zoanthrope

The Tyranid creature that relays the Hive Mind's will to the swarm — zoanthrope is the coordination relay: who works where, what was promised, what happened.

A swarm coordination MCP server. zoanthrope answers one question — who is working on what? — for a workspace of collaborating agents. File claims with read/edit intents, agent presence, a public postboard of intentions, per-path filesystem leases against clobbering, a rejoin path for agents that went away and came back, and a capped audit ring. It keeps a team of agents out of each other's way and leaves a trail of who promised what.

What it is

• a small MCP server (zoanthrope mcp) over stdio JSON-RPC
• a hook-facing CLI (zoanthrope claim / release / rejoin / status)
• a state registry at .zoanthrope/state.json plus filesystem leases at .zoanthrope/locks/ — zero daemons, nothing to install and keep alive

Tool surface

All tools take and return JSON objects. Success bodies carry "ok": true; a refused claim is a successful call whose body has "ok": false plus the conflict shape.

The conflict shape

The heart of the product: a claim that clashes with another agent's recorded intent is refused all-or-nothing and says exactly who is in the way:

{"ok": false, "holder": "beta", "path": "a.rs", "claimed_at": 1791780567.1, "purpose": "fix the parser", "holder_stale": false, "intent": "edit", "caller_stamps": [{"name": "symbol:parse_token", "value": "…"}]}

Claim semantics

A claim carries an intent (read or edit):

• an edit claim conflicts with anything another agent holds on the path — exactly one writer at a time;
• a read claim conflicts with another agent's edit, but readers co-hold happily with each other;
• an agent's recorded intent on a path is the strongest it holds — a later weaker claim never hides a stronger one. Only release (then re-claim) downgrades.

Claims are all-or-nothing per call: one conflict in a batch means nothing is granted. A holder silent past the presence TTL is stale, its claims are held_by_stale, and they still block — claims are never auto-released; forcing one open is a policy this product did not make.

A claim also binds the file's content: at claim (or re-claim) time it records a content fingerprint — mtime+size always, plus sha256 for files ≤ 256 KiB, so same-size rewrites on coarse filesystems are still caught. Every guarded write (the hook family) checks the file against the claim's fingerprint first: if the file changed under the claim — an out-of-band append, a generator run — the write is refused with a stale-claim denial (holder is self) telling the agent to re-read and re-claim. Re-claiming re-binds the fingerprint to the current content; release drops it. No deliberate takeover exists: a claim lock is a content promise, not just a writer lock.

A claim can also carry caller stamps: opaque {name, value} pairs you computed yourself — a symbol fingerprint, a hash, anything — attached with stamps on claim_files or --stamp name=value on the CLI (repeatable). zoanthrope stores caller stamps and re-checks only its own file stamp; verifying caller stamps is the caller's business. They come back on grants, in conflicts (a waiting agent sees what the holder pinned), in list_claims, and in who_am_i. A re-claim that passes stamps replaces the stored set outright; one without keeps the pins already set; release drops them with the claim.

Filesystem leases (anti-clobbering)

A granted edit claim also takes a filesystem lease under .zoanthrope/locks/: a nonce-named sentinel file per path, acquired under a short O_EXCL guard (so concurrent acquirers cannot both pass the holder scan), released by move-aside-and-verify (so a delayed release can never delete a replacement holder's file). A live foreign lease refuses a claim even when the registry is free — a lease-only writer (an editor's save, a stray process) still gets the file to itself. Read claims take no lease.

Leases carry a heartbeat mtime; one older than 60s (ZOANTHROPE_LEASE_TTL_SECS overrides) is stale and may be taken over, again move-aside-and-verify. status prints the lease layer so orphaned sentinels are easy to spot.

Postboard

Posts are the agents' public intentions — one line a human can read while debugging a swarm. post_intention opens one (status: open); update_post moves it through taken → done, or dropped if the plan changed. Posts do not replace each other; an agent may carry several. Only the posting agent may edit or remove its post.

