원클릭으로 Manus에서 모든 스킬 실행

codebase-audit

스타30

포크2

업데이트2026년 7월 10일 00:16

Use when auditing the imp codebase for structural debt, dead code, god-objects/files, duplication, flag sprawl, or deciding whether a cleanup is worth shipping — "structure audit", "tech debt", "is this still used", "dead code", "refactor for clarity", "should we split this file", "remove this flag". Do NOT use for build/test mechanics (building-and-testing), perf/kernel work (benchmark-cuda / sm120-cuda-expert), or output quality (check-degeneration).

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

kekzl

kekzl/imp

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Codebase audit & structural debt — imp

The one hard rule

Never act on a raw finding. Verify every finding against the real code first. Fan-out audits (Explore agents, grep sweeps) systematically over-flag. On the 2026-06-09 pass, the big-ticket findings C1/C2/C4 were all REFUTED on inspection, and several "dead flags/files" were live. Treat a finding as a candidate, not a fact, until a targeted check confirms it. Most of an audit's value is in the verification, not the sweep.

Verification recipes (the load-bearing checks)

Claim	How to actually verify	Trap that fooled a past audit
"flag/symbol X is dead"	grep across `src/ tests/ tools/` — `tools/` IS in scope	`dump_tokens`/`force_bos`/`bench.generate` looked dead in `src/`, but are read in `tools/imp-cli/main.cpp`
"config field never read"	grep the leaf field for a value-read in a conditional, excluding decl/parse/copy	a local var sharing the field name hides the read (`fp8_auto_legacy`)
"kernel/function is dead"	trace the call graph, not just `<<<` sites — reachable from a live dispatch = live	`q4k_imma` cache was dead, but `mmq_q4k_imma_tile`/`_reorder` are LIVE via `q4k_imma_prefill → mmq_q4k_imma_gemm` (called on the fly). Only the ~30-LOC cache struct was dead, not the "~1200 LOC stack"
"this file is a god-file, split it"	split on conflation, not size; check if the shared part is already factored	`gdn.cu`/`gemm.cu`/`sampling.cu`/`json_schema.cpp` are each one cohesive domain. The 6 `attention_paged*` variants already share `attention_paged_common.cuh` (online-softmax loop) — only dequant differs
"rewrite this into a registry/template"	no speculative abstraction — need a concrete bug, not a smell	the `weight_map.cpp` `if (!matched)` ladder works and is readable; a mis-mapped tensor name = garbage output. Risk > cosmetic gain
"imp.conf key is inert"	is it read in `src/` AND `tools/`? Is the live path the C-API instead?	`server.prefix_cache` etc. were parsed but unread — live enablement bridges at engine init (PR #636)

Workflow

Fan out — Explore agents / grep to surface candidates (breadth is fine here).
Verify each with the recipe above. Demote refuted ones immediately.
Write it up — docs/audit/structural_debt_YYYY_MM_DD.md. Counts are evidence, not estimates. Keep a "Refuted (do not re-chase)" section so the next pass doesn't re-flag the same false positives.
Ship cleanups as small PRs (REQUIRED SUB-SKILL: building-and-testing) — branch off main, never stack, Docker make build + make test-unit, prefer batching related nits. Behavior-sensitive removals: also a coherence check (check-degeneration).

File-size gate (god-file findings go through this, not ad-hoc splits)

tools/check_filesize.py (config: tools/filesize_thresholds.toml) measures code LOC (comments/blanks stripped) per category — kernel .cu warn>500/hard>600, normal TU warn>600/hard>800, header warn>500/hard>700. CI job File size: advisory warn step + blocking hard step. The real cost metric is recompile blast radius (one .cu = one ptxas TU), not line count — split on compile-time isolation (kernel def / host wrapper / explicit instantiations), never on size alone. A legitimately monolithic file goes into [allow] in the toml with a reason (empty reason is rejected). Baseline + per-file rationale: docs/audit/AUDIT_FILESIZE.md.

