Skip to content

Fill GPT-6.2 API reasoning profiles in model metadata - #521

Open
johnlanni wants to merge 2 commits into
mainfrom
codex/gpt62-reasoning-efforts
Open

johnlanni wants to merge 2 commits into
mainfrom
codex/gpt62-reasoning-efforts

Conversation

@johnlanni

Copy link
Copy Markdown
Contributor

Summary

The GPT-6.2 Astra and Sol catalog entries recorded reasoning support with an empty effort list. The OpenAI connection form then prefilled a generic thinking switch on both official API protocols.

  • Record low, medium, high, xhigh, and max for both models. The model-specific provider default remains unknown. The existing OpenAI bindings use reasoning.effort on Responses and reasoning_effort on Chat Completions. OpenAI model index · GPT-6.2 migration guide
  • Require cross-checking the official model index or comparison, per-model documentation, and protocol references in model-metadata-maintainer. Verify each affected prefill against the emitted bundle and feed.
  • Regenerate the shared bundle at data revision 21 and the feed at sequence 19. The semantic diff changes the two reasoning records and their source provenance.

Follow-up to #508.

Verification

  • The four Desktop prefill cases reproduced enable_thinking on the previous catalog. They now cover both model IDs and both OpenAI protocols, when creating a check draft and when retaining manual edits. They assert all five ordered profiles and the absence of a provider default.
  • The API default regression checks both protocols. The existing installed-product scenario verifies the prefill and saved capabilities through the same check and save path.
  • Regenerating the feed twice produced identical bytes. The Rust producer passed recorded-reader validation and the website verifier passed. Reading the emitted feed through the production prefill code produced the same five profiles for all four model and protocol combinations.
  • Local verification passed 483 Desktop tests, the Desktop build, 67 metadata tests, two catalog generator tests, catalog validation, the skill validator, and 24 contract convergence tests. The catalog reports 202 existing stale sources outside the scoped correction.

Use the OpenAI model index for exact Sol and Astra efforts. Keep the model-specific default unresolved and validate both registered OpenAI protocols. Require model-level reasoning evidence and consumer checks in the metadata maintenance workflow.

Signed-off-by: 澄潭 <zty98751@alibaba-inc.com>
Signed-off-by: 澄潭 <zty98751@alibaba-inc.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant