Codex App 26.715.9868.0 / CLI 0.145.0: spawn_agent rejects gpt-5.6-luna with multi_agent_v2 enabled

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #34399
• #34301
• #34024

Powered by Codex Action

Additional cross-platform confirmation from native Codex CLI/TUI 0.145.0 on Ubuntu 24.04.4 under WSL2:

• The top-level Select Model and Effort picker lists gpt-5.6-luna.
• Native spawn_agent with fork_turns: "none", model: "gpt-5.6-luna", and reasoning_effort: "medium" is rejected with the same error:

``text
Unknown model gpt-5.6-luna for spawn_agent. Available models: gpt-5.6-sol, gpt-5.6-terra
``

• Positive controls from the same parent thread completed successfully:
• gpt-5.6-sol with reasoning_effort: "max"
• gpt-5.6-terra with reasoning_effort: "low"

This specifically reproduces through native spawn_agent, not top-level model selection or codex exec --model. I originally filed #34909 with the full payload and am closing it as a duplicate of this issue.

Additional Windows confirmation from a newer Codex Desktop package:

• Codex Desktop: 26.721.4979.0
• Codex CLI: 0.145.0
• Windows 10 Pro x64, build 19045
• Feature state:
• multi_agent = true
• multi_agent_v2 = false
• The local model catalog includes gpt-5.6-luna.

A native subagent request using fork_turns: "none", model: "gpt-5.6-luna", and reasoning_effort: "low" is still rejected:

Unknown model 'gpt-5.6-luna' for spawn_agent. Available models: gpt-5.6-sol, gpt-5.6-terra

No Luna child is created. This suggests the spawned-agent model allowlist mismatch remains present in Desktop 26.721.4979.0 / CLI 0.145.0, and is not limited to sessions with the opt-in multi_agent_v2 flag enabled.

I traced this beyond the picker guard. #32751 is masking a real runtime mismatch, not resolving it.

Current V2 spawn first copies the parent config, then applies model = gpt-5.6-luna. But AgentControl::spawn_agent_internal resolves the child runtime from the parent/session path: explicit V2 config wins, otherwise it inherits the parent thread's V2 runtime. It does not resolve the selected child model's ModelInfo.multi_agent_version.

So removing the #32751 filter would start a Luna child on the V2 runtime despite Luna being catalogued V1. That explains why the filter was added, but it also means the user-facing behavior regressed from “selectable” to a misleading “unknown model.”

There are two valid fixes, and they are materially different:

1. If Luna is meant to support the V2 collaboration backend, update the authoritative model metadata/backend rollout to V2. This is the minimal fix.
2. If Luna is intentionally V1, implement a heterogeneous bridge: create the child with its own V1 runtime and adapt V2 spawn/message/completion semantics. V2 currently uses InterAgentCommunication/agent paths; V1 spawn uses UserInput and legacy completion flow. Merely selecting the V1 child runtime is insufficient.

Could a maintainer confirm which of those is the intended product state? Separately, the current error should say “known model, incompatible multi-agent backend” rather than “Unknown model.”

Strong +1. I hit this exact limitation today: spawn_agent accepts gpt-5.6-sol and gpt-5.6-terra, but rejects gpt-5.6-luna. Luna should be available for delegated subagents, including configurable reasoning effort, rather than forcing users onto a different model.

Additional macOS confirmation from a newer Codex Desktop / CLI build:

• Codex Desktop: 26.727.40816 (bundle build 6067, com.openai.codex)
• Bundled Codex CLI: 0.146.0-alpha.9.2
• macOS: 26.5.2 (25F84), Apple Silicon (arm64)
• Feature state:
• multi_agent = true
• multi_agent_v2 = true

I reproduced this in a fresh read-only CLI session, rather than a resumed Desktop thread:

• parent model: gpt-5.6-sol
• parent reasoning effort: max
• spawn_agent arguments:
• agent_type = "default"
• fork_turns = "none"
• model = "gpt-5.6-luna"
• reasoning_effort = "max"

Actual result:

Unknown model `gpt-5.6-luna` for spawn_agent. Available models: gpt-5.6-sol, gpt-5.6-terra

No child agent was created. This confirms that the spawned-agent Luna rejection persists in CLI 0.146.0-alpha.9.2 and is not limited to the 0.145.0 builds previously reported here.

