Agent Loop

Plan → Act → Observe → Verify → Stop

A stateful control cycle for autonomous engineering work. Read docs/agents/loops.md if it exists; if missing, use conservative defaults (iteration cap 5, no push/deploy/CI changes) and suggest /setup-agent-loops.

When exploring the codebase, read CONTEXT.md (if it exists) so changes use the project's domain vocabulary.

When you encounter a knowledge gap (unknown terminology, missing concepts, need to ground a decision), reach for research-from-vault (or run external research first then synthesize into the vault). Goals of the form "research X using the vault until a synthesis note and learning record exist" are valid and encouraged when the vault is configured.

1. Plan

Before the first change:

• State the goal and done signal (what verifier output means success).
• Read docs/agents/loops.md for verification commands, stop rules, and editable scope.
• If building features test-first, reach for the /tdd skill.
• Track iteration count from 1.