agent-context-sync

name	agent-context-sync
description	Periodic, repo-agnostic doctor for AI-context files (CLAUDE.md, AGENTS.md, .cursorrules, .github/copilot-instructions.md). Verifies every concrete claim against the current repo, lints against an opinionated quality spec, surfaces newly-added subsystems the doc misses, and proposes the minimum edits to restore health. Conservative pruning. No-op when files are already healthy — safe to run on a cron.
argument-hint	[--dry-run] [--files=<comma-separated-paths>] [--canonical=<path>] [--since=<git-revision>]
disable-model-invocation	true

AGENT CONTEXT SYNC

Keep a repository's AI-context files accurate, lean, and useful. This skill is user-triggered only — never auto-invoke it. It edits documentation (or opens a PR), which the user must explicitly approve via the slash command.

---

INVARIANTS

These rules apply to every run. Do not restate them, just follow.

1. Conservative pruning. Only delete claims you can programmatically prove are wrong or stale (failed path check, mismatched version, miscount, missing command/script). Anything else — bloat, suspected anti-patterns, unverifiable claims — is flagged in the report but left in place. Humans decide.
2. No-op when healthy. If every claim verifies and lint passes and the recent git delta surfaces no missing content, print the healthy banner and exit. Do not create a branch. Do not open a PR. Do not touch files.
3. One pass, deterministic. Same inputs → same output. No "creative rewriting" — every edit must trace back to a finding in the verification or lint phase.
4. Repo-agnostic. No hardcoded paths, languages, or frameworks. Detect everything from the repo itself.
5. Preserve voice and curated content. If the existing file has a "What Not to Do" section or specific gotchas, keep them. The goal is correctness, not stylistic homogenization.
6. Never invent. If you can't verify a claim, leave it and flag it as unverifiable in the report. Don't replace it with a guess.
7. Reference files are authoritative. Read quality-spec.md, verification-patterns.md, and file-format-conventions.md before phase 2.

Logging format (mandatory):

[ACS-Pn] [ACTION|RESULT|DECISION|COMPLETE] message

where n is the phase number.

---

ARGUMENTS

• --dry-run — print the structured report to stdout. No file edits, no branch, no PR. Default is PR mode.
• --files=<paths> — comma-separated list of AI-context files to check. Overrides discovery. Useful for non-standard layouts.
• --canonical=<path> — designate one file as the canonical source for cross-file divergence comparisons. Defaults to CLAUDE.md if present, else the most recently modified file.
• --since=<git-revision> — git revision to diff against for the augmentation phase. Defaults to the canonical file's last-modifying commit.

---

PHASES

Run in order. Do not skip. Do not reorder.

PHASE 0: Pre-flight

1. Confirm you are inside a git repository: git rev-parse --show-toplevel. If not, print an error and stop.
2. Confirm the working tree is clean (git status --porcelain). If dirty and not --dry-run, refuse: tell the user to stash or commit first. Dirty trees are allowed in --dry-run.
3. Resolve the current branch name. If on the repository's default branch and not --dry-run, you'll be creating a new branch in Phase 5 — that's fine.
4. Read the three reference files in this skill directory: quality-spec.md, verification-patterns.md, file-format-conventions.md.

Log: [ACS-P0] [COMPLETE] preflight ok, repo=<name>, branch=<current>, dry-run=<bool>

PHASE 1: Discover

Find AI-context files. Check in this exact order, recording which exist and their last-modified commit (git log -1 --format=%H:%ct -- <path>):

Path	Format
CLAUDE.md	Markdown
AGENTS.md	Markdown
.cursorrules	Plain text
.github/copilot-instructions.md	Markdown

Also check if --files overrides this list. If yes, only operate on those.

Canonical selection:

• If --canonical provided, use it.
• Else if CLAUDE.md exists, it's canonical.
• Else use the file with the most recent last-modifying commit.

If zero AI-context files exist:

• In --dry-run: print a report recommending creation of CLAUDE.md based on detected stack, with a suggested skeleton. Exit.
• In PR mode: refuse with [ACS-P1] [DECISION] no AI-context files found — bootstrapping is out of scope for this skill. Create a starter CLAUDE.md manually first, then re-run. Exit.

Log each discovered file: [ACS-P1] [RESULT] found=<path> last_modified=<iso8601> bytes=<n> lines=<n>

PHASE 2: Verify claims

For each discovered file, extract verifiable claims using the patterns in verification-patterns.md and run each check. Build a per-file finding list. Each finding has:

{
  file: <path>,
  line: <n>,
  claim: <verbatim quote>,
  pattern: <which extraction pattern matched>,
  check: <what was tested>,
  status: ok | wrong | stale | unverifiable,
  evidence: <command output or file contents proving the status>,
  proposed_fix: <if status is wrong/stale, the corrected text — else null>
}

Status definitions:

• ok — check passed, claim is current.
• wrong — check failed and there is a clearly correct replacement (e.g. file says "version 1.0.4" but plugin.json says "2.0.0").
• stale — check failed and there is no clear replacement (e.g. claim references a deleted directory). Conservative pruning means: flag, but only delete if also referenced by another wrong finding that supersedes it.
• unverifiable — pattern matched but the check can't be run automatically (e.g. claim about runtime behavior, "the cache invalidates every 5 minutes"). Always left in place.

Do not produce findings for content that doesn't match any extraction pattern. Lint handles unstructured concerns separately.

