hf-l2-agents-md-sync

name	hf-l2-agents-md-sync
description	Detects and fixes drift between AGENTS.md documentation and actual codebase state. Scans source files and .opencode/ directories, compares claims against reality, produces a structured drift report, then applies targeted edits. Triggers on: 'sync agents md', 'update AGENTS.md', 'fix agents md drift', 'AGENTS.md out of date', 'check agent instruction drift'. NOT for generic documentation writing or README refreshes.
version	1.0.0
metadata	{"layer":"2","role":"sync","pattern":"scan-diff-apply"}
allowed-tools	["Read","Edit","Bash","Glob","Grep"]

Overview

Detects and fixes drift between AGENTS.md documentation and actual codebase state. Use when documentation claims don't match reality, after refactoring sprints, or when syncing instructions across platforms. Produces a structured drift report with targeted edits to bring documentation back in line with code.

The Iron Law

SCAN REPORTS FACTS. EDIT FIXES DRIFT. NEVER REGENERATES FROM SCRATCH.

This skill reads the codebase, compares claims to reality, and applies targeted edits. It never overwrites entire files, never generates new AGENTS.md files, and never adds sections that don't already exist.

agents-md-sync

Detects and fixes AGENTS.md drift. Scan → Diff → Apply. Never regenerates from scratch.

On Load

1. Verify both AGENTS.md targets exist:
   • AGENTS.md (project root)
   • src/lib/AGENTS.md (library internals)
2. If either is missing → report the missing file and stop. This skill syncs existing documentation — it does not create new AGENTS.md files. For creation, use the create-agentsmd skill instead.
3. Read both files into memory for diff comparison.

Phase 1: Scan — Gather Ground Truth

Run these scans in sequence. The project is small enough that no parallelism is needed.

Scan Targets

Path	Extract	Command
src/lib/*.ts	Module names + approximate LOC	Glob("src/lib/*.ts") then Bash("wc -l src/lib/*.ts")
src/tools/**/*.ts	Tool file names	Glob("src/tools/**/*.ts")
src/hooks/**/*.ts	Hook file names	Glob("src/hooks/**/*.ts")
src/shared/**/*.ts	Shared utility names	Glob("src/shared/**/*.ts")
src/schema-kernel/**/*.ts	Schema file names	Glob("src/schema-kernel/**/*.ts")
.opencode/agents/*.md	Count + file names	Glob(".opencode/agents/*.md")
.opencode/skills/*/SKILL.md	Count + skill names	Glob(".opencode/skills/*/SKILL.md")
.opencode/commands/*.md	Count + command names	Glob(".opencode/commands/*.md")
package.json	Dependencies, scripts, engines	Read("package.json")

Extraction Rules

• Module names: filename without .ts extension (e.g., concurrency.ts → concurrency)
• LOC: wc -l per file, rounded to nearest 5 for estimates (e.g., 78 → ~80, 401 → ~400)
• Counts: exact integer from glob result array length
• Names: sorted alphabetically for stable comparison
• Export functions: Grep("^export (function|async function|const)", path) per module — only extract if AGENTS.md claims specific exports

Phase 2: Diff — Compare Claims to Reality

Read each AGENTS.md file and compare every verifiable claim against the scanned ground truth.

Root AGENTS.md Checks

Category	What to Check	Typical Location
Project Structure tree	All directories and files listed match Glob reality	## Project Structure section
src/lib/ module list	Every .ts file appears in tree, no phantom entries	Tree listing under src/lib/
src/shared/ presence	Directory exists with correct files	Tree listing
src/schema-kernel/ presence	Directory exists with correct files	Tree listing
.opencode/ agent count	Claimed count matches Glob result count	## OpenCode Integration section
.opencode/ skill count	Claimed count matches Glob result count	## OpenCode Integration section
.opencode/ command count	Claimed count matches Glob result count	## OpenCode Integration section
.opencode/ name lists	Listed names match actual files	## OpenCode Integration section
"Where to find things"	Every module/directory has a table entry	Table in that section
Dependency rules	Module list matches src/lib/ contents	### Dependency rules
Runtime features	Feature list matches implemented features	Narrative section
Architecture LOC targets	Total LOC claims match current wc -l totals	### Target architecture

src/lib/AGENTS.md Checks

Category	What to Check
MODULE RESPONSIBILITIES table	Every .ts file has a row, LOC estimates are current, export lists match Grep results
DEPENDENCY GRAPH	Every dependency edge still exists, new modules appear
"Where to Look" table	Every module has an entry
CODE SMELLS	Listed issues still exist (e.g., LOC thresholds still exceeded)

Drift Report Format

After comparison, produce a drift report in this exact format before making any edits:

## AGENTS.md Drift Report — {YYYY-MM-DD}

### Root AGENTS.md
| # | Section | Claimed | Actual | Fix |
|---|---------|---------|--------|-----|
| 1 | [section name] | [claimed value] | [actual value] | [specific edit description] |

### src/lib/AGENTS.md
| # | Section | Claimed | Actual | Fix |
|---|---------|---------|--------|-----|
| 1 | [section name] | [claimed value] | [actual value] | [specific edit description] |

### Summary
- Total drift items: {N}
- Auto-fixable: {N}
- Manual fixes needed: {N}

Quality gate: The drift report must contain specific claimed and actual values with the exact edit description before any edits are applied. No fuzzy edits.

