feat(coding-agent): keep ACP daemon sessions resident

[sethkarten]

Why

ACP hardcoded a client-owned daemon session, so the worker — and with it the live IPython kernel — stopped when the ACP client disconnected. Any ACP consumer that wants state to survive a turn boundary therefore had to keep one long-lived client process alive for the whole rollout, and build keepalive, transport and teardown machinery around it.

resident was already on the wire (daemon-protocol.ts: DaemonSessionLifecycle = "resident" | "client_owned"), so ACP simply never asked for it.

What changed

clientOwned: appMode !== "acp" at the daemon-client call site. ACP gets a resident worker; rpc, print and json keep client-owned lifetimes unchanged.

No protocol or schema change, and no new capability. An earlier draft of this branch bumped DAEMON_PROTOCOL_VERSION and negotiated an acp_resident_sessions capability; both were removed. The bump was actively harmful — startup matches version and schema exactly, so protocol-7 daemons would go stale before any fallback could engage, letting a busy old daemon block ACP startup. A silent fallback to client-owned was also rejected: it would quietly destroy the persistence this change exists to provide.

Re-attach

A resident worker is only useful if a later client can get back to the session. Prime Agent already supports this: relaunching ACP with --continue resolves the resident session by canonical session file and attaches to it. No session/load and no new ACP method are required.

Validation

• npm run check
• test/acp-resident-lifecycle.test.ts and test/daemon-protocol.test.ts: 14 passed, 1 skipped (the skipped one is kernel-heavy, run separately below)
• kernel-heavy re-attach test, run explicitly: passes in ~12s against a live kernel. It sets an object() marker, disconnects ACP stdio, relaunches with --continue, and asserts live namespace identity — object() cannot survive a dill snapshot, so this proves the same interpreter, not a revived one.
• mutation: forcing clientOwned: true fails that test (65s, no kernel survives); restoring passes (12s). Reverting protocol to 8 fails the schema/version assertions.

Two pre-existing failures in acp-features (goal timeout, post-close new session) reproduce on main and are unrelated.

---

Note

High Risk
Changes daemon session lifecycle, worker recovery, and on-disk descriptors plus credential handling; mistakes could leak secrets, break ACP/embedder model access, or leave workers in the wrong lifecycle after disconnect.

Overview
ACP daemon lifecycle now treats session-file ACP runs as resident workers (survive client disconnect; --continue can reattach), while ACP --no-session and other modes stay client-owned. isClientOwnedDaemonSession encodes that rule at the daemon create call site.

Launch environment is always sent on create (launchEnv: collectDaemonLaunchEnv() for both lifecycles) so the first worker spawn gets caller endpoint/credentials/proxy settings. Only non-secret keys are written to resident worker descriptors via filterPersistedDaemonLaunchEnv / DAEMON_PERSISTED_LAUNCH_ENV_KEYS; on load, descriptors are sanitized and rewritten if needed. Supervisor recovery relaunches resident workers from the persisted allowlist (not ambient supervisor secrets); client-owned workers keep transient env from the reconnecting owner. Promoting owned→resident clears launchEnv on the descriptor.

Tests cover lifecycle negotiation, allowlist filtering, supervisor restart + worker recovery without credential persistence, and a kernel-heavy ACP disconnect/--continue reattach integration test.

^{Reviewed by Cursor Bugbot for commit 0626d45. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Keep ACP daemon sessions resident across supervisor restarts

• ACP sessions with a session file are now classified as resident (not client-owned) via the new isClientOwnedDaemonSession utility; --no-session ACP invocations remain client-owned.
• Resident workers persist and recover a safe allowlist of launch environment variables (defined in DAEMON_PERSISTED_LAUNCH_ENV_KEYS) across supervisor restarts, while client credentials and secrets are excluded.
• When a resident worker is auto-recovered, the supervisor builds its environment from a filtered copy of process.env to avoid leaking ambient secrets from the previous worker process.
• Promoting a client-owned worker to resident clears its launchEnv from the descriptor so transient credentials are not persisted.
• Behavioral Change: daemon worker descriptors now include an optional launchEnv field; existing descriptors are sanitized on load to remove any non-allowlisted keys.

^{Macroscope summarized 0626d45.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d4de117. Configure here.}

[sethkarten]

Closing this superseded ACP implementation. Resident recovery coverage continues in #1236, with the exact causal prompt lifecycle successor in #1239. This PR remains available as design and review history.