@bastani/atomic 0.8.26-alpha.1 → 0.8.26-alpha.11

package/dist/builtin/subagents/skills/browser/SKILL.md

@@ -0,0 +1,170 @@
+---
+name: browser
+description: Automate web browser interactions using natural language via CLI commands. Use when the user asks to browse websites, navigate web pages, extract data from websites, take screenshots, fill forms, click buttons, or interact with web applications. Supports remote Browserbase sessions with Browserbase Identity, Verified browsers, automatic CAPTCHA solving, and residential proxies — ideal for protected websites and JavaScript-heavy pages.
+compatibility: "Requires the browse CLI (`npm install -g browse`). Remote Browserbase sessions need `BROWSERBASE_API_KEY`. Local mode uses Chrome/Chromium on your machine."
+license: MIT
+allowed-tools: Bash
+metadata:
+  openclaw:
+    requires:
+      bins:
+        - browse
+    install:
+      - kind: node
+        package: "browse"
+        bins: [browse]
+    homepage: https://github.com/browserbase/skills
+---
+
+# Browser Automation
+
+Automate browser interactions using the browse CLI with Claude.
+
+## Setup check
+
+Before running any browser commands, verify the CLI is available:
+
+```bash
+which browse || npm install -g browse
+```
+
+## Environment Selection (Local vs Remote)
+
+The CLI supports explicit per-command environment flags. If you do nothing, the next session defaults to Browserbase when `BROWSERBASE_API_KEY` is set and to local otherwise.
+
+### Local mode
+- `browse open <url> --local` starts a clean isolated local browser
+- `browse open <url> --auto-connect` attaches to an already-running debuggable Chrome; use `--local` when no debuggable Chrome is available
+- `browse open <url> --cdp <port|url>` attaches to a specific CDP target
+- Best for: development, localhost, trusted sites, and reproducible runs
+
+### Remote mode (Browserbase)
+- `browse open <url> --remote` starts a Browserbase session
+- Without a local flag, Browserbase is also the default when `BROWSERBASE_API_KEY` is set
+- Provides: Browserbase Identity, Verified browsers, automatic CAPTCHA solving, residential proxies, session persistence
+- **Use remote mode when:** the target site has bot detection, CAPTCHAs, IP rate limiting, Cloudflare protection, or requires geo-specific access
+- Get credentials at https://browserbase.com/settings
+
+### When to choose which
+- **Repeatable local testing / clean state**: `browse open <url> --local`
+- **Reuse your local login/cookies**: `browse open <url> --auto-connect`
+- **Simple browsing** (docs, wikis, public APIs): local mode is fine
+- **Protected sites** (login walls, CAPTCHAs, anti-scraping): use remote mode
+- **If local mode fails** with bot detection or access denied: switch to remote mode
+
+## Commands
+
+Most driver commands work across local, remote, and CDP sessions after the daemon starts.
+
+### Navigation
+```bash
+browse open <url>                        # Go to URL
+browse open <url> --local                # Go to URL in a clean local browser
+browse open <url> --remote               # Go to URL in a Browserbase session
+browse reload                            # Reload current page
+browse back                              # Go back in history
+browse forward                           # Go forward in history
+```
+
+### Page state (prefer snapshot over screenshot)
+```bash
+browse snapshot                          # Get accessibility tree with element refs (fast, structured)
+browse screenshot --path <path>          # Take visual screenshot (slow, uses vision tokens)
+browse get url                           # Get current URL
+browse get title                         # Get page title
+browse get text <selector>               # Get text content (use "body" for all text)
+browse get html <selector>               # Get HTML content of element
+browse get value <selector>              # Get form field value
+```
+
+Use `browse snapshot` as your default for understanding page state — it returns the accessibility tree with element refs you can use to interact. Only use `browse screenshot` when you need visual context (layout, images, debugging).
+
+### Interaction
+```bash
+browse click <ref>                       # Click element by ref from snapshot (e.g., @0-5)
+browse type <text>                       # Type text into focused element
+browse fill <selector> <value>           # Fill input; add --press-enter if Enter is needed
