fix(catalog): advertise GPT-5.6 reasoning efforts - #110
Conversation
clawsweeper
Bot
added
rating: 🦪 silver shellfish
|
Codex review: needs maintainer review before merge. Reviewed August 8, 2026, 1:04 PM ET / 17:04 UTC. ClawSweeper reviewWhat this changesAdds optional GPT-5.6 reasoning-effort metadata to provider manifests, the compiled provider catalog, and authenticated Merge readinessKeep open: current main still omits model-specific reasoning efforts, while this clean branch carries the optional metadata from the provider manifest through the compiled and authenticated client catalog. The prior release-note finding is resolved, and the supplied real Worker proof supports normal maintainer merge review. Priority: P2 Review scores
Verification
How this fits togetherProvider manifests are compiled into ClawRouter’s Worker snapshot, then filtered by credential and executable endpoints before clients receive flowchart LR
A[Provider manifest] --> B[Manifest validation]
B --> C[Compiled provider snapshot]
C --> D[Credential and endpoint filtering]
D --> E[Authenticated catalog response]
E --> F[Client reasoning selection]
Before merge
Agent review detailsSecurityNone. Review metrics
Technical reviewBest possible solution: Land the optional provider-native metadata contract and retain the omission behavior for models that do not declare it. Do we have a high-confidence way to reproduce the issue? Not applicable as a bug reproduction: this is an additive catalog capability. The PR supplies high-confidence after-fix evidence from an authenticated local Worker response containing all six declared values. Is this the best way to solve the issue? Yes. A declarative optional manifest field is the narrow provider-neutral solution, and the Worker only carries it through existing compilation and catalog boundaries. AGENTS.md: found and applied where relevant. Codex review notes: model internal, reasoning high; reviewed against 7b0b6f60f7c6. LabelsLabel changes:
Label justifications:
EvidenceWhat I checked:
Likely related people:
Rating scale
Overall follows the weaker of proof and patch quality. Workflow
History |
|
@clawsweeper re-review |
|
🦞🧹 I asked ClawSweeper to review this item again. Re-review progress:
|
clawsweeper
Bot
added
proof: sufficient
|
@clawsweeper re-review |
|
🦞🧹 I asked ClawSweeper to review this item again. Re-review progress:
|
clawsweeper
Bot
added
rating: 🦞 diamond lobster
1 participant
Summary
supportedReasoningEffortsmetadata to provider model manifests and the compiled runtime catalognone,low,medium,high,xhigh, andmax/v1/modelsand request proxying unchangedRoot cause
ClawRouter's credential-scoped catalog exposed model identity, capabilities, and pricing, but omitted model-specific reasoning-effort metadata. Consumers therefore fell back to a stale generic effort profile capped at
high, even though the existing GPT-5.6 route already passes advanced nested reasoning settings through to the provider.The canonical fix records the provider-native contract in the model manifest, validates it during compilation, carries it through runtime types and the generated snapshot, and projects it through
/v1/catalog. The paired OpenClaw consumer is openclaw/openclaw#120631.Evidence
Authenticated local Worker runtime
The runtime candidate
4c69ef6fe5ee22f5541874456f359a016e0b7f4fran through the repository's real local Wrangler Worker, local authority/KV state, and generated local proxy credential using:A temporary uncommitted assertion in the established smoke read the authenticated HTTP response and was removed after the run. Current exact head
c35be57b80500f831656e1ab8682d335f90862cdchanges only the release-owned changelog line; all runtime, manifest, generated-catalog, and test files remain byte-identical to the proven runtime candidate. Redacted observed result:{"proof":"authenticated /v1/catalog","status":200,"model":"openai/gpt-5.6","supportedReasoningEfforts":["none","low","medium","high","xhigh","max"],"sibling":"openai/gpt-5.5","siblingHasSupportedReasoningEfforts":false}The full smoke then completed successfully:
local Worker smoke passed.Consumer compatibility
The companion OpenClaw PR's source-blind validation used the public CLI with a disposable catalog endpoint. It discovered exactly
none/low/medium/high/xhigh/max, rejectedminimalwith allowed choicesoff/low/medium/high/xhigh/max/ultra, and provedxhigh,max, and logical OpenClawultrareached the provider endpoint. It also proved absent metadata preserves the generic fallback. See openclaw/openclaw#120631.The field is optional and additive. Consumers that do not read it continue parsing the pre-existing model fields and retain their prior fallback behavior.
Automated checks
pnpm worker:check: 126/126 passedpnpm check: Worker, admin, and script suites passedpnpm build: provider snapshot, admin production build, Worker typecheck, and Worker tests passedgit diff --check: passedAutoreview
The coordinator ran the global Codex autoreview helper in branch mode against
origin/mainon4c69ef6fe5ee22f5541874456f359a016e0b7f4f. The complete one-pass bundle was 202,469 bytes, including the generated snapshot. Result: clean, with no accepted or actionable findings.After removing the release-owned changelog line, exact-head cleanup review ran:
It reviewed the complete 654-byte follow-up diff in one pass and returned
autoreview clean: no accepted/actionable findings reportedwith overall correctness 0.99.Change size
Source
The existing manifest model and upstream alias are documented by OpenAI at https://developers.openai.com/api/docs/models/gpt-5.6-sol.
Deployment status
Deployment status: not deployed.
Remaining risk
The catalog change is additive and request validation and proxy behavior are intentionally unchanged. The consumer behavior is delivered separately by the paired OpenClaw PR.