runtime/src/manowar. The modules are small enough to inspect directly, but this map gives you the ownership boundaries first.
Runtime Core
| Module | Responsibility |
|---|---|
index.ts | Exposes the single framework id, manowar. |
runtime.ts | Registers agents and workflows, validates cards, derives wallets, warms runtimes. |
framework.ts | Creates agents, model wrappers, managed workers, streams, stop controllers, and Responses calls. |
telemetry.ts | Extracts tokens, records usage, feedback, learning records, and run metadata. |
Agent Loop
| Module | Responsibility |
|---|---|
harness/graph.ts | Native model/tool loop with streaming, tool repair, and turn budgets. |
harness/context.ts | Per-turn execution context via AsyncLocalStorage. |
harness/identity.ts | Agent identity hydration and compact identity rendering. |
harness/memory.ts | Pre-turn retrieval, post-turn persistence, and explicit remember calls. |
harness/memory-scope.ts | Global/local memory scope resolution. |
harness/state.ts | Tool leases and visibility states: base, memory, tool, model, swarm. |
harness/tools/* | Tool factories, tool-call parsing, model/agent/search clients, and surface sentence. |
Memory
| Module | Responsibility |
|---|---|
memory/agent-loop.ts | pre_turn, post_turn, and remember workflow. |
memory/layers.ts | Six-layer retrieval across working, scene, graph, patterns, archives, vectors. |
memory/vector.ts | Atlas vector search, keyword fallback, access-count updates. |
memory/graph.ts | Durable fact extraction and graph-layer retrieval. |
memory/embedding.ts | Voyage embeddings through MongoDB’s AI gateway. |
memory/ranking.ts | Temporal decay, Cloudflare rerank, and MMR ordering. |
memory/mongo.ts | MongoDB connection, collections, and indexes. |
memory/cache.ts | ValKey cache, embedding cache, and scoped invalidation. |
memory/operations.ts | Consolidation, patterns, archives, decay, cleanup, and IPFS sync. |
memory/evals.ts | Memory retrieval evaluation metrics. |
Harness
| Module | Responsibility |
|---|---|
harness/swarm/types.ts | Typed MAL plan, step, budget, result, and subagent contracts. |
harness/swarm/interpreter.ts | Plan parser, validator, deterministic dispatcher, checkpoint/proof orchestration. |
harness/swarm/engine.ts | Subagent execution and budget enforcement. |
harness/swarm/parallel.ts | Fanout concurrency and gather modes. |
harness/swarm/types.ts | Registered-agent validation through the API. |
harness/swarm/scratchpad.ts | Private per-agent/run scratchpad. |
harness/swarm/conclave.ts | Shared swarm bus. |
harness/swarm/checkpoint.ts | ValKey MAL checkpoints. |
harness/swarm/proof.ts | Hash proof bundles and optional Pinata pinning. |
harness/swarm/sandbox.ts | Daytona-backed isolated subagent execution. |
harness/swarm/coordinators.ts | Dynamic coordinator model discovery from the catalog. |
Workflow
| Module | Responsibility |
|---|---|
workflow/orchestrator.ts | Plan -> Act -> Reflect workflow execution. |
workflow/planner.ts | Structured planning, review, reflection, validation, and graph ordering. |
workflow/delegation.ts | HTTP calls to embedded agent runtime. |
workflow/context.ts | Context window lookup and cleanup thresholds. |
workflow/memory.ts | Workflow memory writes, search, reliability, and safe wipe summaries. |
workflow/checkpoint.ts | Workflow observations, decisions, insights, errors, and persistence. |
workflow/run-tracker.ts | In-memory run lifecycle and cron stats. |
workflow/triggers.ts | Natural-language trigger parsing and Temporal schedule management. |
workflow/registry.ts | Workflow card loading and connector tool discovery. |
Comparison Map
| Manowar module | Closest public-frame pattern | Difference that matters |
|---|---|---|
| Native executor | OpenAI Agents SDK runner, ADK runner | Manowar calls the Compose inference gateway and records metering evidence. |
| MAL harness | LangGraph graph, ADK workflow agents | MAL is a typed plan language interpreted by the runtime, not a graph library embedded in user code. |
| Memory | LangGraph store, ADK session/state/memory | Manowar stores layered memory in first-party Mongo/ValKey and injects compact context automatically. |
| Swarm delegation | OpenAI handoffs, AutoGen messages | Delegation requires registered Compose agents with wallet identity. |
| Model store | OpenRouter-style catalog search | Model selection uses the Compose generated catalog and public model ids. |
| Telemetry | OpenAI tracing, LangSmith, AutoGen logs | Telemetry is in-runtime usage and feedback records used by Compose billing and learning surfaces. |