[deanrie]

Summary

Fixes #641. crawlBranch (used by crawl) descends into a subplanner branch when its status is running or handed-off; collectRunningAgentsInTree (used by kill-tree via filterVictimsToSubtree) descended only on running. A subplanner whose own run has ended is handed-off in the parent's state.json, but the state.json on its branch can still hold running rows with agentId/runId (a verifier or respawn it started before returning, workers in flight). crawl listed those agents; kill-tree printed nothing to kill and left them running, while the doc comment on collectRunningAgentsInTree and the kill-tree help text both say the two walk the tree the same way.

Both walkers now share one predicate, descendsIntoSubplanner(t) (running or handed-off subplanner), so they cannot drift apart again. No other behaviour changes: kill-tree still only collects tasks that are running with both ids; it just also looks under handed-off subplanners to find them.

If the orchestrate owners prefer the opposite contract (skip handed-off branches on purpose), the same helper is the one place to say so, and the help text and crawl output should then say it too.

Verification

• New __tests__/kill-tree-handed-off.test.ts builds a temp repo with origin/<branch> refs for a root planner (subplanner row handed-off) and the subplanner's own branch (worker row running), and asserts crawlBranch prints the worker and collectRunningAgentsInTree returns it. Fails on main (0 collected), passes with the fix.
• bun test in orchestrate/skills/orchestrate/scripts: 211 tests, 0 fail; bun run typecheck clean.

[deanrie]

Hey all, for the record: #645 (Ehsan-hash) was opened about an hour before this one and fixes the same handed-off descent for #641, and additionally carries ancestor links through finished planners so agent-scoped kill-tree selection keeps working, with broader fixtures. If the orchestrate owners prefer that one, this can be closed in its favour; the one piece worth keeping from here is the shared descendsIntoSubplanner predicate, so crawlBranch and the kill-tree walker use a single definition and cannot drift apart again. Happy to rebase that onto #645 as a follow-up instead.