fix(browser): fail fast when a new turn offers Retry

[developerisnow]

Fixes #457.

When the submitted ChatGPT turn displays a known English generation failure with a visible Retry control, Oracle returns a structured, body-free browser warning promptly. It preserves manual recovery and never clicks Retry or submits the prompt again. Unknown/localized failure states retain the configured timeout.

The maintainer repair keeps waiting while generation is active, ignores transparent controls, uses the outer conversation turn for busy/progress evidence, stops the pending response observer on failure, and applies the error guard to image output and response-recovery paths. The changelog entry is collected in the final maintenance-notes PR to keep sibling branches independent.

Validation:

• Typecheck, lint, formatting, docs/help consistency, and build pass.
• Complete suite: 1,999 passed, 43 skipped with four workers. An earlier default-concurrency MCP startup timeout passed when isolated and in the complete four-worker run.
• Real Chrome: node scripts/assistant-retry-proof.mjs, opened in an existing Chrome profile, passed nine controlled DOM cases. The production reader/error guard emitted chatgpt-ui-warning; older turns, hidden/transparent Retry, unrelated prose, and active generation did not fail. Zero Retry clicks.
• Built CLI: render/preview and a real OpenAI request completed successfully.
• Independent autoreview through P2 is clean for the final candidate.

The Chrome fixtures are synthetic and do not claim a newly observed live ChatGPT outage. A signed-in Oracle CDP smoke remains blocked by Chrome's separate remote-debugging approval; this is also an outstanding release gate.

Full built-CLI/CDP proof now passes for text and image failure, including persisted recovery metadata and retained Chrome: #458 (comment). The separate signed-in Pro release smoke remains outstanding.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.

ClawSweeper review complete

ClawSweeper finished reviewing this revision. The review result is being finalized.

View the workflow run.

[clawsweeper]

Codex review: blocked before merge. Reviewed September 7, 2026, 5:04 AM ET / 09:04 UTC.

ClawSweeper review

What this changes

Oracle detects stopped ChatGPT turns with known failure text and a visible Retry button, returns a structured warning, and preserves eligible browsers for manual recovery.

Merge readiness

⛔ Blocked before merge - 1 item remains

The fix remains useful, and the owner’s new built-CLI Chrome/CDP evidence resolves the previous proof blocker. No blocking patch defect was found; the signed-in Pro smoke remains a separate release gate.

Priority: P2
Reviewed head: 0ee560837f158fe6af57bef4e98e806d28e331d7

Review scores

Measure	Result	What it means
Overall readiness	🐚 platinum hermit (4/6)	A focused, maintainable fix with production-path proof resolving the prior gap and no blocking findings.
Proof confidence	🐚 platinum hermit (4/6)	Sufficient (live_output): The owner’s exact-head Linux Chrome/CDP report exercises built-CLI submission, text and image polling, error persistence, and browser retention through the production owners. Observed 38 ms/35 ms failures, one send, zero Retry clicks, saved recovery metadata, and reachable retained targets satisfy the previous proof request; the controlled page does not claim a live provider outage.
Patch quality	🐚 platinum hermit (4/6)	No actionable review findings were identified.

Verification

Check	Result	Evidence
Real behavior	Verified	Sufficient (live_output): The owner’s exact-head Linux Chrome/CDP report exercises built-CLI submission, text and image polling, error persistence, and browser retention through the production owners. Observed 38 ms/35 ms failures, one send, zero Retry clicks, saved recovery metadata, and reachable retained targets satisfy the previous proof request; the controlled page does not claim a live provider outage.
Evidence reviewed	7 items	Repository policy and review scope: Verified the origin repository and read the full root AGENTS.md. Only the root AGENTS.md was found; .agents/maintainer-notes is absent. Applied the browser recovery, Pro waiting, and release-smoke guidance. The repository profile prohibits automatic cleanup closure. Complete introduced patch reviewed: Read all 13 files in the pinned main-to-head diff, including the complete proof script and tests omitted from the prompt’s abbreviated patches. Runtime changes total +79/-9 lines, tests +130/-0, the proof script +109/-0, and documentation +1/-0. Scoped failure detection and propagation: The collector requires known failure text, a visible Retry control, and absence of strong generation activity. Existing snapshot filtering rejects older turns; the shared error guard reaches polling, image collection, and recovery paths. The warning contains fixed text rather than assistant content.
Findings	None	None.
Security	None	None.

