Browser mode still advertises GPT-5.2 models that were removed from the ChatGPT model picker

[HidakaKoyo]

Environment: oracle 0.16.1, macOS (darwin), browser engine with manual-login profile, ChatGPT Pro subscription

Problem
oracle --help and the model resolution logic still offer GPT-5.2 variants (gpt-5.2, gpt-5.2-instant, gpt-5.2-pro) for browser runs, but ChatGPT has removed the 5.2 family from the web UI model picker. On a current Pro account the switcher only shows the 5.5 speed tiers (Fastest 5.5 / Medium / High / Very High), Pro, and GPT-5.6 Sol.

Repro

oracle --engine browser -m "5.2 Instant" -p "test"

Result

ERROR: Unable to find model option matching "GPT-5.2 Instant" in the model switcher.
Available: Fastest 5.5, Medium, High, Very High, Pro, GPT-5.6 Sol.

(The error also appends a misleading "No cookies were applied; log in to ChatGPT…" hint even though the session is authenticated and the switcher was reached — a separate minor issue, happy to split it out.)

Suggested fix

• Remove or mark the 5.2 browser labels as deprecated in --help and the model alias table, or
• fail with a clearer "model no longer available in ChatGPT UI" message that lists the detected options (which the error already collects).

[clawsweeper]

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve. Reviewed August 2, 2026, 9:48 PM ET / August 3, 2026, 01:48 UTC.

Summary
This issue remains necessary. Current main still turns the base, Instant, and Thinking GPT-5.2 browser aliases into literal retired picker labels, while the linked merged fix addressed only the misleading cookie/login suffix; choosing explicit rejection versus compatibility remapping is a maintainer product decision.

Reproducibility: yes. at source level: current browser routing preserves the retired GPT-5.2 base, Instant, and Thinking aliases and maps them to literal picker labels. The reporter also supplied an authenticated live-picker reproduction; repeating the full path locally requires a real ChatGPT browser profile.

Root-cause cluster
Relationship: canonical
Canonical: #344
Summary: This issue is the canonical remaining browser-alias contract decision; the linked pull request fixed only the adjacent authenticated-error message.

Members:

• partial_overlap: fix(browser): preserve authenticated picker errors #347 - The merged pull request removed stale cookie/login guidance and preserved picker diagnostics, but intentionally did not alter retired GPT-5.2 browser alias behavior.

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

Maintainer decision needed

Question	Recommendation
When a browser run requests a retired GPT-5.2 base, Instant, or Thinking alias, should Oracle reject it explicitly or remap it to a documented GPT-5.5 picker tier?	Reject retired browser aliases: Keep GPT-5.2 API aliases, but fail browser runs before picker selection with a clear retirement message and the detected available models.

Why: Both choices change an existing browser-facing contract: explicit rejection breaks old commands but preserves caller intent, while remapping changes the requested model without an explicit user choice.

Next step
A maintainer must select the browser alias compatibility policy before a narrow implementation can be safely routed; the issue already carries an explicit no-new-fix-PR signal.

Review details

Best possible solution:

Retain GPT-5.2 API identifiers, then make browser mode explicitly reject retired base/Instant/Thinking GPT-5.2 aliases before picker selection with a clear migration message and detected picker options, unless maintainers deliberately approve and document a specific compatibility remap.

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

Yes, at source level: current browser routing preserves the retired GPT-5.2 base, Instant, and Thinking aliases and maps them to literal picker labels. The reporter also supplied an authenticated live-picker reproduction; repeating the full path locally requires a real ChatGPT browser profile.

Is this the best way to solve the issue?

Unclear until the browser alias contract is chosen. Explicit browser-only rejection is the narrowest maintainable option because it preserves API compatibility and avoids silently substituting a different model.

AGENTS.md: found and applied where relevant.

Remaining risk / open question:

• A silent GPT-5.2-to-GPT-5.5 remap changes the requested model intent for existing browser scripts; immediate removal instead breaks the documented legacy browser alias contract unless the failure is explicit and actionable.
• ChatGPT picker labels continue to drift, so any solution that relies on exact current display strings needs focused regression coverage for the supported compatibility policy.

Codex review notes: model internal, reasoning high; reviewed against 666e01879a80.

Label changes

Label justifications:

• P2: Authenticated browser users can reproducibly fail when selecting retired aliases, but the problem is bounded to a legacy model family and has an available diagnostic path.
• impact:auth-provider: The affected boundary is model routing from Oracle’s browser engine into the authenticated ChatGPT model picker.
• impact:ux-friction: The stale alias contract produces an avoidable failed run and, before the merged split fix, misleading login-recovery guidance.

