Automated summary
Audits only recently changed code from a Git diff, routing each changed file to the project's specialised audit skills in order to catch new bugs and regressions.
Strengths
- Tight, cheap scope: the diff is chosen by argument (working tree, last N commits, revision range, since-date), so untouched code is not re-audited.
- A file-path → owning-audit routing table with a risk floor per area (renderer, ECS, NIF, BSA, and so on).
- Reuses the project's shared conventions (finding format, severity scale) instead of redefining its own checklist.
- Reads the commit log alongside the diff to surface themes, PR numbers and milestone tags.
Limitations
- Tightly bound to one repository layout (renderer/nif/bsa crates, byroredux): the routing table does not transfer to other projects as-is.
- Assumes the referenced audit skills plus `_audit-common.md` and `_audit-severity.md` already exist in the repo.
- By looking only at the diff, it can miss regressions that arise from interaction with unchanged code.
Use it right after a batch of commits or at the end of a branch to quickly confirm the recent changes introduce no regressions.
Do not use it for a full audit of a subsystem or an existing codebase, or in a project lacking the specialised audit skills it dispatches to.
Security analysis
SafeThe skill is a read-only auditing workflow: it instructs running git commands (diff, log, grep) to scope changed files and then writing a single Markdown report under docs/audits/. There are no destructive operations (rm, force-push), no network calls, no piping to shells (curl|sh), no base64/obfuscated payloads, and no attempt to disable safety controls or exfiltrate secrets. The only state change is creating one report file, and shell use is limited to benign git inspection, so execution risk is minimal.
No concerns found
How this analysis is madeExamples
/delta-audit --working/delta-audit --commits 3/delta-audit --range v0.4.0..HEADdescription: "Delta audit — check only recently changed code for regressions and new bugs" argument-hint: "[--working] [--commits <N>] [--range <A>..<B>] [--since <date>]"
Incremental / Delta Audit
Audit only what changed, not a whole subsystem. The goal is to catch new bugs and regressions introduced by recent work — fast — by routing each changed file to the subsystem/per-game audit dimensions that own it and applying their checks to the diff alone.
This is a meta-audit: it does not define dimensions, it dispatches to the
real audit skills (each lives at .claude/commands/audit-<NAME>/SKILL.md).
Use those for the authoritative checklist of any one area.
See .claude/commands/_audit-common.md for project layout, severity,
methodology, deduplication, context rules, and the base finding format.
See .claude/commands/_audit-severity.md for the severity scale + special
rules (the NIFAL / GPU-struct / AS / SSBO rows are the ones a delta most
often trips).
Step 1: Determine the diff scope
Pick the narrowest scope that covers the work under review. Default to the working tree if nothing is specified.
| Argument | Scope | Diff command |
|----------|-------|--------------|
| (none) / --working | uncommitted work | git diff HEAD --name-only (add --staged-less; for staged-only use git diff --staged --name-only) |
| --commits <N> | last N commits | git diff "HEAD~<N>..HEAD" --name-only |
| --range <A>..<B> | explicit revision range | git diff "<A>..<B>" --name-only |
| --since <date> | everything since a date | base=$(git log --since="<date>" --format="%H" \| tail -1); then git diff "${base}^..HEAD" --name-only |
Then pull the actual hunks for the same scope (swap --name-only for
nothing, or -U6 for more context) and the commit log:
git diff HEAD~10..HEAD --stat # changed-file overview (substitute your scope)
git diff HEAD~10..HEAD # the hunks you will actually audit
git log --oneline HEAD~10..HEAD # commit themes, PR numbers, milestone tags
Audit the diff, with just enough surrounding context to confirm each
finding (_audit-common.md § Methodology) — do not re-audit untouched code.
Step 2: Route each changed file to its audit dimension
Map every changed path to the audit skill(s) that own it, then apply that
skill's checks to the diff. A file can hit multiple rows (e.g. a shader +
its #[repr(C)] host struct → renderer and the GPU-struct-sync rule).