+browse select <selector> <values...>     # Select dropdown option(s)
+browse press <key>                       # Press key (Enter, Tab, Escape, Cmd+A, etc.)
+browse mouse drag <fromX> <fromY> <toX> <toY>  # Drag from one point to another
+browse mouse scroll <x> <y> <deltaX> <deltaY>  # Scroll at coordinates
+browse highlight <selector>              # Highlight element on page
+browse is visible <selector>             # Check if element is visible
+browse is checked <selector>             # Check if element is checked
+browse wait <type> [arg]                 # Wait for: load, selector, timeout
+```
+
+### Session management
+```bash
+browse stop                              # Stop the browser daemon
+browse status                            # Check daemon status and resolved mode
+browse tab list                          # List all open tabs
+browse tab switch <index-or-target-id>   # Switch to tab by index or target ID
+browse tab close [index-or-target-id]    # Close tab
+```
+
+### Typical workflow
+If the environment matters, put `--local`, `--remote`, `--auto-connect`, or `--cdp <port|url>` on the first browser command.
+
+1. `browse open <url> --local` or `browse open <url> --remote` — navigate to the page
+2. `browse snapshot` — read the accessibility tree to understand page structure and get element refs
+3. `browse click <ref>` / `browse type <text>` / `browse fill <selector> <value>` — interact using refs from snapshot
+4. `browse snapshot` — confirm the action worked
+5. Repeat 3-4 as needed
+6. `browse stop` — close the browser when done
+
+## Quick Example
+
+```bash
+browse open https://example.com
+browse snapshot                          # see page structure + element refs
+browse click @0-5                        # click element with ref 0-5
+browse get title
+browse stop
+```
+
+## Mode Comparison
+
+| Feature | Local | Browserbase |
+|---------|-------|-------------|
+| Speed | Faster | Slightly slower |
+| Setup | Chrome required | API key required |
+| Reuse existing local cookies | With `browse open <url> --auto-connect` | N/A |
+| Verified browser | No | Yes (Browserbase Verified browser via Identity) |
+| CAPTCHA solving | No | Yes (automatic reCAPTCHA/hCaptcha) |
+| Residential proxies | No | Yes (201 countries, geo-targeting) |
+| Session persistence | No | Yes (cookies/auth persist via contexts) |
+| Best for | Development/simple pages | Protected sites, Browserbase Identity + Verified access, production scraping |
+
+## Best Practices
+
+1. **Choose the local strategy deliberately**: use `browse open <url> --local` for clean state, `browse open <url> --auto-connect` for existing local credentials, and `browse open <url> --remote` for protected sites
+2. **Always `browse open` first** before interacting
+3. **Use `browse snapshot`** to check page state — it's fast and gives you element refs
+4. **Only screenshot when visual context is needed** (layout checks, images, debugging)
+5. **Use refs from snapshot** to click/interact — e.g., `browse click @0-5`
+6. **`browse stop`** when done to clean up the browser session and clear the env override
+
+## Troubleshooting
+
+- **"No active page"**: Run `browse stop`, then check `browse status`. If it still says running, kill the zombie daemon with `pkill -f "browse.*daemon"`, then retry `browse open`
+- **Chrome not found**: Install Chrome, use `browse open <url> --auto-connect` if you already have a debuggable Chrome running, or switch to `browse open <url> --remote`
+- **Action fails**: Run `browse snapshot` to see available elements and their refs
+- **Browserbase fails**: Verify API key is set
+
+## Switching to Remote Mode
+
+Switch to remote when you detect: CAPTCHAs (reCAPTCHA, hCaptcha, Turnstile), bot detection pages ("Checking your browser..."), HTTP 403/429, empty pages on sites that should have content, or the user asks for it.
+
+Don't switch for simple sites (docs, wikis, public APIs, localhost).
+
+```bash
+browse open <url> --local          # clean isolated local browser
+browse open <url> --auto-connect   # attach to existing debuggable Chrome
+browse open <url> --remote         # Browserbase session
+```
+
+Mode flags are applied when a session starts. After `browse stop`, the next start falls back to env-var-based auto detection. Use `browse status` to inspect the resolved mode and target while the daemon is running.
+
+For detailed examples, see [EXAMPLES.md](EXAMPLES.md).
+For API reference, see [REFERENCE.md](REFERENCE.md).

