fix(cross-model): accept provider-qualified codex model ids in cross_model_model

[DennisTraub]

Summary

A codex CLI pointed at a non-default model_provider may require ids in that provider's own namespace. This PR enables compound-engineering users on such a host to run the cross-model pass at all, by pointing cross_model_model at the id their endpoint actually serves.

Today they cannot. The workers pin a bare OpenAI id, and validate_model_override matched only codex:gpt-*|codex:o[0-9]*, so the escape hatch shares the assumption of the default it exists to override. The peer sends the document, the endpoint 404s on the model name, and the pass logs a skip: egress paid, no review back. Measured against bedrock-mantle.<region>.api.aws/openai/v1, gpt-5.6-luna 404s there and openai.gpt-5.6-sol serves.

The codex arm now allows a namespace segment while keeping the family anchor, so bedrock.claude-opus-5 on the codex route still exits 2 and a stock install is unaffected.

codex:gpt-*|codex:o[0-9]*|codex:*[./]gpt-*|codex:*[./]o[0-9]*

The change touches the codex arm in ce-doc-review, ce-code-review, and ce-pov. Fixes #1500.

Two things deliberately not changed

The claude route keeps its arm as it is. The claude binary carries an internal model table pairing each first-party id with its per-provider served id, so --model claude-opus-5 appears to resolve to the provider's own form rather than being forwarded verbatim, which is what codex does with -m. I read that from the shipped binary and could not run it end to end, since the Claude Code install here is on the first-party API. Widening that arm would also mean teaching expected_model_prefix to strip the namespace, or the served-model receipt degrades to a first-key match with a warning.

ce-work is not touched. Its codex route cannot reach a custom provider at all, because the adapter passes --ignore-user-config, which drops the provider definition along with the reasoning tier it is there to control. codex then falls back to its built-in provider and fails on credentials before any model name is checked:

$ codex exec --ignore-user-config -m openai.gpt-5.6-sol --skip-git-repo-check -s read-only "Reply OK"
ERROR: unexpected status 401 Unauthorized: Missing bearer or basic authentication in header,
  url: https://api.openai.com/v1/responses

An earlier revision widened its validator for parity. Per the review note on this PR, that would have left route_model_allowed() and the embedded model_allowed() rejecting the same id at authorization time, so the hunk is gone and ce-work rejects the namespaced form consistently. How that route should reach a custom provider is a separate design question about how much of a user's codex config the unit workspace inherits.

Validation

• bun test tests/skills/ce-doc-review-cross-model-routes.test.ts 54 pass, 0 fail
• bun test tests/skills/ce-pov-cross-model-routes.test.ts tests/skills/ce-code-review-cross-model-routes.test.ts tests/skills/ce-work-cross-model-routes.test.ts 182 pass, 0 fail across the three files
• I ran each new assertion against its pre-fix script first. Each one fails there, so the tests pin the behavior rather than the patch.
• bun run test 3454 pass, 1 skip, 1 fail. The failure is ce-work unit workspace controller > unit and plan-wide verification ignore inherited Git local environment, which reproduces on unpatched origin/main in a clean worktree on this machine, so it does not come from this change.
• bun run release:validate metadata in sync
• bun run plugin:validate both manifests pass
• Per AGENTS.md, the config-option meaning change also lands in skills/ce-setup/references/config-template.yaml, its byte-identical .compound-engineering/config.example.yaml copy (md5 still matches), docs/skills/configuration.md, and the ce-code-review and ce-doc-review pages

The existing model-override validation stays byte-identical across review workers parity test still passes; the doc-review and code-review validators remain identical.

Security Disclosure

This widens a fail-closed allowlist for a model-id token that was already passed as a single argv element and never enters eval. No shell or exec surface is added, no path or input handling changes, no credential, permission, or dependency changes. The value stays bound to its declared target and to the gpt or o-series family, so a cross-family override on the codex route still exits 2.

Agent Disclosure

• Model: Claude Code · claude-opus-5[1m]

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 25118128cf

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 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 address that feedback".

[tmchow]

Nice fix on this @DennisTraub. I'm just going to have an agent rebase against main as I merged in a huge refactor just now.

[tmchow]

I pushed 2 small fixes

[DennisTraub]

Thanks for the quick merge - and for creating this plugin in the first place. I use it for almost everything I do and can't imagine working without it anymore! Happy to keep contributing if I come across anything else.