How does SwarmWeaver stop parallel agents from overwriting the same file?

At compile time, DAGCompiler builds an intra-level conflict graph across normalized read_globs, write_globs, and ::symbol scopes, then colors conflicting tasks into sequential sub-waves. At runtime, FileScopeLeaseManager enforces SHARED_READ and EXCLUSIVE_WRITE locks in sorted canonical order.

How much token cost does CheckpointJournal save on mid-workflow crashes?

Across the 20-workflow crash recovery benchmark at 70% progress, SHA-256 hash-chained replay saves 70.0% of tokens (140,000 tokens saved) at 0.00752 ms/node without re-invoking completed agents.

Does SwarmWeaver require Redis, Postgres, or external Python dependencies?

No. The core package (models, DAGCompiler, FileScopeLeaseManager, CheckpointJournal, TypedBlackboard, SwarmOrchestrator, CLI, Stdio MCP server, and HTTP server) uses pure Python 3.11+ standard library with zero runtime dependencies.

Parallel agent swarms that never clobber files.

Deterministic DAG wave compilation, file-scope read/write leases, typed blackboards, and SHA-256 checkpoint resume in pure Python.

pip install -e . && swarmweaver demo

Open Studio ->

$ swarmweaver plan
SwarmWeaver v1.1.0 — Conflict-Aware Wave Plan
Total Tasks          : 10   | Topological Levels : 5
Conflict-Free Waves  : 7    | Conflicts Auto-Split: 2
Critical Path        : 1170 ms (1.46x speedup vs 1710 ms)
--------------------------------------------------------------------
  Wave 01: 01_ast_audit
  Wave 02: 02_schema_migration, 03_jwt_verifier
  Wave 03: 04_rbac_policy   (split: WRITE_WRITE src/auth/jwt.py)
  Wave 04: 05_api_gateway
  Wave 05: 06_rate_limiter  (split: WRITE_WRITE src/api/deps.py)

Compile locally in Python or simulate right in your browser

Run DAG Studio Read Benchmark

Conflict-free waves.

DAGCompiler runs Kahn's sort across dependencies, detects cycles with full path chains, and splits overlapping read_globs, write_globs, and ::symbol targets into safe sub-waves.

Ordered runtime leases.

FileScopeLeaseManager grants SHARED_READ and EXCLUSIVE_WRITE locks in sorted canonical scope order with monotonic TTL eviction and SHA-256 lease tokens.

Sub-millisecond replay.

CheckpointJournal chains each finished node via prev_hash -> entry_hash SHA-256 digests, replaying 8 completed tasks in 0.077 ms without re-spending LLM tokens.

Typed blackboard handoffs.

TypedBlackboard validates typed contracts (dict, str, predicates) before downstream agents start, cutting silent schema drift and partial corrupt writes.

Zero dependencies, open interfaces

Run SwarmWeaver standalone in Python 3.11+, wire it into Claude Desktop or Cursor over Stdio MCP, or host the Streamable HTTP/SSE portal.

Kahn Sort

Glob Leases

Hash Chain

JSON-RPC 2.0

HTTP Portal

Try the interactive compiler in Studio or review verified numbers and CLI/SDK docs.

Frequently asked questions

How SwarmWeaver keeps concurrent coding agents predictable on real repositories.

Why split waves at compile time if runtime file locks exist?

Runtime locks stop silent overwrites, yet two agents launched blindly in parallel still waste a worker slot waiting or colliding. Compiling conflict-free waves beforehand schedules independent files together and sequences conflicting writes deterministically.

Can two agents edit different functions in the same file concurrently?

Yes. When both tasks narrow their scope using symbol qualifiers such as src/auth/jwt.py::verify_token and src/auth/jwt.py::rotate_keys, normalized scope intersection treats them as disjoint symbol locks.

What happens if someone edits or corrupts a checkpoint journal on disk?

swarmweaver verify-journal and CheckpointJournal.verify_integrity() recompute every canonical SHA-256 link against prev_hash and flag the exact tampered_sequence record.