Documentation Completeness Audit

Determine whether a documentation set covers everything it should by building an inventory of what needs documenting and comparing it to what exists. The output is a prioritized gap report — not new documentation.

When to Use

• After shipping a feature — verify docs cover the new surface area
• Before a release — ensure no undocumented public APIs, CLI flags, or config options
• When users or new hires report "I couldn't find docs for X"
• Periodic health check on doc coverage
• After running doc-maintenance (structural) and doc-claim-validator (accuracy) to go wider

Quick Reference

Workflow Overview

Phase 1: Inventory   → Build the "should exist" list from code and config
Phase 2: Map         → Match inventory items to existing documentation
Phase 3: Classify    → Score each gap by audience impact
Phase 4: Report      → Produce the prioritized gap report

Phase 1: Build the Inventory

Construct a list of everything that should be documented. Use four sources, checking all of them:

Source 1: Public Code Surface

Run the bundled inventory script to extract documentable surface area deterministically:

python3 skills/doc-completeness-audit/scripts/inventory.py --root . --json > inventory.json

# Or human-readable:
python3 skills/doc-completeness-audit/scripts/inventory.py --root .

# Run specific detectors only:
python3 skills/doc-completeness-audit/scripts/inventory.py --root . --detectors env_vars,cli_commands

The script scans source files across Python, JavaScript/TypeScript, Rust, Go, Ruby, Java, and shell, extracting six categories:

Dispatch an Explore agent to scan for these signals. Provide it with the project's primary language and entry points.

Source 2: User-Facing Features

Identify features a user interacts with:

• TUI screens, views, keybindings
• CLI workflows (multi-step operations)
• Integration points (hooks, plugins, extensions)
• Authentication/authorization flows
• Error messages that imply user action

Source 3: Operational Surface

Identify what operators and maintainers need:

• Installation and setup procedures
• Upgrade and migration paths
• Backup and restore procedures
• Troubleshooting common errors
• Environment requirements and dependencies
• CI/CD integration points

Source 4: Existing Docs Cross-References

Check existing docs for promises of documentation that doesn't exist:

• "See [link]" references to pages that don't exist
• "Coming soon" or "TODO" markers
• Table of contents entries without corresponding pages
• Navigation entries without targets

Source 5: Architectural / Operational / Migration Topic Discovery (sonnet)

The first four sources catch code-detectable surface (env vars, CLI flags, endpoints, exported APIs, broken cross-references). They miss topics that exist as architectural patterns, user flows, ops procedures, or migration paths but don't surface as a single greppable symbol. Examples:

• Architectural patterns the system implements (CQRS, event sourcing, saga) — should be documented but won't show up in inventory.py
• User flows implicit across UI surfaces — "how to share a project" may span multiple components and isn't a single CLI command
• Migration paths between versions — typically tribal knowledge until someone needs them
• Operational runbooks (incidents, rollbacks, capacity events)
• Recovery procedures and disaster scenarios

Dispatch one general-purpose + sonnet agent for topic discovery:

subagent_type: "general-purpose"
model: "sonnet"
description: "Architectural/operational topic discovery"

Prompt: read README, top-level docs, and a sample of code (architecture files, integration boundaries, deployment configs, major feature directories). Identify topics that should be documented but aren't captured by the code-surface inventory. For each topic, name:

• topic — what needs documenting (one phrase)
• evidence — what in the codebase implies this topic exists (path:line citations)
• audience — who would read this (operators, contributors, advanced users)
• type — reference, tutorial, guide, explanation, runbook
• confidence — high (clear evidence), medium (inferred), low (speculative)

Append the agent's output to the inventory list before Phase 2.

Output: A structured inventory list. Each item has:

• topic — what needs documenting
• source — where the requirement was discovered (code path, config key, user flow, sonnet inference)
• audience — who needs this (end user, developer, operator)
• type — what kind of doc it needs (reference, tutorial, guide, explanation, runbook)
• confidence — high (deterministic) | medium | low (sonnet-inferred speculative)

Phase 2: Map to Existing Documentation (per-docfile sonnet dispatch)

For each inventory item, determine whether it's documented and how well. "Adequate coverage" requires reading surrounding context — a grep hit doesn't tell you whether the topic is truly explained vs. just mentioned in passing. Orchestrator-side execution would require reading every doc N times (once per inventory item), which strains the context window.

Dispatch strategy

Two-phase mapping:

1. Bulk grep pass (orchestrator) — for each inventory item, grep docs for the topic name. Build a candidate match map: which docs mention each topic.
2. Per-docfile sonnet pass — for each docfile that surfaced as a candidate match for any inventory item, dispatch one general-purpose + sonnet agent. The agent receives the doc + the list of inventory items that grep'd to this doc, and judges each as Documented / Shallow / Misplaced.

This keeps total agent calls ≈ N candidate docfiles (not N inventory items × M docs). For a typical project with 100 inventory items and 50 docs, the candidate map usually has 30–50 docs needing review.

Per-docfile prompt template

subagent_type: "general-purpose"
model: "sonnet"
description: "Coverage mapping for <docfile>"

Prompt:

Read the doc at <DOCFILE_PATH>. The following inventory items grep-matched
this doc — judge each:

<INVENTORY_ITEMS_FOR_THIS_DOC>

For each item, classify as one of:
- Documented: dedicated section or page provides adequate coverage
- Shallow: mentioned but insufficient (missing examples, edge cases,
  parameter listings; flag-in-table without explanation)
- Misplaced: covered, but in the wrong doc type for the audience (API
  reference embedded in a tutorial; user-facing topic in dev-only docs)
- No real match: grep matched but the doc doesn't actually cover the topic
  (incidental mention, different concept with the same word)

Output as YAML:

doc_path: <path>
items_reviewed: N
classifications:
  - item: <topic>
    classification: Documented | Shallow | Misplaced | No real match
    section: <heading or line range where the topic is covered>
    evidence: <quote or paraphrase of the relevant content>
    gap: <if Shallow, what's missing; if Misplaced, where it should live>

Items with no candidate match

Inventory items that grep'd 0 docs go directly to the "Missing" bucket without a