Rejoin

An agent that went away and came back (restart, sleep, disconnect) calls rejoin_agent. It refreshes presence, re-asserts the agent's own claims (filesystem lease refreshed or re-taken), republishes the board, and reports — never rewrites. A claim the registry still credits to the caller but where another agent now holds a live lease comes back as claims_at_risk: reported, not stolen back, not released. Rejoin never breaks claims, own or foreign.

Quickstart

$ cargo build --release
$ ./target/release/zoanthrope mcp

Then, over stdio (one JSON-RPC message per line):

{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"demo","version":"0"}}}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"claim_files","arguments":{"agent_id":"alpha","files":[{"path":"src/a.rs","purpose":"fix the parser"}]}}}

Point another agent at the same file and its call returns the conflict shape above with "holder": "alpha".

Pre-edit hooks use the CLI mirror instead of MCP (see hooks/):

$ zoanthrope claim --files a.rs,b.rs --agent alpha --purpose "fix bug" --intent edit
$ zoanthrope release --files a.rs,b.rs --agent alpha
$ zoanthrope rejoin --agent alpha
$ zoanthrope status [--workspace DIR]

Exit codes: 0 granted/ok, 1 conflict (or ok: false), 2 usage or IO error. --workspace is optional everywhere; without it the root is found by walking up from the current directory for .git.

Wire the hooks (Claude Code)

The guard family claims every file a tool call would mutate and, in blocking mode, turns conflicts into tool-call denials. This is the exact wiring that ran the recorded collision experiments — copy it into your workspace's .claude/settings.json (adjust the paths to where you keep the repo):

{
  "hooks": {
    "PreToolUse": [
      { "matcher": "Edit|Write|MultiEdit|Bash",
        "hooks": [{ "type": "command", "command": "/path/to/zoanthrope/hooks/claude-code/claim-guard.sh" }] }
    ],
    "PostToolUse": [
      { "matcher": "Edit|Write|MultiEdit|Bash",
        "hooks": [{ "type": "command", "command": "/path/to/zoanthrope/hooks/claude-code/release-guard.sh" }] },
      { "matcher": ".*",
        "hooks": [{ "type": "command", "command": "/path/to/zoanthrope/hooks/claude-code/watch.sh" }] }
    ]
  }
}

Each agent needs its identity in the environment (ZOANTHROPE_AGENT=alpha, …) — see hooks/README.md for the identity resolution order and how to port the hooks to other agent frameworks.

One wiring, both modes. Advisory vs blocking is a runtime mode, not a different hook: this wiring serves both. claim-guard.sh resolves the mode per tool call, highest priority wins:

1. ZOANTHROPE_MODE env var (advisory | blocking) — the per-agent override, so a swarm can block its bots and leave one advisory actor;

2. .zoanthrope/config.json at the workspace root (found by walking up to the git toplevel, like the CLI; the file is optional):

   jsonCopy

   {"mode": "advisory"}
   

3. default: blocking (the experiment-proven behavior).

In advisory mode the guard claims the file, prints the claim/conflict JSON to stderr, and always exits 0 — it never emits a permissionDecision: deny. In blocking mode a conflict denies the tool call: exit 2, reason naming the holder (unchanged). An invalid mode value falls back to blocking with one warning line on stderr — a hook never crashes. pre-edit.sh remains as the legacy advisory alias; release-guard.sh and watch.sh never deny, so the mode does not apply to them.

Instructions for agents

Paste this into your swarm's prompt or AGENTS.md — it is the protocol the collision experiments validated:

A coordination service (zoanthrope) is active in this workspace.
- Before touching shared files, post your intention (`post_intention` /
  `zoanthrope claim --files … --agent <you> --purpose "…"`).
- An edit may be DENIED with a reason naming the holder. That means
  another agent owns the file right now: wait briefly, then retry
  (up to 3 times). If still denied, work on something else and record
  it — never work around the lock (no out-of-band writes).
- Prefer append/insert on shared files; if you must rewrite a file
  whole, read it back after writing and merge anything that appeared.
- Post what you plan to do on the postboard before you start, and
  update the post when done (`update_post`).
- If you return after being away, call `zoanthrope rejoin --agent <you>`
  once — it re-asserts your claims and tells you what changed.

