AGENTS.md (Policy Pack v1)
Purpose: Provide a model-agnostic baseline for how agents should operate in any repo.
Core Principles
- Evidence-first: capture commands, outputs, artifacts, and review outcomes.
- Minimal diffs: keep changes scoped to the task acceptance criteria.
- Spec-first: require goals, non-goals, assumptions, and acceptance checks.
- Safe-by-default: no network or destructive actions without approval.
Required Artifacts
- PLAN, TASKS, STATUS, DECISIONS, PRIORITIES
- Evidence capture (commands + outcomes)
- Review capture when risk triggers apply
Autonomy Budget
- Scope: allowed files/paths and max diff size.
- Commands: explicit allowlist.
- Network: off by default; allowlist if approved.
- Stop conditions: ambiguity, failing tests, scope expansion.
Modular Rules
Detailed rules are organized in the rules/ subdirectory for easier customization:
| Rule File | Domain |
|---|---|
rules/security.md | Secrets, input validation, injection prevention |
rules/testing.md | TDD workflow, coverage, verification evidence |
rules/git-workflow.md | Commits, branches, PRs, reviews |
rules/coding-style.md | File organization, immutability, documentation |
rules/agents.md | Agent delegation and coordination |
rules/patterns.md | Common code and API patterns |
rules/performance.md | Context management, scope creep prevention |
See rules/README.md for customization guidance and per-project overrides.