[guanbear]

Summary

Explain ChatGPT/Codex models withheld because their minimal_client_version exceeds OmniRoute's effective HTTP client profile. The warning reports the affected IDs, current version, highest required minimum, and recovery steps: refresh discovery, upgrade OmniRoute, or use a temporary CODEX_CLIENT_VERSION override only when the release is otherwise protocol-compatible.

This applies to the HTTP codex provider. Installing a standalone Codex CLI is relevant only to the separate codex-app-server provider.

Integration with the current release branch

Updated the focused patch on release/v3.8.52 at 4ea24a2f8e1f, following the sequencing advice in this thread:

• Reuse feat(codex): safely discover compatible models #12933's active/candidate classification and its requires-newer-client reason instead of introducing a second compatibility policy.
• Preserve fix(codex): refresh discovery client version automatically #15663's automatic HTTP client-version refresh and fix(codex): route discovered reasoning tiers through catalog and dispatch #13224's reasoning metadata.
• Preserve fix(codex): keep live model inventories authoritative #15132's authoritative account-live inventory. A GitHub-only future model does not generate a warning during successful live discovery.
• Keep an all-filtered live inventory empty, preserve cache-before-GitHub fallback, and recompute GitHub TTL-cache diagnostics against the current version.
• Keep active-only persistence. Existing persisted cache rows do not retain compatibility metadata, so they retain the generic cached-catalog fallback message rather than inventing a minimum-version diagnosis.
• Extract a small Codex route helper; the models route shrinks. No stale migration-count, baseline, or documentation changes remain in the feature diff.

Related to #12761; complements #12827 and #12759.

Validation

• 71 focused tests passed across the new compatibility regression file and existing Codex route, discovery split, live-catalog, GitHub-catalog, and client-version-refresh tests.
• New coverage includes account/public membership separation, unavailable GitHub catalog, all-too-new live inventory, retired/hidden/unsupported candidates, effective-version changes during TTL cache reuse, cache precedence, route-triggered automatic version refresh, numeric maximum version, and bounded model-list output.
• On the October 9 update: full repository ESLint (using the configured frozen suppressions), changed-file ESLint, typecheck:core, dashboard TypeScript ratchet, provider consistency/assets, file-size/test-file-size, mutation-test registration, docs version synchronization, changelog integrity, formatting, and git diff --check passed.
• Expanded verification: 103/104 tests passed. The only failing provider translate-path golden assertion has an exactly identical assertion diff on the unchanged 4ea24a2f8e1f base. The previously failing internal-service-auth and OpenAI→Claude tests now pass after incorporating upstream fixes.
• The repository pre-commit and commit-message gates passed; lint-staged was scoped to the actual five-file PR diff to avoid reformatting unrelated incoming release files. Separate Spec and Standards code reviews have no actionable findings.

Inherited base failures

⚠️ base-red inherited: #15306. Compared with an isolated unchanged checkout at 4ea24a2f8e1f:

• The provider translate-path golden failure is identical on both trees (same assertion diff).
• check:agent-skills-sync reports the same 20 stale generated skills on both trees. The generator sources, OpenAPI/CLI inputs, and generated skills are not changed by this feature PR. No unrelated generated instruction edits are included.
• The scheduled-workflow guard check reports the same five docker-publish.yml jobs on both trees. The complete local workflow audit is unavailable because actionlint/zizmor are not installed; no full local workflow-audit pass is claimed.
• The old full-lint and dashboard type regressions are resolved by the incorporated upstream updates. Both local gates now pass; the dashboard gate still has baselined diagnostics, so this is a no-new-errors result rather than zero TypeScript errors globally.
• The earlier CI build error traced through imageRegistry into server-only modules. The incorporated release contains the upstream image-registry split (fix(imageRegistry): split data from flag-gated logic to avoid client bundle leak (#10692) #15791); production build verification remains with the new CI run.
• The docs/i18n CI matrix remains pending. The sparse-checkout local i18n checker processed zero source entries, so it is not counted as translation validation.

The full unit/Vitest/coverage/build suites remain CI's responsibility under the contribution golden path. This update does not claim the release branch is green.

[guanbear]

The API Route Typecheck failure was not caused by this PR — release/v3.8.51 itself is currently red: #12506 dropped the highWaterMark plumbing that #12179 added, while glm.ts kept the 16-argument call (TS2554). Filed the restoration as #12929 and cherry-picked it onto this branch so CI can go green without waiting; once #12929 lands, the cherry-picked commit reconciles cleanly.

[guanbear]

Pushed the remaining gate fixes on top of the feature commit:

• file-size: route.ts had grown past its frozen 2432-line cap; extracted the compatibility-warning accumulator and the shared models-discovery fetch wrapper into discovery/codexCompatibility.ts (route.ts is back to 2428 lines — extraction rather than a cap bump).
• base-red cherry-picks from fix(pack): allow the opencode-plugin-v2 workspace package in the artifact policy (base-red) #12929: the restored stream.ts highWaterMark plumbing (compacted to its cap), the refreshed migration counts, and the repaired changelog fragment — all currently red on the release tip for every open PR.
• The earlier-reported mutation-coverage, pack-policy, and build-scope failures pass on this tree after the refactor (module→test mapping moved with the extraction).

