菜单
Use when tasks involve tests/benchmark/, sandbox burst load tests, the treadstone CLI against a remote API, scenario YAML under tests/benchmark/scenarios/, or artifacts run.json / events.jsonl / summary.json; also when debugging 429s, poll_timeout, ready stalls, auth failures, or leftover loadtest sandboxes.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/earayu/treadstone --skill benchmark命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | benchmark |
| description | Use when tasks involve tests/benchmark/, sandbox burst load tests, the treadstone CLI against a remote API, scenario YAML under tests/benchmark/scenarios/, or artifacts run.json / events.jsonl / summary.json; also when debugging 429s, poll_timeout, ready stalls, auth failures, or leftover loadtest sandboxes. |
tests/benchmark/ is a standalone Python tool (not pytest) that shells out to the treadstone CLI (--json) to stress sandbox create → poll-until-ready → delete. Each run tags sandboxes with a deterministic label loadtest:run-<run_id>.
It exists for manual or scripted capacity experiments. Do not add it to make test, CI, or PR gates unless the team explicitly decides otherwise.
benchmark.config.yaml.example → tests/benchmark/benchmark.config.yaml (gitignored).lib/binary.py: --binary / TREADSTONE_BINARY → cache tests/benchmark/.bin/ → else install.sh.python tests/benchmark/setup.py — binary + auth whoami (fail fast).python tests/benchmark/runner.py <scenario.yaml> — create → poll → cleanup (unless --no-cleanup).| Action | Command |
|---|---|
| Preflight | python tests/benchmark/setup.py |
| Run | python tests/benchmark/runner.py tests/benchmark/scenarios/pilot_stateless.yaml |
| Plan only | same + --dry-run (matrix expansion, no API calls) |
| Skip deletes | --no-cleanup (sandboxes keep label loadtest:run-<run_id>) |
| Pin binary | --binary PATH or TREADSTONE_BINARY=PATH |
| Pin server | --base-url URL or TREADSTONE_BASE_URL |
base_url resolution (highest wins): --base-url → TREADSTONE_BASE_URL → benchmark.config.yaml → ~/.config/treadstone/config.toml (CLI default).
binary_version resolution: --version → benchmark.config.yaml → "latest".
Validated by runner.load_scenario:
| Field | Rule |
|---|---|
name | Required string |
matrix | Non-empty list; each row needs template, integer count ≥ 1, optional persist (bool). If persist: true, storage_size is required |
concurrency | Optional; default 10 (create + poll pool size) |
ready_timeout_sec | Optional; default 900 |
poll_interval_sec | Optional; default 10 |
auto_stop_interval / auto_delete_interval | Passed through to sandboxes create (defaults 60 / -1) |
The parser is a minimal subset of YAML (no anchors, no complex merges) — keep scenarios simple.
Matrix expansion: each row’s count is repeated; every instance becomes one CreateTask with label loadtest:run-<run_id>.
Cleanup (matches runner.py + executor.py): After poll, cleanup runs unless --no-cleanup. On Ctrl+C, cleanup is skipped only when cleanup.on_failure is false (default true); when true, interrupt still honors --no-cleanup. cleanup.after_run in YAML is not read by the runner.
Deletes: Same persist on every matrix row → persist_filter deletes only that kind; mixed persist rows → delete all with the run label. Cleanup tries sandboxes list --label; on failure, full list (limit 1000) + client-side match on loadtest:<run_id>.
Each run writes under tests/benchmark/results/<run_id>/ (gitignored):
| File | Contents |
|---|---|
run.json | Scenario snapshot, run_label, binary path, base_url, treadstone_version, machine, timestamps |
events.jsonl | One JSON object per line: create_*, ready, poll_timeout, failed_status, delete_* |
summary.json | Aggregates from Reporter (rates, latencies, error counts) — read this for pass/fail summaries |
Large bursts on prod need approval. Do not commit results/, .bin/, or secret config. Auth failures: fix CLI login — no bypass.
| Symptom | Action |
|---|---|
| Auth exit | treadstone auth login or API key per CLI |
429 / create_err | Lower concurrency; check org limits |
poll_timeout | Raise ready_timeout_sec or reduce load; check cluster readiness |
| Stray sandboxes | Manual sandboxes list + filter loadtest; fallback may miss if list truncated |
| YAML errors | Simple YAML only; no anchors |
Stub: tests/benchmark/README.md → this file + AGENTS.md.
First-time Treadstone local development environment setup. Run once after cloning the repo and before starting any development. Covers system dependency installation, Python environment, Neon database connection, migrations, and environment verification. Use this skill when the user/agent just entered the project, needs to rebuild the environment, or encounters setup-related issues like missing dependencies, broken .env, or failed migrations.
Treadstone development lifecycle — feature branches, TDD, ship, PR, CI, merge, GitHub Actions release (workflow_dispatch with version x.y.z), and optional production deploy. Use for any shippable code change, GitHub flow, or release. Includes agreed “codeword” paths (合并代码 / 发版本 / 发生产). Source of truth for agents executing ship, merge, make local / make prod.
Treadstone OpenAPI spec architecture, SDK generation, and sandbox proxy path conventions. Use whenever adding or modifying API routes, changing OpenAPI tags/operationIds, regenerating the Python SDK, updating sandbox runtime paths in docs, or working with treadstone/openapi_spec.py, scripts/export_openapi.py, or scripts/sandbox_openapi_base.json. Also use when someone asks how the OpenAPI docs work, why sandbox routes don't appear in the SDK, or what "three-tier spec" means.