package/dist/builtin/subagents/skills/subagent/SKILL.md

@@ -20,7 +20,7 @@ Use this skill when the parent orchestrator needs to launch a specialized subage
 - **Parallel codebase discovery**: combine `codebase-locator`, `codebase-analyzer`, and `codebase-pattern-finder` to map where code lives, how it works, and what existing conventions look like — concurrently, with fresh context per child.
 - **Local research mining**: pair `codebase-research-locator` with `codebase-research-analyzer` to surface prior decisions in `research/` and `specs/` and extract what still applies.
 - **External research**: use `codebase-online-researcher` for authoritative web sources, with persisted findings in `research/web/`.
-- **Debug and fix**: use `debugger` to reproduce, diagnose, and patch failing behavior with `tdd` and `browser-use` support.
+- **Debug and fix**: use `debugger` to reproduce, diagnose, and patch failing behavior with `tdd` and `browser` support.
 - **Refinement**: use `code-simplifier` to clean up recently changed code without altering behavior.
 - **Adversarial review**: compose read-only specialists (`codebase-analyzer`, `codebase-pattern-finder`, `debugger` in inspect-only mode, `codebase-online-researcher`) into a parallel review pass — there is no generic `reviewer` agent.
 - **Long-running work**: launch async/background runs and inspect them later.
@@ -158,9 +158,9 @@ Builtin agents load at the lowest priority. Project agents override user agents,
 | `codebase-pattern-finder`    | Find similar implementations or conventions                       | `openai/gpt-5.4-mini` | low      | read, grep, find, ls, bash                                                             | Read-only. Returns code snippets with `file:line` references.                                              |
 | `codebase-research-locator`  | Discover prior `research/` and `specs/` docs                      | `openai/gpt-5.4-mini` | low      | read, grep, find, ls, bash                                                             | Read-only. Sorts by date, tiers by recency, flags supersession.                                            |
 | `codebase-research-analyzer` | Extract decisions and constraints from prior docs                 | `openai/gpt-5.5`      | low      | read, grep, find, ls, bash                                                             | Read-only. Filters aggressively for what still applies today.                                              |
-| `codebase-online-researcher` | Web research with authoritative sources                           | `openai/gpt-5.5`      | low      | read, grep, find, ls, bash, write, web_search, fetch_content, get_search_content       | Has the `browser-use` skill. Persists keepers to `research/web/`.                                       |
+| `codebase-online-researcher` | Web research with authoritative sources                           | `openai/gpt-5.5`      | low      | read, grep, find, ls, bash, write, web_search, fetch_content, get_search_content       | Has the `browser` skill. Persists keepers to `research/web/`.                                       |
 | `code-simplifier`            | Clean up recently changed code without changing behavior          | `openai/gpt-5.5`      | low      | read, edit, write, grep, find, ls, bash                                                | **Writer.** Scopes to recently modified code by default; preserves all observable behavior.                |
-| `debugger`                   | Reproduce, diagnose, and fix failing behavior                     | `openai/gpt-5.5`      | high     | read, edit, write, grep, find, ls, bash, web_search, fetch_content, get_search_content | **Writer.** Has the `tdd` and `browser-use` skills. Inspect-only mode requires an explicit instruction. |
+| `debugger`                   | Reproduce, diagnose, and fix failing behavior                     | `openai/gpt-5.5`      | high     | read, edit, write, grep, find, ls, bash, web_search, fetch_content, get_search_content | **Writer.** Has the `tdd` and `browser` skills. Inspect-only mode requires an explicit instruction. |
 
 Each builtin declares an explicit `model` and `fallbackModels` chain (typically `github-copilot/<same>`, then `anthropic/claude-opus-4-8`, then `github-copilot/claude-opus-4.7`). The current user-selected model is automatically appended as the last fallback and de-duplicated. Override per run with inline config:
 
@@ -186,7 +186,7 @@ A strong subagent prompt usually includes:
 - **Output**: the expected summary shape, artifact path, or finding format.
 - **Stop rules**: when to stop after enough evidence, and when not to keep searching.
 
