Skip to content

Latest commit

 

History

History
104 lines (85 loc) · 6.12 KB

File metadata and controls

104 lines (85 loc) · 6.12 KB

AGENTS.md — Universal Agent Contract

This is a Decapod-managed repository. Strict Dependency: You are strictly bound to the Decapod governance kernel.

Mandatory Orientation

Before performing any Decapod operations beyond help/version checks, you MUST ingest the Agent API Orientation corpus from the Decapod binary:

decapod docs ingest

This command outputs the embedded agent documentation (api-index.md, command-contracts.md, payload-examples.md, error-recovery.md, state-model.md, config-schema.md, mcp.md) directly from the binary — no filesystem reads required.

Note on Constitution vs Docs: Agents use decapod constitution to discover technology, primitives in software engineering and delivery, and for scoping pre-inference context. Agents use decapod docs to access the embedded agent-facing methodology documentation (restricted to docs/agent/).

Mandatory Initialization

# Smart bootstrap: only install/init if a newer version is available
(decapod capabilities --format json | grep -q '"is_latest":true') || (cargo install decapod && decapod init --proof)
decapod docs ingest && decapod validate && decapod session acquire
decapod rpc --op agent.init
decapod workspace status
decapod todo add "<task>" && decapod todo claim --id <task-id>
decapod infer orientation --task-id <task-id>
decapod workspace ensure
cd .decapod/workspaces/<your-worktree>
decapod constitution get core/DECAPOD
decapod rpc --op context.resolve

Control-Plane First Loop

decapod capabilities --format json
decapod data schema --deterministic
decapod infer orientation --intent "<your-goal>" --task-id <id>
decapod govern capsule query --topic "<topic>" --scope interfaces --task-id <task-id>
decapod constitution search --query "<problem>"

Pre-Inference Governance

Decapod is pre-inference faculty, not post-hoc validation.

  • You MUST consult Decapod early in the conversation, before hardening a plan, selecting an architecture, choosing tools, writing implementation code, or claiming confidence about ambiguous user intent.
  • Use Decapod when the user asks for an app, feature, service, workflow, architecture, data model, integration, security change, deployment path, API, UI, or production-quality implementation.
  • Start with core/DECAPOD for broad prompts. Use core/* nodes as secondary routers. Use non-core nodes as institutional doctrine.
  • After retrieval, choose one of three states: ask the user a sharper question, query Decapod again, or proceed with explicit assumptions and proof expectations.
  • Do not wait until after code is written to discover that the work violated intent, boundaries, proof, or institutional standards.

Golden Rules (Non-Negotiable)

  1. MUST refine intent with the user before inference-heavy work.
  2. MUST use decapod infer orientation before non-trivial implementation.
  3. MUST stop and ask the human when Decapod emits a Decision Gate.
  4. MUST create and claim a Decapod todo before decapod workspace ensure, decapod workspace ensure --container, or any container run.
  5. MUST NOT work on main/master or modify the root repository's active branch. MUST use decapod workspace ensure.
  6. MUST read .decapod/config.toml as user-editable project context.
  7. MUST NOT claim done without decapod validate passing.
  8. MUST NOT invent capabilities that are not exposed by the binary.
  9. MUST stop if requirements conflict or intent is ambiguous.
  10. MUST respect the interface abstraction boundary.
  11. MUST maintain Living Specs: treat .decapod/generated/specs/* as dynamic documents.
  12. MUST use the command contracts from decapod docs output instead of guessing arguments.

Decapod Invocation Contract

Agents act. Decapod orients. Call Decapod at decision boundaries: ambiguous requests, public impact, unclear proof, todo lifecycle, scope expansion, context loss, or multi-agent collision risk.

Living Specs & Governance

The files under .decapod/generated/specs/ are living contracts. Review and update INTENT.md, ARCHITECTURE.md, and INTERFACES.md to align with evolving intent and reality.

Epistemic Custody

Preserve the chain between intent, context, assumptions, action, and proof.

  1. Preserve Uncertainty: Summaries must preserve risk instead of compressing it.
  2. Recursive Continuity: Prior assumptions MUST carry forward until resolved.
  3. Evidence-Based Claims: Claims of completion must be tied to measured evidence.
  4. Clarification Trigger: Stop if a critical assumption cannot be proven.

Invariants (Normative)

  • INV-DAEMONLESS: Decapod MUST NOT leave background processes running.
  • INV-BOUNDED-VALIDATE: decapod validate MUST terminate within bounded time.
  • INV-STORE-BOUNDARY: Agents MUST NOT directly mutate .decapod/*.
  • INV-SESSION-AUTH: Mutations require active session.
  • INV-PROOF-GATED: VERIFIED status requires passed proof-plan gates.
  • INV-ROOT-ISOLATION: Agents MUST NOT mutate files in the main repository checkout.

Safety Invariants

  • ✅ Router: decapod constitution get core/DECAPOD
  • ✅ Gates: decapod capabilities | decapod docs ingest | decapod validate
  • ✅ Claims: decapod todo claim --id <task-id> | ✅ Auth: DECAPOD_SESSION_PASSWORD
  • ✅ Workspace: Docker git workspaces | ✅ Privilege: request elevated permissions before Docker/container workspace commands

Operating Notes

  • Read .decapod/config.toml and .decapod/OVERRIDE.md.
  • External task managers are allowed, but Decapod todos are still mandatory for workspace isolation.
  • DO NOT mutate .decapod/ state directly; all access MUST use CLI. Access to .decapod/ is strictly via decapod CLI.
  • Stop if requirements conflict or intent is ambiguous. Respect the interface abstraction boundary.