Init or Refactor AGENTS.md

Produce one short, high-signal canonical AGENTS.md.

Workflow

1. Read AGENTS.md, CLAUDE.md, CODEX.md, .cursorrules, and similar repo-local instruction files.
2. Present conflicting instructions with their sources and ask which wins. Flag conflicts with the target policies below, especially speculative, implementation-detail, or fixture-heavy test matrices, unapproved dependencies, blind preservation of local patterns, and rejection of justified rewrites.
3. Prune before drafting, in order: delete vague, obvious, generic, or duplicate advice; collapse repeats; remove tool-wrapper noise and future ceremony; remove speculative, private, branch-matrix, fixture-heavy, or vanity tests and rules that preserve bad structure; retain repo-specific, safety-critical, or repeatedly violated rules. Prefer one bullet to a new section. Group removals as delete, yagni, tool-noise, test-theater, or structure.
4. Summarize existing branch, commit, PR, review, release, and deploy rules. Ask the user to confirm direct-default (work on main or trunk), branch/PR (use a focused branch and open or update its PR), or another workflow; recommend a clear repo convention. For direct-default or branch/PR, require scoped logical commits and early, frequent pushes; for branch/PR, open or update the PR after the first useful slice. Never rewrite shared history without approval. For another workflow, preserve only what the user supplies.
5. Draft AGENTS.md with only a one-line project description, non-obvious commands, the confirmed Git workflow, the target policies, and critical safety or approval rules. Use terse bullets and no linked detail docs unless explicitly asked.
6. Present the sources, conflicts, Git choice, proposed AGENTS.md, and tagged removals. Get confirmation before writing.
7. Write AGENTS.md, then delete the other contributing instruction files after consolidating their retained instructions. Create no tool-specific instruction files unless explicitly asked.

Target policies