ln-614-docs-fact-checker（ln 614 文档事实检查器）

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root. If shared/ is missing, fetch files via WebFetch from https://raw.githubusercontent.com/levnikolaevich/claude-code-skills/master/skills/{path}.

Documentation Fact-Checker (L3 Worker)

Type: L3 Worker

Specialized worker that extracts verifiable claims from documentation and validates each against the actual codebase.

Purpose & Scope

• Prioritize canonical and high-claim docs, then extract verifiable claims from markdown documentation
• Verify each claim against codebase (Grep/Glob/Read/Bash)
• Detect cross-document contradictions (same fact stated differently)
• Includes docs/reference/, docs/tasks/, tests/ in scope
• Single invocation (not per-document) -> cross-doc checks require global view
• Does NOT check scope alignment or structural quality

Inputs

MANDATORY READ: Load shared/references/audit_worker_core_contract.md, shared/references/docs_quality_contract.md, shared/references/docs_quality_rules.json, shared/references/markdown_read_protocol.md, shared/references/mcp_tool_preferences.md, and shared/references/mcp_integration_patterns.md.

Receives contextStore with: tech_stack, project_root, output_dir.

Workflow

Phase 1: Parse Context

Extract tech stack, project root, output_dir from contextStore.

Phase 2: Discover Documents

Glob markdown docs in project. Exclude:

• node_modules/, .git/, dist/, build/
• docs/project/.audit/ (audit output, not project docs)
• CHANGELOG.md (historical by design)

If docs/project/.context/doc_registry.json exists:

• load it first
• prioritize doc_role=canonical
• prioritize files with dense claim types (paths, endpoints, versions, commands)
• de-prioritize navigation hubs unless contradictions point back to them

Phase 3: Extract Claims (Layer 1)

MANDATORY READ: Load shared/references/two_layer_detection.md for detection methodology.

For each prioritized document, use section-first reads to extract verifiable claims using Grep/regex patterns.

For code files referenced by docs, use outline() and discovery-first read_file() before built-in reads. Only request edit_ready=true, verbosity="full" if verification turns into a follow-up edit. Use hex-graph only when entity identity or reference resolution remains ambiguous after direct manifest/file checks.

MANDATORY READ: Load references/claim_extraction_rules.md for detailed extraction patterns per claim type.

9 claim types:

Output per claim: {doc_path, line, claim_type, claim_value, raw_context}.

Phase 4: Verify Claims (Layer 2)

For each extracted claim, verify against codebase:

False positive filtering (Layer 2 reasoning):

• Template placeholders ({placeholder}, YOUR_*, <project>, xxx) -> skip
• Example/hypothetical paths (preceded by "e.g.", "for example", "such as") -> skip
• Future-tense claims ("will add", "planned", "TODO") -> skip or LOW
• Conditional claims ("if using X, configure Y") -> verify only if X detected in tech_stack
• External service paths (URLs, external repos) -> skip
• Paths in SCOPE/comment HTML blocks describing other projects -> skip
• .env.example values -> skip (expected to differ from actual)

Phase 5: Cross-Document Consistency

Compare extracted claims across documents to find contradictions:

Algorithm:

claim_index = {}  # key: normalized(claim_type + entity), value: [{doc, line, value}]
FOR claim IN all_verified_claims WHERE claim.verified == true:
  key = normalize(claim.claim_type, claim.entity_name)
  claim_index[key].append({doc: claim.doc_path, line: claim.line, value: claim.claim_value})

FOR key, entries IN claim_index:
  unique_values = set(entry.value for entry in entries)
  IF len(unique_values) > 1:
    CREATE finding(type=CROSS_DOC_*_CONFLICT, severity=HIGH,
      location=entries[0].doc + ":" + entries[0].line,
      issue="'" + key + "' stated as '" + val1 + "' in " + doc1 + " but '" + val2 + "' in " + doc2)

Phase 6: Score & Report

MANDATORY READ: Load shared/references/audit_worker_core_contract.md and shared/references/audit_scoring.md.

Calculate score using penalty formula. Write report.

Audit Categories (for Checks table)

Severity Mapping

Output Format

MANDATORY READ: Load shared/references/audit_worker_core_contract.md and shared/templates/audit_worker_report_template.md.

Write JSON summary per shared/references/audit_summary_contract.md. In managed mode the caller passes both runId and summaryArtifactPath; in standalone mode the worker generates its own run-scoped artifact path per shared contract.

Write report to {output_dir}/ln-614--global.md with category: "Fact Accuracy" and checks: path_claims, version_claims, count_claims, endpoint_claims, config_claims, command_claims, entity_claims, line_ref_claims, cross_doc.

Return summary per shared/references/audit_summary_contract.md.

When summaryArtifactPath is absent, write the standalone runtime summary under .hex-skills/runtime-artifacts/runs/{run_id}/evaluation-worker/{worker}--{identifier}.json and optionally echo the same summary in structured output.

Report written: .hex-skills/runtime-artifacts/runs/{run_id}/audit-report/ln-614--global.md
Score: X.X/10 | Issues: N (C:N H:N M:N L:N)

Critical Rules

MANDATORY READ: Load shared/references/audit_worker_core_contract.md.

• Do not auto-fix: Report violations only; coordinator aggregates for user
• Code is truth: When docs contradict code, document is wrong (unless code is a bug)
• Evidence required: Every finding includes verification command used and result
• No false positives: Better to miss an issue than report incorrectly. When uncertain, classify as LOW with note
• Location precision: Always include file:line for programmatic navigation
• Broad scope: Scan ALL.md files -- do not skip docs/reference/, tests/, or task docs
• Targeted depth: Spend the deepest verification effort on canonical and high-claim docs first
• Cross-doc matters: Contradictions between documents erode trust more than single-doc errors
• Batch efficiently: Extract all claims first, then verify in batches by type (all paths together, all versions together)
• Shared placeholder policy: Respect allowlisted setup placeholders from docs_quality_rules.json; do not escalate them in task setup docs
• Use hex-graph only for semantic ambiguity: For code entities and references, prefer graph queries over repeated grep only when direct manifest/file checks leave symbol identity or reference resolution ambiguous

Definition of Done

MANDATORY READ: Load shared/references/audit_worker_core_contract.md.

• contextStore parsed successfully (including output_dir)
• All .md files discovered (broad scope)
• Claims extracted across 9 types
• Each claim verified against codebase with evidence
• Cross-document consistency checked
• False positives filtered via Layer 2 reasoning
• Score calculated using penalty algorithm
• Report written to {output_dir}/ln-614--global.md (atomic single Write call)
• Summary written per contract

Reference Files

• Audit output schema: shared/references/audit_output_schema.md
• Detection methodology: shared/references/two_layer_detection.md
• Claim extraction rules: references/claim_extraction_rules.md

Version: 1.0.0 Last Updated: 2026-03-06