fix(browser): match German Intelligence effort labels; keep current tier when none match

[Jonasdero]

Summary

• Match German Intelligence effort labels (Sofort / Mittel / Hoch / Sehr hoch, Intelligenz, Pro Erweitert) — LEVEL_TOKENS was English + Chinese only, and normalize() stripped umlauts before matching could see them, so a German workspace could never have its effort set at all
• Match Latin effort labels as whole words/phrases instead of substrings, so row descriptions (Hoch – für sehr komplexe Aufgaben) and unrelated UI words (Hochladen, Ermitteln) can no longer decide the tier; sehr hoch / extra high now require the adjacent phrase rather than two independent tokens
• Stop hard-wiring Pro for GPT-5.6 + heavy: a level with no matching row leaves the tab's current effort alone instead of silently switching the model to Pro, and the log now says so
• Document why launched Chrome is pinned to --lang=en-US and why localized matching is still required despite it

Follow-up to #354 / #353, which fixed extra-high vs heavy for the English UI only.

Why the German path exists at all

--lang=en-US --accept-lang=en-US,en in buildChromeFlags() (src/browser/chromeLifecycle.ts) applies only to Chrome that Oracle launches. --browser-attach-running and --remote-chrome never build those flags (src/browser/controlPlan.ts) and inherit the user's own Chrome locale, and a ChatGPT account language setting can localize the UI even on the launch path. Both are now stated in a comment above the flags so the next reader doesn't conclude English is the only case to handle.

Whole-menu vs per-row text

countEffortLevels() / isProEffortMenu() read a whole menu's textContent, where adjacent row labels concatenate without a separator ("Pro StandardPro Extended"), so word-boundary matching does not hold there and substring matching is kept deliberately. Per-row classification goes through a single matchesTokens() used by both matchesLevel and matchesAnyEffortLevel.

Proof

pnpm vitest run — 1676 passed, 43 skipped (169 files); pnpm exec tsc --noEmit and oxlint clean; pnpm run build OK.

New unit coverage drives a German Intelligence menu table-style: light→Sofort, standard→Mittel, extended→Hoch, extra-high→Sehr hoch, plus negative cases proving Sehr hoch never satisfies extended, Hoch never satisfies extra-high, Hoch – für sehr komplexe Aufgaben is still extended, and Hochladen matches nothing. Two tests that asserted GPT-5.6 + heavy switches to Pro now assert option-not-found.

Live browser smoke, built CLI, GPT-5.6 Sol, manual-login profile temporarily launched with --lang=de-DE so ChatGPT rendered German (flag reverted before commit; verified git status clean):

flag	session	[browser] Thinking time:
--browser-thinking-time light	de-smoke-light	Sofort5.5
--browser-thinking-time standard	de-smoke-standard	Mittel
--browser-thinking-time extended	de-smoke-extended	Hoch
--browser-thinking-time extra-high	de-smoke-extra-high	Sehr hoch
--browser-thinking-time heavy	de-smoke-heavy	option not found (requested Heavy); keeping the effort already selected in ChatGPT.

All five runs report resolvedLabel=GPT-5.6 Sol; verified=yes — notably the heavy run, which on 0.17.0 would have clicked Pro. A control run on the normal en-US launch profile still reports Extra High (already selected) for extra-high, so English behaviour is unchanged.

Docs: docs/browser-mode.md notes the supported label languages and the keep-current-tier fallback; CHANGELOG updated under 0.17.1.

@clawsweeper re-review

[clawsweeper]

Codex review: needs changes before merge. Reviewed August 5, 2026, 4:51 PM ET / 20:51 UTC.

ClawSweeper review

What this changes

This PR adds German and phrase-aware matching for ChatGPT browser effort menus and retains the current tier when no requested row can be found.

Regression provenance

Possible regression — probable (reviewed change; known regression link). No predecessor PR is attributed.

Merge readiness

⚠️ Needs maintainer review before merge - 3 items remain

Keep open: the German-label work is useful, but the branch removes the established GPT-5.6 heavy → explicit Pro selection when a Pro row is available. The same P1 blocker remains at the unchanged reviewed head.

