timeoutMs: 120_000,
 			tools: ["read", "grep", "bash"],
 			isolatedSkills: [getBrowserQaSkillPath()],
 		},
@@ -219,6 +228,8 @@ const BUILTIN_CONFIG: SubagentConfig = {
 		},
 		tests: {
 			description: "Use for tests: locate coverage, find gaps, run/check targeted test commands, diagnose failing tests.",
+			model: "openai-codex/gpt-5.6-terra",
+			fallbackModels: ["zai/glm-5-turbo"],
 			thinking: "medium",
 		},
 		review: {

package/external/pi-tools-suite/src/async-subagents/core/process.ts

@@ -13,3 +13,22 @@ export function terminateProcess(pid: number, signal: ProcessSignal): void {
 	}
 	process.kill(pid, signal);
 }
+
+/** Signal a launcher-owned process tree. POSIX callers must create the child as
+ * a process-group leader before using this helper. */
+export function terminateProcessTree(pid: number, signal: ProcessSignal): void {
+	if (process.platform === "win32") {
+		terminateProcess(pid, signal);
+		return;
+	}
+
+	try {
+		process.kill(-pid, signal);
+	} catch (error) {
+		const code = typeof error === "object" && error && "code" in error
+			? String((error as { code?: unknown }).code)
+			: undefined;
+		if (code !== "ESRCH") throw error;
+		// The owned group is already gone. Never fall back to a possibly reused PID.
+	}
+}

package/external/pi-tools-suite/src/async-subagents/core/spawn.ts

@@ -6,7 +6,7 @@ import { selectSuitableToolsForModel } from "../../lib/tool-args.js";
 import { validateBasename } from "./paths.js";
 import { getPiInvocation } from "./pi-invocation.js";
 import { writePromptFile } from "./prompt.js";
-import { terminateProcess } from "./process.js";
+import { terminateProcessTree } from "./process.js";
 import { getAgentSessionDir, SUBAGENT_PARENT_SESSION_FILE, SUBAGENT_RETURN_SESSION_FILE, SUBAGENT_SESSION_FILE, writeParentSessionLink, writeSessionFileLink } from "./sessions.js";
 import { getAgentState } from "./state.js";
 import { writeStructuredResult } from "./structured-result.js";
@@ -31,6 +31,7 @@ const AGENT_TIMEOUT_KILL_GRACE_MS = 5_000;
 const AGENT_SETTLED_TERMINATE_GRACE_MS = 50;
 const AGENT_SETTLED_COMPLETION_FALLBACK_MS = 1_000;
 const EXIT_STDIO_FLUSH_GRACE_MS = 10;
+const PROGRESS_LOG_MAX_BYTES = 1024 * 1024;
 
 export function shouldPersistSubagentSessions(env: NodeJS.ProcessEnv = process.env): boolean {
 	return isTruthyEnv(env.ASYNC_SUBAGENTS_ENABLE_SESSIONS);
@@ -136,11 +137,16 @@ export function spawnAgent(
 	const invocation = getPiInvocation(piArgs);
 	const stderrStream = createDeferredFileWriter(stderrFile, logLimits.stderrMaxBytes, "stderr.log");
 	const transcriptStream = createBoundedFileWriter(transcriptFile, logLimits.eventsMaxBytes, "events.jsonl");
+	const progressStream = createBoundedFileWriter(path.join(agentDir, "progress.jsonl"), PROGRESS_LOG_MAX_BYTES, "progress.jsonl");
+	const writeProgress = (stage: string, details: Record<string, unknown> = {}) => {
+		progressStream.write(serializeJsonLine({ at: isoNow(), stage, ...details }));
+	};
 
