Anthropic/ChatCompletions clients → ChatGPT Codex via Responses API

[Jiliac]

Stacked on katanemo#988 (adil/fix-responses-identity-passthrough) — merge that first; this PR's diff will shrink to just the commits below once it lands.

Problem

Claude Code speaks the Anthropic Messages API. The ChatGPT Codex backend only serves /backend-api/codex/responses (the Responses API) — any client hitting it with a /chat/completions-shaped or Messages-shaped payload gets a hard 404 before this PR. There were two independent root causes:

1. The ChatGPT provider was grouped together with the OpenAI-compatible providers in providers/id.rs, so it never got routed to the Responses API in the first place.
2. Even once routed, the Anthropic client fell through to the /chat/completions default in clients/endpoints.rs instead of /responses.

What was added

• Dedicated ChatGPT → Responses API routing for all client APIs (Anthropic Messages, ChatCompletions, and Responses passthrough).
• Direct MessagesRequest ⇄ ResponsesAPIRequest and ChatCompletionsRequest → ResponsesAPIRequest transforms, including mapping reasoning into Anthropic thinking blocks with multi-turn encrypted_content replay.
• Streaming transforms that turn Responses SSE events into Anthropic and ChatCompletions stream events.
• Anthropic streaming buffer: multi-block balancing and [DONE] synthesis for the ChatCompletions leg.
• InputItem untagged-enum fix: replaced with a manual type-dispatching Deserialize impl. This fixes a bug where ItemReference was silently swallowing function_call/reasoning items — it also benefits the pre-existing identity passthrough path, not just this feature.
• Codex payload normalization: strip parameters Codex rejects (max_output_tokens, truncation, metadata, user, temperature, top_p), migrate the nested reasoning object handling, and add the ReasoningEffort::None/Minimal variants Codex can return.
• Provider-prefix routing precedence fix in common/llm_providers.rs: an explicit openai/ prefix no longer loses to another provider's wildcard alias. This was a pre-existing bug, exposed by the new gpt-5.6 registry entries added under the chatgpt provider.

Verification

• 215+ hermesllm lib tests, including full-stream fixture ladders and chunk-boundary replay scenarios.
• cargo fmt, cargo clippy -D warnings, and the wasm build target all green.
• Live-verified against real Codex across 3 rounds of testing:
  • Text, tool-call, and thinking ladders on gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, and gpt-5.4.
  • ChatCompletions leg (single [DONE], correct finish_reason).
  • Identity passthrough left intact.
  • Routing precedence confirmed (openai/* explicit prefix beats chatgpt/* wildcard).
  • Culminated in a real headless Claude Code session completing a Write + Read tool loop end-to-end against Codex with zero 400s.

Relationship to katanemo#958

This is orthogonal to katanemo#958. It deliberately does not adopt that PR's JSON post-pass adapter approach — all Codex-specific quirks here live on the typed ResponsesAPIRequest inside normalize_for_upstream, not as a generic JSON patch layer. Conflict surface should be minimal: provider_models.yaml is untouched by this PR.

Known follow-ups (minor, listed for transparency)

• message_start emits model: "unknown" for Codex streams.
• The incomplete → Length stop-reason mapping on the ChatCompletions leg is broader than strictly specced.
• Codex only ever produced empty-summary reasoning items during testing, so thinking blocks currently observed are signature-only (empty thinking text + a signature_delta). Claude Code accepts this shape fine, but a non-empty reasoning summary has not been exercised end-to-end.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 5473b344-d081-4561-84b9-5fffc2b8bf19

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch adil/anthropic-client-codex-responses

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[Jiliac]

Closing as superseded for our local deployment by a chained architecture: Claude Code → Plano routing → CLIProxyAPI → Codex. The direct Anthropic/ChatCompletions → Codex Responses bridge remains incomplete for client-nonstream/upstream-stream and tool/subagent flows, so the PR’s current end-to-end compatibility claim is too broad. Preserving the branch for reference and possible extraction of independently useful fixes.