oma-docs - Documentation Drift Detector

Scheduling

Goal

Detect broken references in repo markdown — default glob **/*.md (verify mode), propose LLM-generated patch proposals for docs affected by recent code changes (sync mode), detect structural drift between English source docs and their translations (i18n mode), and lint translated docs for CJK style anti-patterns (lint mode). All modes run on-demand; sync is always interactive.

Intent signature

• User asks to check if docs are up to date, find broken doc links, verify file paths referenced in docs, or detect documentation drift.
• User asks to update docs after a code change, propose doc patches for a git diff, or sync affected docs.
• User asks whether translations are stale, which i18n docs drifted from the English source, or to lint translated docs for style issues (em-dashes in CJK targets, etc.).
• A workflow hook checks docs.auto_verify: true and runs oma docs verify --json at completion.

When to use

• After a refactor, rename, or file deletion, to find stale references in docs.
• Before a release, to confirm that CLI commands, file paths, and config keys in docs still exist.
• After a significant git diff, to discover which docs reference the changed files and may need updating.
• After updating English source docs (web/docs), to find translations (web/i18n/{lang}/...) that drifted or went missing.
• Before a docs release, to lint CJK translations for content-level anti-patterns.
• Routine drift check on any docs-heavy repo.

When NOT to use

• Generating docs from scratch for undocumented features → v2 create mode.
• Actually translating or restructuring docs → use oma-translator (oma docs i18n / oma docs lint only detect; they never edit translations).
• Symbol-level semantic drift (function signature changes not reflected in prose) → v2 L3 mode.
• CI-blocking enforcement → v2 block mode (v1 is warn-only).
• Explaining a code change as an educational document → use oma-explainer (this skill detects drift; it does not author explainers).

Expected inputs

verify mode: Optional glob path (default **/*.md), optional --json flag, optional --report-file <path>.

sync mode: Optional git diff range (default --cached, fallback HEAD~1..HEAD).

i18n mode: Optional --min-severity <CRITICAL|HIGH|MEDIUM|LOW> (default MEDIUM), optional --json.

lint mode: Optional --locales <list> (comma-separated CJK locales for the cjk-em-dash rule, default ko,ja,zh), optional --json. The default wrong-language rule always scans every locale under web/i18n/ and is not narrowed by --locales.

Expected outputs

verify mode:

• Markdown drift report to stdout (default), or raw JSON with --json, or full markdown written to file with --report-file.
• Exit code 0 if clean, 1 if any broken refs found.

sync mode:

• Per-doc patch proposals drafted by the host LLM from the CLI's candidate-doc list, confirmed per doc ([y] apply [n] skip [d] diff [s] full proposal style).
• Docs modified only on explicit user approval; doc-refs.json regenerated after applies.

i18n mode:

• Per-pair structural drift signals (line count, heading count, EN-newer recency flag) with severity, as markdown summary or --json. Report-only: drifting pairs are handed to oma-translator in diff-sync mode.

lint mode:

• Style issues grouped by rule / language / file, as markdown summary or --json. Report-only: restructuring goes through oma-translator.

Dependencies

• cli/commands/docs/extract.ts: markdown AST + L2 pattern extractor.
• cli/commands/docs/resolve.ts: deterministic broken-ref checker.
• cli/commands/docs/reporter.ts: deterministic markdown/JSON report renderer (no LLM call; host LLM does narrative synthesis).
• cli/commands/docs/sync-propose.ts: git diff intake, reverse lookup, candidate-doc selector with secret redaction (no LLM call; host LLM drafts patches).
• cli/commands/docs/i18n-drift.ts: EN↔translation structural drift detector (line/heading counts, last-commit recency; no LLM call).
• cli/commands/docs/lint-i18n.ts: translated-doc linter for selected-locale CJK em-dash style and all-locale wrong-language placeholders (no LLM call, no auto-fix).
• docs/generated/doc-refs.json: single-direction reference index, regenerated on every verify run. Gitignored — the CLI force-adds docs/generated/ to .gitignore so generated artifacts are never committed.
• docs/generated/url-drift.json: lychee-produced URL drift report (written by background lychee spawn; gitignored under the same docs/generated/ rule).
• lychee: external Rust tool for URL link checking. Detected on PATH; install via brew install lychee or see https://github.com/lycheeverse/lychee#installation. Optional but recommended.
• .agents/oma-config.yaml: docs.auto_verify (workflow hook opt-in), docs.check_urls (URL checking on/off, default true), and docs.exclude (glob list of markdown trees the walker must not scan — benchmark artifacts, translation mirrors, etc.; default []) toggles.

