v0.31.0

• Coding-agent integration (issue #15): two new commands wire perfscale into
  AI coding agents. perfscale init-skills [--loop claude|agents] [-g]
  installs embedded skill folders (perfscale-cli with YAML/actions/
  load-shapes/thresholds/metrics/reporters/protocols references,
  perfscale-library-sdk, perfscale-migration-k6) into ./.claude/skills,
  ./.agents/skills, or the home directory with -g; re-running overwrites.
  perfscale init-agents [--loop claude|codex|copilot|opencode|vscode| vscode-legacy] [-c config] [--prompts] (default loop: claude) generates
  the perfscale-test-{planner,generator,healer} agent definitions per loop,
  registers the @perfscale/mcp server in the loop's MCP config
  (npx -y @perfscale/mcp), and initializes the repo: specs/README.md, a
  found-or-created tests/seed.yaml, and a minimal config. --prompts
  installs the perfscale-{plan,generate,heal,threshold} prompt templates —
  the threshold prompt replaces Playwright's coverage with a HOLD /
  NOT HOLD verdict at a confirmed full-load run. Agents verify with small
  1-VU runs by default, and the healer never edits thresholds:. See
  docs/agents.md.

• Streaming reporters (issue #12): report.reporters can now stream live,
  cumulative metrics during the run to pluggable sinks — std/report/serve@v1
  (the historical report.url path as a reporter), std/report/prometheus-rw@v1
  (Snappy + protobuf remote write; classic _bucket{le} histograms by default,
  native histograms opt-in via with.histogram: native) and
  std/report/otel@v1 (OTLP/gRPC with cumulative exponential histograms).
  Engine-owned keys sit next to use:: during_run, interval_ms, labels
  and max_series (default 10000, deterministic drop counted in
  perfscale_report_dropped_series_total{reporter}); max_samples moved out
  of with: (the legacy key still reads). Labels merge step > reporter > run,
  interpolate ${{ inputs.* }} / ${{ env.* }} once before the run, and steps
  accept their own labels:; run_id/worker are engine-set and rejected in
  user config. No client-side quantile series in the new exporters — backends
  compute correct cross-worker quantiles from the buckets
  (histogram_quantile(0.95, sum by (le) (rate(http_req_duration_bucket[1m])))).
  Each reporter gets an independent bounded queue (with.max_pending), batched
  delivery (with.batch_size, never exceeded — oversized snapshots split in
  order), retry with exponential backoff (with.retry, Retry-After
  honored, other 4xx poisoned) and a synchronous final flush bounded by
  with.final_timeout (default 30s) that coalesces the backlog to the
  terminal cumulative snapshot; overflow drops the OLDEST snapshot (counted in
  perfscale_report_dropped_snapshots_total{reporter}), and a slow or dead
  backend can never block the VUs, another reporter, or fail the run. The
  delivery counters are per-run, registered as the report metric family and
  exported to each backend as perfscale_report_*_total{reporter="…"}. A during_run: true on a final-only reporter (json/junit/html/wasm) is
  now a lint/run error.

• Benchmarks: a renaissance suite (in the default suites and the daily CI
  run) executes the official Renaissance MIT v0.16.2 jar
  (renaissance-mit-0.16.2.jar, sha256-pinned and cached by CI under
  target/renaissance/) via scripts/bench_renaissance.sh — default
  workload finagle-http (RENAISSANCE_BENCHMARKS), bounded to
  RENAISSANCE_RUN_SECONDS (default 60) measured seconds per benchmark
  through the harness's own --run-seconds. The report and results JSON
  record benchmark operation throughput (ops/s, never HTTP RPS) and
  per-operation time plus CPU/RSS/duration; a missing/invalid jar, a Java
  failure or empty results fail the run rather than reporting fake numbers.
  Console log and native JSON are preserved under renaissance-artifacts/
  and uploaded with the bench artifact; bench_compare.py compares only
  same-benchmark rows.

• Benchmarks: new opt-in specjbb suite (SUITES=specjbb ./scripts/bench.sh)
  runs a caller-supplied commercial SPECjbb2015 kit (SPECJBB_HOME /
  SPECJBB_JAR, plus SPECJBB_JAVA / SPECJBB_MODE, default COMPOSITE)
  via the official java -jar specjbb2015.jar -m <MODE> — never bundled,
  never downloaded, never shortened — and reports native max-jOPS /
  critical-jOPS (Java operations/s, not HTTP RPS) with mode, duration, CPU
  and peak RSS in the report and results JSON. It is excluded from the
  default suite list and CI; with an invalid kit path the bench fails loudly
  instead of reporting a fake success. Console log and the SPEC report tree
  are preserved as artifacts.

• Pub/Sub drivers can report asynchronous delivery failures through
  PubSubOutcome.errors. The engine adds them to the synchronous failure flag
  in pubsub_errors, so Kafka delivery errors, MQTT ack timeouts and Redis
  pipeline error replies remain visible with flush: none; the counter can
  exceed 1 per exchange without failing the step.

• k6-style mid-run threshold abort plus failure diagnostics: an after:
std/thresholds@v1 gate with abort_on_fail: true is now monitored while
  the VUs run (re-evaluated every 250 ms against the metrics collected so
  far), and the first violated severity: fail gate stops all load schedules
  early — between steps, like Ctrl-C — then runs the usual after: teardown,
  reporters and summary (the after: evaluation stays the single recorded
  gate result). The monitor starts only after a successful before: and never
  records intermediate results; a normal finish or Ctrl-C prints no abort
  line. Adding http_debug: error (valid only with abort_on_fail: true;
  lint/run reject other combinations) emits one secret-masked
  http_debug: {…} JSON line per failed step action and failed check:,
  carrying the VU id, iteration, interpolated request params (payload
  included) and the response output (body included) — k6 #6618-style
  diagnostics with zero output for successful exchanges.

What's Changed

• fix(release): idempotent npm publish, re-runnable for an existing tag by @vitalics in #28
• feat: streaming metric reporters and JVM benchmark suites by @vitalics in #29
• feat: report asynchronous pubsub errors by @vitalics in #30
• feat: add agent and skill initialization commands by @vitalics in #32
• feat: add aborting thresholds and failure diagnostics by @vitalics in #31
• chore(release): v0.31.0 by @vitalics in #33

Full Changelog: v0.30.0...v0.31.0