-Avoid carrying over old prompt habits that over-specify every step. Use `must`, `always`, and `never` for real invariants; for judgment calls, give decision rules. For example, tell `codebase-analyzer` to trace the staged diff directly and report only evidence-backed findings, rather than prescribing every file or command. Tell `codebase-online-researcher` the retrieval budget: start with broad targeted searches, fetch the strongest sources via `fetch_content`, fall back to `browser-use` only when JS execution is required, and stop when the question is answered.
+Avoid carrying over old prompt habits that over-specify every step. Use `must`, `always`, and `never` for real invariants; for judgment calls, give decision rules. For example, tell `codebase-analyzer` to trace the staged diff directly and report only evidence-backed findings, rather than prescribing every file or command. Tell `codebase-online-researcher` the retrieval budget: start with broad targeted searches, fetch the strongest sources via `fetch_content`, fall back to `browser` only when JS execution is required, and stop when the question is answered.
 
 For implementation handoffs to `debugger` or `code-simplifier`, name the approved scope and success criteria more clearly than the process. Good prompts say what to change, what not to change, where the evidence lives, how to validate, and when to escalate. They should not ask the child to create another subagent plan or continue the parent conversation.
 

package/dist/builtin/subagents/src/runs/background/subagent-runner.ts

@@ -48,7 +48,12 @@ import { outputEntryFromAsyncResult, resolveOutputReferences } from "../shared/c
 import { createStructuredOutputRuntime, readStructuredOutput } from "../shared/structured-output.ts";
 import { collectDynamicResults, DynamicFanoutError, materializeDynamicParallelStep, validateDynamicCollection } from "../shared/dynamic-fanout.ts";
 import { nestedSummaryFromAsyncStatus, writeNestedEvent } from "../shared/nested-events.ts";
-import { formatModelAttemptNote, isRetryableModelFailure } from "../shared/model-fallback.ts";
+import { formatModelAttemptNote, isRetryableModelFailure, modelFailureMessage } from "../shared/model-fallback.ts";
+import {
+	assistantStopReason,
+	isAssistantFailureStopReason,
+	shouldStartSubagentFinalDrain,
+} from "../shared/final-drain.ts";
 import { attachPostExitStdioGuard, trySignalChild } from "../../shared/post-exit-stdio-guard.ts";
 import { detectSubagentError, extractTextFromContent, extractToolArgsPreview, getFinalOutput } from "../../shared/utils.ts";
 import { evaluateCompletionMutationGuard } from "../shared/completion-guard.ts";
@@ -230,6 +235,7 @@ interface RunPiStreamingResult {
 	finalOutput: string;
 	interrupted?: boolean;
 	observedMutationAttempt?: boolean;
+	modelFailureSignal?: unknown;
 }
 
 function runPiStreaming(
@@ -270,6 +276,7 @@ function runPiStreaming(
 		let model: string | undefined;
 		let error: string | undefined;
 		let assistantError: string | undefined;
+		let assistantFailureSignal: unknown;
 		let interrupted = false;
 		let observedMutationAttempt = false;
 		const rawStdoutLines: string[] = [];
@@ -330,7 +337,6 @@ function runPiStreaming(
 
 				if (event.type !== "message_end" || event.message.role !== "assistant") return;
 				if (event.message.model) model = event.message.model;
-				if (event.message.errorMessage) assistantError = event.message.errorMessage;
 				const eventUsage = event.message.usage;
 				if (eventUsage) {
 					usage.turns++;
@@ -340,12 +346,21 @@ function runPiStreaming(
 					usage.cacheWrite += eventUsage.cacheWrite ?? 0;
 					usage.cost += eventUsage.cost?.total ?? 0;
 				}
-				const stopReason = (event.message as { stopReason?: string }).stopReason;
-				const hasToolCall = Array.isArray(event.message.content)
-					&& event.message.content.some((part) => (part as { type?: string }).type === "toolCall");
-				if (stopReason === "stop" && !hasToolCall) {
-					if (!event.message.errorMessage && extractTextFromContent(event.message.content).trim()) assistantError = undefined;
-					cleanTerminalAssistantStopReceived ||= !event.message.errorMessage;
+				const stopReason = assistantStopReason(event.message);
+				if (event.message.errorMessage) {
+					assistantError = event.message.errorMessage;
+					assistantFailureSignal = event.message;
+				}
+				if (isAssistantFailureStopReason(stopReason)) {
+					assistantError = modelFailureMessage(event.message);
+					assistantFailureSignal = event.message;
+				}
+				if (shouldStartSubagentFinalDrain(event.message)) {
+					if (extractTextFromContent(event.message.content).trim()) {
+						assistantError = undefined;
+						assistantFailureSignal = undefined;
+					}
+					cleanTerminalAssistantStopReceived = true;
 					startFinalDrain();
 				}
 			}
@@ -448,6 +463,9 @@ function runPiStreaming(
 				finalOutput,
 				interrupted,
 				observedMutationAttempt,
+				...(assistantFailureSignal !== undefined && finalError === assistantError
+					? { modelFailureSignal: assistantFailureSignal }
+					: {}),
 			});
 		});
 
