fix(coding-agent): send a supported codex client_version in model discovery

[MingTeer]

Summary

ModelRegistry.getExecutableModels() passed the prime-agent package version as client_version when discovering ChatGPT codex models. The backend gates the catalog by codex CLI version, so 0.7.0 read as an ancient client and returned {"models": []}, hiding every openai-codex/* model from find_models and rlm() subagent spawns, which then failed with the misleading "unavailable, unauthenticated, or expired" error.

Changes

• Send a pinned, known-good codex CLI version (OPENAI_CODEX_CLIENT_VERSION = "0.146.1") as client_version instead of the prime-agent package version, with a comment noting the floor must be bumped if the catalog ever comes back empty.
• Treat an empty catalog as a discovery failure instead of caching it as a valid result for 5 minutes, so a later non-empty response can recover within the same session.
• Add regression tests covering both behaviors (test/suite/regressions/702-codex-client-version.test.ts).

Verification

• New regression tests pass (2/2), adjacent 4649-subagent-model-selection suite passes (12/12)
• npm run check clean
• Reproduced end-to-end on macOS arm64: with the fix, openai-codex/* models appear in find_models and subagent spawns succeed with a valid OAuth credential

Fixes #702

Note

Fix OpenAI Codex model discovery to send a supported client_version

• Replaces the prime-agent package VERSION with a hardcoded Codex CLI version ("0.146.1") in the client_version query parameter of openAICodexModelsUrl, since the backend gates access by Codex CLI version.
• Treats an empty model catalog (200 response with models: []) as a discovery failure rather than caching it as a successful empty result, causing callers to fall back to excluding Codex models until a non-empty discovery succeeds.
• Adds regression tests in 702-codex-client-version.test.ts covering both fixes.

^{Macroscope summarized c9d1377.}

[ProDrifterDK]

Validated this fix independently against current main (a18809e0) by reapplying it in ProDrifterDK/prime-agent@a2f910e3.

Environment:

• Prime Agent source: 0.7.1
• Pop!_OS 24.04 LTS, kernel 7.0.11-76070011-generic, x86_64
• OpenAI Codex subscription OAuth

Evidence:

• client_version=0.7.1 returned HTTP 200 with models: [].
• client_version=0.146.1 returned 9 models, including gpt-5.6-sol and gpt-5.6-luna.
• The issue-specific regression passed: 2/2 tests.
• The Codex discovery and subagent-selection regressions passed together: 14/14 tests.
• npm run check passed.
• An isolated live smoke test on a separate daemon socket succeeded: a Codex Sol parent discovered openai-codex/gpt-5.6-luna, admitted a Luna RLM child, and received SMOKE LUNA MESSAGE OK through agent_message.

This confirms both the catalog fix and the real Sol-to-Luna RLM path without modifying the globally installed runtime.

[sethkarten]

Thank you for the report and proposed work. This root cause is now covered by maintainer-owned stacked PR #1164, authored independently from upstream/main.

We did not inspect or reuse this PR's diff, branch, commits, implementation code, or tests; its public description/comments were used only as a bug report. To keep one review surface, this PR is superseded by #1164 and is being closed.

The complete review stack is #1158–#1165. It is being left unmerged for human review after CI and review-bot findings are cleared.