fix(delivery): reap stale subagent freeze so busy parents keep receiving

[zane668]

Problem

A Task subagent that dies without a clean SubagentStop leaves the parent instance's running_tasks.active stuck at true. While it is stuck, every parent PostToolUse short-circuits down the "in subagent context" branch and delivers nothing, so hcom messages addressed to a busy parent are withheld until the user happens to type a prompt — UserPromptSubmit was the only hook running the dead-subagent cleanup.

Seen in the wild as a parent whose running_tasks still tracked agent_ids that were no longer present in the instances table: the subagent had gone, but the parent stayed frozen and silently stopped receiving.

Fix

On the parent's own PostToolUse, reconcile the freeze before honoring it: reap dead subagents (reusing check_dead_subagents) and re-evaluate in_subagent_context. A live subagent keeps its instances row, so it is never reaped — active stays true and parent messages remain deferred to end_task rather than being mis-injected into the subagent's own context. This is scoped to PostToolUse because that is the only suppressed hook that carries a parent message delivery.

Follow-up: only act on definitive death signals

Reusing check_dead_subagents wholesale pulled the transcript-mtime staleness heuristic into the reconciliation path. A live subagent that simply hasn't written its transcript for a while (a long tool call, waiting on I/O) can cross the stale threshold without being dead; reaping it would unfreeze the parent mid-Task and let the next subagent tool call inject the parent's messages into the subagent's context — a new cross-delivery vector.

To avoid that, DeathSignals scopes which signals a sweep may act on:

• All (DB row missing + stale mtime + interrupt marker) stays in effect at UserPromptSubmit, preserving existing behavior there.
• DefinitiveOnly (DB row missing or interrupt marker) is what the new PostToolUse reconciliation uses; the mtime heuristic is excluded.

Tests

Regression tests cover: stale freeze cleared and delivery resumes to the parent; a live subagent kept frozen; and a live subagent whose transcript mtime is aged past the stale threshold — DeathSignals::All must flag it (fixture sanity / UserPromptSubmit behavior preserved) while reconciliation keeps the parent frozen and running_tasks intact.

Verified locally with cargo fmt --all -- --check, cargo clippy --all-targets --locked -- -D warnings, and cargo test --locked (all green).

Only src/hooks/claude.rs is touched.

🤖 Generated with Claude Code

[zane668]

Heads-up for whoever picks these up 👋

The three fork PRs I opened (#82, #83, #84) all show the same red checks — rust-tests, windows-build, and plan. I dug into the logs, and these look like pre-existing / toolchain-drift issues on main rather than anything the PRs introduce (each PR is a clean single-file diff that doesn't touch the files below):

1. rust-tests + windows-build — both fail at the clippy step on main's own code:

error: redundant reference in `println!` argument
  --> src/commands/list.rs:834:38
   |
834|         println!("  Time:       {}", &entry.timestamp);
   |                                      ^^^^^^^^^^^^^^^^^ help: remove the redundant `&`: `entry.timestamp`
   = note: `-D clippy::useless-borrows-in-formatting`

clippy::useless_borrows_in_formatting was tightened in Rust 1.97, and CI installs the latest stable — so the & on src/commands/list.rs:834 now trips -D warnings. Dropping the & clears it. (Local clippy on 1.96 still passes, which is why it slipped through.)
Logs: rust-tests · windows-build

2. plan — dist plan reports .github/workflows/release.yml has out of date contents and needs to be regenerated (the newer cargo-dist no longer emits the swatinem/rust-cache@v2 step). A dist init regen should clear it.
Log: plan

Just flagging to save some triage time — no rush at all, and happy to rebase once main is green. Thanks for the project! 🙏

[aannoo]

thanks, did it different, should be fixed with these:

26287c2 — Routes hooks by Claude agent_id instead of running_tasks.active, so an uncleanly terminated child cannot freeze parent message delivery.
9fbdd78 — Prevents concurrent duplicate SubagentStop hooks from racing teardown and delivery.
d6b4ac5 — Makes child teardown exactly-once and recoverable if a stop-hook process crashes.
2a841f9 — Makes failed teardown transactionally retryable instead of leaving inconsistent task tracking.
38cc5fb — Verifies the exact child actor for hcom commands, preventing parent/child/sibling misattribution.