Hivemind

Hivemind: Claude Code as Orchestrator, opencode as Free Worker Swarm

Claude Code = brain (plans, reviews, merges). opencode = disposable workers on free models (opencode/mimo-v2.5-free default; verified $0.00 per run).

Prerequisites (external dependency)

This skill is a thin orchestration layer over opencode, a third-party CLI. It is not bundled — install and authenticate it yourself first:

Nothing here calls the Anthropic API on the worker side; worker traffic goes to opencode's endpoints. Do not delegate secrets or private code you would not send there.

Setup

1. Put this skill folder wherever your agent loads skills from (e.g. ~/.claude/skills/hivemind).
2. Export HIVEMIND_HOME pointing at that folder — the bundled slash commands use it:

   export HIVEMIND_HOME="$HOME/.claude/skills/hivemind"
   

3. Copy the bundled assets into place:
   • commands/*.md -> ~/.claude/commands/ (the /hive, /oc, /swarm, ... entry points)
   • assets/agents/*.md -> ~/.config/opencode/agent/ (the scout / coder / tester worker personas)

Both copies are optional: everything the commands do can be driven by invoking scripts/oc-worker.mjs directly, and any opencode agent name works with --agent.

Runtime state (.runs/*.jsonl) is written inside this folder and is gitignored.

Components

Slash commands (registered by the plugin; manual installs copy commands/*.md into ~/.claude/commands/):

• /hive <task> - AUTO-ROUTER. Classifies task -> single worker, generic swarm, or template. Default entry point; prefer this over manual routing.
• /oc <task> - single worker delegation
• /swarm <task> - generic parallel swarm
• /review-panel <diff> - 4-lens parallel review (correctness/security/performance/style) + consensus aggregation
• /research-sweep <question> - 3-5 parallel research angles, synthesized
• /migration <task> - batched per-worktree migration workers + sequenced merge
• /test-fleet <target> - partitioned parallel test runs with safety checks

Worker agents (ship in assets/agents/, copy to ~/.config/opencode/agent/):

• scout - read-only research (no write/edit/bash)
• coder - implements one subtask in its worktree
• tester - runs tests only, never edits source

Invocation contract

node "<skill-dir>\scripts\oc-worker.mjs" [--agent scout|coder|tester] [--dir <path>] [--model <p/m>] [--timeout 900] [--run <id> --label <name>] "TASK TEXT"

Returns exactly ONE compact JSON line: { ok, result, tokens:{total,input,output,cache}, cost_usd, duration_ms, label, agent, model }

On failure: { ok:false, stage:"args"|"exec"|"api"|"parse"|"empty", error } with stderr capped at 300 chars.

--run <id> + --label <name> append lifecycle events (start/done/fail) to .runs/<id>.jsonl inside this skill dir. Use them for EVERY swarm worker so progress is recoverable via oc-status.mjs even after orchestrator context loss.

The script auto-manages the shared server: health-checks 127.0.0.1:4096, spawns opencode serve if dead, waits 5s, falls back to cold start. Workers are idempotent against their --dir; re-run once on ok:false before giving up.

HIVEMIND_SERVER_URL overrides that address (default http://127.0.0.1:4096). It must be a valid URL with a numeric port; anything else fails fast with a single stage:"args" JSON line rather than reaching the spawned process.

Verify instead of re-reading (do this before you trust a claim)

Re-checking every worker claim yourself costs more than the delegation saved. Send the claims to a verifier instead:

node scripts/oc-verify.mjs --dir <same dir the worker saw> --claims findings.json --run <id>

Returns { confirmed, refuted, unsupported, unverified, needs_review[] }. Read needs_review only — those are the refuted and unsupported claims. Confirmed claims are counts, not prose, and do not come back into your context.

HARD RULES for this stage:

• The verifier MUST run on a different model from the worker. Same model = same blind spots = a rubber stamp. Default is nemotron-3.5-lightning-free against a mimo worker.
• The verifier sees the FILES and the claims, never the worker's reasoning.
• It defaults to REFUTED when uncertain, because a false CONFIRMED is acted on unchecked.
• unverified > 0 means the verifier skipped claims. Treat those as unverified, not passed.
• Still spot-check a sample yourself. Two free models agreeing is evidence, not proof.

