Why Decapod 🧠
AI coding agents are strong at generating code. Most failures happen before and after generation: unclear intent, fuzzy boundaries, and weak completion checks.
Decapod is the missing layer in that loop. Agents call it mid-run to lock intent, enforce boundaries, and prove completion with explicit gates. It shapes inference without doing inference.
Decapod is daemonless. There is no long-lived service. The binary starts when an agent calls it and exits immediately after the call.
"Just use Decapod" is literal:
cargo install decapoddecapod init
Then continue with Claude Code, OpenAI Codex, Gemini CLI, Cursor, or any tool that can invoke a CLI command. Decapod is agent-agnostic and safe for concurrent multi-agent execution.
State is local and durable in .decapod/: shared context, decisions, and traces persist across sessions and remain retrievable over time.
Related: Evaluating AGENTS.md (ETH SRI, 2026) on context-file quality and agent cost/performance.
Assurance Model ✅
Decapod centers execution around three outcomes:
Advisory: clear next actions that tighten intent and reduce wasted loops.Interlock: hard policy boundaries that block unsafe or out-of-contract flow.Attestation: durable, structured proof that completion criteria actually passed.
Operating Model ⚙️
Human Intent
|
v
AI Agent(s) <----> Decapod Runtime <----> Repository + Policy
| | |
| | +-- Interlock (enforced boundaries)
| +------- Advisory (guided execution)
+------------ Attestation (verifiable outcomes)
Features ✨
- ✅ Daemonless runtime: invoked on demand, exits immediately after each call.
- ✅ Agent-agnostic CLI/RPC surface for Claude, Codex, Gemini, Cursor, and more.
- ✅ Multi-agent-safe task/workspace coordination in the same repo.
- ✅ Work Unit manifests (
intent/spec/state/proof) with governed status transitions. - ✅ VERIFIED completion gating backed by explicit proof-plan pass results.
- ✅ Publish-time and validate-time workunit enforcement.
- ✅ Deterministic context capsule query (
core|interfaces|pluginsscoped). - ✅ Optional persisted context capsules in
.decapod/generated/context/*.json. - ✅ Knowledge promotion firewall ledger in
.decapod/data/knowledge.promotions.jsonl. - ✅ Procedural knowledge writes require event-backed promotion provenance.
- ✅ Repo-native durable state and traces under
.decapod/(local-first, auditable). - ✅ Authoritative
decapod validategate for promotion readiness.
And dozens more. For the full high-level and data-level surface area, see decapod docs show core/INTERFACES.md and the override template at .decapod/OVERRIDE.md.
Getting Started 🚀
cargo install decapod
decapod init
Then keep using your agents normally. Decapod is called from inside those agent runs when control-plane decisions are needed.
Agent integration: If you use Claude Code / Codex / Gemini / Cursor / similar tools, see AGENTS.md and the tool-specific entrypoint files (CLAUDE.md, CODEX.md, GEMINI.md) for the exact operational contract.
Learn more about the embedded constitution via the CLI:
Override constitution defaults with plain English in .decapod/OVERRIDE.md.
Contributing 🤝
Documentation 📚
- Development guide: CONTRIBUTING.md
- Security policy: SECURITY.md
- Release history: CHANGELOG.md
Support 💖
License 📄
MIT. See LICENSE.