AGENTS.md

Ref: main worktree-web-client-spec   Size: 1.7 KiB   History

# Agent instructions

## Plan first; main agent coordinates

These rules apply to the top-level agent. Delegated workers may implement
their approved briefs.

- Begin change requests with read-only inspection and a proposed plan.
  Do not edit files or run builds/tests before the user approves it.
- Plans must specify behavior, scope, file ownership, acceptance checks,
  and the model and reasoning effort assigned to each agent.
- After approval, delegate implementation, test changes, validation, and
  review fixes. The main agent coordinates and assesses returned evidence;
  it does not implement or run tests itself.
- Give each worker a self-contained brief, including existing dirty work,
  constraints, required checks, and its stopping condition.
- Explicitly select worker models; do not inherit the main agent's model.
  Use a separate agent for independent review.
- “Do it” or “implement” approves a presented plan; it does not waive
  delegation. Small changes and one-line fixes are not exceptions.
- If delegation is unavailable or a worker is blocked, report that and
  ask for direction. Do not silently take over implementation.
- Only an explicit user instruction may waive planning or authorize the
  main agent to implement directly.

## Context
Read [CLAUDE.md](CLAUDE.md) for the repository's toolchain, architecture,
invariants, and working rules; those instructions apply to every coding agent.

When the user requests sprint delivery, read and follow the repository's
[sprint-delivery skill](docs/skills/sprint-delivery/SKILL.md). It is ordinary
Markdown and can be followed without a skill loader or a personal installation.
Use this repository copy for mux work. Ordinary isolated edits do not require
the sprint workflow.