Mk:benchmark

mk:benchmark — Experimental Harness Canary Suite

Measures harness performance against a small set of ground-truth tasks. Provides the empirical signal that the dead-weight audit (per .claude/rules/dead-weight-audit-rules.md) consumes to make load-bearing decisions about each harness component.

When to Use

Activate when:

• User runs /mk:benchmark run (default = quick tier, 5 tasks, ≤$5)
• User runs /mk:benchmark run --full (quick tier + 1 heavy task, ≤$30)
• User runs /mk:benchmark compare <run-id-a> <run-id-b> (delta table)
• Before applying a harness change (baseline)
• After applying a harness change (verify delta)
• During the dead-weight audit playbook (component enable/disable cycles)

Skip when:

• The harness has been run end-to-end manually within the last hour (use that data instead)
• Budget cap is hit before the suite finishes (record partial result, alert)

Hard Constraints

1. Quick tier ≤$5 total cost. Hard block if projected cost exceeds.
2. Full tier ≤$30 total cost. Hard block if projected cost exceeds.
3. --full is opt-in. The heavy task (06-small-app-build) requires explicit --full flag because it triggers mk:autobuild which can run for hours. Refuses to run without the flag.
4. NOT a replacement for unit tests. This is harness-level measurement only.
5. Results recorded in trace-log.jsonl as event=benchmark_result records, tagged with benchmark_version + harness_version + model_version.

Subcommands

Tier Layout

.claude/benchmarks/
├── README.md                                  ← how to use + add tasks
├── canary/
│   ├── quick/                                 ← default tier (5 tasks, ≤$5)
│   │   ├── 01-react-component-spec.md
│   │   ├── 02-api-endpoint-spec.md
│   │   ├── 03-bug-fix-spec.md
│   │   ├── 04-refactor-spec.md
│   │   └── 05-tdd-feature-spec.md
│   └── full/                                  ← --full only (1 task, ~$25)
│       └── 06-small-app-build-spec.md
└── results/                                   ← per-run JSON dumps

Usage

Run quick tier

/mk:benchmark run

Outputs:

• Per-task verdict + score
• Total cost + duration
• Run ID written to .claude/benchmarks/results/{run-id}.json AND trace-log.jsonl

Run full tier

/mk:benchmark run --full

Same as quick, plus the heavy 06-small-app-build task. Refuses to run without --full to prevent accidental cost burn.

Compare two runs

/mk:benchmark compare 260408-1430 260408-1530

Outputs a delta table:

| Task | Run A score | Run B score | Δ |
|---|---|---|---|
| 01-react-component | 0.92 | 0.88 | -0.04 |
| 02-api-endpoint    | 0.85 | 0.91 | +0.06 |
| 03-bug-fix         | 1.00 | 1.00 |  0.00 |
| ... | ... | ... | ... |
| TOTAL              | 0.89 | 0.91 | +0.02 |

Output Schema

Each benchmark run writes a JSON dump to .claude/benchmarks/results/{run-id}.json:

{
  "run_id": "260408-1430-bench",
  "tier": "quick",
  "started": "2026-04-08T14:30:00Z",
  "ended": "2026-04-08T14:42:00Z",
  "harness_version": "3.0.0",
  "model": "claude-opus-4-6",
  "total_cost_usd": 4.20,
  "total_duration_seconds": 720,
  "tasks": [
    {
      "spec": "01-react-component-spec.md",
      "verdict": "PASS",
      "weighted_score": 0.92,
      "duration_seconds": 145,
      "cost_usd": 0.85,
      "rubric_preset": "frontend-app"
    },
    ...
  ],
  "summary": {
    "passed": 4,
    "warned": 1,
    "failed": 0,
    "average_score": 0.89
  }
}

Audit Mode

scripts/git-index-audit.sh records a reproducible git tracked-state fingerprint as a JSON artifact — independent of the canary benchmark loop. Use it to verify two checkouts of a repo are identical, or to snapshot tracked state over time.

