메뉴
Full benchmark cycle for devloop — collect criterion results, check against budgets, commit the run to SQLite, surface regressions. Use after running cargo bench or when managing performance budgets.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Deep codebase analysis — architecture health, test coverage gaps, tech debt, and actionable findings
Run pre-commit checks (fmt, clippy, test) then stage, commit, and push with a conventional message
Review staged or recent changes for correctness, style, and Rust conventions before committing
Use when SSHing into remote machines, looking up credentials, renaming Tailscale devices, or accessing services on the tailnet. Symptoms - SSH auth failures, too many authentication failures, command not found tailscale, need to find credentials for remote machines.
Use when mixing Tokio async code with synchronous blocking I/O libraries (tar, flate2, zip, csv), or when seeing "cannot block the current thread from within the async context" panics, or needing to stream async bytes into a sync reader.
Structured BAML schema edit → validate → test loop for the devloop project. Use when editing any *.baml file in crates/baml/baml_src/ — covers version parity gotchas, uvx invocation, regeneration workflow, and the analyze verify step.
| name | devloop-bench-cycle |
| description | Full benchmark cycle for devloop — collect criterion results, check against budgets, commit the run to SQLite, surface regressions. Use after running cargo bench or when managing performance budgets. |
Full workflow for collecting, checking, and committing benchmark results in devloop.
devloop binary installed at ~/.local/bin/devloop--repo)cd ~/dev/devloop
cargo bench --workspace 2>&1 | tee /tmp/bench-output.txt
Criterion outputs JSON results to target/criterion/. The bench cycle reads from there.
devloop bench collect 2>&1
This reads criterion output, parses mean_ns, median_ns, std_dev_ns, ci_lower_ns, ci_upper_ns per benchmark, and saves a run to the SQLite bench store.
If collect fails with "no criterion data": ensure cargo bench was run first and target/criterion/ exists.
devloop bench check 2>&1
Output shows regression status per benchmark:
✓ STABLE — within CI bounds of previous run⚠ WARNING — above previous CI upper bound (statistical regression)✗ REGRESSION — over budget (max_ns exceeded)↓ IMPROVED — below previous CI lower boundIf regressions are found: do not commit the run. Investigate root cause first.
devloop bench history --last 10 2>&1
Shows trend per benchmark. Useful for distinguishing noise from real regressions.
Set a new budget:
devloop bench budget set <benchmark-name> --max 500ms --warn 400ms
List all budgets:
devloop bench budget list
Remove a budget:
devloop bench budget remove <benchmark-name>
Budget values use time units: 100ns, 1ms, 500ms, 2s.
After a clean check (no regressions):
git add -A # bench results are stored in SQLite, not git-tracked directly
git commit -m "bench: collect run on <branch>"
If there were regressions that were investigated and accepted:
git commit -m "bench: collect run — <benchmark-name> regression accepted, reason: <explanation>"
| Status | Action |
|---|---|
| IMPROVED | Note in commit message, check it's not a measurement artifact |
| WARNING (+5–20%) | Investigate — check for recent changes to the hot path |
| WARNING (>20%) | Treat as regression — find root cause before proceeding |
| REGRESSION (over budget) | Block — must fix or explicitly raise the budget with justification |
Benchmarks are auto-categorized:
micro — < 1ms mean (e.g., pure computation, data structure ops)integration — >= 1ms mean (e.g., full pipeline runs, I/O-bound ops)Integration benchmarks get budget enforcement; micro benchmarks use statistical comparison only.