Audit Suite Orchestrator

Read _audit-common.md and _audit-severity.md for shared protocol.

This skill owns no audit logic: it selects audits and fans them out as background agents, then merges their reports. Audits are .claude/commands/audit-<name>/SKILL.md, invoked /audit-<name>. /audit-publish (post-processing) and /audit-sync (skill maintenance) are not audits and never appear in a suite. Who owns which path: .claude/commands/_audit-owners.md.

Modes

1. --preset <name> — a curated list (below).
2. --area <path…> — derive the set: git ls-files <path…> | .claude/commands/_audit-route.sh → owners of those files, plus the neighbors of each owner from the table below. No hand-maintained preset to drift. A per-game owner expands to all six per-game audits, so a byroredux/src/cell_loader/ area fans out ~17 audits (2026-10-09 streaming-deep). Say so before launching.
3. --changed <scope> — same, from git diff <scope> --name-only (scope forms as in /audit-incremental). Fans out whole owner audits; use /audit-incremental when you only want hunk-level checks.

Each audit delta-scopes itself (_audit-common.md § Delta-first scoping), so a suite over a code area costs roughly what changed there. Pass --focus <dimensions> to one audit only when the user names it; never bake dimension numbers into a preset — they drift on every renumbering.

Curated presets

Preset	When	Audits
quick	after any change, <10 min	incremental --commits 5
pre-release	before tagging	safety · renderer · ecs · regression --limit 20
comprehensive	monthly / pre-milestone	every audit below, then runtime --game all
tech-debt-deep	after a milestone closes	tech-debt · incremental --commits 30
per-game-all	compat sweep, reference title first	fnv · fo3 · skyrim · oblivion · fo4 · starfield
nif-all-games	NIF parser vs every corpus	nif --game each of fnv fo3 skyrim oblivion fo4 starfield
runtime-regression	telemetry vs baselines	runtime --game all

comprehensive covers every owner in _audit-owners.md: renderer · performance · ecs · concurrency · safety · nif · nifal · esm · parsers · physics · character · gameplay · exterior · ui · audio · speedtree · scripting · papyrus · save · tooling · legacy-compat · tech-debt · the six per-game audits · regression. If an owner row names an audit missing from this list, add it here — _audit-validate.sh fails on owners that have no skill, and reading the map is the coverage check.

Neighbors (added by --area / --changed)

Owner	Always add
renderer	performance · concurrency · safety
nif	nifal · safety
nifal	nif · renderer
esm	the per-game audit, only when a title is the target
physics	concurrency · safety
character	esm · ecs
gameplay	ecs · save · physics
exterior	renderer · physics (water half) · esm (WATR/WTHR/GRAS decode)
parsers	safety · the format's per-game audit
audio	concurrency · safety
ui	safety · concurrency · tech-debt
scripting	ecs · papyrus (when .pex/.psc inputs are involved)
papyrus	scripting · safety (decompiler recursion/panic bounds)
save	ecs · gameplay
tooling	concurrency (debug-server threads) · safety

Aliases for the old *-deep presets

Each is just --area over these paths (neighbors apply):

Alias	--area paths
renderer-deep	crates/renderer/ byroredux/src/render/
rt-deep	crates/renderer/src/vulkan/acceleration/ crates/renderer/src/vulkan/{svgf,gbuffer,composite}.rs
material-deep	crates/renderer/src/vulkan/material.rs crates/renderer/src/vulkan/scene_buffer/ byroredux/src/material_translate.rs
texture-roles-deep	crates/nif/src/import/material/ byroredux/src/asset_provider/material/ crates/bgsm/ crates/sfmaterial/
upscaler-deep	crates/fsr3-sys/ crates/renderer/src/vulkan/{frame_upscaler,upscaling,presentation,exposure,exposure_meter}.rs crates/renderer/src/tonemap.rs (run with BYRO_VALIDATION=1)
water-deep	crates/renderer/src/vulkan/water.rs crates/physics/ byroredux/src/cell_loader/water.rs byroredux/src/systems/water.rs
volumetrics-deep / bloom-deep / skin-deep	crates/renderer/src/vulkan/volumetrics.rs / bloom.rs / skin_compute.rs + acceleration/blas_skinned.rs
nif-deep / nifal-deep	crates/nif/ / byroredux/src/material_translate.rs crates/nif/src/import/
esm-deep / character-deep / physics-deep	crates/plugin/ / crates/core/src/character/ / crates/physics/ byroredux/src/ragdoll.rs
ui-deep / audio-deep / save-deep	crates/ui/ crates/menuxml/ byroredux/src/{hud,scaleform_hud,ui_input,objectives}.rs / crates/audio/ / crates/save/ byroredux/src/save_io/
scripting-deep / speedtree-deep	crates/{scripting,pex,papyrus}/ / crates/spt/
streaming-deep	byroredux/src/streaming/ byroredux/src/npc_spawn/ byroredux/src/cell_loader/
exterior-deep / gameplay-deep / parsers-deep / tooling-deep	the new owners' Paths: lines
legacy-deep	legacy-compat alone

Execution

1. Parse the mode. Unknown preset → list the tables above and stop. mkdir -p /tmp/audit.
2. Launch each audit as a background agent, max 3 concurrent — they read the tree and write distinct reports, so there is no ordering dependency.
3. Orchestration hazard: an agent that fans out its own sub-agents cannot receive their completion notices and will drop real findings from its report (seen on the 2026-08-03 comprehensive run: three real findings — incl. a HIGH ragdoll double-compose — were reported "clean"). Tell every launched audit up front to analyse dimensions synchronously or to write /tmp/audit/<name>/dim_N.md scratch files. Before accepting "no findings" for a dimension, check that scratch file against the written report. Also tell each audit to skip its final rm -rf /tmp/audit/<name>; the suite deletes those directories itself once the reconciliation is done. Put these rules in one shared file (/tmp/audit/SUITE_RULES.md, as on the 2026-10-08/09 runs) together with one open-issue snapshot, and forward each finished audit's IDs to the agents still running (SendMessage) so they dedup mid-run. Cargo targets: have agents build in the repo target/ and name any private target dir in their final message. Per-agent targets reached ~16 GB on the 2026-10-09 run.
4. Each audit writes docs/audits/AUDIT_<TYPE>_<TODAY>.md (_audit-common.md § Report Finalization). Verify each exists with real content, not a placeholder.
5. Merge into a summary; warn prominently if any CRITICAL:

# Audit Suite Summary — <mode + args> — <date>

| Audit | Findings | CRITICAL | HIGH | MEDIUM | LOW | Report |
|---|---|---|---|---|---|---|

Total: X findings (C critical, H high, M medium, L low). Unchanged-since-baseline dimensions skimmed: <list>

6. For each report with findings suggest /audit-publish docs/audits/AUDIT_<TYPE>_<TODAY>.md.