agentware AGENTS.md

AGENTS.md — Canonical Methodology (Deterministic Steering Format)

Single source of truth for agent behaviour. Loaded as always-on context because the auto-loaded CLAUDE.md imports it with @AGENTS.md. The agentware subagents reference it too. Rationale and examples live in docs/methodology.md (NOT agent-loaded). Every rule is atomic, carries a stable ID, and is enforced by scripts/agentware steering lint.

agentware ships ZERO personal data. The knowledge base lives in an EXTERNAL directory the operator chose at onboarding. Resolve it with scripts/agentware config --knowledge-dir-only; NEVER hardcode its path.

Execution loop

Every task runs this loop. Do not skip steps; do not stop until all subtasks are done.

1. PLAN — break the task into verifiable subtasks; save the plan.
2. EXECUTE — run commands / write code for one subtask.
3. VERIFY — confirm it works with the project's own build/test/health command.
4. NEXT — move to the next subtask.
5. REPEAT — until every subtask is complete.
6. COMPLETE — update the knowledge base, then present results.

Build and complete

• MUST treat building and completing tasks as the primary job. [R-EXEC-01]
• NEVER mark a task complete while any subtask is in progress or unverified. [R-EXEC-02]
• ALWAYS ask "is the original goal fully achieved?" before presenting results. [R-EXEC-03]
• IF something is bootstrapping THEN wait and verify; NEVER defer it to the user. [R-EXEC-04]
• IF a step failed THEN fix it and retry; NEVER present failure as completion. [R-EXEC-05]
• IF blocked by a dependency not resolvable under R-AUTO-01 THEN document the blocker and ask the user. [R-EXEC-06]
• ALWAYS track ongoing work in the worklog and plan status markers. [R-EXEC-07]
• NEVER pivot to a different approach mid-task without user approval; in-scope judgment is licensed by R-AUTO-01. [R-EXEC-08]
• ALWAYS update the knowledge base with what was built, what works, and gotchas after completion. [R-EXEC-09]

Planning