Run Benchmark

Invocation contract

Run this skill only after an explicit user request. Implementing, fixing, or reviewing performance-adjacent code does not by itself authorize a benchmark run.

All timing in this workflow assumes one dedicated, quiet local machine accumulating a series over time. Numbers from CI runners, other machines, or a loaded host are not comparable and must not enter the baseline series. Never run other builds, tests, or heavy agents concurrently with a timed run.

Background and design decisions live in docs-ai/057-performance-benchmark-suite/; the measurement methodology (why workload context matters, why load kills comparability) in docs-ai/032-performance-hardening/004-agent-detection-steady-state.md.

Step 0 — Preflight

1. git status --short — an A/B run switches the checkout to the baseline commit, so the tree must be clean and HEAD committed. If dirty, tell the user what is blocked and ask whether to commit/stash or fall back to a branch-only run.
2. Quiet-host check: compare uptime load1 against sysctl -n hw.ncpu. Above roughly half the cores, warn that starvation inflates numbers unevenly and ask whether to wait or proceed; above one runnable process per core, refuse the timed run and say why.
3. Note what is available for the live layer: is a Prowl instance running (prowl list --json), and are any agents working (animating titles)?

Step 1 — Classify intent, then always confirm

Gather evidence before guessing:

• The user's own words in the invocation — these override everything below.
• git branch --show-current; BASE=$(git merge-base HEAD origin/main); git diff --name-only $BASE and git log --oneline $BASE..HEAD.
• Map changed files through the table in Reference to see which benchmark suites the branch can plausibly affect.

Classify into one of three modes:

Then confirm with the user before running anything, even when the evidence is clear: present the guessed mode with its one-line evidence and the alternatives (AskUserQuestion with the recommended option first). If the evidence is contradictory or absent, do not guess — ask. Benchmark runs take minutes and switch the checkout; a wrong mode wastes both.

Step 2 — Run the confirmed mode

Every make bench appends records to ~/Library/Logs/Prowl/measurements/bench/bench.jsonl keyed by git rev-parse --short HEAD. Record the exact SHA before each run so the comparison step can find its rows.

hot-path — A/B against the merge-base

1. Check whether the baseline already exists on this machine: jq -c "select(.gitSHA == \"<base-short-sha>\")" bench.jsonl. Reuse it if present.
2. If absent: git switch --detach $BASE → make bench → git switch -. The first optimized build after a switch takes minutes; later ones are incremental.
3. make bench on the branch HEAD.
4. Compare per Step 3.

new-optimization — pin the claim before measuring it

1. Check the Reference inventory: does any suite already cover the optimized path?
2. If not, the primary deliverable is a new benchmark following the 057 pattern, in the same PR as the optimization:
   • dig the replaced implementation out of git history and keep it verbatim as the reference;
   • assert output equivalence before timing;
   • assert a floor ratio 2–4x below the measured Release ratio;
   • gate Swift-vs-Swift ratios on optimized builds (BenchmarkMeasurement.isOptimizedBuild); C-call/filesystem ratios assert everywhere.
3. Then run the hot-path A/B: the new suite quantifies the win, the existing suites catch collateral damage.
4. The measured ratio goes into the report as the claim's durable evidence — this is exactly what the #644–#665 wave lacked.

release — full sweep plus series

1. make test — the behavioral + ratio layer (Debug).
2. make bench — compare against the trailing series (last ~5 records per case across recent SHAs), not a single baseline.
3. Live layer, best-effort, report each as run/inconclusive/skipped:
   • make measure-titles — needs a running app with at least one animating agent tab; exit 3 means nothing was animating (inconclusive, not a pass).
   • make measure-cpu — needs a running Prowl Debug instance; set PROWL_PID when several run. Records agent mix and load next to the attribution, so quote them together.
4. make capture-spike is a standing watch for long-running instances, not a step here; mention it if the user is chasing an intermittent spike.

Step 3 — Compare and judge

Extract both sides:

jq -r 'select(.gitSHA == "SHA") | [.suite, .name, .referenceMilliseconds, .shippedMilliseconds, .ratio] | @tsv' \
  ~/Library/Logs/Prowl/measurements/bench/bench.jsonl

Judgment rubric (single machine, warm cache, medians of 15):

• A ratio below its asserted floor already failed the test — confirmed regression, no further interpretation needed.
• shippedMilliseconds against baseline: within ±20% is noise; 1.2–1.5x growth is a watch — rerun make bench once to confirm before claiming anything; above 1.5x sustained is a regression; above 2x is major. Name suspect commits via the file→suite map.
• Use referenceMilliseconds as the built-in host canary: the reference workload never changes, so if it moved by a similar factor the host moved, not the code; if shipped moved alone, the code moved.
• Never average across load regimes. If load was elevated during a run, discard the run and redo it — do not "correct" the numbers.

Step 4 — Persistent report

Write a markdown report to ~/Library/Logs/Prowl/measurements/reports/$(date +%Y%m%d-%H%M)-<branch-or-tag>.md (mkdir -p the directory; it lives with the rest of the measurement home, survives reboots, and is never cleaned by the system). Reports are append-only history — never overwrite or delete earlier ones.

Required sections:

• Intent — the mode the user confirmed, in one line.
• Context — date, branch, baseline and branch SHAs, load average and core count at run time, app/agent state if the live layer ran.
• Results — one table per layer: case, baseline shipped ms, branch shipped ms, delta, ratio vs floor.
• Verdict — no-regression / watch / regression / major, with suspect commits for anything above noise, and the host-canary observation.
• Not run — layers skipped or inconclusive, and why.
• Raw refs — the bench.jsonl SHAs and any script run directories.

Reply to the user with the verdict and the report path. The report, not the chat, is the durable record.

Reference

Benchmark inventory (all under supacodeTests/, suite root PerformanceBenchmarks):

File → suite map for intent classification and suspect naming:

Caveats:

• referenceMilliseconds values measured inside the Release test bundle are not comparable to standalone -O binaries (~10x apart for Data.reduce); ratios are unaffected. Compare bench.jsonl rows only with other bench.jsonl rows.
• The suite is .serialized and make bench disables parallel testing and pins the destination arch deliberately — do not "speed it up" by removing either.
• bench.jsonl is append-only; the same SHA can have several rows (reruns). Prefer the newest row per (suite, name, SHA) and note reruns in the report.