chore: update model catalog from bot issues - #843
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Codex (@codex) review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ab5ad4ab4e
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "Codex (@codex) review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "Codex (@codex) address that feedback".
| "openai.gpt-5.5": ["bedrock"], | ||
| "openai.gpt-5.4": ["bedrock"], | ||
| "nvidia/nemotron-3-ultra-550b-a55b": ["together"], | ||
| "moonshot.kimi-k2-thinking": ["bedrock"], | ||
| "xai.grok-4.3": ["bedrock"], |
There was a problem hiding this comment.
These new Bedrock endpoint entries make getModelEndpointTypes() select Bedrock for openai.gpt-5.5, openai.gpt-5.4, and xai.grok-4.3, but their catalog specs are still format: "openai". Requests are dispatched by modelSpec.format in fetchModel(), so they go through fetchOpenAI(), which explicitly rejects secret.type === "bedrock" with ProxyBadRequestError: Bedrock does not support OpenAI format; any user routed to these newly advertised Bedrock endpoints will fail until the specs use the Bedrock/converse path or that path supports these OpenAI-format Bedrock models.
Useful? React with 👍 / 👎.
ab5ad4a to
5ee0a31
Compare
Codex Review ResponseIssue: Bedrock models with incorrect
|
…rmat to converse
Both model-sync workflows have a 'Respond to Codex review with Claude Code'
step that auto-applies Codex's suggestions to the catalog. When the index.ts
Bedrock mappings for openai.gpt-5.5 / openai.gpt-5.4 / xai.grok-4.3 are present,
Codex posts a P1 ('openai format on a Bedrock model fails') because the TS
proxy's fetchOpenAI rejects Bedrock secrets, and the auto-apply step 'fixes' it
by flipping format openai->converse. That is wrong: these are Bedrock Mantle
models served only via the OpenAI-compatible bedrock-mantle endpoint, which does
not support the Converse API or InvokeModel, so converse breaks invocation at
AWS. This recurred across #834/#836/#840/#843 (all reverting the validated #822
openai fix). The LiteLLM field sync never touches format (no checkAndUpdateFormat),
so this guardrail on the codex-response prompt is the correct lever.
Adds a 'Bedrock Mantle exception' instruction to the codex-response prompt in
both sync-models.yaml and fix-missing-model-bot-issues.yaml: keep format=openai
for bedrock-only openai-format models and explain in the summary that this is an
intentional Mantle case to be handled proxy-side, not a catalog change.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
1 participant
Automated daily batch of model catalog updates from bot issues.
Included issues
Summary
gpt-oss-120bgpt-oss-120b2
3
r1-1776r1-17762
Verified metadata
#841: [BOT ISSUE] Cerebras: update gpt-oss-120b max_output_tokens from 32768 to 40960
Verification notes
Verification
sync_models vs proposed update
sync_models cross-check found differences. Official provider verification was used for the applied values, and sync_models discrepancies are listed below for review.
#842: [BOT ISSUE] Perplexity: mark r1-1776 as deprecated (retired August 2025)
Verification notes
Verification
sync_models vs proposed update
sync_models cross-check found differences. Official provider verification was used for the applied values, and sync_models discrepancies are listed below for review.