How this fits together

Oracle’s browser engine submits prompts to ChatGPT and reads responses through Chrome’s debugging connection. Its response collector feeds text, image results, or structured errors into CLI output and saved session metadata.

flowchart TD
  A[Submitted prompt] --> B[ChatGPT assistant turn]
  B --> C[Response collector]
  C --> D{Visible Retry and known failure with generation stopped?}
  D -->|No| E[Continue normal response capture]
  D -->|Yes| F[Structured browser warning]
  F --> G[Save recovery metadata and retain eligible browser]

Loading

Before merge

• Resolve merge risk (P1) - Controlled Chrome proof establishes the failure and recovery path, but does not establish current signed-in ChatGPT Pro behavior; the owner explicitly retains that smoke as a release gate.

Agent review details

Security

None.

Review metrics

Metric	Value	Why it matters
Production and validation growth	Production +79/-9; tests +130/-0; proof script +109/-0	Runtime growth supports shared detection and propagation across existing capture paths, with larger supplemental validation coverage.

Root-cause cluster

Relationship: fixed_by_candidate
Canonical: #457
Summary: This PR is the explicit candidate fix for the open paired timeout report.

Members:

• canonical: Browser mode waits for the full timeout after ChatGPT offers Retry #457 - The report describes the same stopped-generation Retry state and requests prompt failure with retained manual recovery.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge-risk options

Maintainer options:

1. Decide the mitigation before merge
   Use the shared turn-scoped error guard and existing recovery lifecycle, preserving normal generation waits and completing the separately tracked signed-in smoke before release.
2. Pause or close
   Do not merge this PR until maintainers decide whether the risk is worth taking.

Technical review

Best possible solution:

Use the shared turn-scoped error guard and existing recovery lifecycle, preserving normal generation waits and completing the separately tracked signed-in smoke before release.

Do we have a high-confidence way to reproduce the issue?

Yes, source inspection supports the reported path: a stopped new assistant turn with failure text and Retry lacks an immediate terminal guard on current main. No current-main runtime reproduction was executed during this read-only review.

Is this the best way to solve the issue?

Yes. The patch extends the existing response collector and structured-error lifecycle, guards active generation, and avoids automatic retries or a competing recovery mechanism.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning medium; reviewed against 1c5ac1b38d64.

Labels

Label justifications:

• P2: This bounded browser improvement avoids consuming the full response timeout after a known terminal generation failure.
• rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The owner’s exact-head Linux Chrome/CDP report exercises built-CLI submission, text and image polling, error persistence, and browser retention through the production owners. Observed 38 ms/35 ms failures, one send, zero Retry clicks, saved recovery metadata, and reachable retained targets satisfy the previous proof request; the controlled page does not claim a live provider outage.
• proof: sufficient: Contributor real behavior proof is sufficient. The owner’s exact-head Linux Chrome/CDP report exercises built-CLI submission, text and image polling, error persistence, and browser retention through the production owners. Observed 38 ms/35 ms failures, one send, zero Retry clicks, saved recovery metadata, and reachable retained targets satisfy the previous proof request; the controlled page does not claim a live provider outage.