@@ -459,7 +477,20 @@ function runPiStreaming(
 			outputStream.end();
 			const finalOutput = getFinalOutput(messages) || rawStdoutLines.join("\n").trim();
 			const spawnErrorMessage = spawnError instanceof Error ? spawnError.message : String(spawnError);
-			resolve({ stderr, exitCode: 1, messages, usage, model, error: error ?? assistantError ?? spawnErrorMessage, finalOutput, observedMutationAttempt });
+			const finalError = error ?? assistantError ?? spawnErrorMessage;
+			resolve({
+				stderr,
+				exitCode: 1,
+				messages,
+				usage,
+				model,
+				error: finalError,
+				finalOutput,
+				observedMutationAttempt,
+				...(assistantFailureSignal !== undefined && finalError === assistantError
+					? { modelFailureSignal: assistantFailureSignal }
+					: {}),
+			});
 		});
 	});
 }
@@ -650,6 +681,7 @@ async function runSingleStep(
 	const attemptedModels: string[] = [];
 	const modelAttempts: ModelAttempt[] = [];
 	const attemptNotes: string[] = [];
+	const pendingAttemptNotes: string[] = [];
 	const eventsPath = path.join(path.dirname(ctx.outputFile), "events.jsonl");
 	let finalResult: RunPiStreamingResult | undefined;
 	let finalFastMode: boolean | undefined;
@@ -768,9 +800,20 @@ async function runSingleStep(
 		finalFastMode = attemptFastMode;
 		finalOutputSnapshot = outputSnapshot;
 		finalResult = { ...run, exitCode: effectiveExitCode, model: candidate ?? run.model, error, structuredOutput } as RunPiStreamingResult & { structuredOutput?: unknown };
-		if (attempt.success || completionGuardTriggered) break;
-		if (!isRetryableModelFailure(error) || index === candidates.length - 1) break;
-		attemptNotes.push(formatModelAttemptNote(attempt, candidates[index + 1]));
+		if (attempt.success) break;
+		const retrySignal = run.modelFailureSignal ?? error;
+		if (
+			!completionGuardTriggered
+			&& structuredError === undefined
+			&& hiddenError?.hasError !== true
+			&& isRetryableModelFailure(retrySignal)
+			&& index < candidates.length - 1
+		) {
+			pendingAttemptNotes.push(formatModelAttemptNote(attempt, candidates[index + 1]));
+			continue;
+		}
+		attemptNotes.push(...pendingAttemptNotes);
+		break;
 	}
 
 	const rawOutput = finalResult?.finalOutput ?? "";

package/dist/builtin/subagents/src/runs/foreground/execution.ts

@@ -17,6 +17,7 @@ import {
 	type AgentProgress,
 	type ArtifactPaths,
 	type ControlEvent,
+	type Details,
 	type ModelAttempt,
 	type RunSyncOptions,
 	type SingleResult,
@@ -53,7 +54,13 @@ import {
 	buildModelCandidates,
 	formatModelAttemptNote,
 	isRetryableModelFailure,
+	modelFailureMessage,
 } from "../shared/model-fallback.ts";
+import {
+	assistantStopReason,
+	isAssistantFailureStopReason,
+	shouldStartSubagentFinalDrain,
+} from "../shared/final-drain.ts";
 import {
 	createMutatingFailureState,
 	didMutatingToolFail,
@@ -75,6 +82,7 @@ import { acceptanceFailureMessage, evaluateAcceptance, formatAcceptancePrompt, r
 
 const artifactOutputByResult = new WeakMap<SingleResult, string>();
 const acceptanceOutputByResult = new WeakMap<SingleResult, string>();
+const modelFailureSignalByResult = new WeakMap<SingleResult, unknown>();
 
 function emptyUsage(): Usage {
 	return { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, cost: 0, turns: 0 };
@@ -139,6 +147,29 @@ function snapshotResult(result: SingleResult, progress: AgentProgress): SingleRe
 	};
 }
 
