carla-simulator/carla/audit-carla-docs

Audit CARLA documentation (static)

Reports what the repository documents badly or not at all, as a markdown list a human triages. It only reads files — so it is fast, repeatable and safe, and it is blind to anything that exists only at runtime. The companion runtime audit covers that; this skill will not let you forget it.

Paths. scripts/… and references/… are relative to the directory holding

this SKILL.md, not to the user's working directory. Prefix them with its

absolute path.

The deliverable is a report, never an edit. Deciding what to document, and writing it, are separate jobs — see document-carla-feature.

Instructions

Progress:
- [ ] Step 1: Check prerequisites (bash scripts/check_env.sh), clear FAILs
- [ ] Step 2: Run the audit, writing the report to a file
- [ ] Step 3: Read the summary table and triage by signal, not top to bottom
- [ ] Step 4: Tell the user a runtime (Tier B) audit is still owed

Step 1: Check prerequisites

bash scripts/check_env.sh

Only two things are hard blockers: no CARLA checkout, and no python3. A missing input file downgrades its signal to skipped rather than failing the run.

Step 2: Run the audit

python3 scripts/audit_docs.py --root "${CARLA_ROOT:?}" --out

Writes Docs/audit_carla_docs.md. Drop --out entirely to print to stdout instead; pass --out <path> to write somewhere else.

--signal <name> narrows the run, repeatable: changelog, binding, stale, catalogue-stale, sensor, catalogue, link, orphan. Narrow when iterating on one area; run them all before a release.

Step 3: Triage

The report opens with a count per signal. Work by signal, not in file order — the five differ enormously in effort per finding, and mixing them wastes time:

Two cautions the report cannot resolve for you, both in references/signals.md:

• The changelog section **makes no claim about whether anything is

documented.** It lists every unreleased feature as an unticked checkbox; you decide. Judging coverage automatically was tried and was too generous to trust — see the reference.

• `Docs/python_api.md` is generated from PythonAPI/docs/*.yml by

PythonAPI/docs/doc_gen.py. Fix a binding finding in the .yml and regenerate. Editing the .md is silently reverted by the next run.

Step 4: Say what this audit could not see

Always close by telling the user a runtime audit is still owed, and why. The script prints this too, but say it in your own summary — it is the single easiest thing to forget, and the gap it leaves is invisible.

A static pass cannot see the blueprint library, the available maps, or per-blueprint attributes: new vehicles, walkers and props are content, so they appear in no source diff.

Examples

Example 1: preparing a branch for release

User says: "audit the docs before we make ue58-dev the default"

bash scripts/check_env.sh
python3 scripts/audit_docs.py --root "${CARLA_ROOT:?}" --out

Result: Docs/audit_carla_docs.md with a count per signal. Report the totals, name the two or three heaviest items, and state that a runtime audit is still owed.

Example 2: checking one area while fixing it

User says: "are there still broken links in the docs?"

python3 scripts/audit_docs.py --root "${CARLA_ROOT:?}" --signal link

Result: the link findings on stdout. No file is written; nothing is changed.

Troubleshooting

Error: `not a carla checkout (no CHANGELOG.md): <path>` Cause: --root or CARLA_ROOT points somewhere else. Solution: pass --root /path/to/carla, or export CARLA_ROOT= before sourcing.

A signal reports 0 findings and you expected some Cause: its input is missing, so it was skipped — check_env.sh prints a WARN per missing input. Solution: re-read the check_env.sh output; a skipped signal is not a clean one.

`link` findings inside fenced code or HTML comments Cause: none — the audit strips both before scanning. A link reported as broken is in live prose.

Outputs

• Docs/audit_carla_docs.md (or stdout) — findings grouped by signal, with a

count table and the runtime-audit warning. The audit excludes this file from its own link and orphan signals.