Control-flow features

• Mode is selected from the first argument: verify, sync, i18n, or lint.
• verify: extract → resolve → report (fully deterministic CLI; host LLM adds narrative summary on top of the JSON/markdown output).
• sync: git diff → reverse lookup → candidate list (CLI) → host-LLM patch proposals → interactive accept/reject.
• i18n / lint: fully deterministic CLI reports; host LLM routes drifting pairs / style issues to oma-translator.
• Branches on --json, --report-file, LLM availability, and network reachability.
• Never blocks workflow completion in v1 (warn-only hook policy).

Structural Flow

Entry

1. Read first argument to select mode (verify | sync | i18n | lint). If absent, print help and exit.
2. Load oma-config.yaml to check docs.auto_verify when invoked from a workflow hook.
3. Confirm required CLI dependencies (oma docs verify, oma docs sync) are on PATH.

Scenes

1. PREPARE: Determine mode, resolve path/diff-range arguments, confirm tool availability.
2. ACQUIRE: Run extractor (extract.ts) to regenerate doc-refs.json from repo markdown (**/*.md, verify) or build in-memory reverse index from doc-refs.json (sync; a cached index newer than 5 minutes is reused, otherwise regenerated).
3. REASON: Resolve each reference deterministically (verify) or correlate changed files to candidate docs via reverse lookup (sync).
4. ACT: Render the deterministic drift report (verify) or list candidate docs with matched refs (sync). Host LLM does any natural-language synthesis or patch drafting on top of this output.
5. VERIFY: Confirm output shape is valid (JSON schema check for --json; structured candidate list for sync).
6. FINALIZE: Print to stdout, write report file if requested, emit exit code.

Transitions

• verify mode: PREPARE → ACQUIRE (extract) → REASON (resolve) → ACT (report) → FINALIZE.
• sync mode: PREPARE → ACQUIRE (reverse index) → REASON (candidate matching) → ACT (LLM proposals) → VERIFY (interactive) → FINALIZE (apply approved).
• If LLM is unavailable in verify: skip reporter summary, emit raw JSON drift report.
• If LLM is unavailable in sync: emit candidate-list-only output (no patch proposals); user reviews manually.
• If doc-refs.json is missing or older than 5 minutes in sync: regenerate via the extractor, then continue.

Failure and recovery

• Extractor parse error on a single doc: skip doc + warn, continue with remaining docs.
• lychee unavailable or URL check incomplete: print install hint, skip URL checking, continue (core check is unaffected).
• Host-LLM context limit exceeded while drafting patches: process candidate docs in smaller batches.
• oma docs CLI not found: skip with installation hint (workflow hook: skip silently).
• doc-refs.json write failure: abort and report the write error; do not emit partial index.

Exit

• Success (verify): drift report emitted; exit 0 if clean, exit 1 if broken refs found.
• Success (sync): approved patches applied; doc-refs.json regenerated; session summary printed.
• Partial success: extractor or resolver errors are explicit in the report; no silent failures.

Logical Operations

Actions

Tools and instruments

• cli/commands/docs/extract.ts: remark + unified markdown AST, L2 pattern extraction, escape hatch filter, docs/generated/doc-refs.json writer.
• cli/commands/docs/resolve.ts: case-sensitive file existence, which for CLI tokens, package.json scripts lookup, ripgrep/git grep for env vars, oma-config.yaml deep-path check. URL kind is filtered out by the verify command and delegated to lychee. (Internal caching strategy: see design doc.)
• cli/commands/docs/reporter.ts: deterministic markdown + JSON renderer. No LLM call. Friendly summary, severity tagging, fix prioritization are the host LLM's responsibility.
• cli/commands/docs/sync-propose.ts: git diff intake, reverse index build, secret-pattern + gitignore file exclusion. Returns candidate docs with matched refs only. No LLM call. Patch synthesis is the host LLM's responsibility.
• cli/commands/docs/i18n-drift.ts: pairs web/docs English sources with web/i18n/{lang}/... translations, emits line/heading/recency drift signals with severity. No LLM call. Translation patching is oma-translator's responsibility.
• cli/commands/docs/lint-i18n.ts: content-level linter with cjk-em-dash for selected CJK locales and wrong-language placeholder detection across all locales. No LLM call, no auto-fix. Restructuring is the host LLM's responsibility via oma-translator.
• External: lychee (background URL link checking; install via brew install lychee).