+type RunSyncUpdate = import("@earendil-works/pi-agent-core").AgentToolResult<Details>;
+
+function extractUpdateText(update: RunSyncUpdate): string | undefined {
+	const text = update.content
+		.map((item) => item.type === "text" ? item.text : undefined)
+		.filter((item): item is string => Boolean(item?.trim()))
+		.join("\n");
+	return text || undefined;
+}
+
+export function shouldSuppressIntermediateRetryableFailureUpdate(update: RunSyncUpdate): boolean {
+	const result = update.details?.results?.[0];
+	if (!result) return false;
+	const progress = update.details?.progress?.[0];
+	const status = result.progress?.status ?? progress?.status;
+	if (status !== "failed") return false;
+	const failureText = result.error
+		?? result.progress?.error
+		?? progress?.error
+		?? extractUpdateText(update);
+	return isRetryableModelFailure(failureText);
+}
+
 async function runSingleAttempt(
 	runtimeCwd: string,
 	agent: AgentConfig,
@@ -269,6 +300,7 @@ async function runSingleAttempt(
 		let detached = false;
 		let intercomStarted = false;
 		let assistantError: string | undefined;
+		let assistantFailureSignal: unknown;
 		let removeAbortListener: (() => void) | undefined;
 		let removeInterruptListener: (() => void) | undefined;
 		let activityTimer: NodeJS.Timeout | undefined;
@@ -520,16 +552,25 @@ async function runSingleAttempt(
 						progress.tokens = result.usage.input + result.usage.output;
 					}
 					if (!result.model && evt.message.model) result.model = evt.message.model;
-					if (evt.message.errorMessage) assistantError = evt.message.errorMessage;
 					const assistantText = extractTextFromContent(evt.message.content);
 					appendRecentOutput(progress, assistantText.split("\n").slice(-10));
-					// Final assistant message: start the exit drain window.
-					const stopReason = (evt.message as { stopReason?: string }).stopReason;
-					const hasToolCall = Array.isArray(evt.message.content)
-						&& evt.message.content.some((part) => (part as { type?: string }).type === "toolCall");
-					if (stopReason === "stop" && !hasToolCall) {
-						if (!evt.message.errorMessage && assistantText.trim()) assistantError = undefined;
-						cleanTerminalAssistantStopReceived ||= !evt.message.errorMessage;
+					// Clean final assistant stops start the exit drain window; provider error/aborted
+					// stop reasons remain failure evidence so pi-ai can auto-retry before exit.
+					const stopReason = assistantStopReason(evt.message);
+					if (evt.message.errorMessage) {
+						assistantError = evt.message.errorMessage;
+						assistantFailureSignal = evt.message;
+					}
+					if (isAssistantFailureStopReason(stopReason)) {
+						assistantError = modelFailureMessage(evt.message);
+						assistantFailureSignal = evt.message;
+					}
+					if (shouldStartSubagentFinalDrain(evt.message)) {
+						if (assistantText.trim()) {
+							assistantError = undefined;
+							assistantFailureSignal = undefined;
+						}
+						cleanTerminalAssistantStopReceived = true;
 						startFinalDrain();
 					}
 				}
@@ -609,6 +650,9 @@ async function runSingleAttempt(
 			processClosed = true;
 			if (buf.trim()) processLine(buf);
 			if (!result.error && assistantError) result.error = assistantError;
+			if (assistantFailureSignal !== undefined && result.error === assistantError) {
+				modelFailureSignalByResult.set(result, assistantFailureSignal);
+			}
 			const forcedDrainAfterFinalSuccess = forcedTerminationSignal && cleanTerminalAssistantStopReceived && !result.error;
 			if (code !== 0 && stderrBuf.trim() && !result.error && !forcedDrainAfterFinalSuccess) {
 				result.error = stderrBuf.trim();
@@ -875,6 +919,7 @@ export async function runSync(
 	const modelAttempts: ModelAttempt[] = [];
 	const aggregateUsage = emptyUsage();
 	const attemptNotes: string[] = [];
+	const pendingAttemptNotes: string[] = [];
 	let totalToolCount = 0;
 	let totalDurationMs = 0;
 