Yup, this is happening to me too! Can only spawn with the CLI

Its kinda crazy this is still an open issue...

Its kinda crazy this is still an open issue...

Yup, this staying around so long, plus other issues over the past few weeks with codex-cli made me switch to pi instead. Not a 1-to-1 transition by far, but it doesn't have all these issues that codex introduced, _and_ it's easy to fix issues yourself/using an LLM since pi is very extensible & open source. It also supports codex subscriptions out-of-the-box, (thank you OpenAI for explicitly allowing this!), so it's easy to get started with pi and a codex subscription :)

Reproduced on macOS as well, with the current npm release.

Environment:

• macOS 26.5 (25F71), arm64
• codex-cli 0.146.0
• ChatGPT-authenticated Codex App / native subagent surface
• multi_agent = true
• multi_agent_v2 = true
• [agents] default_subagent_model = "gpt-5.6-luna"
• Dedicated ~/.codex/agents/luna.toml declaring model = "gpt-5.6-luna" and model_reasoning_effort = "high"

After fully restarting Codex, the custom luna agent type was successfully discovered and appeared in the native agent catalog. Spawning it with agent_type = "luna" and fork_turns = "none" (with no explicit model override, so the profile owns model selection) still failed immediately with:

Unknown model `gpt-5.6-luna` for spawn_agent. Available models: gpt-5.6-sol, gpt-5.6-terra

No child was created. I also reproduced the same rejection from a fresh codex exec process after confirming codex features list reports both multi_agent and multi_agent_v2 as enabled.

This suggests profile discovery/config parsing is working, while the child dispatcher model catalog remains limited to Sol/Terra. It is therefore not Windows-specific and persists in CLI 0.146.0.

Confirming the same model-catalog versus subagent-allowlist mismatch from a Codex Desktop task attached to a Linux host.

Environment:

• Codex CLI: 0.146.0
• Host: Linux 6.8.0-106-generic x86_64
• ChatGPT authentication
• multi_agent: stable/enabled
• multi_agent_v2: stable/disabled in local config
• models_cache client version: 0.146.1, fetched 2026-08-05

The freshly fetched local model catalog lists gpt-5.6-luna with visibility=list, and Luna is selectable as the primary agent in the ChatGPT/Codex model selector. However, the spawn_agent contract generated for this task exposes only gpt-5.6-sol and gpt-5.6-terra as valid model overrides; gpt-5.6-luna is absent.

This indicates that primary-model availability and the subagent worker allowlist are diverging. There is no local config key in this environment that controls that allowlist. We are using Terra workers as a temporary fallback.

Same here. Just commenting for visibility, as it's been 3 weeks and this is still open.

I can reproduce the same model-catalog mismatch on a newer Windows Codex Desktop build.

Environment:

• Codex Desktop package: 26.803.8161.0 (read from the installed Windows package path)
• Platform: Windows x64
• Subscription tier: not exposed to the agent, so I am not guessing
• Bundled CLI version: not independently verified because the executable under C:\Program Files\WindowsApps\... is access-restricted from the tool shell

Current behavior in the same Desktop session:

• The parent-thread model picker offers GPT-5.6 Luna, including Max reasoning.
• The native collaboration.spawn_agent tool schema exposes only these explicit model overrides: gpt-5.6-sol and gpt-5.6-terra.
• gpt-5.6-luna is absent from that schema/allowlist, so a Luna child agent cannot be selected even though Luna is selectable for the parent thread.

This is therefore still reproducible on Desktop package 26.803.8161.0, newer than the originally reported 26.715.9868.0. The mismatch appears to be between the Desktop model catalog/picker and the model allowlist supplied to native spawn_agent, rather than a user prompt or reasoning-effort error.

Official model guidance also lists gpt-5.6-luna and says GPT-5.6 supports Max reasoning: https://developers.openai.com/api/docs/guides/latest-model

same problem

+1. please get this done.