 	const proc = spawn(invocation.command, invocation.args, {
 		cwd,
 		env: subagentEnvironment(process.env, task.subagentType === "browser-qa" ? agentDir : undefined),
 		stdio: ["pipe", "pipe", "pipe"],
+		detached: process.platform !== "win32",
 	});
 	proc.stdin.on("error", (error: NodeJS.ErrnoException) => {
 		if (error.code === "EPIPE") return;
@@ -156,18 +162,33 @@ export function spawnAgent(
 	let timeoutKillTimer: NodeJS.Timeout | undefined;
 	let agentSettledKillTimer: NodeJS.Timeout | undefined;
 	let agentSettledCompletionFallbackTimer: NodeJS.Timeout | undefined;
+	let processExitTreeKillTimer: NodeJS.Timeout | undefined;
 	let exitFinalizationTimer: NodeJS.Timeout | undefined;
+	let processTermination: { code: number | null; signal: NodeJS.Signals | null } | undefined;
+	let stdoutEnded = false;
 	const suppressedRpcEventCounts = new Map<string, number>();
+	const scheduleProcessTreeKill = (reason: string) => {
+		if (processExitTreeKillTimer) return;
+		processExitTreeKillTimer = setTimeout(() => {
+			try {
+				writeProgress("shutdown_signal", { reason, signal: "SIGKILL" });
+				terminateChildProcessTree(proc, "SIGKILL");
+			} catch {
+				/* process group may already be gone */
+			}
+		}, AGENT_SETTLED_COMPLETION_FALLBACK_MS);
+		processExitTreeKillTimer.unref?.();
+	};
 
 	const notifyComplete = (exitCode: number) => {
 		if (completionNotified) return;
 		if (exitCode !== 0) shouldKeepStderr = true;
 		completionNotified = true;
 		if (timeoutTimer) clearTimeout(timeoutTimer);
-		if (timeoutKillTimer) clearTimeout(timeoutKillTimer);
 		if (agentSettledKillTimer) clearTimeout(agentSettledKillTimer);
-		if (agentSettledCompletionFallbackTimer) clearTimeout(agentSettledCompletionFallbackTimer);
 		if (exitFinalizationTimer) clearTimeout(exitFinalizationTimer);
+		writeProgress("completed", { exitCode });
+		progressStream.end();
 		if (!fs.existsSync(agentDir)) {
 			onComplete?.({
 				runDir,
@@ -207,6 +228,7 @@ export function spawnAgent(
 
