[xz-dev]

Summary

• Preserve native max reasoning effort for GPT-5.6 Sol, Terra, and Luna through the Chat Completions → Responses → Codex request path while retaining the legacy max → xhigh compatibility mapping for older models.
• Align the Codex GPT-5.6 catalog context window with the 372K OpenAI Codex model contract.
• Publish exact native reasoning-effort metadata and defaults: Sol defaults to low and supports through ultra; Terra defaults to medium and supports through ultra; Luna defaults to medium and supports through max.
• Keep Responses API text-generation models in the VS Code supplemental metadata routes and point their imports at /responses, while retaining the existing non-text and specialty-model exclusions.
• Preserve the pinned 372K/128K GPT-5.6 token contract when merging live Codex discovery metadata: remote-only models, names, and capabilities remain live-driven, while legacy Codex models retain their existing live-limit precedence.
• Keep pricing and the chat/image catalog-collision fix out of this PR.

First-party reference: OpenAI Codex rust-v0.144.1 model catalog.

Related Issues

• Related to fix(models): preserve chat-capable image model rows #7004
• Related to fix(providers): Codex GPT-5.6 models rejected as unsupported #6778

Validation

• npm run lint
• npm run test:unit
• npm run test:coverage
• Coverage is still >= 60% for statements, lines, functions, and branches
• SonarQube PR analysis is green or any remaining issues are explicitly documented below

Completed locally:

• Original GPT-5.6 focused affected suites: 66 tests passed, 0 failed.
• Responses metadata regression: 1 test passed, 0 failed; existing VS Code route suite: 36 tests passed, 0 failed.
• Codex discovery merge regression was developed TDD: before the fix it failed at live 272000 versus pinned 372000; after the fix its focused suite passed 28/28.
• Full affected catalog/discovery/GPT-5.6/Responses set after the discovery fix: 96 tests passed, 0 failed.
• npm run typecheck:core: passed after the Responses metadata and discovery follow-ups.
• Targeted ESLint with repository suppressions for changed source and focused test files: passed.
• Final-head full npm run lint:json -- --max-warnings 0: passed locally with 0 errors and 0 warnings. The corresponding CI job exited 1 silently after its full-tree run with no reported diagnostic; a failed-job rerun was unavailable because the fork author lacks upstream repository admin permission.
• npm run check:file-size: passed.
• npm run check:dead-code: passed (225 vs baseline 227).
• npm run check:cognitive-complexity: passed (890 vs baseline 890).
• git diff --check: passed.
• Full npm run test:unit: 22,799 passed, 6 failed, 14 skipped. The same six failures reproduce on the unmodified origin/release/v3.8.47 tree: three platform-specific doas -S DNS tests, two existing i18n integrity tests, and the existing Cloud Code Gemini request-default assertion.
• Direct ESLint of tests/unit/vscode-token-routes.test.ts reports its existing 65 no-explicit-any errors on both this branch and the same unmodified release tree; changed source and other focused tests are clean.

Tests Added Or Updated

• tests/unit/executor-codex-gpt56.test.ts
• tests/unit/codex-gpt56-catalog.test.ts
• tests/unit/vscode-shared-metadata.test.ts
• tests/unit/vscode-token-routes-gpt56.test.ts
• tests/unit/vscode-token-routes.test.ts
• tests/unit/vscode-responses-models.test.ts
• tests/unit/provider-models-discovery-split.test.ts

Coverage Notes

The cross-layer executor test starts with a Chat Completions request, passes it through openaiToOpenAIResponsesRequest, then through CodexExecutor.transformRequest. It catches the original max → xhigh loss while retaining a GPT-5.5 compatibility assertion. Existing catalog and VS Code route tests cover the context, supported-effort, and default-effort metadata exposed to clients.

The Responses metadata regression seeds a future Responses-only text model, asserts it survives both raw and grouped supplemental endpoints, and verifies the grouped import URL targets /responses.

Coverage was not run locally; CI should provide the authoritative coverage result.