Priors — settled, do NOT re-flag

From the structural-debt audit chain (D1–D4 / C1–C8, archived in docs/archive/README.md; current verdicts in docs/audit/housekeeping_2026_06_13.md):

exec/ god-object (GraphExecutor) is intrinsically forward-pass-coupled; the D2 runner-extraction track ENDED (cosmetic friend-backdoor only). Header split done (1188→584 lines). Don't re-attempt runner classes.
ModelProfile (D1) centralized arch facts — don't reintroduce scattered cfg.arch==.
gdn.cu is domain-cohesive ("don't touch"); the compute/ files are not god-files.
The two-config-systems split (RuntimeConfig global vs ModelConfig::Overrides) is known and deliberate; config keys use the typed-binder B/I/F/S(...) ladder (PR #626).
Audit #5 (2026-07-07, docs/audit/structural_debt_2026_07_07.md): server layer is the dominant debt source → issues #888–#897 filed. Before a new pass, read that report's NOT-flagged list — re-flagging its verified negatives wastes the sweep. Companion: docs/audit/vram_audit_2026_07_07.md (note: a VramOwned type does NOT exist — past audits hallucinated it).

Red flags — you are about to over-flag

About to delete code because a sweep said "no references" → did you grep tools/?
About to split a file because it's big → is it conflated or just one large domain?
About to rewrite a working ladder/loader for elegance → where's the concrete bug?
Reporting LOC of "dead" kernels → did you trace the call graph, or just <<<?

이 저장소의 다른 Skills

같은 저장소

server-api

kekzl/imp

Use when working on or testing imp-server and its HTTP APIs — OpenAI/Anthropic endpoints, /v1/chat/completions, /v1/messages, SSE streaming, tool calling, json_schema/constrained decoding, thinking/reasoning_content, cache_control/prefix cache, model loading semantics, "server returns 404/garbage", API compliance. Do NOT use for CLI-only inference (imp-cli), kernel work (sm120-cuda-expert), or pure output-quality checks (check-degeneration).

2026-07-1930

benchmark-cuda

kekzl/imp

Use when benchmarking, profiling, or A/B-testing CUDA kernels or end-to-end perf in the imp inference engine on RTX 5090 (sm_120), including refreshing tests/perf_baseline.json or publishing numbers to docs/BENCHMARKS.md and the README. Triggers on "benchmark kernel", "profile cuda", "ncu", "nsys", "kernel timing", "occupancy", "bandwidth bound", "compute bound", "roofline", "perf baseline", "is this regression real", "decode dropped". Do NOT use for writing/optimizing kernel code (sm120-cuda-expert) or output-quality checks (check-degeneration).

2026-07-1530

sm120-cuda-expert

kekzl/imp

Use when writing, reviewing, or optimizing CUDA kernels targeting sm_120a (RTX 5090 / Consumer Blackwell, GB202) in the imp inference engine. Triggers on CUDA/PTX kernel code, shared-memory layout, tensor-core MMA, GEMV/GEMM/attention/quantization kernels, decode tok/s under expected, kernel emitting HMMA instead of mxf4nvf4, occupancy or register-pressure questions. Pair with `benchmark-cuda` for measurement and `check-degeneration` after hot-path changes.

2026-07-1330

quant-formats

kekzl/imp

Use when working on imp's quantization formats, loaders, or dequant paths — GGUF Q4_0…Q8_0/Q*_K/IQ4/MXFP4, NVFP4 two-level scaling, FP8 E4M3, StorageTier, decode cache, KV-cache dtypes, "which quant should I use", scale-factor layout, dequant kernel wiring. Do NOT use for writing/optimizing GEMM/GEMV kernels (sm120-cuda-expert) or measuring quant perf (benchmark-cuda).

2026-07-1130

building-and-testing

kekzl/imp

Use when building imp, running its test suite, checking CI status, or debugging build/test failures — "make build", "run the tests", "test-gpu", "verify-fast", GTEST_FILTER, Docker/CUDA toolchain, dependency bumps, "CI is red/blocked", determinism or perplexity checks. Do NOT use for benchmarking/profiling (benchmark-cuda) or output-quality batteries (check-degeneration).