 	const finalizeCompletion = (code: number | null, signal: NodeJS.Signals | null) => {
 		if (completionNotified) return;
+		if (!timeoutKillTimer && !agentSettledCompletionFallbackTimer) scheduleProcessTreeKill("process_exit");
 		writeSuppressedRpcEventSummary(transcriptStream, suppressedRpcEventCounts);
 		const exitCode = resolveAgentExitCode({
 			timedOut,
@@ -233,19 +255,21 @@ export function spawnAgent(
 		agentSettledKillTimer = setTimeout(() => {
 			agentSettledKillTimer = undefined;
 			try {
-				terminateChildProcess(proc, "SIGTERM");
+				writeProgress("shutdown_signal", { reason: "agent_settled", signal: "SIGTERM" });
+				terminateChildProcessTree(proc, "SIGTERM");
 			} catch {
 				/* process may have exited before the graceful termination timer fired */
 			}
 		}, AGENT_SETTLED_TERMINATE_GRACE_MS);
 		agentSettledKillTimer.unref?.();
 		agentSettledCompletionFallbackTimer = setTimeout(() => {
-			if (completionNotified) return;
 			try {
-				terminateChildProcess(proc, "SIGKILL");
+				writeProgress("shutdown_signal", { reason: "agent_settled", signal: "SIGKILL" });
+				terminateChildProcessTree(proc, "SIGKILL");
 			} catch {
 				/* process may already be gone */
 			}
+			if (completionNotified) return;
 			proc.stdin.destroy();
 			proc.stdout?.destroy();
 			proc.stderr?.destroy();
@@ -261,6 +285,7 @@ export function spawnAgent(
 			if (completionNotified) return;
 			timedOut = true;
 			const timeoutMessage = `Sub-agent timed out after ${Math.round(timeoutMs / 1000)} seconds.`;
+			writeProgress("timeout", { timeoutMs });
 			if (fs.existsSync(agentDir)) {
 				fs.writeFileSync(path.join(agentDir, "timeout_ms"), String(timeoutMs), "utf-8");
 				fs.writeFileSync(path.join(agentDir, "timed_out_at"), isoNow(), "utf-8");
@@ -269,14 +294,15 @@ export function spawnAgent(
 				stderrStream.write(`${timeoutMessage}\n`);
 			}
 			try {
-				terminateChildProcess(proc, "SIGTERM");
+				writeProgress("shutdown_signal", { reason: "timeout", signal: "SIGTERM" });
+				terminateChildProcessTree(proc, "SIGTERM");
 			} catch {
 				/* process may have exited between the timer and signal */
 			}
 			timeoutKillTimer = setTimeout(() => {
-				if (completionNotified) return;
 				try {
-					terminateChildProcess(proc, "SIGKILL");
+					writeProgress("shutdown_signal", { reason: "timeout", signal: "SIGKILL" });
+					terminateChildProcessTree(proc, "SIGKILL");
 				} catch {
 					/* process may have exited after SIGTERM */
 				}
@@ -297,6 +323,8 @@ export function spawnAgent(
 			const event = JSON.parse(line) as RpcEventRecord;
 			const storedEvent = compactRpcEventForTranscript(event, Buffer.byteLength(line, "utf8"));
 			if (storedEvent) transcriptStream.write(serializeJsonLine(storedEvent));
+			const progressEvent = compactRpcEventForProgress(event);
+			if (progressEvent) writeProgress("rpc_event", progressEvent);
 			onRpcEvent?.(event);
 			const sessionFile = extractSessionFileFromEvent(event);
 			if (sessionFile) writeSessionFileLink(agentDir, sessionFile);
@@ -307,8 +335,13 @@ export function spawnAgent(
 			if (event.type === "response" && event.command === "prompt" && event.success === false) {
 				const errorText = typeof event.error === "string" ? event.error : "RPC prompt failed";
 				fs.writeFileSync(path.join(agentDir, "result.md"), errorText, "utf-8");
+				try {
+					terminateChildProcessTree(proc, "SIGTERM");
+				} catch {
+					/* process may have exited immediately after emitting the failure */
+				}
+				scheduleProcessTreeKill("prompt_failed");
 				notifyComplete(1);
-				terminateChildProcess(proc, "SIGTERM");
 				return;
 			}
 			if (event.type === "agent_end") {
@@ -359,10 +392,29 @@ export function spawnAgent(
 		},
 	});
 
-	proc.once("exit", (code, signal) => {
-		exitFinalizationTimer = setTimeout(() => finalizeCompletion(code, signal), EXIT_STDIO_FLUSH_GRACE_MS);
+	// Bun on Windows may emit the ChildProcess `close` event before the stdout
+	// Readable has delivered its final buffered `data`/`end` events. Gate process
+	// finalization on stdout itself so a trailing RPC failure cannot be mistaken
+	// for a clean exit.
+	const finalizeAfterStdout = () => {
+		if (!processTermination || !stdoutEnded || completionNotified || exitFinalizationTimer) return;
+		const { code, signal } = processTermination;
+		exitFinalizationTimer = setTimeout(
+			() => finalizeCompletion(code, signal),
+			EXIT_STDIO_FLUSH_GRACE_MS,
+		);
 		exitFinalizationTimer.unref?.();
+	};
+	const recordProcessTermination = (code: number | null, signal: NodeJS.Signals | null) => {
+		processTermination ??= { code, signal };
+		finalizeAfterStdout();
+	};
+	proc.stdout.once("end", () => {
+		stdoutEnded = true;
+		finalizeAfterStdout();
 	});
+	proc.once("exit", recordProcessTermination);
+	proc.once("close", recordProcessTermination);
 
 	proc.once("error", (error) => {
 		const message = String(error);
@@ -376,6 +428,11 @@ export function spawnAgent(
 		notifyComplete(1);
 	});
 
