2c05206907
Hivemind delegates grunt work to headless opencode workers (scout/coder/tester) while the orchestrating agent stays the planner, reviewer, and merger. - placed under engineering/ per repo layout; targets dev per CONTRIBUTING - frontmatter is name + description only; SKILL.md is 165 lines - anti-patterns and cross-references sections included - external opencode CLI dependency and free-model caveats documented up front - runtime state (.runs/) gitignored, not committed
1.4 KiB
| description | allowed-tools |
|---|---|
| Migration coordinator - parallel per-file/chunk workers with sequenced merge and validation | Bash(git:*), Bash(node:*), Read(*), Edit(*) |
You are the ORCHESTRATOR of a migration. Task: $ARGUMENTS
PLAN
- Identify the file set or chunks (e.g., all files importing old API). Cap at 5 workers; if more files exist, group them into <=5 batches by dependency/similarity.
- Generate run id: hive-.
- Create ONE worktree + branch PER batch: git worktree add ../-mig-N -b swarm/mig-N
SPAWN (PARALLEL - one Bash call per batch in ONE message): node "$HIVEMIND_HOME/scripts/oc-worker.mjs" --agent coder --run --label mig-N --dir --timeout 1200 ". Touch ONLY these files: . Do not commit."
The transformation spec must be IDENTICAL across workers except the file list - consistency beats cleverness.
MERGE (sequentially, YOU only): For each N in order: review git diff main...swarm/mig-N yourself -> merge -> resolve any conflicts YOURSELF (never delegate conflict resolution) -> delete worktree.
VALIDATE (single tester worker after ALL merges): node "...\oc-worker.mjs" --agent tester --run --label validate --dir "Run . Report PASS/FAIL + failures."
FALLBACK: a batch fails twice -> do that batch's migration yourself inline. REPORT: table (batch | files | tokens | outcome), your conflict resolutions, test verdict.