Priority: P1
Reviewed head: 794b46893299dfff6c5826126f117f759d596412

Review scores

Measure	Result	What it means
Overall readiness	🦐 gold shrimp (3/6)	The localization patch has strong live proof and focused coverage, but one compatibility-critical selection path remains incorrect.
Proof confidence	🦞 diamond lobster (5/6)	Sufficient (live_output): The PR body supplies after-fix live browser output for all German effort levels plus an English control run; it is sufficient proof for the localization behavior, though it also exposes the remaining heavy contract mismatch.
Patch quality	🦐 gold shrimp (3/6)	1 actionable review finding remain.

Verification

Check	Result	Evidence
Real behavior	Verified	Sufficient (live_output): The PR body supplies after-fix live browser output for all German effort levels plus an English control run; it is sufficient proof for the localization behavior, though it also exposes the remaining heavy contract mismatch.
Evidence reviewed	4 items	Current documented contract: Current main documents that GPT-5.6 Sol maps extra-high to Extra High and heavy to Pro. Current implementation behavior: Current main explicitly finds the standalone Pro row for a GPT-5.6 heavy request in the Intelligence menu before generic effort matching. Merged contract provenance: The currently documented distinction between Extra High and explicit Pro was introduced by the merged follow-up for the related report; no release tag contains that commit, so it is current-main behavior.
Findings	1 actionable finding	[P1] Preserve standalone Pro selection for heavy
Security	None	None.

How this fits together

Oracle browser mode translates the CLI thinking-time setting into a selection in ChatGPT’s rendered effort menu before submitting a prompt. The selected tier is then logged and retained in session metadata for the browser run.

flowchart LR
  A[CLI thinking-time option] --> B[Browser automation]
  B --> C[ChatGPT effort menu]
  C --> D[Localized label matching]
  D --> E[Selected effort tier]
  E --> F[Browser session log]

Loading

Before merge

• Preserve standalone Pro selection for `heavy` (P1) - Restore the guarded GPT-5.6 heavy lookup for a standalone Pro row. Current main and the documented contract reserve heavy for explicit Pro; changing this path to option-not-found leaves an existing user on whatever lower tier was already selected even though Pro is available.
• Resolve merge risk (P1) - Merging would make existing GPT-5.6 users with --browser-thinking-time heavy or browser.thinkingTime: "heavy" retain a lower current tier even when the UI exposes the documented standalone Pro choice.
• Complete next step (P2) - A single source-path repair can restore the established heavy behavior without changing the localized matching design.

Findings

• [P1] Preserve standalone Pro selection for heavy — src/browser/actions/thinkingTime.ts:461-476

Agent review details

Security

None.

Review metrics

Metric	Value	Why it matters
Patch surface	5 files; production +93/-72, tests +179/-3	The behavior change is concentrated in the browser effort selector with substantial focused regression coverage.

Merge-risk options

Maintainer options:

1. Restore explicit Pro for heavy (recommended)
   Restore the GPT-5.6 heavy branch that selects a standalone Pro row, while retaining the keep-current fallback only when that row is absent.
2. Accept changed heavy semantics
   Intentionally change heavy to mean best-effort no-op when its exact token is absent, then obtain maintainer approval and revise the documented contract.

Copy recommended automerge instruction

@clawsweeper automerge

Special instructions:
Restore guarded GPT-5.6 `heavy` selection of a standalone Pro row; retain the no-match fallback only when no Pro row exists, and add regression coverage for both cases.

Technical review

Best possible solution:

Keep localized matching and the no-match fallback, but retain the guarded standalone-Pro selection for GPT-5.6 heavy; only leave the current tier unchanged when no explicit Pro row exists.

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

Yes, from source: current main documents and implements heavy as explicit Pro for GPT-5.6, while this branch removes that selection and changes its regression tests to expect no matching option.

Is this the best way to solve the issue?

No. The localization and phrase-boundary work is a narrow solution, but the unmatched-row fallback must not replace the existing standalone-Pro path for heavy.

