Deep Audit

Deep Repository Audit

When to Use

• After broad changes to .claude/, CLAUDE.md, or either package's public surface
• Before a CRAN or PyPI release
• When the user asks to "find inconsistencies", "audit", or "check everything"

Workflow

PHASE 0: Mechanical checks (run FIRST — cheap and deterministic)

There is no check script in this repo, so run these directly. Each is a class of bug that agent prompts historically miss because attention drifts across a long checklist.

# 1. Dead rule globs — a rule whose paths: match nothing never fires. This is THE bug
#    that made the inherited config inert; check it every time.
for f in .claude/rules/*.md; do
  echo "== $f"
  awk '/^paths:/{p=1;next} /^[a-z]/{p=0} p&&/^ *- /{gsub(/^ *- "|"$/,"");print}' "$f" |
    while read -r g; do
      n=$(ls $g 2>/dev/null | wc -l)
      [ "$n" -eq 0 ] && echo "   DEAD GLOB: $g"
    done
done

# 2. Referenced-but-missing files across all config
grep -rhoE '\]\(\.\./[^)]+\)|\]\([a-z0-9_./-]+\.md\)' .claude CLAUDE.md | sort -u

# 3. Commands named in config that do not exist on this machine
grep -rhoE '(python3?|uv|gh|quarto|Rscript) [a-z-]+' .claude CLAUDE.md | sort -u

# 4. Skills/agents/rules on disk vs. the inventory in CLAUDE.md
ls .claude/skills/ ; ls .claude/agents/ ; ls .claude/rules/
grep -n "Rules\|Skills\|Agents" CLAUDE.md

Fix P0/P1 findings here before launching agents — the mechanical layer is cheaper and more precise than a prompt for these classes.

PHASE 1: Launch parallel audit agents

Launch simultaneously via Task with subagent_type=general-purpose. Every agent prompt must include the claim-vs-reality lens: does every documented command, path, count, and cross-reference actually resolve in this repo?

Agent 1: Configuration integrity

Focus: .claude/rules/*.md, .claude/skills/*/SKILL.md, .claude/agents/*.md

• Valid YAML frontmatter everywhere
• allowed-tools covers every tool the body invokes. If the body says "spawn X via Task", Task must be in allowed-tools — otherwise a runtime permission error or a silent bypass.
• Rule paths: globs match real files (Phase 0 catches this; confirm the fix)
• Rule ↔ skill parity: if a rule names a skill, that skill actually implements the protocol
• No contradictions between rules
• No references to skills, agents, hooks, or scripts that do not exist here

Agent 2: Documentation accuracy

Focus: CLAUDE.md, MEMORY.md, README.md, r-package/README.md, python-package/README.md

• Every file path mentioned exists on disk
• Every command is runnable, or is explicitly marked as needing an absent toolchain
• The surface inventory in CLAUDE.md matches ls .claude/{skills,agents,rules} exactly
• Architecture claims match the code (line numbers in CLAUDE.md and MEMORY.md still point at what they say they do — these rot whenever utils.R / utils.py is edited)
• No stale counts, no placeholder text ([YOUR ...]), no lecture/Beamer/Quarto leftovers

Agent 3: R ↔ Python drift

Focus: r-package/R/, python-package/geobr/

• The parity contract in cross-language-parity.md
• Release tag consistent across its five hardcoded copies
• A public reader present on one side but not the other, with no changelog record
• Do not report native-idiom differences as drift (sf vs GeoDataFrame, arrow/dplyr vs duckdb/pandas). See the false-alarm list below.
• Prefer running /parity-check and auditing its report over re-deriving the comparison

Agent 4: Package hygiene

Focus: r-package/DESCRIPTION, r-package/NAMESPACE, python-package/pyproject.toml, python-package/uv.lock, .github/workflows/

• DESCRIPTION Imports/Suggests match what R/ actually uses
• pyproject.toml dependencies match what geobr/ actually imports
• pyproject.toml and uv.lock are coherent
• CI workflow paths still match the repo layout
• Metadata agrees across surfaces (URLs, versions, license, maintainer)

PHASE 2: Triage

Categorize each finding as genuine bug (fix) or false alarm (discard, and record why so the next round does not re-raise it).

Known false alarms — do not "fix" these:

PHASE 3: Fix

Apply fixes in parallel where safe. For each: read the file, apply, verify (grep for the stale value). Configuration and documentation only — a finding in r-package/R/ or python-package/geobr/ is reported, not silently patched. Package source changes need the maintainer, because they change what users get.

PHASE 4: Loop-until-dry

Relaunch a fresh set of agents to verify. This is the loop-until-dry primitive (orchestrator-protocol.md):

• Converge when a round surfaces 0 new genuine issues (deduped on file+issue).
• Otherwise fix and loop.
• Fallback cap: 5 rounds. A finding surviving rounds N and N+2 is escalated to the user, not patched a third time (summary-parity.md).

Output Format

## Round N Audit Results

### Issues: X genuine, Y false alarms

| # | Severity | File | Issue | Status |
|---|----------|------|-------|--------|
| 1 | Critical | .claude/rules/foo.md | Dead glob: matches 0 files | Fixed |
| 2 | Medium | CLAUDE.md:47 | utils.R:266 line ref is stale | Fixed |

### Reported, not fixed (package source — needs maintainer)
| File | Issue |
|---|---|

### Verification
- [ ] Every rule glob matches >= 1 file
- [ ] Every documented command exists or is marked as unavailable
- [ ] CLAUDE.md inventory matches disk
- [ ] No references to absent skills/agents/scripts

### Result: [CLEAN | N issues remaining]