SwarmWeaver

Deterministic Multi-Agent DAG Orchestrator, File-Scope Lease Conflict Guard & Crash-Safe Checkpoint Resume Engine (with MCP Server).

Why swarmweaver?

Engineering teams running parallel multi-agent swarms (subagents editing code, running analyses, and synthesizing reports concurrently) repeatedly hit three critical production failures:

1. File Clobbering & Race Conditions: Two parallel subagents edit overlapping files, directory globs, or AST symbols simultaneously, silently overwriting each other's patches.

2. Crash Amnesia on Long-Running DAGs: A 15-node multi-agent workflow fails at node 12 due to a transient rate limit or timeout, forcing the orchestrator to re-run nodes 1–11 from scratch and burn duplicate tokens.

3. Unvalidated Handoffs & Cyclic Deadlocks: Agents pass unstructured prose between steps ("telephone game" drift) or declare cyclic task dependencies that deadlock execution.

SwarmWeaver solves all three problems in a zero-dependency Python 3.11+ engine and JSON-RPC 2.0 Model Context Protocol (MCP) server:

• Conflict-Aware Parallel Wave Compiler (DAGCompiler): Groups tasks by topological level, detects cycles with exact cycle path reporting (A -> B -> C -> A), and automatically splits tasks with write-write or read-write file-glob overlaps (fnmatch, directory prefix, and file.py::symbol scope) into sequential sub-waves so parallel subagents never clobber the same file.

• Pessimistic/Optimistic File-Scope Lease Guard (FileScopeLeaseManager): Enforces SHARED_READ and EXCLUSIVE_WRITE locks at runtime with deadlock-free canonical acquisition order, TTL expiration, cryptographic SHA-256 lease tokens, and an immutable LeaseConflict audit log.

• Hash-Chained Crash-Safe Checkpoint Journal (CheckpointJournal): Content-addressable JSONL journal (prev_hash -> entry_hash) that replays completed nodes whose input_digest matches in 0.007 ms/node with 0 tokens spent when resuming a crashed workflow (engine.execute(resume=True)).

• Schema-Enforced Shared State (TypedBlackboard): Validates required input_contract keys and inline/custom type schemas before a downstream agent starts, and verifies output_contract artifacts before a task can mark itself COMPLETED.

Related MCP server: Foremerge

System Architecture

flowchart TD
    Specs["TaskSpec Graph\n(depends_on, read_globs, write_globs, contracts)"] --> Compiler["DAGCompiler\nCycle & Contract Validator"]
    Compiler --> WaveSplitter["Conflict-Aware Wave Planner\n(Splits W/W & R/W Glob Collisions into Sub-Waves)"]
    WaveSplitter --> Engine["SwarmOrchestrator\nWave-Synchronous Execution Loop"]

    Engine <--> Journal["CheckpointJournal\nSHA-256 Hash Chain (prev_hash -> entry_hash)"]
    Engine <--> Blackboard["TypedBlackboard\nSchema & Lineage Contract Guard"]
    Engine <--> Leases["FileScopeLeaseManager\nSHARED_READ / EXCLUSIVE_WRITE Tokens"]

    Journal -- "Hit (input_digest match)" --> Replay["Deterministic Zero-Token Replay\n(0.007 ms/node, 0 tokens)"]
    Leases -- "Grant Acquired" --> Workers["Parallel Subagent Pool\n(Retry Budget & Fault Isolation)"]
    Workers --> Report["DAGExecutionReport\n(Critical Path, Saved Tokens, Audit Log)"]
    Replay --> Report

Verified Benchmark Results (v1.0.0)

Reproducible via swarmweaver benchmark (65 unit & concurrency tests, zero external runtime dependencies):

Quickstart

Installation

git clone https://github.com/codesphere-u/swarmweaver.git
cd swarmweaver
pip install -e ".[dev]"