+	const pid = proc.pid!;
+	fs.writeFileSync(path.join(agentDir, "pid"), String(pid), "utf-8");
+	if (process.platform !== "win32") fs.writeFileSync(path.join(agentDir, "process_group"), String(pid), "utf-8");
+	writeProgress("spawned", { pid });
+
 	proc.stdin.write([
 		serializeJsonLine({
 			id: "sub_get_state",
@@ -388,6 +445,7 @@ export function spawnAgent(
 			...(promptImages ? { images: promptImages } : {}),
 		}),
 	].join(""));
+	writeProgress("prompt_sent");
 	// Keep stdin open while the RPC prompt is running. pi RPC mode treats stdin
 	// EOF as a shutdown request, while the prompt command itself is handled
 	// asynchronously after preflight. Closing stdin here can therefore terminate
@@ -395,9 +453,6 @@ export function spawnAgent(
 	// producing exit 0 with no result.md. The child is terminated explicitly after agent_settled,
 	// timeout, stop, or process error.
 
-	const pid = proc.pid!;
-	fs.writeFileSync(path.join(agentDir, "pid"), String(pid), "utf-8");
-
 	return { pid, agentDir, process: proc };
 }
 
@@ -423,9 +478,9 @@ function getAntigravityAuthExtensionPath(): string {
 	return path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..", "..", "antigravity-auth", "index.ts");
 }
 
-function terminateChildProcess(proc: ChildProcess, signal: NodeJS.Signals): void {
+function terminateChildProcessTree(proc: ChildProcess, signal: NodeJS.Signals): void {
 	if (proc.pid) {
-		terminateProcess(proc.pid, signal as "SIGTERM" | "SIGINT" | "SIGKILL");
+		terminateProcessTree(proc.pid, signal as "SIGTERM" | "SIGINT" | "SIGKILL");
 		return;
 	}
 	proc.kill(signal);
@@ -603,6 +658,31 @@ function compactRpcEventForTranscript(event: RpcEventRecord, originalBytes: numb
 	return { type: event.type, bytes: originalBytes };
 }
 
+function compactRpcEventForProgress(event: RpcEventRecord): Record<string, unknown> | undefined {
+	if (event.type === "message_update" || event.type === "tool_execution_update") return undefined;
+	if (event.type === "response") {
+		return stripUndefined({
+			type: event.type,
+			command: typeof event.command === "string" ? event.command : undefined,
+			success: typeof event.success === "boolean" ? event.success : undefined,
+		});
+	}
+	if (event.type === "tool_execution_start" || event.type === "tool_execution_end") {
+		return stripUndefined({
+			type: event.type,
+			toolName: typeof event.toolName === "string" ? event.toolName : undefined,
+		});
+	}
+	if (event.type === "message_end") {
+		return stripUndefined({
+			type: event.type,
+			role: isRecord(event.message) && typeof event.message.role === "string" ? event.message.role : undefined,
+			stopReason: isRecord(event.message) && typeof event.message.stopReason === "string" ? event.message.stopReason : undefined,
+		});
+	}
+	return { type: event.type };
+}
+
 function suppressedRpcEventType(line: string): string | undefined {
 	if (line.includes('"type":"message_update"') || line.includes('"type": "message_update"')) return "message_update";
 	if (line.includes('"type":"tool_execution_update"') || line.includes('"type": "tool_execution_update"')) return "tool_execution_update";

package/external/pi-tools-suite/src/async-subagents/core/stop.ts

@@ -1,7 +1,7 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
 import { getAgentState, getRunState } from "./state.js";
-import { terminateProcess } from "./process.js";
+import { terminateProcess, terminateProcessTree } from "./process.js";
 import { writeStructuredResult } from "./structured-result.js";
 import type { AgentState } from "./types.js";
 import { isoNow } from "./utils.js";
@@ -70,7 +70,9 @@ function stopAgent(runDir: string, agent: AgentState, signal: StopSignal): StopA
 	}
 
 	try {
-		terminateProcess(agent.pid, signal);
+		const ownsProcessGroup = fs.existsSync(path.join(runDir, agent.id, "process_group"));
+		if (ownsProcessGroup) terminateProcessTree(agent.pid, signal);
+		else terminateProcess(agent.pid, signal);
 		markStopped(runDir, agent.id, signal);
 		return {
 			...result,