feat: add agent execution slot accounting

[jif-oai]

Summary

Separates logical agent registration from active execution-slot accounting for multi-agent sessions. This keeps spawned agents addressable after they become idle while allowing V2 pending turns to reserve capacity only when they actually need to run.

The V2 configured concurrency limit continues to include the root thread, so spawned-agent execution uses the non-root capacity derived from that limit.

Changes

• Track active execution slots separately from logical agent metadata.
• Add execution reservations for idle agents that wake up from queued trigger-turn mail.
• Release V2 execution slots once a thread is idle and has no trigger-turn mailbox work.
• Preserve existing V1 spawned-thread accounting.

Stack

1. refactor: split agent control modules #26610 - Mechanical AgentControl split: extracts spawn and V1 legacy code without behavior changes.
2. This PR - Execution slot accounting: separates logical agents from active execution slots.
3. feat: add agent residency reload runtime #26611 - Residency and reload runtime: adds resident-agent LRU, eviction/reload, durable lookup, and V2 delivery through reload.
4. feat: make v2 close_agent interrupt only #26612 - V2 tool semantics: narrows close_agent to interrupt-only and updates V2 tool coverage.

[chatgpt-codex-connector]

💡 Codex Review

codex/codex-rs/core/src/agent/control/spawn.rs

Lines 578 to 580 in e92893d

	let mut agent_metadata = agent_metadata;
	agent_metadata.agent_id = Some(resumed_thread.thread_id);
	reservation.commit(agent_metadata.clone());

Release idle slots after resuming agents

Resuming a closed V2 agent now commits a spawn reservation, which inserts the thread into active_thread_ids, but resume_agent does not start a turn or call the idle release path afterward. In a session that resumes an idle agent (or its open descendants) and then leaves it idle, that agent keeps consuming max_concurrent_threads_per_session capacity indefinitely, so subsequent runnable agents can be rejected even though nothing is executing. Please release the execution slot for resumed V2 threads unless they immediately have work to run.

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".

[parasol-aser]

agent/registry.rs — release_execution_slot / reserve_spawn_slot (shared total_count)

This routes idle releases through the same counter the spawn path checks, so an idle-but-alive sub-agent frees spawn budget. The cap now bounds concurrently executing threads rather than live threads: a parent agent can spawn A → A finishes and idles → spawn B → … and accumulate an unbounded set of long-lived idle sub-agent threads, each still holding a Session, a rollout file handle, and possibly MCP connections. Spawn depth still bounds the tree, but nothing bounds breadth over time, and I don't see an idle-thread reaper anywhere in codex-rs/core/src (release_spawned_thread only fires on death/close). This looks like the intended design of the feature, so this is mostly a question of confirming the resource model rather than a clear bug.

Fix: if a bound on total live sub-agents is still wanted, add a separate live-thread cap or an idle-thread reaper; otherwise it's worth documenting the new model and confirming the per-idle-thread footprint is negligible.