Cross-file divergence pass: for each verifiable fact appearing in 2+ files (e.g. a version number, a build command), if the files disagree, emit a divergence finding scoped to the non-canonical files with the canonical value as proposed_fix. Note: divergence findings are advisory — they are reported but NOT auto-fixed even in PR mode, because the non-canonical file might intentionally lag (e.g. .cursorrules is rebuilt by a separate tool).

Log a summary per file: [ACS-P2] [RESULT] file=<path> ok=<n> wrong=<n> stale=<n> unverifiable=<n>

PHASE 3: Lint against quality spec

For each file, run the checks in quality-spec.md:

• Required sections present? Use the section list appropriate to the file format (markdown vs plain-text). Missing sections produce lint:missing-section findings.
• Length within budget? Soft cap, hard cap, and minimum from quality-spec.md. Produce lint:length finding if outside soft cap; lint:length-critical if outside hard cap.
• Anti-pattern scan. Run each detector in quality-spec.md (directory catalogue, fluff, redundant style rules, manifest duplication, stale TODOs). Produce lint:anti-pattern findings with the matched lines.

Anti-pattern findings are advisory only under the conservative pruning policy. They surface in the report but are not auto-removed.

Log: [ACS-P3] [RESULT] file=<path> lint_findings=<n> antipatterns=<n>

PHASE 4: Augment

Look for material additions to the repository that the canonical file does not mention.

1. Determine the diff base: --since argument, or the canonical file's last-modifying commit (git log -1 --format=%H -- <canonical>).
2. Compute git log --since-as-of=<sha> --name-status to get changed paths. Group by top-level directory.
3. For each significantly-changed area (new top-level directory, new manifest file, new build script, ≥10 changed files in a subsystem), check whether the canonical file mentions it. If not, emit a missing-coverage finding with:
   • The area (e.g. "new plugin plugins/github-maintenance/")
   • Evidence (commit list, file count)
   • Suggested addition (1–3 sentence summary, written in the file's existing voice)

Filter: do NOT propose additions for content an AI could discover by grep in 30 seconds. Examples to skip: a new test file, a renamed variable, a CSS tweak, a single new endpoint within an already-documented subsystem.

Log: [ACS-P4] [RESULT] missing_coverage_findings=<n>

PHASE 5: Output

If --dry-run:

Print a single Markdown report to stdout with this structure:

# agent-context-sync report

**Repo:** <owner>/<name>
**Canonical file:** <path>
**Run mode:** dry-run
**Result:** <healthy | findings>

## Summary

| File | OK | Wrong | Stale | Lint | Missing coverage |
|------|----|----|------|------|------------------|
| ... | ... | ... | ... | ... | ... |

## Findings

### <file>

#### Wrong (will fix in PR mode)
- **Line N:** `<claim>` — <reason>. Proposed: `<fix>`.

#### Stale (will fix in PR mode)
- ...

#### Lint (advisory, not auto-fixed)
- ...

#### Missing coverage (will add in PR mode)
- ...

#### Cross-file divergence (advisory, not auto-fixed)
- ...

## Recommended actions

- <bulleted list, prioritized>

If there are zero findings of any kind, print only:

[agent-context-sync] healthy — no changes proposed.

Exit.

If PR mode and zero findings:

Print the healthy banner (same as above). Do not create a branch. Do not open a PR. Exit.

If PR mode and findings exist:

1. Create branch: claude/agent-context-sync-<YYYYMMDD-HHMM> from the current branch.
2. For each wrong and stale finding with a proposed_fix, apply the edit. Use exact-string replacement; preserve surrounding context including indentation and whitespace.
3. For each missing-coverage finding, insert the suggested addition into the most appropriate existing section. If no section fits, append a new section in the file's existing style.
4. Do NOT apply lint findings or divergence findings (conservative policy).
5. Stage and commit with message:

   docs: refresh AI-context files (agent-context-sync)
   
   - <count> wrong claims corrected
   - <count> stale references removed
   - <count> new subsystems documented
   
   Lint findings and cross-file divergences reported but not auto-fixed (conservative policy).
   

6. Push the branch.
7. Open a PR against the repo's default branch with:
   • Title: docs: refresh AI-context files (agent-context-sync)
   • Body: the same Markdown report from the dry-run output, prefixed with a one-paragraph summary.
8. Print the PR URL and exit.

Log: [ACS-P5] [COMPLETE] mode=<dry-run|pr> findings_applied=<n> pr_url=<url-or-null>

---

OUTPUT CONTRACT

Always print exactly one report or one healthy banner per run. Never print partial progress to stdout — use the [ACS-Pn] log lines for that, on stderr if your environment separates streams.

The report is the deliverable. The PR (when in PR mode) is a delivery mechanism for the same report plus the corresponding edits. A human reading the PR must be able to understand every change by reading only the report.

---

FAILURE MODES TO AVOID

• Rewriting the whole file. You are a doctor, not a ghostwriter. If you find yourself touching more than 20% of the file in a single run, stop and reconsider — you are probably regenerating instead of repairing.
• Generating prose to fill perceived gaps. Missing-coverage findings should be terse and trace to specific git history. If you can't point to a commit or directory, don't add the content.
• Touching files outside the AI-context set. This skill edits only the files discovered in Phase 1. Never edit code, manifests, or other docs.
• Auto-fixing lint or divergence findings. These are advisory. Surface them; let humans decide.
• Running without a clean working tree in PR mode. Refuse, don't merge your edits with the user's in-flight work.