Risk is the floor severity for an un-disproven finding in that area.
| Changed path | Owning audit(s) | Risk |
|--------------|-----------------|------|
| crates/renderer/src/vulkan/** (pipeline, sync, descriptors, context/) | /audit-renderer, /audit-safety, /audit-concurrency | HIGH |
| crates/renderer/src/vulkan/acceleration/**, svgf.rs, gbuffer.rs, composite.rs (RT / denoise / G-buffer) | /audit-renderer | HIGH |
| crates/renderer/src/vulkan/scene_buffer/**, material.rs (#[repr(C)] GPU structs) | /audit-renderer, /audit-nifal | HIGH |
| crates/renderer/src/vulkan/volumetrics.rs + shaders/volumetrics_*.comp (M55) | /audit-renderer | HIGH |
| crates/renderer/src/vulkan/bloom.rs + shaders/bloom_*.comp (M58) | /audit-renderer | HIGH |
| crates/renderer/src/vulkan/water.rs, shaders/water.vert/water.frag, byroredux/src/systems/water.rs, byroredux/src/cell_loader/water.rs (M38) | /audit-renderer, /audit-fnv | HIGH |
| crates/renderer/shaders/** (any .comp/.vert/.frag) | /audit-renderer (+ GPU-struct-sync rule) | HIGH |
| crates/core/src/ecs/** | /audit-ecs, /audit-concurrency | HIGH |
| crates/nif/src/blocks/**, crates/nif/src/import/**, crates/nif/src/anim/** | /audit-nif; per-game /audit-<game> | HIGH |
| crates/bsa/src/** (BSA / BA2 / CSG) | /audit-nif (archive feed), per-game /audit-<game> | HIGH |
| byroredux/src/material_translate.rs, crates/core/src/ecs/components/material.rs, crates/nif/src/import/collision/mod.rs (NIFAL boundary) | /audit-nifal | HIGH |
| byroredux/src/env_translate.rs (EXAL boundary) | /audit-nifal (mirror), /audit-renderer | MEDIUM |
| byroredux/src/ragdoll.rs, crates/physics/src/** (PHYSAL / Rapier bridge) | /audit-safety, per-game /audit-<game> | MEDIUM |
| crates/spt/src/**, byroredux/src/cell_loader/refr.rs (.spt route) | /audit-speedtree | MEDIUM |
| crates/plugin/src/esm/** (incl. records/misc/{water,character,world,ai,magic,effects,equipment}.rs) | per-game /audit-<game>, /audit-legacy-compat | MEDIUM |
| crates/core/src/animation/** | /audit-nif (anim import), /audit-ecs | MEDIUM |
| byroredux/src/cell_loader/** | per-game /audit-<game> | MEDIUM |
| byroredux/src/systems/**, byroredux/src/render/** | /audit-ecs, /audit-renderer, /audit-performance | MEDIUM |
| byroredux/src/scene/** | per-game /audit-<game> | MEDIUM |
| byroredux/src/main.rs, byroredux/src/commands/** | /audit-ecs | MEDIUM |
| crates/scripting/**, crates/pex/**, crates/papyrus/** | /audit-scripting | MEDIUM |
| crates/save/** | /audit-save | MEDIUM |
| byroredux/src/asset_provider/archive.rs (sibling-BSA auto-load, AE path strip) | per-game /audit-<game> | MEDIUM |
| crates/audio/src/{lib,tests}.rs | /audit-audio | MEDIUM |
| crates/sfmaterial/src/** (Starfield CDB) | /audit-starfield | MEDIUM |
| crates/bgsm/src/** (FO4+ BGSM/BGEM) | /audit-fo4, /audit-nifal | MEDIUM |
| crates/facegen/src/** | per-game /audit-<game> | MEDIUM |
| crates/debug-ui/src/**, crates/renderer/src/vulkan/egui_pass.rs (egui overlay → touches draw_frame) | /audit-renderer, /audit-concurrency | MEDIUM |
| **/tests/**, **/*_tests.rs, byroredux/tests/golden_frames.rs | /audit-regression | LOW |
| *.md, docs/** | /audit-tech-debt (doc rot) | LOW |
Layout shifts that often surprise a delta audit:
render.rs,systems.rs,scene.rs, andcell_loader.rsare all directories now (byroredux/src/render/,systems/,scene/,cell_loader/), each with a thinmod.rsdispatch + topic submodules +*_tests.rssiblings.crates/renderer/src/vulkan/acceleration/andscene_buffer/are likewise split. The authoritative tree is in_audit-common.md§ Project Layout — route against it, not against memory.
