Align Docs

Align Docs — one canonical agent context

Claude Code, Codex CLI, and every other agent read the same canonical context from the repo, without duplication. Convert structure where missing, capture what the session learned, then reduce every docs/agents file and project memory to the most critical, non-trivial domain · infrastructure · company · project-specific facts in the fewest words that stay clear. Everything else is noise — removed. The repo is the only context teammates share: anything load-bearing outside it is invisible to them and lost.

Target architecture

repo/
├── CLAUDE.md                  # single line: @AGENTS.md
├── AGENTS.md                  # canonical context, ≤70 lines, references rules by path
├── .agents/rules/             # single source of truth — path-scoped rule files
├── .claude/
│   ├── rules → ../.agents/rules   # symlink; Claude auto-loads
│   └── CLAUDE.md                  # gitignored — personal per-project (Claude)
└── .gitignore                 # ignores .claude/CLAUDE.md and AGENTS.override.md

Step 0 — Root & mode

root=$(git rev-parse --show-toplevel 2>/dev/null) || root=. — all paths below resolve under $root. $ARGUMENTS contains --clean / clean → read references/clean-mode.md and follow it (runs Steps 1-7, then deep consolidation).

Step 1 — Inventory

ls -la CLAUDE.md AGENTS.md .claude/rules .agents/rules MEMORY.md 2>&1
wc -l AGENTS.md CLAUDE.md 2>&1
grep -nE "\.claude/CLAUDE\.md|AGENTS\.override\.md" .gitignore 2>&1
ls ~/.claude/projects/*/memory/ 2>&1   # pick the dir matching this project

Report what exists, what's missing, what's over budget. Then diff docs against disk: structure trees vs ls, versions vs manifests, README tables (skills/agents/shared) vs directories, every referenced path exists, links resolve.

Step 2 — Structure (idempotent)

• CLAUDE.md — overwrite with exactly one line, no trailing prose: @AGENTS.md
• .agents/rules/ — create if missing. One topic per file, descriptive name. Every file starts with scope frontmatter (else it loads into Claude's context every session):

---
paths:
  - "src/**"
---

• Symlink — mkdir -p .claude; [ -e .claude/rules ] || ln -s ../.agents/rules .claude/rules; commit it (git stores symlinks natively on Unix).
• .gitignore — append .claude/CLAUDE.md and AGENTS.override.md if missing. Tell the user: personal notes → .claude/CLAUDE.md (Claude, auto-loaded) or ~/.codex/AGENTS.md (Codex, user-global); AGENTS.override.md replaces shared AGENTS.md — avoid unless scoped to a subdirectory.

Step 3 — AGENTS.md (≤70 lines)

Prepend references/agents-template.md verbatim at the top (replace a similar existing block). Then, in this order and nothing else — no section headings or decoration:

• Project scope (1-2 sentences)
• Single fewest-words list of the most critical, non-trivial domain·infra·company·project facts (safety rules, test commands; tables/commands beat prose)
• Rules index table at the bottom — one row per file in .agents/rules/: Rule | Scope (paths:) | Topic

If it's in a rule file, reference — never duplicate.

Step 4 — Capture session discoveries

Cold invocation (no prior work this session) → state "CAPTURE: NONE". Else harvest what this session learned that it lacked at the start:

• Non-inferable only — domain·infra·company·project-specific, not derivable from the code. No structure/architecture overviews.
• Idempotent — write only facts not already on disk.
• Critical always-read fact → AGENTS.md line; topic with depth → .agents/rules/<topic>.md + index row. Never memory.

Step 5 — Memory sweep

Memory is per-machine: teammates share only the repo, so facts parked there diverge per person and get lost. Read every project-memory file (Claude auto-memory ~/.claude/projects/<project>/memory/, repo MEMORY.md if present):

• Critical, non-trivial domain·infra·company·project fact → move to AGENTS.md or .agents/rules/<topic>.md, delete from memory.
• Machine-specific (env paths, personal tooling) → gitignored .claude/CLAUDE.md / ~/.codex/AGENTS.md.
• Leave memory ≈ empty: only what fits neither home.

Step 6 — Total cleanup

Not just AGENTS.md ≤ 70 lines: audit every docs/agents file — AGENTS.md, CLAUDE.md, .agents/rules/**, READMEs, docs/**. Every line passes one gate: most critical, non-trivial fact the next agent must read — domain·infra·company·project-specific — in the fewest words? Keep it at its right home (always-read → AGENTS.md · topic depth → rules file · WHY → chronicle · HOW → plan), else delete.

• Fix misalignments found in Step 1 (structure, versions, paths, README tables, index rows — including files added in Steps 4-5). Chronicle bodies are immutable — touch only if they reference incorrect content.
• Remove noise: entries referencing nonexistent things · duplicates (keep once, link from others) · empty sections/placeholders · generic content the agent already knows from general training.
• Preserve all load-bearing content (safety rules, domain glossary, test commands, rules index) — shrink the expression, never the function.

Step 7 — Verify & gates

wc -l AGENTS.md CLAUDE.md
[ "$(cat CLAUDE.md | tr -d '[:space:]')" = "@AGENTS.md" ] && echo "CLAUDE.md OK"
readlink .claude/rules && diff <(ls .claude/rules/) <(ls .agents/rules/) && echo "symlink OK"
git check-ignore -v .claude/CLAUDE.md AGENTS.override.md

Report: AGENTS.md line count · rule files vs index rows · symlink status · gitignore entries · memory residue. STOP and fix if any invariant fails:

• CLAUDE.md is not exactly @AGENTS.md.
• AGENTS.md exceeds 70 lines — trim further.
• A rule file lacks paths: frontmatter (else it loads every session).
• A rule file is absent from the AGENTS.md index table — index by textual row, never @import (Codex has no imports).
• A domain·infra·company·project fact survives only in memory / MEMORY.md — move it into AGENTS.md or .agents/rules/.
• AGENTS.override.md committed beside AGENTS.md (two sources of truth) — or .claude/CLAUDE.md committed. Never commit either.