菜单
Scan Claude Code JSONL session transcript files for behavioral patterns — loops, errors, expensive turns, trial-and-error signatures. Returns structured findings for the improvement-analyst agent.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run a Criterion benchmark on a Rust DSP crate and assert against a checked-in baseline. Establishes whether the implementation fits the real-time budget and catches performance regressions — the Rust equivalent of `perf stat` against the audio buffer time.
Regenerate a C header from a Rust crate via cbindgen and diff it against the committed header. Fails if the generated output differs — catches "Rust signature changed but the C header was not regenerated" before it reaches the C side.
Run `cargo +nightly miri test` against a Rust DSP crate to detect undefined behaviour the compiler missed — out-of-bounds access, use-after-free, uninitialised memory reads, invalid pointer arithmetic. Miri is the cheapest sound UB detector for Rust; this skill wraps it with the right isolation flags for audio DSP crates.
Audit a Rust DSP crate for real-time-safety violations — verifies `#![no_std]`, no `alloc` dependency, `panic = "abort"`, runs targeted clippy lints, greps for banned constructs in the hot path, and scans the compiled LLVM-IR for allocator symbols. Designed to be run on any Rust DSP crate before merge.
Opinionated wasm-pack wrapper that builds a Rust crate to wasm32-unknown-unknown with project-pinned flags (SIMD, LTO, panic=abort), applies a wasm-opt post-pass, and auto-invokes wasm-module-inspect on the result. Single entry point so the agent never reaches for raw wasm-pack and forgets a flag.
FFT-based sanity check for audio output files — verifies fundamental frequency, THD, SNR, clipping, and silence. Use after offline C tests or WASM OfflineAudioContext captures to confirm a signal chain is producing correct output.
| name | log-pattern-scan |
| description | Scan Claude Code JSONL session transcript files for behavioral patterns — loops, errors, expensive turns, trial-and-error signatures. Returns structured findings for the improvement-analyst agent. |
| allowed-tools | Bash |
| argument-hint | <jsonl-file-or-project-path> [all|errors|loops|expensive|trialerror|summary] |
Scan one or more Claude Code JSONL session transcript files for behavioral inefficiency patterns.
Arguments are in $ARGUMENTS:
.jsonl file, or a project root path (scans all sessions for that project)all (default), errors, loops, expensive, trialerror, summaryIf no arguments are provided, ask the user before proceeding.
When given a project root, the scan recurses into subagent sidechains
(<session>/subagents/agent-*.jsonl, marked isSidechain: true) as well as the
top-level session transcripts. This matters: the orchestrator/main transcript records
only ~15% of tool activity — the bulk of tool calls, errors, and loops happen inside
subagents. Scanning only the top level can undercount errors by an order of magnitude.
Each finding is headed by its path relative to the project folder, so sidechain
findings are attributable to the agent that produced them.
The scan script lives alongside this file. Run it directly:
bash "$(dirname "$0")/scan.sh" <first-arg> <second-arg>
| Type | What it detects |
|---|---|
summary | Total lines, tool calls, errors, token totals per session |
errors | Tool calls that returned is_error: true |
loops | Same tool called 3+ times consecutively, or same tool+input repeated 3+ times across the session |
expensive | Turns with >2000 output tokens (EXPENSIVE-OUTPUT) or >10000 new cache tokens (EXPENSIVE-CONTEXT) |
trialerror | 3+ edits to the same file (CHURN), 3+ identical bash commands (REPEAT-CMD), sessions >200 messages (LONG SESSION) |
all | All of the above |
Each finding is prefixed for easy parsing:
| Prefix | Meaning |
|---|---|
ERROR | Failed tool call |
LOOP | Consecutive same-tool repetition |
REPEAT | Same tool+input repeated across session |
EXPENSIVE-OUTPUT | High output token turn |
EXPENSIVE-CONTEXT | Large new context loaded in one turn |
CHURN | Repeated edits to same file |
REPEAT-CMD | Repeated identical bash command |
LONG SESSION | Session exceeds 200 messages |
OK | No issues found for that scan type |
Return the full script output. The improvement-analyst agent will interpret the findings and write suggestions to the project's self-improvement log.