Step 3: Regression-focused checks on each changed file
For every changed file, read the hunk + minimal context and ask:
- [ ] New bug — logic error, off-by-one, wrong byte width, missing version/era gate (B-splines reach FNV/FO3, not just Skyrim+).
- [ ] Contract break — did a public signature change without all
call sites updating? (
git grepthe symbol across the workspace.) - [ ] Silent divergence — was a value built at two sites and only one
edited? The classic NIFAL leak: the two
Materialload paths. - [ ] Unsafe delta — new
unsafeblock, changed safety invariant, or a safety comment that no longer matches the body (_audit-severity: MEDIUM floor for unsafe-without-comment). - [ ] Lock / query delta — changed RwLock scope or a new multi-component query? Verify TypeId-sorted acquisition (deadlock → HIGH floor).
- [ ] Vulkan delta — new pipeline/barrier/sync, AS build/refit, or descriptor write? Missing barrier or wrong AS geometry → see the severity special-rules table (HIGH/CRITICAL floors).
- [ ] GPU-struct lockstep — a touched
#[repr(C)]struct (GpuInstance/GpuCamera/GpuMaterial/GpuLight) and its mirror in every shader that reads it. Size/offset drift → HIGH. - [ ] Missing test — a changed code path with no corresponding test update. Flag in the "Missing Tests" section even if the code is right.
Rust-specific deltas
- [ ] Drop ordering — Vulkan destruction order still reverse of build?
- [ ] Error handling — new
unwrap()/expect()on a recoverable path? - [ ] Lifetimes — a borrow whose scope changed (temporary outliving / new dangling borrow)?
- [ ] Trait impls — a new impl consistent with the existing family (Component storage decl, Send+Sync)?
NIFAL boundary delta (when the diff touches the material rows)
The canonical material contract is resolve-once at the translation boundary, no render-time fallback — so a wrong value there is silently wrong across every game.
Material::metalness/roughnessare plain resolvedf32(crates/core/src/ecs/components/material.rs), notOption. They are finalized byMaterial::resolve_pbr, which clamps and — only when the upstream override arrivedNaN— falls back toclassify_pbr_keyword(the surviving keyword classifier; it is a sentinel-backstop for non-pre-classified sources, not a per-draw safety net).- NIF-imported content is pre-classified at import (
classify_legacy_pbr) soresolve_pbronly clamps; BGSM/BGEM also arrive pre-classified. Confirm a diff did not leave either scalarNaNat draw time. - Both load paths must still route through
translate_material:byroredux/src/cell_loader/spawn.rs(REFR spawn) andbyroredux/src/scene/nif_loader.rs(loose-NIF). If a diff adds a field to one and not the other, that is the divergence this layer exists to prevent. Defer to/audit-nifalfor the full single-boundary checklist.
NIFAL particle / collision chain (multi-file translation surfaces)
These are not single-file changes — a diff to one tier is incomplete without the others:
- Particle emitter: typed blocks
(
crates/nif/src/blocks/particle.rs:NiPSysEmitter/NiPSysEmitterCtlr/NiPSysEmitterCtlrData/NiPSysGrowFadeModifier) → extraction (crates/nif/src/import/walk/mod.rs:extract_emitter_params/extract_emitter_rate) → system (byroredux/src/systems/particle.rs:apply_emitter_params). Audit all three. - Collision shape: a new
Bhk*Shapeparser (crates/nif/src/blocks/collision/) is also a translation surface —crates/nif/src/import/collision/mod.rsmust map it toCollisionShape, or it is silently dropped (MEDIUM floor; HIGH if it removes visible game content). PHYSAL now consumes ragdoll constraints, so a collision diff can ripple intobyroredux/src/ragdoll.rs+crates/physics/.