Full review comments:

• [P1] Preserve standalone Pro selection for heavy — src/browser/actions/thinkingTime.ts:461-476
  Restore the guarded GPT-5.6 heavy lookup for a standalone Pro row. Current main and the documented contract reserve heavy for explicit Pro; changing this path to option-not-found leaves an existing user on whatever lower tier was already selected even though Pro is available.
  Confidence: 0.99

Overall correctness: patch is incorrect
Overall confidence: 0.99

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against d0898f4bd765.

Labels

Label justifications:

• P1: The patch would regress the documented heavy browser workflow for GPT-5.6 users with an available Pro row.
• merge-risk: 🚨 compatibility: Existing CLI and persisted heavy settings would stop selecting their documented explicit Pro tier after upgrade.
• rating: 🦐 gold shrimp: Overall readiness is 🦐 gold shrimp; proof is 🦞 diamond lobster and patch quality is 🦐 gold shrimp.
• status: ⏳ waiting on author: ClawSweeper has contributor-facing work open and is waiting for author action. Sufficient (live_output): The PR body supplies after-fix live browser output for all German effort levels plus an English control run; it is sufficient proof for the localization behavior, though it also exposes the remaining heavy contract mismatch.
• proof: sufficient: Contributor real behavior proof is sufficient. The PR body supplies after-fix live browser output for all German effort levels plus an English control run; it is sufficient proof for the localization behavior, though it also exposes the remaining heavy contract mismatch.

Evidence

Acceptance criteria:

• [P1] pnpm vitest run tests/browser/thinkingTime.test.ts.
• [P1] pnpm exec tsc --noEmit.
• [P1] Redacted live GPT-5.6 browser smoke confirming heavy selects a visible Pro row.

What I checked:

• Current documented contract: Current main documents that GPT-5.6 Sol maps extra-high to Extra High and heavy to Pro. (docs/browser-mode.md:104, d0898f4bd765)
• Current implementation behavior: Current main explicitly finds the standalone Pro row for a GPT-5.6 heavy request in the Intelligence menu before generic effort matching. (src/browser/actions/thinkingTime.ts:461, d0898f4bd765)
• Merged contract provenance: The currently documented distinction between Extra High and explicit Pro was introduced by the merged follow-up for the related report; no release tag contains that commit, so it is current-main behavior. (src/browser/actions/thinkingTime.ts:461, a1825d09aede)
• Re-review continuity: The current head is the same SHA as the prior completed review, whose P1 finding required preserving explicit Pro selection for heavy; the PR body and changed tests still intentionally expect option-not-found instead. (tests/browser/thinkingTime.test.ts:612, 794b46893299)

Likely related people:

• ozeron: Authored the merged follow-up that separated extra-high from explicit Pro and established the current behavior this patch must preserve. (role: introduced current selection contract; confidence: high; commits: a1825d09aede; files: src/browser/actions/thinkingTime.ts, docs/browser-mode.md)
• dragonfsky: Introduced the GPT-5.6 unified effort-picker support that owns the affected browser menu path. (role: GPT-5.6 picker contributor; confidence: high; commits: 220fbd18bfc7; files: src/browser/actions/thinkingTime.ts)
• Peter Steinberger: Has the strongest sustained history in the browser thinking-time implementation and documentation. (role: recent area contributor; confidence: medium; commits: 467b41805961, f7efb31d41e3; files: src/browser/actions/thinkingTime.ts, docs/browser-mode.md)

Rank-up moves

Optional improvements that raise the rating; they are not merge blockers.

• Restore standalone-Pro selection for GPT-5.6 heavy and keep the fallback only when Pro is absent.
• Re-run the existing live browser smoke with a visible Pro row and record that heavy selects Pro; redact account details before posting output.

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 (3 earlier review cycles)

• reviewed 2026-08-05T14:06:38.810Z sha 794b468 :: needs changes before merge. :: [P1] Preserve heavy selecting the available Pro row
• reviewed 2026-08-05T16:33:24.885Z sha 794b468 :: needs changes before merge. :: [P1] Preserve heavy selecting the available Pro row
• reviewed 2026-08-05T19:40:56.509Z sha 794b468 :: needs changes before merge. :: [P1] Preserve heavy selecting the available Pro row