Evidence

What I checked:

• Repository policy and review scope: Verified the origin repository and read the full root AGENTS.md. Only the root AGENTS.md was found; .agents/maintainer-notes is absent. Applied the browser recovery, Pro waiting, and release-smoke guidance. The repository profile prohibits automatic cleanup closure. (AGENTS.md:1, 0ee560837f15)
• Complete introduced patch reviewed: Read all 13 files in the pinned main-to-head diff, including the complete proof script and tests omitted from the prompt’s abbreviated patches. Runtime changes total +79/-9 lines, tests +130/-0, the proof script +109/-0, and documentation +1/-0. (0ee560837f15)
• Scoped failure detection and propagation: The collector requires known failure text, a visible Retry control, and absence of strong generation activity. Existing snapshot filtering rejects older turns; the shared error guard reaches polling, image collection, and recovery paths. The warning contains fixed text rather than assistant content. (src/browser/actions/assistantResponse.ts:1306, 0ee560837f15)
• Browser recovery uses the existing lifecycle: assistant-ui-error joins the existing reattachable-capture classification. The lifecycle retains eligible headful browsers, emits runtime metadata, and preserves copied-profile cleanup and non-reattachable handling. (src/browser/index.ts:2394, 0ee560837f15)
• Updated production-path proof satisfies prior request: Owner evidence at fix(browser): fail fast when a new turn offers Retry #458 (comment) explicitly verifies the reviewed head on Linux using the built CLI and real headful Chrome/CDP against a controlled ChatGPT-shaped page. Text and image warnings persisted after 38 ms and 35 ms; each run sent once, clicked Retry zero times, saved error and recovery metadata, and retained a reachable browser and target. This covers the actual wait, serialization, and cleanup owners requested by the previous review. The signed-in Pro smoke is explicitly separate. (0ee560837f15)
• Current-main necessity and related issue: Current main normalizes assistant text without the new terminal Retry error guard. Existing page-warning collection recognizes alerts, banners, and similar containers, which does not establish prompt-scoped immediate failure for ordinary assistant turns. The open paired report Browser mode waits for the full timeout after ChatGPT offers Retry #457 describes exactly this remaining behavior; this PR is its implementation candidate. No merged replacement is supplied. (src/browser/actions/assistantResponse.ts:863, 1c5ac1b38d64)

Likely related people:

• steipete: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)
• Ellis Nieuwpoort: Suggested for follow-up; no historical authorship or introduction is verified. (role: unverified routing candidate; confidence: low)

Rating scale

Score	Internal tier	Crab rank	Meaning
6/6	S	🦀 challenger crab	Exceptional readiness
5/6	A	🦞 diamond lobster	Very strong readiness
4/6	B	🐚 platinum hermit	Good normal PR; ordinary maintainer review
3/6	C	🦐 gold shrimp	Useful, but confidence is limited
2/6	D	🦪 silver shellfish	Proof or implementation needs work
1/6	F	🧂 unranked krab	Not merge-ready
N/A	NA	🌊 off-meta tidepool	Rating does not apply

Overall follows the weaker of proof and patch quality.
Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

Workflow

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

History

Review history (2 earlier review cycles)

• reviewed 2026-09-06T04:59:21.906Z sha 4f681e4 :: needs real behavior proof before merge. :: none
• reviewed 2026-09-07T08:39:58.286Z sha 0ee5608 :: needs real behavior proof before merge. :: none

[steipete]

Verified on head 0ee560837f158fe6af57bef4e98e806d28e331d7.

The full built CLI now has real headful Chrome/CDP proof on a clean Linux machine, using a synthetic local ChatGPT-shaped page and no provider credentials. Both ordinary text and --generate-image calls went through the production submission, response polling, error serialization, and browser cleanup owners.

• Text: structured chatgpt-ui-warning persisted 38 ms after submission.
• Image: the same structured warning persisted after 35 ms.
• Both: one prompt send, zero Retry clicks; session status error; target/port/submission metadata persisted; Chrome and the target remained reachable after the CLI exited.
• The 30-second response deadline was not consumed. The task-owned Chrome processes were closed after verification.

Local suite: 1,999 passed, 43 skipped. Typecheck/lint/build/docs checks pass. Final branch autoreview through P2 is scoped-clean. CI: https://github.com/steipete/oracle/actions/runs/34101387198 — all four jobs succeeded.

This proves the synthetic failure and recovery path. The signed-in Pro smoke remains a separate release gate.