Non-worker draft discard: unobserved closeSession rejection can exit the process; pendingAttaches window discards a draft mid-attach

[snimu]

Two latent bugs in the non-worker daemon draft-discard path (packages/coding-agent/src/modes/daemon/daemon-mode.ts), found by Macroscope while reviewing #1897 but pre-dating that stack (both verified present on main). They only bite when AgentDaemon runs with options.worker unset (isDiscardableDraft returns false in worker mode), i.e. in-process/SDK-style sessions.

1. Unhandled rejection on draft discard can kill the process. Both discard sites call void this.closeSession(state, "killed") (broadcastToSession, and the queueMicrotask in detachClientFromSession). The daemon installs an unhandledRejection handler that exits the process, so a rejection during best-effort teardown terminates it instead of failing the discard. Fix: observe the rejection (.catch + log), matching other best-effort teardown calls in the file.

2. isDiscardableDraft ignores pendingAttaches. attach increments state.pendingAttaches (daemon-mode.ts:3698) before the client joins state.clients, but isDiscardableDraft checks only clients.size. A turn_end/compaction_end/bash_end in that window discards a draft that has an in-flight attach, and the attach then fails its liveness check. Fix: return false when state.pendingAttaches > 0.

Both fixes are one-liners plus a regression test each; kept out of #1897 because the behavior pre-exists the roster stack.

[snimu]

From PR #1926 review (macroscope): get_rlm_children (daemon-mode.ts, worker command handler) returns only resident children via getRlmChildSnapshots(), while the attach/snapshot path merges passivated children via buildRlmChildSnapshotsWithPassiveRlmSubagents — so a client querying the command gets an incomplete roster after passivation, inconsistent with the advertised authoritative_child_roster capability. Pre-existing on main / the roster stack (byte-identical handler at ae22984); not introduced by #1926. Likely fix: use the merging builder in the handler.

[bounce12340]

Working on this.

[bounce12340]

Fix is in #2029 (f98109c6). Both non-worker discard sites now .catch + log closeSession failures, and isDiscardableDraft returns false while pendingAttaches > 0. Regression tests cover the unhandledRejection path and the in-flight attach window. packages/coding-agent vitest --run test/daemon-mode.test.ts: 189 passed.

The later get_rlm_children roster comment is a separate bug and is not in this PR.

[aarmens702-hub]

Picking up the get_rlm_children roster item from the review comment above (separate from the draft-discard fix in #2029).

Reproduced at d74a75f: after a child completes and its runtime is closed, buildRlmChildSnapshotsWithPassiveRlmSubagents (the attach path) still lists it, but the get_rlm_children handler returns [] because it reads session.getRlmChildSnapshots() directly. ACP reads this command at turn end and on cancel, so embedding clients get the shorter roster.

Fix on my fork: main...aarmens702-hub:prime-agent:fix/1922-rlm-children-passive-roster

• Handler now uses the same merging builder as attach, so passivated children and the daemon activeSessionId decoration come through. Response shape is unchanged (backward-compatible; no protocol bump).
• Regression test in test/daemon-mode.test.ts next to the existing passive-discovery test: red against the main handler (expected [] to deeply equal ['child-1']), green with the fix. Full daemon-mode.test.ts and the ACP/agent-connection suites pass; npm run check passes.
• Changelog fragment included.

If you'd rather keep the handler resident-only and change the capability doc instead, I can rework it.