agentleFS
mikeparcewskimikeparcewski/wicked-garden/skills/platform/gate-benchmark-rebaseline/SKILL.md

wicked-garden-platform-gate-benchmark-rebaseline

Skill9 starsChanged 7 months agoMITView on GitHub

Re-baseline procedure for the AC-11 gate-result benchmark lane (`tests/crew/test_gate_result_benchmark.py`). The benchmark enforces a 2× p95 SLO on `gate-result.json` ingestion. When a deliberate perf change lands on main (validator hardening, cache tuning, schema expansion), the baseline needs updating. Never re-baseline to silence a regression.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do not re-baseline to silence a regression: If the benchmark fails on a PR that didn't intend perf work, treat it as a real regression and …

    line 33
  2. Very specific:

    Missing-baseline behavior: if benchmark_baseline.json is deleted or malformed, the test soft-skips with a directive to record one.

    line 45
  3. Distinctive:

    If the benchmark lane is broken or disabled on main:

    line 87
  4. Very specific:

    All flags auto-expire at WG_GATE_RESULT_STRICT_AFTER.

    line 102

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Gate-Result Benchmark Re-baseline

Operational procedure for updating tests/crew/benchmark_baseline.json after a legitimate perf change on main.

When to re-baseline

Re-baseline only when a deliberate perf change lands on main. Examples:

  • Validator hardening (added / removed gate-result schema checks)
  • Cache eviction tuning (changes to gate-result ingestion memoization)
  • Schema expansion (new required fields in gate-result.json)
  • Sanitizer pattern changes (scripts/qe/content_sanitizer.py)

Do not re-baseline to silence a regression. If the benchmark fails on a PR that didn't intend perf work, treat it as a real regression and find the cause.

How the lane works

  • Workflow: .github/workflows/benchmark.yml
  • Trigger: PRs touching scripts/crew/phase_manager.py, the gate-result schema/sanitizer (scripts/qe/content_sanitizer.py), or the benchmark test/baseline themselves. Other PRs skip the lane to keep default CI cost flat.
  • Test: tests/crew/test_gate_result_benchmark.py::test_load_gate_result_p95_within_2x_baseline
  • Marker: benchmark — opt-in. Local uv run pytest deselects it (see pyproject.toml addopts = "-m 'not benchmark'").
  • Explicit invocation: uv run pytest -m benchmark
  • Fixtures: 50 deterministic gate-result.json files cycling through 1 KB / 4 KB / 16 KB / 60 KB (bounded by MAX_SUMMARY_BYTES). Each round clears the memoization cache so the full validate + sanitize + cache-insert path is measured. Cache-hit timing is not part of the SLO.
  • Baseline: tests/crew/benchmark_baseline.json — p95_ns in nanoseconds.
  • Assertion: p95_current ≤ slo_multiplier × p95_baseline (default slo_multiplier = 2.0). The workflow comments the delta on every triggered PR.
  • Missing-baseline behavior: if benchmark_baseline.json is deleted or malformed, the test soft-skips with a directive to record one. The SLO is not enforced until a valid baseline is present — so a re-baseline PR can merge without a circular dependency.

Re-baseline procedure

  1. Check out the target main commit.

    git checkout main && git pull
  2. Run the benchmark three times. The SLO is against p95, and CI noise varies — taking the highest of three local runs adds margin.

    uv run pytest -m benchmark tests/crew/test_gate_result_benchmark.py -s
    uv run pytest -m benchmark tests/crew/test_gate_result_benchmark.py -s
    uv run pytest -m benchmark tests/crew/test_gate_result_benchmark.py -s
  3. Record the highest p95_current reading from the three runs.
  4. Update tests/crew/benchmark_baseline.json:

    FieldNew value
    p95_nsThe highest p95_current from step 3, rounded up
    recorded_onToday's date (YYYY-MM-DD)
    recorded_from_commitShort SHA of the target main commit
    slo_multiplierLeave at 2.0 unless the AC-11 contract changes
    rebaseline_procedureLeave at wicked-garden:platform:gate-benchmark-rebaseline
  5. Commit the baseline update in a dedicated PR. Title convention:

    chore(benchmark): re-baseline AC-11 after {change-summary}
  6. The benchmark workflow runs on the PR and must pass (p95 should be well under 2× the new baseline since you just measured it).

Pre-flip dependency (semantic-gate condition)

Strict mode for gate-result ingestion (WG_GATE_RESULT_STRICT_AFTER, default 2026-06-18) requires the AC-11 benchmark lane to be active in CI. Rationale: once strict-mode activates, a silent 2×+ perf regression would push every approve_phase call over the SLO without a guard.

If the benchmark lane is broken or disabled on main:

  • Before the flip date: push WG_GATE_RESULT_STRICT_AFTER out (env var or default-constant update). Do not let strict-mode activate without benchmark enforcement.
  • After the flip date: fix the benchmark lane before touching the ingestion path.

Env-var rollback (preferred over revert)

For production rollback of a specific ingestion check, prefer env-var soft-disable over git-revert:

VariableEffect
WG_GATE_RESULT_SCHEMA_VALIDATION=offSkip schema validator
WG_GATE_RESULT_CONTENT_SANITIZATION=offSkip content sanitizer
WG_GATE_RESULT_DISPATCH_CHECK=offSkip dispatch-log orphan check

All flags auto-expire at WG_GATE_RESULT_STRICT_AFTER.

References

  • Test: tests/crew/test_gate_result_benchmark.py
  • Baseline: tests/crew/benchmark_baseline.json
  • Workflow: .github/workflows/benchmark.yml
  • Ingestion path: scripts/crew/phase_manager.py (gate-result load + memoization)
  • Sanitizer: scripts/qe/content_sanitizer.py

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.