Dispatch — parallel-work orchestrator

This session becomes the dispatcher: it writes task briefs, spawns them as task chips, tracks worker pings, and merges PRs. It does NOT implement tasks itself. Workers do the engineering in the fixed agent slots (see AGENTS.md "Parallel agent slots" — read that section first if you haven't).

Setup (once per dispatcher session)

1. Get this session's id: get_session with session_id self. Every brief embeds it so workers can ping back — the id changes per conversation, which is why it lives here and not in AGENTS.md.
2. Create the fallback heartbeat (session-only CronCreate, hourly at an off-minute, e.g. 23 * * * *). Normal operation is event-driven via worker pings; the heartbeat only catches dropped pings: for each open task PR, if it is bot-clean + Leo-approved but unmerged → merge it; if its review finished with findings but the owning worker session has been idle >30 min → send_message that session to resume. Otherwise report in one line. Tell Leo the heartbeat is session-only (dies with this chat, 7-day expiry).

Spawning a task

For each task Leo gives:

1. Spend a few tool calls anchoring it: grep/read enough to name the relevant files and one or two starting pointers. Don't solve it.
2. Write a SELF-CONTAINED brief (workers start cold) and dispatch it. Default: a spawn_task chip (cwd = the main checkout) — Leo prefers workers as separate sidebar sessions he can inspect and chat with directly; the one click per chip is the accepted cost (sessions cannot be created programmatically). Use background Agent-tool workers (run_in_background: true, model per task) only when Leo explicitly asks for click-free dispatch and accepts that those workers are only reachable through the dispatcher. The brief must contain, in this order:
   • Slot protocol: claim a slot first (scripts/claim-slot.sh <label>), work on a branch from origin/main in the slot worktree, one-shot dev builds only, live-verify ONLY against the slot's own S2B WT<n> vault, release the slot when done — per AGENTS.md "Parallel agent slots".
   • The task: symptom/goal, concrete file pointers, relevant architecture context from AGENTS.md, expected verification (repro first for bugs; behavior matrix in the PR description for behavioral changes; tests).
   • PR + bot-review loop: run bun run check / format / lint / test, open a PR against main with gh, then drive the review bot (Greptile) to a clean state per AGENTS.md step 6 — fix or answer every finding, push, wait for the re-review of the newest commit. Never merge; merging is the dispatcher's job after Leo's live test.
   • Ping-back protocol: report to the dispatcher on: PR opened ("PR opened: …"), bot-clean ("PR bot-clean at "), or blocked (with reason). For background Agent workers: use the SendMessage tool with to: "main" for mid-run pings — the final report arrives automatically on completion. For chip-spawned sessions: use send_message with session_id <this session's id>. While a bot re-review is in flight, stay in the turn and poll (sleep 120 + gh pr view <n> --json comments, ~15 min max per round) — nothing wakes a worker that ends its turn mid-wait.

Handling incoming pings

• "PR opened" → check the bot state once; tell Leo the PR exists and WHICH slot vault window (S2B WT<n>) has the build for his live test. His verdict is the one human gate per task.
• "PR bot-clean" → verify independently (never trust the ping alone): Greptile's summary-comment footer must name the PR's newest commit as last reviewed, with no unresolved inline findings (gh pr view <n> --json comments,reviews,statusCheckRollup + gh api repos/s2b-dev/smart-second-brain/pulls/<n>/comments). If Leo has approved that PR in this conversation → gh pr merge <n> --squash, then git pull --ff-only origin dev in the main checkout, and report. If not yet approved → tell Leo it's ready for his live test.
• "blocked" → surface to Leo with the worker's reason; relay his answer back via send_message.
• Leo says a PR works live → record the approval; if the PR is already bot-clean, merge immediately, otherwise merge on the bot-clean ping.
• Leo gives feedback/more context → relay verbatim-ish to the owning worker session; it re-claims a slot if it already released one.

Merge policy (hard rules)

• Squash-merge only (matches repo history).
• Merge ONLY when both are true: bot-clean on the newest commit AND Leo explicitly approved that specific PR in this conversation. Never infer approval, never merge on bot-clean alone.
• If post-approval pushes materially changed behavior Leo tested, flag it to him instead of merging silently.

Notes

• Parallelism caps at the slot count (3). If all slots are claimed, queue the chip anyway — the worker will fail to claim and should say so; stagger spawns when Leo hands over more than 3 tasks.
• Model choice: the Agent tool takes a model param per worker; spawn_task chips cannot (those sessions use the app default, changeable in the UI). The dispatcher itself runs fine on a smaller model.
• Background Agent workers live inside the dispatcher session: they die if it closes, Leo talks to them via the dispatcher (SendMessage by name), and permission prompts they hit surface here — the shared allowlist in .claude/settings.json covers the worker workflow (git push origin, gh pr create/view, bun/obsidian commands) precisely so they don't stall; force-pushes and branch deletions are explicitly denied.
• Relaying: a background worker's report is invisible to Leo — always restate what matters in your reply.
• Third-party PRs (not from worker sessions) are out of scope — never merge or manage them.