Evidence reviewed

What I checked:

• Retired aliases remain in browser label mapping: The current browser label table maps gpt-5.2-thinking, gpt-5.2-instant, and gpt-5.2 to literal GPT-5.2 picker text. The normalizer leaves the first two unchanged, so browser runs still reach the picker with labels the reporter showed are absent. (src/cli/browserConfig.ts:36, 666e01879a80)
• Current error preserves detected picker choices: When selection fails, the browser action adds the detected Available: options when the DOM probe returned them; it no longer adds the misleading cookie/login suffix on this authenticated picker path. (src/browser/actions/modelSelection.ts:86, 666e01879a80)
• Merged split fix is on current main but not the remaining alias fix: The merged commit for fix(browser): preserve authenticated picker errors #347 is an ancestor of current main; its current regression test preserves a missing-picker error without cookie guidance. The owner’s linked comment explicitly states that alias retirement/remapping was intentionally left undecided. (tests/browser/index.test.ts:76, f5c087616b55)
• Documentation still promises GPT-5.2 browser availability: The browser-mode documentation says existing GPT-5.5, GPT-5.4, and GPT-5.2 variants remain available, which defines an existing user-facing browser alias contract that now conflicts with the current picker evidence. (docs/browser-mode.md:120, 5daa6ce8c352)
• Release and maturity checks: git tag --contains f5c087... returned no release tag, so the linked cookie-hint fix is current-main-only after v0.16.1. The required maturity helper could not run because this repository has no qa/maturity-scores.yaml, so no maturity label is supported. (package.json, 666e01879a80)

Likely related people:

• steipete: The history shows the dominant contribution to the browser-model files, and the owner reproduced this failure, authored the adjacent merged fix, and identified the remaining compatibility choice. (role: feature owner and recent browser contributor; confidence: high; commits: f2f20b565a86, f5c087616b55; files: src/cli/browserConfig.ts, src/browser/actions/modelSelection.ts, src/browser/index.ts)
• Dongwoo Jeong: Recent history includes several browser model/Pro-effort compatibility changes, making this person a useful technical routing candidate if the owner chooses a browser-alias contract. (role: recent adjacent browser-model contributor; confidence: medium; commits: ea8b1b57f140, 34127975acd1; files: src/cli/browserConfig.ts, src/browser/actions/modelSelection.ts)

How this review workflow works

• 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.

[HidakaKoyo]

Re-verified today (2026-07-30) on oracle 0.16.1, macOS, browser engine with the manual-login profile (~/.oracle/browser-profile), authenticated ChatGPT Pro account. Adding the expected-vs-actual the review asked for, plus two new data points.

Expected vs actual

Command	oracle --engine browser -m "5.2 Instant" -p "test"
Expected	Either the run proceeds, or it fails with a message saying the model is no longer offered in the ChatGPT UI — and --help does not advertise 5.2 aliases for browser runs in the first place.
Actual	Fails at the model switcher, and the message blames cookies/login even though the session is authenticated and the switcher was reached.

Actual error from today's run:

ERROR: Unable to find model option matching "GPT-5.2 Instant" in the model switcher. No cookies were applied; log in to ChatGPT in Chrome or provide inline cookies (--browser-inline-cookies[(-file)] or ORACLE_BROWSER_COOKIES_JSON).
User error (browser-automation): Unable to find model option matching "GPT-5.2 Instant" in the model switcher. No cookies were applied; log in to ChatGPT in Chrome or provide inline cookies (--browser-inline-cookies[(-file)] or ORACLE_BROWSER_COOKIES_JSON).

The live picker, read straight from the DOM

Options actually present in the model menu on a Pro account today (innerText of the open menu container, verified over CDP against the same profile oracle drives):

Intelligence
Instant     5.5   ✓
Medium
High
Extra High
Pro
GPT-5.6 Sol   >

No 5.2 entry exists anywhere in the menu.

Two things that changed since the original report

1. The UI labels moved again in one day. The original report quoted Fastest 5.5, Medium, High, Very High, Pro, GPT-5.6 Sol. Today the same picker reads Instant 5.5 / Medium / High / **Extra High** / Pro / GPT-5.6 Sol — "Fastest" is now "Instant" and "Very High" is now "Extra High". So this is not only a stale-5.2 problem: an alias table matching ChatGPT's display strings exactly will keep drifting. Resolution probably wants to be tolerant (or at least fail with the detected list) rather than pinned to literal labels.
2. Today's error did not include the Available: ... list, while yesterday's did. So the branch that enumerates detected options is not always the one that reports the failure — worth checking before building the "clearer message that lists the detected options" fix on top of it, since that suggestion assumed the list is already collected on every failure path.

The misleading "No cookies were applied" hint reproduces on every one of these runs on a fully authenticated session, so it is not incidental to the 5.2 case. Happy to split it into its own issue if you prefer.

I can attach a screenshot of the open picker if it adds anything beyond the DOM text above — say the word and I'll drop it in.

[HidakaKoyo]