Expand description
Decapod: A Project OS for AI Agents
Decapod is a local-first control plane for agentic software engineering.
This is NOT a tool for humans to orchestrate. This IS a tool for AI agents to coordinate. Humans steer via intent; agents execute via this orchestration layer.
§Core Principles
- Local-first: All state is local, versioned, and auditable
- Deterministic: Event-sourced stores enable reproducible replay
- Agent-first: Designed for machine consumption, not human UX
- Constitution-driven: Embedded methodology enforces contracts
- Proof-gated: Validation harness ensures methodology adherence
§For AI Agents
You MUST:
- Read the constitution first:
decapod docs show core/DECAPOD.md - Use the CLI exclusively: Never bypass
decapodcommands - Validate before completion:
decapod validatemust pass - Record proofs:
decapod proof runfor executable claims - Track work:
decapod todo addbefore multi-step tasks
§Architecture
§Dual-Store Model
- User Store (
~/.decapod/data/): Agent-local, blank-slate semantics - Repo Store (
<repo>/.decapod/data/): Project-scoped, event-sourced, deterministic
§The Thin Waist
All state mutations route through DbBroker for:
- Serialization (in-process lock)
- Audit logging (
broker.events.jsonl) - Intent tracking
§Subsystems (Plugins)
todo: Task tracking with event sourcinghealth: Proof-based claim status trackingknowledge: Structured knowledge with provenancepolicy: Approval gates for high-risk operationswatcher: Read-only constitution compliance monitoringarchive: Session archival with hash verificationcontext: Multi-modal context packing for agentscron: Scheduled recurring tasksreflex: Event-triggered automationfeedback: Agent-to-human proposal systemtrust: Trust score tracking for agentsheartbeat: Liveness monitoring
§Examples
# Initialize a Decapod project
decapod init
# Read the methodology
decapod docs show core/DECAPOD.md
# Add a task
decapod todo add "Implement feature X"
# Run validation harness
decapod validate
# Run proof checks
decapod proof run