Host-LLM contract

This skill follows the OMA pattern (mirroring oma-scholar): the CLI emits structured data; the host LLM (the agent runtime that invoked the skill) does any natural-language synthesis or judgment.

After oma docs verify --json:

1. Read the JSON drift report.
2. Group findings by severity / urgency (host-LLM judgment).
3. Suggest fixes per finding, prioritizing files most central to the project.
4. If the user asks for natural-language summary, host LLM produces it from the JSON, never from cached prose.

After oma docs sync <range> --json:

1. Read the candidate doc list (each entry: { doc, changedFiles, matchedRefs }).
2. For each candidate doc: read the doc itself, read git diff for changedFiles, draft a unified-diff patch reflecting the code change.
3. Present patches to the user for review. Never auto-apply.
4. On user approval, apply via git apply or by writing the doc directly.

After oma docs i18n --json:

1. Read the drift pairs (severity, line/heading diff, EN-newer flag per translation).
2. Prioritize CRITICAL/HIGH pairs; pass each to oma-translator in diff-sync mode (see that skill's § Diff-Sync Mode).
3. Never bulk-retranslate; patch only the drifted sections, with user confirmation.

After oma docs lint --json:

1. Read the style issues grouped by rule / language / file.
2. Restructure flagged sentences via oma-translator (e.g. § Stage 4-A em-dash rule); confirm edits with the user before writing.

Canonical command path

verify mode runs a drift check against the current codebase:

# Default: scan all repo markdown (**/*.md, gitignored files excluded),
# render markdown to stdout.
# URL link checking is delegated to lychee in the background
# (install: `brew install lychee`). Core check ~8s on a 1k-doc repo.
oma docs verify

# Narrow to a path or glob (uses minimatch)
oma docs verify "docs/**/*.md"
oma docs verify cli/README.md

# Machine-readable output for CI / hooks
oma docs verify --json

# Persist full markdown report to a file (works alongside --json too)
oma docs verify --report-file ./drift-report.md

# Skip URL checking entirely (when lychee is run separately, or as a
# one-off override of docs.check_urls=true in oma-config.yaml)
oma docs verify --no-urls

# Block until lychee finishes (CI scenarios needing complete URL data)
oma docs verify --urls-sync

# Exit code: 0 = clean, 1 = broken refs found in core check.
# URL drift, if any, is reported separately at docs/generated/url-drift.json
# and does NOT affect this exit code.

sync mode proposes patches for docs affected by a git diff (always interactive, never auto-applies):

# Default: staged changes (--cached), fallback HEAD~1..HEAD
oma docs sync

# Explicit range
oma docs sync HEAD~5..HEAD
oma docs sync main..feature-branch

# The CLI emits the candidate-doc list; the host LLM drafts patches and
# confirms per doc ([y] apply / [n] skip / [d] diff / [s] full proposal).
# Sync regenerates docs/generated/doc-refs.json after applying any patches.

i18n mode detects structural drift between English source docs (web/docs) and translations (web/i18n/{lang}/...); report-only, never edits translations:

# Default: severity ≥ MEDIUM, markdown summary to stdout
oma docs i18n

# Machine-readable, custom threshold
oma docs i18n --json --min-severity HIGH

# Output: per-pair drift signals (line/heading diff, EN-newer flag).
# Hand CRITICAL/HIGH pairs to `oma-translator` diff-sync mode.

lint mode checks translated docs for content-level style anti-patterns (report-only, no auto-fix):

# Default CJK locales: ko,ja,zh
oma docs lint
oma docs lint --json --locales ko,ja

Workflow hook (opt-in) runs verify automatically at workflow completion when docs.auto_verify: true in oma-config.yaml:

# Hook command emitted by /scm, /work, /ultrawork
oma docs verify --json
# Hook policy: warn-only in v1; non-zero exit does NOT block workflow completion

Resource scope

Preconditions

• Markdown files exist in the repo; docs/generated/ is created on demand for the index.
• cli/commands/docs/ is built and oma binary is on PATH (or invoked directly via bun run).
• For sync mode: a git diff is available (--cached stage or recent commits).
• For i18n / lint modes: web/docs (EN source) and web/i18n/{lang} trees exist.