@@ -897,7 +942,18 @@ export async function runSync(
 		const candidate = modelsToTry[i];
 		if (candidate) attemptedModels.push(candidate);
 		const outputSnapshot = captureSingleOutputSnapshot(options.outputPath);
-		const result = await runSingleAttempt(runtimeCwd, agent, taskWithAcceptance, candidate, options, {
+		let attemptOptions = options;
+		if (i < modelsToTry.length - 1 && options.onUpdate) {
+			const forwardUpdate = options.onUpdate;
+			attemptOptions = {
+				...options,
+				onUpdate: (update) => {
+					if (shouldSuppressIntermediateRetryableFailureUpdate(update)) return;
+					forwardUpdate(update);
+				},
+			};
+		}
+		const result = await runSingleAttempt(runtimeCwd, agent, taskWithAcceptance, candidate, attemptOptions, {
 			sessionEnabled,
 			systemPrompt,
 			resolvedSkillNames: resolvedSkills.length > 0 ? resolvedSkills.map((skill) => skill.name) : undefined,
@@ -928,10 +984,13 @@ export async function runSync(
 		if (attemptSucceeded) {
 			break;
 		}
-		if (!isRetryableModelFailure(result.error) || i === modelsToTry.length - 1) {
-			break;
+		const retrySignal = modelFailureSignalByResult.get(result) ?? result.error;
+		if (isRetryableModelFailure(retrySignal) && i < modelsToTry.length - 1) {
+			pendingAttemptNotes.push(formatModelAttemptNote(attempt, modelsToTry[i + 1]));
+			continue;
 		}
-		attemptNotes.push(formatModelAttemptNote(attempt, modelsToTry[i + 1]));
+		attemptNotes.push(...pendingAttemptNotes);
+		break;
 	}
 
 	const result = lastResult ?? {

package/dist/builtin/subagents/src/runs/shared/acceptance.ts

@@ -1,6 +1,7 @@
 import { spawn } from "node:child_process";
 import { spawnSync } from "node:child_process";
 import * as path from "node:path";
+import { createGitEnvironment } from "@bastani/atomic";
 import type {
 	AcceptanceConfig,
 	AcceptanceEvidenceKind,
@@ -399,7 +400,7 @@ function reportEvidencePresent(report: AcceptanceReport, kind: AcceptanceEvidenc
 }
 
 function checkNoStagedFiles(cwd: string): AcceptanceRuntimeCheck {
-	const result = spawnSync("git", ["status", "--short"], { cwd, encoding: "utf-8" });
+	const result = spawnSync("git", ["status", "--short"], { cwd, env: createGitEnvironment(), encoding: "utf-8" });
 	if (result.status !== 0) {
 		return { id: "no-staged-files", status: "not-applicable", message: "git status unavailable; no staged-files check skipped" };
 	}

package/dist/builtin/subagents/src/runs/shared/final-drain.ts

@@ -0,0 +1,34 @@
+export interface SubagentAssistantDrainMessage {
+	readonly role?: unknown;
+	readonly content?: unknown;
+	readonly stopReason?: unknown;
+	readonly errorMessage?: unknown;
+}
+
+export function assistantStopReason(message: SubagentAssistantDrainMessage): string | undefined {
+	return typeof message.stopReason === "string" ? message.stopReason : undefined;
+}
+
+export function isAssistantFailureStopReason(stopReason: string | undefined): boolean {
+	return stopReason === "error" || stopReason === "aborted";
+}
+
+export function assistantMessageHasToolCall(message: SubagentAssistantDrainMessage): boolean {
+	return Array.isArray(message.content)
+		&& message.content.some((part) => part !== null
+			&& typeof part === "object"
+			&& (part as { readonly type?: unknown }).type === "toolCall");
+}
+
+function assistantMessageHasError(message: SubagentAssistantDrainMessage): boolean {
+	const errorMessage = message.errorMessage;
+	if (typeof errorMessage === "string") return errorMessage.trim().length > 0;
+	return errorMessage !== undefined && errorMessage !== null;
+}
+
+export function shouldStartSubagentFinalDrain(message: SubagentAssistantDrainMessage): boolean {
+	if (message.role !== undefined && message.role !== "assistant") return false;
+	return assistantStopReason(message) === "stop"
+		&& !assistantMessageHasError(message)
+		&& !assistantMessageHasToolCall(message);
+}