2026-07-1030

check-degeneration

kekzl/imp

Use when verifying that a model in the imp inference engine produces coherent output without repetition loops, token-stuck states, or state corruption across turns. Triggers on "degenerates", "check degeneration", "repetition loop", "own own own", "stuck token", "multi-turn regression", "does it still work", and after enabling CUDA graphs / changing forward pass / MoE routing / KV cache / GDN state.

2026-07-1030

name

codebase-audit

description

Codebase audit & structural debt — imp

The one hard rule

Verification recipes (the load-bearing checks)

Claim	How to actually verify	Trap that fooled a past audit
"flag/symbol X is dead"	grep across `src/ tests/ tools/` — `tools/` IS in scope	`dump_tokens`/`force_bos`/`bench.generate` looked dead in `src/`, but are read in `tools/imp-cli/main.cpp`
"config field never read"	grep the leaf field for a value-read in a conditional, excluding decl/parse/copy	a local var sharing the field name hides the read (`fp8_auto_legacy`)
"kernel/function is dead"	trace the call graph, not just `<<<` sites — reachable from a live dispatch = live	`q4k_imma` cache was dead, but `mmq_q4k_imma_tile`/`_reorder` are LIVE via `q4k_imma_prefill → mmq_q4k_imma_gemm` (called on the fly). Only the ~30-LOC cache struct was dead, not the "~1200 LOC stack"
"this file is a god-file, split it"	split on conflation, not size; check if the shared part is already factored	`gdn.cu`/`gemm.cu`/`sampling.cu`/`json_schema.cpp` are each one cohesive domain. The 6 `attention_paged*` variants already share `attention_paged_common.cuh` (online-softmax loop) — only dequant differs
"rewrite this into a registry/template"	no speculative abstraction — need a concrete bug, not a smell	the `weight_map.cpp` `if (!matched)` ladder works and is readable; a mis-mapped tensor name = garbage output. Risk > cosmetic gain
"imp.conf key is inert"	is it read in `src/` AND `tools/`? Is the live path the C-API instead?	`server.prefix_cache` etc. were parsed but unread — live enablement bridges at engine init (PR #636)

Workflow

Fan out — Explore agents / grep to surface candidates (breadth is fine here).
Verify each with the recipe above. Demote refuted ones immediately.
Write it up — docs/audit/structural_debt_YYYY_MM_DD.md. Counts are evidence, not estimates. Keep a "Refuted (do not re-chase)" section so the next pass doesn't re-flag the same false positives.
Ship cleanups as small PRs (REQUIRED SUB-SKILL: building-and-testing) — branch off main, never stack, Docker make build + make test-unit, prefer batching related nits. Behavior-sensitive removals: also a coherence check (check-degeneration).

File-size gate (god-file findings go through this, not ad-hoc splits)

Priors — settled, do NOT re-flag

From the structural-debt audit chain (D1–D4 / C1–C8, archived in docs/archive/README.md; current verdicts in docs/audit/housekeeping_2026_06_13.md):

exec/ god-object (GraphExecutor) is intrinsically forward-pass-coupled; the D2 runner-extraction track ENDED (cosmetic friend-backdoor only). Header split done (1188→584 lines). Don't re-attempt runner classes.
ModelProfile (D1) centralized arch facts — don't reintroduce scattered cfg.arch==.
gdn.cu is domain-cohesive ("don't touch"); the compute/ files are not god-files.
The two-config-systems split (RuntimeConfig global vs ModelConfig::Overrides) is known and deliberate; config keys use the typed-binder B/I/F/S(...) ladder (PR #626).
Audit #5 (2026-07-07, docs/audit/structural_debt_2026_07_07.md): server layer is the dominant debt source → issues #888–#897 filed. Before a new pass, read that report's NOT-flagged list — re-flagging its verified negatives wastes the sweep. Companion: docs/audit/vram_audit_2026_07_07.md (note: a VramOwned type does NOT exist — past audits hallucinated it).

Red flags — you are about to over-flag

About to delete code because a sweep said "no references" → did you grep tools/?
About to split a file because it's big → is it conflated or just one large domain?
About to rewrite a working ladder/loader for elegance → where's the concrete bug?
Reporting LOC of "dead" kernels → did you trace the call graph, or just <<<?