Local validation: api-typecheck OK, open-sse-typecheck OK, file-size OK, docs-sync/docs-counts/changelog-integrity OK, pack-policy OK, mutation-test-coverage OK, codex/glm targeted suites 59/59.

[guanbear]

Rebased onto the current release tip and slimmed the branch to three commits:

• feat(codex): report client catalog incompatibilities — the feature, now merged cleanly with the dispatch-time live-catalog changes (fix(backend): chat dispatch ignores model-picker additions (customModels vs syncedAvailableModels) #12597/fix(catalog): union picker customModels into the dispatch-time live catalog (#12597) #12934); the GitHub-catalog fallback response keeps upstream's new buildResponse({source: "github_catalog"}) shape with the compatibility warning appended.
• refactor(codex): keep route.ts within the frozen size cap — warning accumulator + shared discovery-fetch wrapper extracted to discovery/codexCompatibility.ts (route.ts at 2427 lines, cap 2432).
• fix(pack): allow the opencode-plugin-v2 workspace package in the artifact policy — the one base-red still unfixed on the tip (also filed separately as fix(pack): allow the opencode-plugin-v2 workspace package in the artifact policy (base-red) #12929).

The earlier base-red cherry-picks (stream highWaterMark, chatCore narrowing, doc counts) are dropped — the maintainers landed equivalents in #12963/#12970/#12990. Full gate suite on this tree: api-typecheck, open-sse-typecheck, file-size, docs-sync, changelog-integrity, pack-policy, mutation-coverage all green; codex suites 40/40.

[guanbear]

Merged the current release tip into the branch and re-synced the migration count (171→172, all 41 llm.txt mirrors) so CI evaluates a fresh tree. On this tree the only remaining red I can see is tests/unit/volcengine-plan-binding-upsert.test.ts (4 no-explicit-any ESLint errors) — that file arrived on the tip with #12950 and is not touched by this PR; flagging it here in case the next quality wave wants to pick it up.

[diegosouzapw]

Nice diagnostic — the "why is this model missing" message is genuinely useful, and I ran both
test files at the PR head: 40/40 pass, including the "every remote model is too new" case.
Two notes: (1) your branch's AGENTS.md/llm.txt diff is just a stale migration-count bump
(171→172) picked up from an older base — the tip is already at 175, so that hunk should be
dropped on rebase and left to the docs-sync generator rather than hand-edited. (2) this overlaps
textually with #13224 and #12933 (same discovery file) — the maintainer will sequence the three
rather than merge independently; your compatibility-warning logic may end up reusing #12933's
candidate/active classification once that lands, so hold off rebasing until that's decided.

[guanbear]

Thank you for the review — glad the diagnostic message lands well, and appreciate you running the suites at the head.

Understood on both notes: the 171→172 docs hunks are stale base leftovers and will be dropped (left to the docs-sync generator) whenever the branch next moves, and we'll hold off rebasing until the sequencing with #13224/#12933 is decided. Happy to rework the compatibility-warning logic onto #12933's candidate/active classification if that's the shape it takes — just ping here.

[diegosouzapw]

Re-homed to release/v3.8.52: v3.8.51 entered its release freeze, so the branch now belongs to the release captain and development continues on the next cycle. Nothing is wrong with this PR — it just needed a live base. No action needed from you; CI will re-run against the new base.

[guanbear]

Thanks for the guidance on sequencing this with the Codex discovery changes. I have now adapted the patch to the current release/v3.8.52 branch and reused the merged active/candidate classification rather than keeping a separate parser or policy.

The warning now follows the selected account/catalog candidates and the effective refreshed HTTP client version. It preserves the new reasoning metadata, live account membership, empty all-filtered results, and cache fallback ordering. I also removed the obsolete baseline/docs changes.

The focused regressions pass (71 tests), and the changed-file and provider gates pass. I reproduced the existing reasoning/golden failures on the unchanged base and documented those, together with the inherited full-lint/typecheck limitations, in the PR description. The update is ready for your review when it fits the release queue. Thank you!

[guanbear]

Thanks again for the sequencing guidance. I have now incorporated the current release/v3.8.52 tip (4ea24a2f8e1f) and kept the feature diff limited to the same five files. This preserves the new provider discovery paths and continues to reuse the merged active/candidate classification and effective HTTP client version.

The former lint, dashboard typecheck, internal-service-auth, and OpenAI→Claude failures are resolved after bringing in the upstream fixes. Full lint, the core/dashboard type gates, provider and file-size gates, mutation-test registration, docs version sync, and commit hooks pass locally. Expanded tests pass 103/104; the remaining golden failure has the exact same assertion diff on an isolated unchanged base.

I also checked the generator/workflow reds against that exact base: the stale-skills report and five scheduled-job guard findings are identical, so they remain inherited #15306 work rather than extra fixes in this feature PR, as requested by the contribution guide. The description records these boundaries and leaves broad build/i18n verification to the new CI run. The update is ready for review when it fits the release queue. Thank you!

[diegosouzapw]

Thanks @guanbear — Codex model discovery names the models that need a newer client profile and says which version the catalog requires. This is in the upcoming release.