Wake parents for parked director questions

[TheGreatAxios]

Summary

• An idle TUI parent receives one coalesced wake when a top-level worker parks on ask_director.
• Wakes wait through live parent turns, operator gates, and parent-idle fleet holds, then reply with send_input target set to the worker session ID, not a shared catalog name.
• Authoritative snapshots drop resolved, cancelled, replaced, and nested asks. Session rotation suppresses stale publications. Synthetic wakes use the idle delivery path, so they skip composer /feedback capture and leave queued user follow-ups alone.

Verification

• bun run typecheck, bun run build, and bun run test pass
• bun run check pass (6315 pass, 0 fail, 447 files)

Fixes CL-7524

[linear-code]

CL-7524

[TheGreatAxios]

Primary · Comment

Idle TUI parent wakes on top-level ask_director parks; replies use send_input target with the worker session ID.

GitHub will not accept an approve from this account (own PR). Merge verdict from this read: no blocking defects in origin/main...HEAD (14 files). Snapshot-before-fleet, rotation barrier, internal-origin quota replay, and idle-worker gates match the stated contract.

Checks

• git log --oneline origin/main..HEAD: three in-scope commits (dab784c0, 7205e47e, 482db17d).
• git diff origin/main...HEAD --stat: 14 source/doc files, no Bin.
• git log origin/main..HEAD --format='%s' / '%b' plus awk 'length > 72': subjects are plain English, no prefixes/paths/trailing punctuation; bodies empty.
• git show --stat per commit: subjects match their trees. 482db17d also uses the same lanes snapshot for live fleet count (one line, same concern).
• bun run check at this SHA (482db17d): exit 0, 6315 pass / 0 fail / 447 files (prior session, HEAD unchanged).

Gaps

Did not re-run live-provider or manual TUI this pass. Nested-orchestrator asks are excluded by parentSessionId and covered in unit tests, not a live nested session.

[TheGreatAxios]

Greybeard · Request changes

Idle TUI parents receive one coalesced ask_director wake through an authoritative store snapshot and idle-send, rather than only wait_agents pull.

GitHub refused --request-changes on this PR (own-PR limitation); the lens verdict is still revise.

Findings

• src/agent/directors/skywalker/package.ts:23 — Parent contract still says worker questions arrive only when wait_agents returns status running plus a question. This branch adds a second inbound (synthetic idle-send) the parent must answer with send_input target = session id. That contract belongs on the director, not only in pendingAskWakeText. The package test at src/agent/directors/skywalker/package.test.ts:188 locks the stale running wording.
• docs/PRODUCT.md:171 — Same pull-only contract and stale running status. Public product surface now disagrees with docs/ARCHITECTURE.md.

Notes

• Snapshot/coalesce, rotation barrier, session-id targeting, and unmixing worker gates from isProcessing own the TUI-idle constraint at the right layer.
• Nested asks staying pull-only is accepted: idle-orchestrator is Skywalker/TUI, not nested directors.
• Reusing the user idle-send pipe is acceptable only if the director prompt names that inbound so it is not treated as operator chat.

[TheGreatAxios]

Verdict: approve (cannot GitHub-approve own PR).

CL-7524 holds at 3f62f0a4. Idle TUI parent wakes once for parked ask_director; send_input target is the worker session id; snapshots drop resolved/cancelled/replaced/nested asks; synthetic wakes skip /feedback and leave queued follow-ups.

Should-fix: leftover delivery still runs worker-authored wake text through ingestOperatorPrompt (src/tui/runtime-bridge.ts sendInternalText → src/tui/runner/submit.ts leftover send). A parked question that contains @src/foo.ts or an inline image path can be rewritten and miss the echo, so the transcript paints a second user row and the parent model sees ingest output instead of the parked question. Common "Which port?" asks are fine.

Could leftover send skip ingest for ask_director wake text the same way it already skips sent-message history?

[TheGreatAxios]

Verdict: approve (GitHub cannot approve own PR).

CL-7524 holds at 3f62f0a4. Idle TUI parent wakes once for parked ask_director; send_input target is the worker session id; snapshots drop resolved/cancelled/replaced/nested asks; synthetic wakes skip /feedback and leave queued follow-ups.

Should-fix: leftover delivery still runs worker-authored wake text through ingestOperatorPrompt (src/tui/runtime-bridge.ts sendInternalText → src/tui/runner/submit.ts leftover send). A parked question that contains @src/foo.ts or an inline image path can be rewritten and miss the echo, so the transcript paints a second user row and the parent model sees ingest output instead of the parked question. Common "Which port?" asks are fine.

Could leftover send skip ingest for ask_director wake text the same way it already skips sent-message history?