RLM child model selection fails for all non-parent models under subscription OAuth; rlm.find_models returns empty

[jonaowen]

Summary

With a ChatGPT Plus/Pro (Codex Subscription) OAuth login as the only credential, RLM child agents can only inherit the parent session's model. Explicit model= selection fails admission for every model other than the parent's, and rlm.find_models returns [] for all queries — even though prime-agent model list shows the full openai-codex catalog and direct one-shot runs on those same models work fine.

Environment

• prime-agent 0.7.0 (stable channel), installed via npm tarball
• Windows 11 (Git Bash present at the documented path), Node v24.13.0
• Kernel via PRIME_AGENT_KERNEL_PYTHON pointing at a Python 3.11 venv (ipykernel + prime-agent-runtime), bootstrapped per docs
• Auth: /login → ChatGPT Plus/Pro (Codex Subscription) OAuth only; no API keys in env or auth.json

Repro

1. prime-agent model list → shows openai-codex/gpt-5.6-sol, gpt-5.6-luna, gpt-5.5, etc.
2. Direct one-shots work on any of them, e.g. prime-agent -p --no-session --provider openai-codex --model gpt-5.6-luna "Reply OK" → OK.
3. Inside a session (parent model openai-codex/gpt-5.6-sol), in IPython:
   • await rlm.find_models("gpt", limit=20) → [] (same for "", "luna", etc.)
   • await rlm("Reply DONE", name="p1", model="openai-codex/gpt-5.6-sol") → succeeds (parent model)
   • await rlm("Reply DONE", name="p2", model="openai-codex/gpt-5.5") → RuntimeError: Requested subagent model "openai-codex/gpt-5.5" is unavailable, unauthenticated, or expired
   • await rlm("Reply DONE", name="p3", model="openai-codex/gpt-5.6-luna") → same error

Expected

rlm.find_models surfaces models available under the active subscription credential, and explicit child model= admission succeeds for any model that same credential demonstrably serves (per the working one-shots).

Observed

RLM-side model discovery is empty and admission rejects all non-parent models. Behavior is identical in fresh -p --no-session runs, in long-running daemon sessions, and after a worker stop/resume, so it does not appear to be stale session state. Docs (rlm-runtime.md) say model search is "bounded to active, non-expired credentials" — subscription OAuth credentials appear to be excluded from that bound on the RLM path while working everywhere else.

Impact

Mixed-tier delegation (e.g. cheap gpt-5.6-luna children for mechanical work under a gpt-5.6-sol parent) is impossible for subscription-auth users; children silently only ever run the parent model. Current workaround: shell out to prime-agent -p --provider openai-codex --model gpt-5.6-luna one-shots from the kernel.

[ayushnangia]

Likely resolved by the pair #731 + #717. Root cause chain: with subscription OAuth, the model registry does live Codex catalog discovery — and it sends prime-agent's own version as client_version, which the backend reads as an ancient client and returns {"models": []} (probes in #702). The empty catalog then (a) filters every non-parent openai-codex model out of rlm.find_models and child admission — exactly this symptom — while (b) the parent session keeps working because it's already running. #731 fixes the discovery request (pinned Codex CLI client_version; restores the full catalog with the same bearer token), #717 makes an empty catalog fail open instead of silently disabling the provider. prime-agent model list showing the full catalog while RLM sees none is consistent: the static list vs the discovery-gated path. If you can test against those branches, this issue should close with them.

[liebero3]

Additional cross-provider reproduction from a subscription/OAuth setup: both Anthropic and openai-codex are linked via /login and work as lead sessions. From an Anthropic lead session, however, an explicit rlm(..., model="openai-codex/<model>") child cannot be created.

I have not confirmed the reverse direction (openai-codex parent → Anthropic child), so I am only reporting Anthropic → Codex here.

[kevinjosethomas]

Hi, thanks for taking the time to report this! Since open sourcing Prime Agent, we’ve received far more Issues than we can reliably triage and maintain. Rather than leave a large backlog that we cannot meaningfully work through, we’re closing the current Issue queue and moving to a discussion-first process.

We have established new contribution guidelines to help us continue iterating on Prime Agent and better manage contributions from the community. Going forward, GitHub Issues will be our maintainer work queue rather than the initial place for public reports and requests. Bug reports, feature requests, questions, and other feedback should begin in GitHub Discussions. Please read the full process documented in our contribution guidelines. Security vulnerabilities should continue to be reported privately according to our security policy.

While we’re closing this backlog, we’re still reviewing it at a high level to identify recurring bugs, useful ideas, and important problems that we should address ourselves. Thanks again for the time you put into this!