agentic-jujutsu

Agentic Jujutsu Skill

Coordinates concurrent multi-agent writes to a single git repository without traditional file locking. Agents register soft claims through ClaimsManager, fan out into isolated worktree branches, and reconverge via semantic 3-way merge. Benchmarked at ~87% auto-conflict resolution on the EVOKORE corpus; the remaining 13% route to manual escalation with structured context.

Trigger

Use this skill when:

• Orchestrating 3+ agents writing to overlapping code paths in the same session
• Fan-out / fan-in workflows (e.g. architect → 5 implementers → reviewer) are hitting .git/index.lock collisions
• Agents are stomping on each other's commits mid-squash
• A merge wave has produced more conflicts than it resolved and you need a replayable strategy

Core Principle

Claim first, write second, merge semantically. Locking is replaced with cooperative soft claims that record intent. If two agents claim overlapping regions, the later claim is told upfront (not after a wasted write) and can either wait, re-scope, or route to a different branch.

Claim Protocol

Every agent invokes claim_acquire before editing. Claims are scoped to {path, range?, intent} and held by session ID.

// Agent A: claim exclusive write on a function range
{
  "tool": "claim_acquire",
  "arguments": {
    "path": "src/ProxyManager.ts",
    "range": { "startLine": 120, "endLine": 210 },
    "intent": "refactor:hashmap-lookup",
    "ttlMs": 900000
  }
}

// Agent B: overlapping claim returns conflict with holder info
{
  "tool": "claim_acquire",
  "arguments": {
    "path": "src/ProxyManager.ts",
    "range": { "startLine": 180, "endLine": 260 },
    "intent": "add:connection-pool"
  }
}
// → { ok: false, conflict: { holder: "agent-a", intent: "refactor:hashmap-lookup", expiresAt: ... } }

Release on completion (success or abort):

{ "tool": "claim_release", "arguments": { "claimId": "clm_7h2..." } }

List active claims to debug stalls: