vmafx

Reconcile fork master with Netflix/vmaf master. Detects the fork's port-only topology (no shared merge-base) and emits a coverage report; falls back to merge-based sync when the histories are actually connected.

Scaffold a new feature extractor (e.g. a novel metric) with C source+header, registry entry, doc stub, and a smoke test. Does not produce a SIMD or GPU path — those come via /add-simd-path and /add-gpu-backend.

Scaffold a complete new GPU backend (hip, rocm, metal, opencl, etc.) with runtime, feature kernel stubs, public header, Meson options, CI workflow, doc stub, and smoke test. Flagship scaffolding skill. Established precedents on the fork — CUDA (legacy), SYCL (legacy), Vulkan (T5-1, ADR-0175 scaffold + ADR-0186 image-import).

Scaffold a new SIMD implementation for an existing feature. Creates intrinsics source + header + a bit-exact-vs-scalar comparison test; wires into runtime dispatch. Supports kernel-spec flags that pull recurring patterns from simd_dx.h.

Scaffold a new Kubernetes CRD + kubebuilder controller + RBAC + helm chart entry for the vmafx-operator. Follows the VmafxJob / VmafxNode / VmafxModelTraining precedent in cmd/vmafx-operator/ (ADR-0714, ADR-0709 parent).

Scaffold a new VMAFX MCP tool handler with byte-compatible Go (cmd/vmafx-mcp) and Python (mcp-server/vmaf-mcp) implementations, registration, smoke tests, and docs/mcp/ page. Companion to /add-gpu-backend and /add-feature-extractor for the MCP surface.

Register a new VMAF model (.json / .pkl / .onnx) with the build, validate its schema, and add a loader smoke test.

Replay the project modernization audit on the current tree and produce /tmp/modernization-audit-YYYY-MM-DD.md. Wraps scripts/dev/project_modernization_audit.py with sensible defaults and timestamped output.

Binary-search a timeline of ONNX checkpoints for the first one that falls below a PLCC / SROCC / RMSE gate on a held-out set. Companion to /bisect-regression (which bisects code commits).

Run git bisect against a user-defined failure predicate (numeric diff, perf threshold, test failure). Outputs the first-bad commit with context.

Clone (or update) ffmpeg, apply our fork's patches under ffmpeg-patches/, configure with --enable-libvmaf against our local libvmaf.so, build, and run smoke tests on both the 'libvmaf' and 'vmaf_pre' filters.

Build the libvmaf library + CLI with the requested backend combination (cpu, cuda, sycl, hip, all) and build type. Wraps meson setup + ninja. Reports wall time + output artifact paths.

Core debugging skill. Runs the same (ref, dist) through every enabled backend (cpu, cuda, sycl, vulkan), reports per-frame per-feature ULP divergence, flags anything beyond configurable tolerance. Use after any SIMD / GPU change. Mirrors the T6-8 GPU-parity CI gate (ADR-0214) locally.

Draft a Conventional-Commits message from the current `git diff --staged` using a local LLM (Ollama). User edits before commit.

Draft a Doxygen @brief/@param/@return block for a named function in a C/C++/CUDA file via local LLM (Ollama).

Draft a Markdown model-card for a shipped ONNX tiny-AI model by collecting hard facts (graph, sidecar, op allowlist, optional live PLCC/SROCC) and handing them to the local LLM.

Run a local LLM code review via vmaf-dev-llm (Ollama-backed) against a specific file. Dev-time self-review pass before PR.

Apply clang-format, black, isort, and shfmt across the whole repo. Idempotent; safe to run repeatedly.

Run clang-tidy, cppcheck, include-what-you-use, ruff, and semgrep in parallel; produce a single merged report.

Cherry-pick a single upstream Netflix/vmaf commit onto the fork's master, auto-adapting for SIMD/GPU paths where the commit touches a feature we have multiple implementations of.

Dry-run release-please locally, preview the CHANGELOG diff, verify signing prerequisites (Sigstore/OIDC), and surface any blockers before a release PR merges.

Profile a hot path (feature + backend) using the appropriate profiler (perf / ncu / Vtune / rocprof), produce flamegraph + top-N hot functions, and suggest concrete optimizations. Delegates to the perf-profiler agent.

Rebase our ffmpeg-patches/ series onto the latest ffmpeg master (or a specified ref), resolve trivial conflicts, regenerate .patch files, and surface unresolved hunks for human attention.

Regenerate the mkdocs-material site, validate cross-references, surface stale or broken links.

Regenerate fork-added test snapshot JSONs under testdata/ (scores_cpu_*.json, netflix_benchmark_results.json) after an intentional numerical change. Requires justification committed to the message.

Run the Netflix benchmark suite (testdata/bench_all.sh) and diff the output against testdata/netflix_benchmark_results.json. Reports any delta per (resolution, feature, backend).

Run identical (ref, dist) through all enabled backends and report per-backend score + pairwise ULP diffs. Use to verify bit-exactness of a new SIMD path, new GPU backend, or any hot-path change.

Add or audit replay-manifest sidecars for AI scripts that create datasets, feature tables, model artifacts, evaluation reports, or fetch/download roots.