parallel-huntlisted

# Parallel hunt One thin orchestrator session runs the whole round. Workers are background subagents with fresh contexts, never extra chat sessions. Nobody waits on the user after launch. Settled design decisions and the incidents behind them live in `decisions.md`, next to this file. Read it if you are tempted to change how the round works — not to run one. ## Who runs what Each role is a registered agent type carrying its own brief, model and effort. Spawn by `subagent_type`; the orchestrator never pastes a brief. | Role | Agent type | Effort | Lives for | |---|---|---|---| | Finder | `parallel-hunt-finder` | xhigh | one sweep group | | Fixer | `parallel-hunt-fixer` | high | one batch of fixes | | Claim gate | `parallel-hunt-claim-gate` | high | one review, then dies | | Fix gate | `parallel-hunt-fix-gate` | high | one batch, then dies | | Fix gate, critical | `parallel-hunt-fix-gate-critical` | high | one batch, then dies | Use the critical fix gate for `severity: critical` entries, or any fix whose diff touches money, auth or security. Spawn prompts carry **only** what varies — the sweep group, the register IDs, the batch size. Everything stable already lives in the agent file, where it caches. ## The register — all state in files In **the main checkout**, not any worktree, so every agent reads and writes the same absolute path with no git sync: ``` <main-repo-root>/.scratch/<feature>/register.md # the index table <main-repo-root>/.scratch/<feature>/bugs/<ID>.