[codex] Select multi-agent version from model info

[aibrahim-oai]

Why

#25351 locks the selected multi-agent runtime per thread. Feature flags remain the fallback for fresh threads, but backend model metadata also needs to select the runtime for models that opt into a specific system.

What changed

• Add optional multi_agent_version metadata to ModelInfo. Omitted or unrecognized metadata behaves as absent, preserving feature-flag fallback and allowing older clients to tolerate future selector values.
• Resolve selected model metadata after persisted and inherited thread locks, but before feature flags.
• Keep ordinary child inheritance authoritative, so a V2 parent remains V2 when spawning a child whose selected model advertises V1.
• Keep explicit Disabled inheritance authoritative for review and guardian delegates, so a model advertising V2 cannot re-enable multi-agent tools for them.
• Leave bundled model catalog entries unset.

Test plan

CI only, per repository workflow.

Added end-to-end remote-model integration coverage using the real /models mock and real agent flows:

• A root model advertises V2 while only the V1 feature is enabled; its child model advertises V1, but the spawned child inherits the parent V2 lock.
• A guardian delegate remains locked to Disabled when its selected model advertises V2.

Added deserialization coverage that treats an unknown multi_agent_version catalog value as omitted.

Stack

Depends on #25351.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b52a0177ac

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c3c94a7d86

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Record model-selected agent version in config locks

When config-lock export/replay is enabled and the remote catalog selects V2 here without the local Feature::MultiAgentV2 flag, the selected version is kept only in the session OnceLock; checked session/config_lock.rs, and the lockfile still serializes features.multi_agent_v2.enabled from config.features only. That means a run that actually exposed V2 tools can export a lock saying V2 is disabled, so replay can still validate after the catalog selector changes and silently run with V1/no multi-agent tools instead of reproducing the original behavior.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Avoid rejecting legacy agent limits on catalog V2 rollout

If /models starts returning multi_agent_version: v2 for a model while a user still has the legacy agents.max_threads setting, this line now feeds V2 into effective_agent_max_threads, which returns InvalidRequest whenever agent_max_threads is set. That makes session startup fail purely because server metadata changed, even though the same config remains valid for V1/Collab and the user did not opt into the local V2 feature; the catalog selector should either tolerate/translate the legacy limit or avoid hard-failing existing configs during rollout.

Useful? React with 👍 / 👎.

[aibrahim-oai]

This change is part of the following stack:

• [codex] Lock multi-agent runtime version per thread #25351
  • [codex] Select multi-agent version from model info #25427 ◀

_{Change managed by git-spice.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 28d732dd4d

ℹ️ About Codex in GitHub

Your team has set up Codex to 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 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Keep legacy rollouts on their original runtime

When resuming a thread whose rollout predates persisted SessionMeta.multi_agent_version, conversation_history.get_multi_agent_version() is None, so this newly added catalog fallback can lock the thread to whatever /models returns today. If the backend starts advertising multi_agent_version: v2 for a model, an older thread that was created with the V1 multi_agent_v1 tool surface will resume as V2 and expose the plain V2 tools, even though its existing context can contain V1 tool calls/outputs; for these legacy rollouts the safer fallback is the feature-derived runtime that was in effect before model metadata existed, or another compatibility inference from the stored history.

Useful? React with 👍 / 👎.