prooflisted

# Proof — Stress Test (Proofing a blade = deliberate overload to find where it breaks.) Read `${CLAUDE_PLUGIN_ROOT}/references/creed.md`, `${CLAUDE_PLUGIN_ROOT}/references/stacks.md`, and `${CLAUDE_PLUGIN_ROOT}/references/dispatch.md` first. Log: `bash ${CLAUDE_PLUGIN_ROOT}/scripts/ledger.sh append temper <slug> proof STARTED -` ## Preconditions — all three, or stop 1. **A runnable service.** Run command + target URL from spec.md/STATE.md, or ask. No service → tell the user proof doesn't apply; suggest `/smithy:hone` for non-service performance work. 2. **A safe target.** LOCAL or explicitly user-approved environments only. Never load-test a production URL or any host the user hasn't named in writing this session. 3. **Thresholds from the user.** Never invent SLOs. Offer this default and get explicit confirmation: p99 < 500ms, 0 errors (5xx), at 50 concurrent connections for 60s. The user may change any number. ## Process 1. **Detect tooling** (stacks.md matrix): the load client is language- agnostic — default `npx autocannon` for any HTTP target; locust for Python; k6/wrk/hey/vegeta only if already installed. The per-stack playbook (`references/{ts,python,go,java,rust}.md`) sets the build/run discipline (release builds, JVM warm-up) and the runtime monitoring hooks (pprof, jcmd/JFR, RSS/fd tracking). 2. **Write the test brief** (`briefs/proof.md`): `## Persona`: masters/sre.md (test lens — failure modes + monitoring hooks); tool +