Effects and side effects

• verify: regenerates docs/generated/doc-refs.json (always overwrites).
• sync: modifies docs files only on user approval; regenerates doc-refs.json after applies.
• i18n / lint: stdout report only; no file writes.
• All modes: stdout output (summary or full report).
• No .agents/ files are ever modified.

Guardrails

1. Never modify .agents/: CLAUDE.md SSOT protection applies in all modes.
2. Never auto-apply sync patches: sync is always interactive; [y] confirm required per doc.
3. LLM unavailable → graceful degradation: verify falls back to raw JSON; sync falls back to candidate-list-only (no proposals). Neither mode blocks on LLM availability.
4. Response language follows oma-config.yaml language: user-facing report text is localized; code, paths, JSON keys, and CLI commands stay in English.
5. Secret-bearing files excluded from sync output: .env*, *.pem, *.key, id_rsa*, and gitignored files never appear in candidate changedFiles lists. Host LLM never sees secret file paths.
6. URL link checking delegated to lychee: when docs.check_urls=true (default), URL refs are checked by lychee running in the background; results land in docs/generated/url-drift.json. If lychee is missing, an install hint is printed and URL checking is skipped (no internal HEAD fallback).
7. No direct LLM API calls from the CLI: the CLI never imports vendor SDKs, never reads API keys, never makes outbound LLM requests. All synthesis, patch drafting, and natural-language framing is the host LLM's responsibility (mirrors oma-scholar's pattern). This makes oma-docs vendor-agnostic: works identically under Claude Code / Codex / Gemini / Qwen / Antigravity.
8. Hook is warn-only in v1: broken refs never block workflow completion; docs.auto_verify: false by default (explicit opt-in required).
9. Escape hatch respected: <!-- oma-docs:ignore-start --> / <!-- oma-docs:ignore-end --> blocks and frontmatter oma-docs: skip are honored; no ref extraction from ignored regions. Use this for illustrative example paths in tutorials (hypothetical project files in inline code) that intentionally do not resolve.
10. Gitignored targets are generated, not broken: a file ref whose target does not exist but matches the project's gitignore rules (git check-ignore) is classified as skipped (a documented runtime/generated output such as .agents/results/result-*.md, .agents/state/memories/*, .serena/memories/*, .agents/state/*.json), never as broken. gitignore is the single source of truth for "produced at runtime" — gitignore an output path and it stops being flagged. The skipped count is surfaced (markdown summary + JSON skippedCount) so nothing is silently dropped.
11. Non-prose trees excluded via docs.exclude: committed-but-non-prose markdown (benchmark run artifacts, translation mirrors validated separately by oma docs i18n) is dropped from the scan by the docs.exclude globs rather than producing unactionable broken refs. An explicit single-file path argument bypasses docs.exclude.
12. i18n / lint modes never write: both emit reports only. Translation patches go through oma-translator with per-file user confirmation; the CLI never edits translations.

v1 scope note

v1 covers verify, sync (broken-only classification, L2 ref extraction), i18n (structural translation drift), and lint (CJK style anti-patterns). The following are explicitly deferred to v2: create mode (generate missing docs), semantic (content-level) translation drift, L3 symbol-level extraction (Tree-sitter/LSP), GitHub Action wrapper, block hook mode.

References

• Design doc: docs/plans/designs/008-oma-docs.md (full architecture, schema spec, decision log, edge cases).
• Schema spec: doc-refs.json v1 schema defined in design doc § doc-refs.json Schema.
• Workflow hook integration: design doc § Workflow Hook Integration.
• Migration: deepinit Step 6 retirement, design doc § Migration: deepinit Step 6.
• Adjacent skills: oma-translator (v2 multilingual), oma-skill-creator (SSL-lite validation).