القائمة
WDBX compute backend superpower. CPU/GPU/NPU/TPU backend selector with deterministic CPU SIMD fallback. ANE detection, remote TPU dispatch.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Synchronize ABI Markdown and docs claims with executable behavior. Use before editing README, AGENTS.md, CLAUDE.md, GEMINI.md, docs/**/*.mdx, walkthroughs, changelogs, or any old ABI Markdown that mentions WDBX, MCP, GPU, clustering, compression, FHE, agents, or benchmarks.
Validate ABI docs and external collateral against repo evidence and docs/contracts/external-claims-audit.mdx. Use when auditing claims, scanning docs for unsupported QPS/sharding/FHE/AES language, checking one claim before publish, or re-confirming the external-claims audit. Agent procedure only — not a CLI or slash command. Pair with abi-doc-claims-sync when editing docs.
Plan and execute ABI repo work from the TODO board, roadmap/spec docs, Zig 0.17 constraints, and validation gates. Use when the user asks to find ABI todos/roadmaps, organize a large implementation from tasks/todo.md, or decide the next safe work slice. Distinct from the goals skill — that skill owns the persistent ledger tasks/goals.md; this skill plans from the active board and north-star specs.
TUI/dashboard superpower. Launch interactive agent REPL, diagnostics dashboard panes, and slash commands via real abi CLI paths. Use when working on abi tui / dashboard / agent tui. There is no /abi-superpower-tui binary.
Build the abi CLI and non-interactively smoke the diagnostics dashboard — render `abi dashboard` one-shot (stdin from /dev/null forces the non-interactive fallback), assert exit 0 and all 5 panels (System, Plugins, WDBX Storage, Scheduler, Memory). Use to smoke-test `abi dashboard`/`abi tui`/`abi --tui` in CI or headless without hanging in the interactive TUI. Complements run-tui (which drives the interactive path via tmux).
Capture, track, and execute goals for the abi Zig project. Use when the user wants to record a goal, see outstanding goals, update goal status, or break a goal into a verified work slice. Distinct from abi-goal-orchestrator (which plans from the TODO board) — this skill owns a persistent goal ledger and drives goal-to-slice execution.
| name | abi-superpower-wdbx-compute |
| description | WDBX compute backend superpower. CPU/GPU/NPU/TPU backend selector with deterministic CPU SIMD fallback. ANE detection, remote TPU dispatch. |
| superpower | {"command":"execute","parameters":[{"name":"action","type":"string","enum":["info","select","backend","remote"],"description":"Compute action"},{"name":"backend","type":"string","enum":["cpu","gpu","npu","tpu","auto"],"description":"Target backend"},{"name":"endpoint","type":"string","description":"Remote compute endpoint (TPU)"}]} |
Exposes the WDBX compute backend selector as a superpower. Dynamic backend selection across CPU (scalar/AVX2/AVX-512/NEON), GPU (CUDA/Metal/Vulkan), NPU (ANE detection), and TPU (remote dispatch transport). Always degrades to deterministic CPU SIMD path — native accelerator kernels are NOT linked.
Show compute backend matrix (availability + dispatch status):
/abi-superpower-wdbx-compute info
Output includes:
std.simd.suggestVectorLength)ABI_REMOTE_COMPUTE_ENDPOINT) plus a
separately tested reference TCP transportSelect and initialize a backend:
/abi-superpower-wdbx-compute select --backend auto
Options: cpu, gpu, npu, tpu, auto (prefers Metal on macOS, else CPU SIMD)
Get details for a specific backend:
/abi-superpower-wdbx-compute backend --backend metal
Show or plan remote TPU reference transport configuration:
/abi-superpower-wdbx-compute remote --endpoint http://tpu-server:8080
| Backend | Available | Dispatches | Reality |
|---|---|---|---|
| CPU SIMD | Always | ✅ Always | Vectorized @Vector with host-matched width |
| Metal (macOS) | On macOS | ⚠️ Only if g_metal_context.init() succeeds | Pure-Zig ObjC FFI, runtime MSL compile |
| CUDA | Never | ❌ No | Capability reported, not linked |
| Vulkan | Never | ❌ No | Needs loader/ICD + SPIR-V; not linked |
| ANE (NPU) | On Apple Silicon | ❌ No | compute.aneHardwarePresent() truthfully detects; needs CoreML/ObjC |
| TPU (remote) | Reference only | ❌ Not production-wired | remote_compute.zig has a loopback-tested DOT transport; the environment endpoint is report-only |
| Component | Source | Role |
|---|---|---|
| Backend Selector | src/features/wdbx/compute.zig | Dynamic CPU/GPU/NPU/TPU selection |
| CPU SIMD | src/features/wdbx/hnsw_distance.zig | @Vector cosine, std.simd.suggestVectorLength |
| GPU | src/features/gpu/vector_ops.zig + compute_api.zig | cosineSimilarity() via Metal, CPU fallback |
| NPU Detection | src/features/wdbx/compute.zig | aneHardwarePresent() — truthful |
| Remote TPU | src/features/wdbx/remote_compute.zig | Reference DOT transport; no production caller wires ABI_REMOTE_COMPUTE_ENDPOINT |
| Parity Test | src/features/gpu/compute_api.zig | CPU/GPU dot-product parity |
abi wdbx compute info
Requires feat-wdbx=true and feat-gpu=true (both default). When disabled, returns FeatureDisabled.
Per docs/spec/wdbx-north-star.mdx §3.3 and docs/contracts/external-claims-audit.mdx: