Why Decapod 🧠
AI coding agents can write code fast. Shipping it safely is the hard part.
Decapod gives agents a consistent operational contract: guided execution, enforceable boundaries, and auditable completion signals. It replaces "looks done" with explicit outcomes.
Decapod is invoked by agents; it never runs in the background. It is a single executable binary that provides deterministic primitives:
- Retrieve canon (constitution .md fragments) as context.
- Provide authoritative schemas for structured state (todos, knowledge, decisions).
- Run deterministic validation/proof gates to decide when work is truly done. Example gate: forbid direct pushes to protected branches — fails if the agent has unpushed commits on main.
AGENTS.md stays tiny (entrypoint). OVERRIDE.md handles local exceptions. Everything else is pulled just-in-time.
Traces: .decapod/data/traces.jsonl. Bindings: context.bindings. Architecture-agnostic (not coupled to a specific OS or CPU).
Recent independent research confirms this design direction: Evaluating AGENTS.md (Gloaguen et al., ETH SRI, 2026; AgentBench repo) found that LLM-generated context files tend to reduce agent performance while increasing cost by over 20 %; human-written minimal requirements can help slightly. Decapod was built independently and without knowledge of ETH SRI's AgentBench research or this paper.
Assurance Model ✅
Decapod is built around three execution outcomes:
Advisory: guidance toward the next high-value move.Interlock: hard stops for unsafe or out-of-policy flow.Attestation: structured evidence that completion criteria were met.
Operating Model ⚙️
Human Intent
|
v
AI Agent(s) <----> Decapod Runtime <----> Repository + Policy
| | |
| | +-- Interlock (enforced boundaries)
| +------- Advisory (guided execution)
+------------ Attestation (verifiable outcomes)
Features ✨
- Agent-native CLI and RPC surface for deterministic operation.
- Guided project understanding through structured prompting.
- Standards-aware execution aligned with project policy.
- Workspace safety for isolated implementation flow.
- Validation and completion gates with explicit pass/fail outcomes.
- Multi-agent-ready orchestration surface for tooling integrations.
Getting Started 🚀
Install Decapod with Cargo, initialize it in your repository, and let your agent operate through the Decapod contract instead of direct ad-hoc repo mutation.
$ decapod validate
✅ constitution/core PASS
✅ workspace isolation PASS
❌ uncommitted changes FAIL — 2 files modified outside worktree
For command details and full usage, use decapod --help.
Contributing 🤝
Documentation 📚
- Development guide: CONTRIBUTING.md
- Security policy: SECURITY.md
- Release history: CHANGELOG.md
Support 💖
License 📄
MIT. See LICENSE.