Step 4: Deduplicate
Run the dedup pass from _audit-common.md § Deduplication for every
finding (existing-issue search + prior-report scan) before recording it.
A regression of a closed issue is reported as "Regression of #NNN".
Extra Per-Finding Field
In addition to the base format in _audit-common.md:
- Changed in:
<file-path>(commit<hash>/ working tree)
Output
Write to: docs/audits/AUDIT_INCREMENTAL_<TODAY>.md (YYYY-MM-DD).
Report structure
- Change summary — scope (range/commits/since), files changed, themes.
- Routing map — each changed file → dimension(s) it was audited under.
- Findings — new bugs + regressions (base format +
Changed in). - Missing tests — changed code paths with no test update.
Then suggest:
/audit-publish docs/audits/AUDIT_INCREMENTAL_<TODAY>.md
From the file to a first result
The download contains exactly the content displayed on this page. This copy may differ from the latest version in the source repository.
View the source and supporting filesThis download contains SKILL.md only. Any scripts, references, templates and dependencies are not included. Get the required files from the source, preserving their directory structure, before continuing.
1. Download the file
Download SKILL.mdFrom the root of a test project, save the file as .skill-downloads/delta-audit-mrx3y0cn/SKILL.md, or use the command below. This folder is for inspecting the skill before installation.
Command for macOS/Linux or WSL: requires a POSIX shell and curl. On Windows without WSL, use the download button and create the folder manually.
mkdir -p '.skill-downloads/delta-audit-mrx3y0cn' && curl --fail --show-error --location 'https://skills-guides.com/api/skills/delta-audit-mrx3y0cn/download' --output '.skill-downloads/delta-audit-mrx3y0cn/SKILL.md.part' && mv '.skill-downloads/delta-audit-mrx3y0cn/SKILL.md.part' '.skill-downloads/delta-audit-mrx3y0cn/SKILL.md'2. Check prerequisites
Read SKILL.md before using it. If it references scripts or external services, inspect them and prepare the required access in your test environment.
No recognized prerequisites section in this file. This does not mean it is self-contained: libraries, tools and access requirements may be mentioned elsewhere in the content or source repository.
The platforms on this page are catalog indications. The paths below follow tool documentation; they do not prove this skill has been run in those tools. No minimum tool version is confirmed for this listing.
3. Load the skill
This file does not provide a usable name/description header for this native procedure. It remains available to download and inspect; use the manual method below.
Manual method or another tool: attach SKILL.md to a conversation, or open it in an agent that can read files. Include any required supporting resources. No universal folder is recognized by every tool.
4. Try a first task
Use a short, synthetic input: an API endpoint, a small document or an interface brief. In a new conversation, attach the file or give the agent its path, then adapt this request:
Read .skill-downloads/delta-audit-mrx3y0cn/SKILL.md and apply its instructions to this example: [describe a test input]. Identify any missing files, dependencies or access. Produce the result requested by the skill without inventing data or running unreviewed scripts.Expected result: output matching the skill instructions, or a precise list of missing prerequisites. A read confirmation alone does not validate a first use. If an external resource is unavailable, that path remains untested.
ClickHouse Performance Comparison
Testing
Evaluate ClickHouse performance test results, check PR performance changes, compare with history, and assess whether a result is actionable.
Analyze CI Performance Reports
Testing
Analyze CI performance comparison reports for a ClickHouse PR. Cross-references master history to distinguish real regressions from flaky tests.
Santa Method
Testing
Multi-agent adversarial verification with two independent reviewers. Output is checked twice and iteratively fixed until both pass before shipping.