Reviewer Notes

• The translator change is intentionally model-aware. Unknown and legacy models keep the prior max → xhigh behavior.
• The discovery merge exception is likewise narrow: only the pinned GPT-5.6 Sol/Terra/Luna contract (including their supported effort suffixes) keeps local token limits when live metadata is stale; model existence and non-limit live fields still come from discovery.
• ultra remains a client/catalog preset for Sol and Terra; CodexExecutor intentionally sends upstream wire effort max. Luna advertises no ultra support.
• Reasoning-summary catalog metadata remains absent; this PR does not add an implicit summary default.
• No pricing files or pricing behavior are changed.
• PR fix(models): preserve chat-capable image model rows #7004 independently fixes the same-ID chat/image catalog collision needed for the text base row to appear in /v1/models.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[steveepreston]

Thanks for merge request

[steveepreston]

sad to see this still not merged

[diegosouzapw]

Really thorough validation write-up. This is orthogonal to the other Codex-reasoning PRs in this batch (#7176/#7095 touch response-side reasoning-summary synthesis; this is request-side effort mapping + catalog contract), so no reconciliation needed there — it merges independently. Confirmed locally: the executor/catalog regression tests reproduce the pre-fix max→xhigh collapse and pass after your fix, and the 372K context-window correction matches the cited upstream rust-v0.144.1 catalog. One thing we'll double check on our end before merge: your PR notes CI's full-tree lint exited 1 with no reported diagnostic and you couldn't rerun it (no admin perms on the fork) — we'll rerun that job ourselves.

[diegosouzapw]

Babysit summary

Drove the CI reds down from 4 to a single root cause. 10/12 checks now green; the 2 remaining reds are the same assertion and need a maintainer decision (see below).

Fixed

Red	Cause	Fix
Fast Quality Gates (file-size ratchet)	Branch was 57 commits behind release/v3.8.49, carrying a stale config/quality/file-size-baseline.json (translator-openai-to-gemini.test.ts frozen at 1541 vs the tip's 1553). Not caused by this PR.	Merged the base tip in (bbc131646, clean merge).
No new ESLint warnings	Same staleness — ESLint crashed on tests/unit/ui/provider-plan-config.test.tsx, deleted upstream.	Same merge. ✅ now passing.
Unit Tests fast-path (3/4)	check-test-masking-selfref-6634 resolves git show origin/main, unavailable in the fork checkout. Environmental, not a defect.	Same merge. ✅ now passing.
Fast Quality Gates (complexity ratchet)	Real regression from this PR: isUsableChatModel reached cyclomatic complexity 16 (max 15), pushing the ratchet 2056 → 2057.	a52a1b696 — extracted excludesChatAndResponsesEndpoints() / excludesTextOutputModality(). Behavior identical; ratchet back to exactly 2056.

No test was weakened, no CI/baseline edited. All 76 codex/vscode tests pass locally.

---

🛑 Remaining red — needs a maintainer decision (not auto-fixable)

Both Unit Tests fast-path (1/4) and Fast Quality Gates fail on one assertion:

tests/unit/provider-models-route-codex.test.ts:181
AssertionError: 372000 !== 999999

This is not a broken test — it is this PR's intent colliding with an existing contract.

• The test (from Discover live Codex models #6776/feat(providers): add GPT-5.6 model family #6862) encodes: "Live payload wins on overlapping fields" — live discovery reports max_input_tokens: 999999.
• This PR deliberately inverts that for GPT-5.6 via mergeLiveAndLocalCodexModel(), so the pinned GPT_5_6_CODEX_CAPABILITIES (maxInputTokens: 372000) wins — which is precisely the PR's stated goal.

Verified this is caused by the PR, not a base-red: 6/6 pass on the release/v3.8.49 tip, 5/6 on this head.

Updating that assertion means ratifying the policy change ("live is source of truth" → "pinned local contract wins for GPT-5.6"), so I've deliberately left it to a maintainer rather than editing the guard myself.

⚠️ Policy conflict with #7101

#7101 ("strip reasoning_effort when GPT-5.x models carry function tools") does the opposite of this PR for all gpt-5*. Same root cause, opposite strategies, disjoint files — git merges both cleanly and CI goes green on both, but the runtime outcome depends on merge order. Recommend deciding these two together. No reconciliation attempted here.

Not merging — handing off for human review.

[diegosouzapw]

Merged into release/v3.8.49 — thank you for the contribution, @xz-dev! 🎉 Validated in today's full-suite merge-train (33 PRs, 19k+ tests green) before landing.

[steveepreston]

Please merge this as soon as possible. Thanks