CLI Usage

# 1. Compile the 10-agent full-stack refactoring swarm into conflict-free parallel waves
swarmweaver plan

# 2. Run the 10-agent swarm demo with an injected crash at node 8 + instant zero-token resume
swarmweaver demo

# 3. Run the reproducible 4-part benchmark suite (human-readable or JSON)
swarmweaver benchmark
swarmweaver benchmark --json

# 4. Verify the SHA-256 hash chain of a persisted checkpoint journal
swarmweaver verify-journal ./path/to/checkpoint.jsonl

Python SDK Usage

from swarmweaver import SwarmOrchestrator, TaskSpec

tasks = [
    TaskSpec(
        task_id="scan_ast",
        title="Audit auth module AST",
        agent_role="auditor",
        read_globs=["src/auth/**/*.py"],
        output_contract=["ast_report:dict"],
        estimated_tokens=800,
    ),
    TaskSpec(
        task_id="patch_jwt",
        title="Harden JWT token verifier",
        agent_role="auth-coder",
        depends_on=["scan_ast"],
        write_globs=["src/auth/jwt.py"],
        input_contract=["ast_report:dict"],
        output_contract=["jwt_patch:dict"],
        estimated_tokens=1200,
    ),
    TaskSpec(
        task_id="patch_rbac",
        title="Update RBAC claims in jwt.py",
        agent_role="policy-coder",
        depends_on=["scan_ast"],
        # Overlaps with patch_jwt on src/auth/jwt.py -> automatically split into a sequential sub-wave!
        write_globs=["src/auth/jwt.py", "src/auth/rbac.py"],
        input_contract=["ast_report:dict"],
        output_contract=["rbac_patch:dict"],
        estimated_tokens=950,
    ),
]

orchestrator = SwarmOrchestrator(
    tasks,
    workflow_id="auth-hardening-v1",
    journal_path=".swarmweaver/auth_journal.jsonl",
    validate_contracts_at_compile=True,
)

# If a prior run crashed midway, resume=True replays completed nodes in <0.01ms with 0 tokens!
report = orchestrator.execute(resume=True)
print(f"Status: {report.status} | Waves: {report.wave_plan.waves}")
print(f"Tokens Spent: {report.tokens_spent} | Tokens Saved: {report.tokens_saved_by_replay}")

MCP Server Integration

swarmweaver includes a built-in JSON-RPC 2.0 Stdio MCP server (swarmweaver mcp) exposing 4 tools to Claude Code, Cursor, Antigravity, and custom MCP clients:

Claude Code / Cursor / Antigravity mcpServers Config

{
  "mcpServers": {
    "swarmweaver": {
      "command": "swarmweaver",
      "args": ["mcp"]
    }
  }
}

Part of the Codesphere Agentic Infrastructure Ecosystem

SwarmWeaver is built by Akandwanaho Ivan Alvin (@codesphere-u), Founder & Managing Director of Codesphere Innovations (U) Limited (Kampala, Uganda), as part of a modular open-source stack for production-grade AI engineering:

• swarmweaver — Deterministic Multi-Agent DAG Orchestrator, File-Scope Lease Conflict Guard & Crash-Safe Checkpoint Resume Engine.

• patchpilot — Autonomous Code Repair Engine: Stack-Trace & AST Fault Localizer, Anti-Reward-Hacking Patch Verifier & Transactional Sandbox.

• synapse-router — Complexity-Aware Multi-Model LLM Router, LinUCB Contextual Bandit, Semantic Cache & Budget Guard.

• agentflight — Deterministic Flight Recorder, Cryptographic Action Ledger & Chaos Replay Engine for Autonomous AI Agents.

• cortex-mcp — AST-Native Codebase Graph, Hybrid RRF Semantic Search & Token-Budgeted Context Engine.

• agentshield — Zero-Trust Runtime Security Firewall, Prompt-Injection Detector & Tool-Call Policy Enforcer for AI Agents.

• ARGUS-AI — Multi-Modal Enterprise Intelligence & Observability Platform.

License

Released under the MIT License. Copyright (c) 2026 Akandwanaho Ivan Alvin (Codesphere Innovations (U) Limited).