Providers: opencode is not always up

opencode/* free models return provider 404s and empty bodies often enough that an unattended swarm cannot depend on them. or-worker.mjs speaks the identical one-JSON-line contract over OpenRouter, so every downstream script works unchanged:

node scripts/or-worker.mjs --model google/gemini-2.5-flash --dir <path> --json "TASK"

Key comes from OPENROUTER_API_KEY, else ~/.claude/.openrouter_key. Never paste a key into a prompt, a committed file, or a transcript. In CI it comes from a repository secret. Unlike opencode's free tier, OpenRouter bills — check the model's price before a fan-out.

Choosing a model (measured, not guessed)

Free tiers failed six times out of six across both providers, for five different reasons:

But openrouter/free WITHOUT --json answered correctly at zero cost. So the split is capability, not availability:

• Structured output (--json) — use a paid model. Free models cannot reliably emit parseable JSON, and every hivemind contract downstream depends on it. A verified google/gemini-2.5-flash call cost $0.0000506. Fractions of a cent beat a pipeline that fails at 3am.
• Prose, summaries, drafting — free is fine, when a human is present to retry. Do not put a free model on a schedule.
• Never hardcode a free model id. They stop being free without warning. List what is actually free today with the public models endpoint (no key required): curl.exe -s https://openrouter.ai/api/v1/models and filter pricing.prompt == 0.

Running workers in CI

.github/workflows/hivemind-worker.yml runs a worker on a GitHub runner via workflow_dispatch, so work continues when the laptop is closed. Inputs reach the script through the environment, never interpolated into a shell command line.

Golden Rule (non-negotiable)

Raw opencode NDJSON streams must NEVER enter your context. All output arrives via the script's single JSON line. Never pipe opencode run --format json directly into this conversation; never re-implement what the script does.

Single worker flow (/oc)

For one read-only question or small delegation: run oc-worker.mjs without worktrees. Read-only tasks may omit --agent/--dir. Summarize result for the user. If files were written: show git diff before letting the user commit.

Swarm flow (multi-worker)

1. Decompose task into 2-5 INDEPENDENT subtasks (no shared files).
2. Writing workers get isolated worktrees FIRST: git worktree add ../<repo>-wt-N -b swarm/N.
3. Issue ALL worker invocations as PARALLEL Bash tool calls in ONE message.
4. Run oc-verify.mjs over the workers' claims, then review needs_review plus every diff yourself (git diff main...swarm/N). YOU are the only merger.
5. Merge approved branches, remove worktrees, run tests.
6. Report table: subtask | agent | tokens | outcome + total worker tokens.

HARD RULES: workers never share directories; never delegate merging/reviewing; escalate to your own Sonnet only when a free-model worker demonstrably fails twice.

Benchmarking

node scripts\bench\run-bench.mjs --repo <project> [--configs a,b,c] [--task 1-5]

Appends JSONL records (ts, config, tokens, cost, duration) to bench-results.jsonl. Grade artifacts blind with grader-prompt.md (grader sees only task spec + output). Configs: A=claude solo baseline, B=opencode solo, C=claude orchestrating 2 workers.

Fallback ladder (all flows)

1. Worker ok:false -> re-invoke once against the same dir.
2. Still failing -> orchestrator performs that subtask inline, marks it [orchestrator-sourced].
3. opencode entirely down (exec/api twice) -> announce, abandon workers, do the task directly. Never let a swarm fail a task that Claude could have done itself.

Fleet patterns

Four reusable topologies ship as slash commands (see table above). Shared invariants: parallel spawns in one message; --run/--label on every worker; aggregation via oc-aggregate.mjs when 3+ workers produce findings; consensus beats single-lens claims; worktree isolation whenever any worker writes.

Windows notes (hard-won)

• Requires OPENCODE_GIT_BASH_PATH=C:\Program Files\Git\bin\bash.exe (set persistently).
• The script resolves the REAL opencode.exe by parsing the npm .cmd shim — Node's EINVAL policy blocks spawning .cmd directly. Do not "simplify" resolver back to where.exe first-line.
• Free models: opencode/mimo-v2.5-free, opencode/nemotron-3.5-lightning-free, opencode/hy3-free. NOTE: opencode-go/* models require workspace billing — avoid.

Known limits

• Free-tier rate limits can 429 under heavy swarms; space out retries.
• Worker quality varies; always review diffs. Scout answers are evidence-cited.
• Bench config C consumes real Claude tokens for orchestration (~1-2k/task).

Anti-patterns