Desktop/buzz-acp: idle pool teardown is a hidden 15-minute constant; expose it per agent and refresh the idle clock on turn completion #7

TL;DR: Desktop hard-codes BUZZ_ACP_IDLE_POOL_SLEEP=900 for agents and forbids overriding the key; after 15 minutes of "idle" time the harness tears down the whole worker pool together with all of the agent's sessions. The idle clock is refreshed only when a turn is dispatched, so a turn longer than 15 minutes loses its session right after it completes. We propose exposing the value in the agent card, porting upstream PR block#7173 and unifying the lazy flag across all launch paths.

Problem

Desktop launches every managed buzz-acp with BUZZ_ACP_LAZY_POOL=true and BUZZ_ACP_IDLE_POOL_SLEEP=900 (desktop/src-tauri/src/managed_agents/agent_env.rs:16, const IDLE_POOL_SLEEP_SECS: &str = "900"). The key is on RESERVED_ENV_KEYS (reserved_env_keys.rs:65), so Edit agent → Advanced → Environment variables rejects it, and there is no UI field. After 900 s without a dispatched turn the reaper tears the whole pool down (crates/buzz-acp/src/lib.rs, inactivity_expired at lib.rs:2196, "idle pool sleep bound reached — tearing pool back to lazy state") and with it the sessions map of every thread and channel this agent serves.

Two aggravating details:

1. The idle clock only moves on dispatch. last_activity is set at lib.rs:2894 (start), lib.rs:3636 (teardown) and lib.rs:4903 (dispatch). Turn completion does not refresh it. A turn that runs longer than 15 minutes is "idle" the moment it finishes and is reaped on the next 30-second tick. Observed 2026-09-29: dispatch 07:11:01, last reply 07:25:13, teardown 07:26:16 — one minute after the agent stopped talking.
2. Behaviour depends on how the agent was started. Agents spawned by the mention preflight path run with BUZZ_ACP_LAZY_POOL=false, IDLE_POOL_SLEEP=0 (observed in /proc/<pid>/environ on 2026-09-30: one agent eager, the four agents started via Start/Restart/launch-restore lazy with 900). So the same agent either never sleeps or sleeps after 15 minutes, undocumented, decided by which button spawned it. Note also that BUZZ_ACP_LAZY_POOL itself is not reserved, which upstream feat(buzz-acp): make BUZZ_ACP_IDLE_POOL_SLEEP configurable per agent, or document why it must not be block/buzz#6381 points out as a side door.

Why it matters

Session state is unrecoverable once the pool is gone (see the companion resume issue). For a human who steps away for 20 minutes, or for an agent that runs a 20-minute test suite, the reset is silent: the next mention gets a fresh session with 12 thread messages and no memory of what it was doing. The team's current workaround is a verbal protocol ("pause, we'll continue in the next session") plus hand-written handoff files.

Proposal

1. Expose the value. Remove BUZZ_ACP_IDLE_POOL_SLEEP from RESERVED_ENV_KEYS, or better, add an "Idle pool sleep (seconds, 0 = never)" field next to Parallelism in Edit agent → Advanced and in global agent defaults, the way Parallelism is exposed while BUZZ_ACP_AGENTS stays reserved. Keep 900 as the default.
2. Refresh the idle clock on turn completion and steer ack. Port upstream PR fix(buzz-acp): refresh idle clocks on turn completion and steer ack block/buzz#7173 (lib.rs +191, head 91957a54, OPEN since 2026-09-01).
3. One lifetime policy for all spawn paths. Start, Restart, launch restore and the mention-preflight spawn should pass the same lazy flag (today runtime_commands.rs:227 and restore.rs pass lazy=true, commands/agents.rs preflight passes lazy=false).
4. Log the teardown as an agent activity event so it is visible in the UI, not only in the harness log.

Cost to document in the UI: a warm pool is BUZZ_ACP_AGENTS (10) worker subprocesses at ~100 MB RSS each, ≈1 GB per awake agent. With 0 the pool lives until Restart or Desktop exit.

Benefit

Owners choose the trade-off (RAM vs. context survival) per agent instead of a hidden 15-minute constant; long turns stop killing their own session; identical behaviour regardless of how the agent was spawned.

References

• upstream: feat(buzz-acp): make BUZZ_ACP_IDLE_POOL_SLEEP configurable per agent, or document why it must not be block/buzz#6381 (make the value configurable, OPEN), fix(buzz-acp): refresh idle clocks on turn completion and steer ack block/buzz#7173 (PR, OPEN), buzz-acp: idle pool teardown kills an in-flight session waiting on background subagent notifications (queued wakeup lost, no stop bookkeeping) block/buzz#6378 (teardown kills turns waiting on background subagents), buzz-acp: idle-pool teardown race silently swallows a concurrently arriving message (no log trace beyond routine steer ack) block/buzz#7720 (teardown race swallows a message)
• fork: desktop/src-tauri/src/managed_agents/agent_env.rs:16, reserved_env_keys.rs:65, crates/buzz-acp/src/lib.rs:2196,3636,4903