fix(coding-agent): keep Codex models on empty discovery

[sethkarten]

Summary

• Treat an empty successful OpenAI Codex model catalog as inconclusive instead of disabling every locally known Codex model.
• Keep non-empty account catalogs authoritative, preserving entitlement filtering.
• Add registry coverage for fresh, cached-empty, and non-empty subset catalogs.
• Add an end-to-end RLM regression proving a Sol parent can discover and explicitly spawn a Luna child.
• Document the user-visible fix in the coding-agent changelog.

Fixes #1011.
Addresses the executable-model behavior in #696; warning and error-message improvements remain separate follow-up work.

This consolidates and supersedes #717 and #1012, with co-author credit retained in the commit.

Evidence

On Prime Agent 0.7.1 with the same authenticated Codex account:

• direct openai-codex/gpt-5.6-luna execution succeeded;
• rlm.find_models("luna") returned no result;
• explicit Luna child admission failed until the parent model was temporarily switched to Luna.

The inconsistent behavior comes from treating HTTP 200 {"models":[]} as an authoritative no-model entitlement response. The inference request remains the authoritative check when discovery is empty.

Safety

• Missing or invalid authentication still excludes Codex models.
• Missing account IDs still exclude Codex models.
• Malformed, non-2xx, and failed discovery requests still fail closed without a fresh cache.
• Non-empty discovery catalogs still restrict execution to their returned model IDs.

Verification

• env -u GEMINI_API_KEY npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/696-empty-codex-catalog.test.ts test/suite/regressions/1011-luna-subagent-selector.test.ts test/suite/regressions/4649-subagent-model-selection.test.ts
  • 3 files passed, 16 tests passed.
• npm run check
  • formatting, linting, type checking, installer rendering, and browser smoke checks passed.

Related

• RLM excludes OpenAI Codex models because discovery sends Prime Agent version as client_version #639
• openai-codex models invisible to rlm subagents: model discovery sends prime-agent VERSION as codex client_version, backend returns empty catalog #702
• RLM child model selection fails for all non-parent models under subscription OAuth; rlm.find_models returns empty #740

---

Note

Medium Risk
Changes Codex model gating used by RLM and model selection; behavior is narrower (less aggressive filtering) but still fails closed on bad auth, missing account ID, and failed discovery without a usable cache.

Overview
OpenAI Codex executable-model filtering no longer treats a successful discovery response with an empty models list as “no entitlements.” In getExecutableModels, a new filterByCatalog helper skips Codex ID filtering when the catalog set is empty, so locally known authenticated models (e.g. GPT-5.6 Luna) stay selectable; non-empty catalogs still restrict Codex models to returned slugs, including cache hits and error fallbacks.

Adds regression tests for empty, cached-empty, and subset catalogs in the model registry, plus an RLM harness test that finds and spawns a Luna child from a Sol parent when discovery returns []. Changelog entry documents the user-visible RLM/subagent fix (#696, #1011).

^{Reviewed by Cursor Bugbot for commit 5bc67f4. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix ModelRegistry.getExecutableModels to keep Codex models when discovery returns empty catalog

• getExecutableModels previously filtered out all openai-codex models when the discovered model ID set was empty, effectively hiding authenticated Codex child models.
• Introduces a filterByCatalog helper in model-registry.ts that treats an empty discovered set as inconclusive, skipping the filter entirely.
• The fix applies to both the cache-hit path and the fresh-fetch path, so cached empty catalogs also no longer hide Codex models.
• Regression tests added in 696-empty-codex-catalog.test.ts and 1011-luna-subagent-selector.test.ts covering the empty-catalog and Luna subagent discovery scenarios.
• Behavioral Change: a non-empty catalog remains fully authoritative and still restricts executable models to only the discovered subset.

^{Macroscope summarized 5bc67f4.}

[sethkarten]

Closed in favor of #1153. This PR aggregated implementation and tests from external contributions, which was not the intended requirement. #1153 is an independent implementation written from upstream/main without importing outside commits, source code, or test files. Do not merge this PR.