buzz-acp: resume agent sessions via session/load (SessionStore keyed by channel + thread root) #6

TL;DR: buzz-acp never restores an agent session: after the pool sleeps (15 min), Stop/Start, Restart or a Desktop restart, the agent gets a new session and loses all accumulated context. The adapters (claude-agent-acp, codex-acp) support session/load, and the transcripts are on disk. We propose porting upstream PR block#6088 and extending the session key to the thread.

Problem

buzz-acp only ever creates sessions: the single ACP session call is session/new (crates/buzz-acp/src/acp.rs:758). The map of live sessions sessions: HashMap<SessionScope, String> lives inside the AgentPool (crates/buzz-acp/src/pool.rs:133) and is wiped by SessionState::invalidate_all (pool.rs:219) whenever the pool is torn down. That happens on:

• idle pool sleep (BUZZ_ACP_IDLE_POOL_SLEEP=900, see the companion issue on idle teardown);
• Stop → Start or Restart of the agent in Desktop;
• Desktop quit/restart (shutdown: waiting for in-flight prompts → grace period → abort);
• worker crash / cold respawn after a failed turn (upstream buzz-acp: failed/steered turns respawn cold sessions with ~2.4s capped backoff (token-burn storm); ephemeral event kinds bypass respond_to gate block/buzz#5327).

The next mention opens a fresh session ([session/query] … resume=none in the agent log) and the harness injects only the thread root plus the last BUZZ_ACP_CONTEXT_MESSAGE_LIMIT (12) replies as <thread-context>. Everything the model had reasoned about, intermediate decisions, file paths it discovered, subagent results — gone.

Evidence

• 2026-09-29, one agent (session policy thread), harness log agents/logs/<agent>__<session>.log: turn dispatched 07:11:01, last message 07:25:13, idle pool sleep bound reached — tearing pool back to lazy state at 07:26:16, new session 59aac60b… with resume=none at 07:44:54 on the next mention. Old transcript ~/.claude/projects/<workspace>/<session>.jsonl (2.4 MB) was intact on disk.
• The owner's expectation (2026-09-29): "I set each thread to start its own session with the agent and expected it to persist until I close Buzz." Since then the team runs a manual ritual in every long thread: "let's pause, we'll continue in the next session", "write a prompt to start the next session" — handoff files and restart prompts written by hand because the harness cannot resume.

What already exists

• @agentclientprotocol/claude-agent-acp ≥ 0.79 implements loadSession → createSession(..., { resume: sessionId }), i.e. native Claude Code resume from the on-disk transcript. codex-acp also advertises loadSession: true in its initialize response.
• Upstream PR buzz-acp: resume channel sessions across a restart, and name the channel on session/new block/buzz#6088 ("buzz-acp: resume channel sessions across a restart, and name the channel on session/new", head 313d5fd4, +1753/−52, files acp.rs, config.rs, lib.rs, pool.rs, new session_store.rs): a SessionStore at <state-dir>/sessions-<pubkey>.json, session/load when there is no live session, fallback to session/new on JSON-RPC error, flags BUZZ_ACP_STATE_DIR / BUZZ_ACP_NO_RESUME_SESSIONS. Maintainer confirmed the basic scenario and requested changes (2026-08-18); author replied 2026-08-23; no activity since, mergeable_state=dirty. Still OPEN as of 2026-10-09.

Proposal

1. Port buzz-acp: resume channel sessions across a restart, and name the channel on session/new block/buzz#6088 onto fork/main as a registered fork patch (scripts/fork/patches.json, per docs/fork/MAINTENANCE.md).
2. Extend the store key from channel_id to the full SessionScope (crates/buzz-acp/src/scope.rs:71): Conversation { channel_id } and Thread { channel_id, root_event_id }. Thread sessions landed in upstream main on 2026-09-18, after the PR was written.
3. Cover the teardown path explicitly: lazy-pool sleep → empty AgentPool::from_slots → next dispatch must call session/load with the stored id, not session/new. Add a test for it.
4. Gate on the runtime capability from initialize (loadSession); runtimes without it keep today's behaviour.
5. On successful load, skip the full <thread-context> re-injection and deliver only the delta (the delivered_event_ids bookkeeping would also need to be persisted, or the harness re-sends the window once and accepts the duplication).
6. Document the limitation: resume restores conversation history, not in-flight turn state (background subagents, scheduled wakeups do not come back).

Benefit

• Agents survive the 15-minute idle teardown, Desktop restarts and Restart buttons with their context intact; no more hand-written handoff prompts between "sessions" of the same thread.
• Fewer tokens: a resumed session does not re-read the thread, the repo and the feature directory from scratch.
• Thread-scoped sessions (the policy the owner relies on) finally mean what the UI promises.

References

• upstream: buzz-acp: resume channel sessions across a restart, and name the channel on session/new block/buzz#6088, [Bug] buzz-acp loses channel sessions on restart and creates duplicate Codex cloud threads block/buzz#5342 (loses channel sessions on restart), buzz-acp: failed/steered turns respawn cold sessions with ~2.4s capped backoff (token-burn storm); ephemeral event kinds bypass respond_to gate block/buzz#5327, buzz-acp: bound ACP session + provider-resource lifecycle (no eviction today) block/buzz#6958 (session lifecycle)
• fork: crates/buzz-acp/src/acp.rs:758, pool.rs:133,219, scope.rs:71