For consultation and debugging (who holds what, stale agents, orphaned claims), give your agents skills/zoanthrope/SKILL.md.

State

Everything lives in .zoanthrope/state.json at the workspace root: claims, agents, posts, the published board (who holds what, stamped with a generation epoch), and an append-only audit ring capped at 1000 events. Writes are atomic (temp file + rename — the rename is the generation publish) and serialized by a lock file. A state file from a different format version is refused with instructions; move .zoanthrope/ aside to start fresh.

Presence TTL defaults to 15 minutes; ZOANTHROPE_TTL_SECS overrides.

Performance

Decide in nanoseconds, persist in microseconds, scale to 128 concurrent agents without losing a write. The guarantee covers write paths that go through the guarded hooks (Edit/Write/MultiEdit and recognized Bash mutations); unparseable commands are audited as unclaimed-mutation rather than blocked, and generator scripts (node script.js) are not automatically covered unless they write through recognized idioms — the audit trail remains the safety net. All numbers measured on this machine; each family reproduces with the command named.

Latency — cargo run --release --bin zoo-bench:

Persistence is the cost, the decision is not. The contended-8 p95 is 7.3 ms; under the earlier fixed 20 ms poll the same tail reached 101 ms, and lock acquisition now backs off exponentially to bring it down. Postboard creation runs at 614 posts/s.

Clobber chaos — cargo test --release --test clobber_chaos -- --ignored (scale with ZOO_CHAOS_WORKERS):

Prevention holds as the swarm grows. One documented limit, plainly: the post-hoc stomp detector samples faster than it can observe when writes exceed ~10⁵/s — prevention (the lock) is the guarantee; the audit trail is best-effort at machine-gun rates.

The A/B with real agents — 3 concurrent agents on shared files, locked arm vs an unguarded control (full protocol and per-run data in experiments/zoo-ab-2026-10-12.md): the locked arm finished with 0 lost work, 3 denials that named the holder, all recovered on retry, at equal wall time (53.1 s vs 52.9 s). Scope, framed as coverage: tool-mediated edits (Edit/Write) ride the host harness's own CAS/OCC — verified 0 lost work in 24 agent runs — so the lock is engineered for the vector that escapes those tools: Bash mutations, codemods, generators, scripts. On that vector the round-3 table reads 0 vs 1 lost, with +6 in-flight clobbers self-healed in the control.

Formal guarantees

The coordination core is checked against TLA+ models in formal/:

• PresenceClaims — claim arbitration: no conflicting holders, read co-holding, and strongest-intent recording;
• FsLease / FsLeaseGuard — filesystem lease exclusion, the anti-clobbering core (move-aside-and-verify release, O_EXCL acquire guard);
• RejoinProtocol — returning agents re-assert without breaking claims, own or foreign.

Verified with TLC via formal/check.sh: 11/11 configurations — 4 invariant configurations green, and 7 negative-control configurations reproduce their counterexamples (the specs demonstrably catch the failure modes they claim to: broken arbitration, unguarded leases, leaky rejoin). One command to re-verify:

$ cd formal && ./check.sh

(TLA2TOOLS_JAR points at a tla2tools.jar, TLA+ tools v1.8.0; the spec→code mapping lives in formal/README.md.)

Development

$ cargo test
$ cargo fmt --check
$ cargo clippy --all-targets -- -D warnings

tests/mcp_smoke.rs spawns zoanthrope mcp and runs the full initialize → tools/list → claim → conflict → postboard round trip → rejoin → release over stdio. npm-shim/ is the @spuentesp/zoanthrope node launcher (thin, no download machinery — it forwards to a binary on PATH or ZOANTHROPE_BIN).

License

MIT — see LICENSE.