# single-repo snapshot
bash .claude/skills/benchmark/scripts/git-index-audit.sh [repo-path]
# comparison (adds local/remote-only counts + recursive diff status)
bash .claude/skills/benchmark/scripts/git-index-audit.sh <local> <remote>

Artifact location: .claude/benchmarks/audits/{YYMMDD-HHMMSS}-audit.json — a SIBLING of results/, NOT inside it. compare-runs.sh prefix-globs results/*.json and assumes a tier key; an audit artifact placed in results/ would crash it. Every audit artifact carries a top-level "type": "audit" discriminator. Override the output dir with MEOWKIT_AUDIT_OUT_DIR.

Artifact schema: type, run_id, ts, repo, tracked_file_count, directory_count_excl_git, tracked_path_sha256, tracked_index_sha256, comparison (null in single-repo mode), working_tree_clean. A best-effort audit_result trace event is appended via append-trace.sh.

Index-hash definition (toolkit canonical): tracked_index_sha256 = sha256(sort(git ls-files -s)). The -s flag includes mode + blob hash + stage, so the index hash captures tracked CONTENT, not just paths; tracked_path_sha256 = sha256(sort(git ls-files)) captures paths only. The source method did not specify an index-hash command — this definition is the toolkit's, documented here and in the script header so future comparisons are reproducible.

Gotchas

• run-canary.sh is an orchestrator-driven runner for the model-in-loop canary. It writes a manifest with PENDING tasks then prints orchestrator instructions. The script CANNOT itself invoke mk:autobuild per task because each invocation requires a fresh subagent context, which only an orchestrator agent can spawn — not a shell process. The agent invoking this skill MUST follow the printed instructions to fill in each task's results. Failure to do so leaves the manifest as a stub.
• The cost cap is now enforced (not just recorded). After each task the orchestrator appends a {costUsd,…} receipt to the run's .ledger.jsonl and runs run-canary.sh check-cap <ledger> <cap>; exit 2 means the cap was reached and the run STOPS. Thresholds follow harness-rules.md Rule 6 (warn at $30, halt at the effective cap; --budget N / MEOWKIT_BUDGET_CAP override the tier cap). This mirrors the TypeScript cost-ledger the deferred live backends inherit.
• The cross-harness journey (J10) IS automated. Its deterministic layer runs offline in CI via the TypeScript journey runner (packages/mewkit/src/journey-validation) — migration → target validation → route/artifact/denied-token/side-effect oracles — with no model calls. Only the model-in-loop (live) canary above still needs the orchestrator handoff.
• Circular dependency with mk:autobuild. This skill invokes mk:autobuild per task. If a harness bug is exactly what the dead-weight audit is trying to find, the audit can fail to even start. The manual fallback is documented in .claude/rules/dead-weight-audit-rules.md Rule 8 — run individual canary specs via /mk:cook <spec.md> and score by hand.
• Don't treat 100% pass as "harness is perfect." Canary tasks are intentionally simple. Real-world failures live in the long tail; canary catches regressions, not all bugs.
• Don't skip --full for the dead-weight audit. The audit needs the heavy task to detect issues that only manifest in real product builds.
• Don't compare runs across different model versions without noting it in the delta table — model upgrade is a confounding variable.
• Don't auto-rerun on FAIL. Investigate FAILs manually; rerun only after a code change.

References

Start

For run: scripts/run-canary.sh [--full]. For compare: scripts/compare-runs.sh <run-id-a> <run-id-b>.

Memory Write

After each completed benchmark run, append the baseline to .meowkit/telemetry/cost-log.json (top-level array). Create the file with [] if it does not exist.

{"run_id": "{id}", "date": "{ISO-date}", "tier": "quick|full", "pass_rate": N, "avg_score": N, "total_cost_usd": N}

Use mkdir -p .meowkit/telemetry before the append. This persists baselines for compare-runs.sh and the dead-weight audit.