Phase 3: Apply — Edit AGENTS.md Files

For each auto-fixable drift item in the report:

1. Locate exact text — Use Grep to find the exact claimed value in the target AGENTS.md file. Note the surrounding context to ensure uniqueness.
2. Edit — Use Edit(oldString=<exact-wrong-text>, newString=<correct-text>). The oldString must be a verbatim copy from the file — character-for-character match.
3. Verify — After each edit, Read the edited section to confirm the change applied correctly.
4. Log — Record the result (success or failure) for the summary.

Apply Rules

• Always use Edit tool — never Write on existing files. Both AGENTS.md files already exist and must be preserved.
• Never add new sections — only fix content in existing sections. If a section is missing entirely, flag it as a manual fix.
• Skip and flag — if the exact text is not found (perhaps already fixed, or the structure changed), log it as "MANUAL FIX NEEDED" and proceed to the next item.
• One edit at a time — each drift item gets its own Edit call. Never batch multiple changes into one edit.

Quality Gates

Gate	Rule	Enforcement
Edit-only	Never use Write tool on existing AGENTS.md files	Use Edit tool exclusively
Exact match	oldString must be verbatim from the file	Grep first, copy exact text
No new sections	Only fix existing content, don't create new sections	Skip missing sections, flag as manual
Verify after edit	Re-read the edited section to confirm change applied	Read immediately after Edit
Report before apply	Always show drift report to user before any edits	Two-pass: Phase 2 report → Phase 3 apply

Anti-Patterns

Anti-Pattern	Detection	Correction
The Regenerator	Using Write tool on existing AGENTS.md files	STOP. Use Edit tool only. Write destroys context.
The Hallucinator	Adding content not verified by filesystem scan	STOP. Only edit what the scan confirms is wrong.
The Skipper	Applying edits without producing a drift report first	STOP. Always run Phase 2 before Phase 3.
The Creator	Generating new AGENTS.md files for directories that don't have one	STOP. This skill syncs existing files. Use create-agentsmd skill instead.
The Over-Scoper	Scanning node_modules/, dist/, .git/, or other non-source directories	STOP. Only scan the paths listed in Phase 1 Scan Targets table.

Self-Correction

When the Task Keeps Failing

[Detection] Drift report shows same items across 3+ sync attempts. Edit fails because oldString not found despite appearing in the drift report. Phase 2 Diff over-corrects (reports drift that doesn't exist because file was already fixed). [Recovery] STOP making edits. Re-run Phase 1 Scan to get fresh ground truth. Rebuild the drift report from current state (not cached results). If an edit's oldString is not found, the text may have already changed — skip that item and move to the next. If the same drift items persist, the scan extraction rules may be incorrectly matching — verify the extraction GLobs and Grep patterns.

When Unsure About the Next Step

[Detection] AGENTS.md structure has changed significantly since last scan. Claimed values in AGENTS.md are in a different section than expected. Unclear whether a discrepancy is drift or intentional divergence. [Recovery] Read the relevant AGENTS.md section in full to understand the current structure. If the section was renamed or reorganized: update the "Typical Location" column in the scan table before proceeding. For intentional divergences: flag them as "MANUAL FIX NEEDED — possible intentional divergence" in the drift report rather than auto-fixing.

When the User Contradicts Skill Guidance

[Detection] User says "just regenerate the whole file" or "use Write instead of Edit" or "add a new section for the new modules." User wants to violate the "NEVER REGENERATES FROM SCRATCH" iron law. [Recovery] Acknowledge: "This skill's iron law is SCAN REPORTS FACTS. EDIT FIXES DRIFT. NEVER REGENERATES FROM SCRATCH. Regenerating destroys hand-written context, cross-references, and narrative sections." If the user insists, warn that regeneration will lose narrative content. Proceed only with explicit confirmation, and note in the drift report that regeneration was user-authorized.

When an Edge Case Is Encountered

[Detection] AGENTS.md references files that don't exist (phantom entries) but also has file entries not in the scan targets table. Multiple AGENTS.md files exist at different paths. File was renamed but AGENTS.md still uses old name. [Recovery] For phantom entries: flag for removal in the drift report with the note "file does not exist." For new files not in scan targets: add them to the scan targets list dynamically and re-scan. For renames: report both the old claimed name and the new actual name with a "rename" fix type. For multiple AGENTS.md files: process them one at a time, starting with the root AGENTS.md.