AGENT SKILL

codespine-workload

Benchmark a program in a controlled environment and attribute the cost with the code graph. Use when asked to profile under a CPU/memory cap, measure server capacity ("how many requests per second before p99/latency breaks an SLO"), decide "when do I need more servers", find where runtime is spent under load, or run a workload in a constrained "one box" container. Two modes: cpu-profile (where the time goes) and loadtest (how much load it takes). Requires the codespine CLI; Docker is optional (falls back to running on the host).

License 许可
MIT
最近更新
2026-06-29
Stars
★ 7
Forks
0

01 前言 Front Matter

name
codespine-workload
description
Benchmark a program in a controlled environment and attribute the cost with the code graph. Use when asked to profile under a CPU/memory cap, measure server capacity ("how many requests per second before p99/latency breaks an SLO"), decide "when do I need more servers", find where runtime is spent under load, or run a workload in a constrained "one box" container. Two modes: cpu-profile (where the time goes) and loadtest (how much load it takes). Requires the codespine CLI; Docker is optional (falls back to running on the host).

02 资源文件 4

03 SKILL.md 正文

codespine-workload

Exercise a program under a chosen environment, then attribute the cost back onto the code graph. This turns vague runtime questions ("why is it slow?", "how many requests can it take?", "when do I add a server?") into measured, graph-grounded answers.

The capability is one idea in three parts — workload kind × environment × attribution:

  • kind — cpu-profile (loop the hot path) or loadtest (ramp a running server).
  • environment — --host (uncapped baseline) or a container under an enforced --cpus/--memory cap (a realistic "one server" box).
  • attribution — cpu-profile → enrich → hotspots/cost (graph-grounded where the time goes); loadtest → a latency-vs-throughput curve and the knee (how many requests/second before the SLO breaks).

When to use this skill

  • "Where is the time spent / why is this slow under load?" → cpu-profile.
  • "How many requests/second can my server handle before p99 crosses my SLO?" → loadtest.
  • "When do I need another server / how do I size capacity?" → loadtest.
  • "Will it hold up on a small box (0.5 CPU / 512 MB)?" → either mode, docker environment.
  • The agent is mid-task in /codespine-interview or /codespine-optimize and hits a needs-workload situation (it must measure a runtime claim, not guess) → use this.

For deciding which symbol matters from an existing profile, use the codespine-query skill (hotspots, cost). This skill is about producing the measurement in a controlled environment in the first place.

Realism vs. determinism — the honesty rule

There are two tracks, and they answer different questions. Keep them separate:

  • Determinism track (the benchmark command): low-variance before/after deltas for "did my edit make X faster?". Stable median, advisory.
  • Realism track (this skill, under a cap): "will it hold up on a constrained box?". It intentionally introduces a cap and accepts scheduler noise. These numbers are a capacity estimate, not a deterministic gate. Never diff two realism runs as a pass/fail, and never use a realism number to approve an optimization — that is what benchmark/verify are for.

State this caveat whenever you report realism-track numbers. See reference/two-track.md for the full framing.

Prerequisites

  • The codespine CLI (codespine …, or npx codespine …).
  • For cpu-profile attribution: a built graph at ./.codespine/graph.kuzu. If it is missing, build it once: codespine extract . --semantic then codespine load.
  • For the docker environment: a container runtime (Docker Desktop / OrbStack / podman). If absent, use --host.
  • For loadtest: the server must be runnable locally and expose a readiness endpoint (something that returns 2xx when ready).

Where artifacts live, and what this skill ships

All per-project artifacts go in the user's repo under ./.codespine/workload/ (gitignorable) — never in the codespine package. This skill ships starting-point templates next to this file under templates/:

Run codespine workload scaffold --kind cpu-profile|loadtest|both to copy the relevant template(s) into ./.codespine/workload/, then fill in the // EDIT FOR YOUR PROJECT block. The rest is generic.

Mode A — cpu-profile: where does the time go?

  1. Build the graph if needed: codespine extract . --semantic && codespine load.
  2. Scaffold: codespine workload scaffold --kind cpu-profile writes the driver + Dockerfile to ./.codespine/workload/.
  3. Author the driver: in the EDIT block of ./.codespine/workload/cpu_profile_driver.ts, import the project's hot module(s) and exercise the real public API in a deterministic loop (no timing, just work). It must live outside the extracted source root so it never becomes a graph node; imports are module-relative.
  4. Run it — one command profiles, enriches the graph, and prints the top self time. On the host (simplest; uses the project's existing node_modules):
    codespine workload run --driver ./.codespine/workload/cpu_profile_driver.ts --root .
    
    …or under an enforced cap for the "constrained box" answer — --project is the directory mounted at /work, and the driver must live inside it:
    codespine workload run --driver ./.codespine/workload/cpu_profile_driver.ts --root . \
      --docker --cpus 0.5 --memory 512m --project .
    
    --root must be the directory you passed to extract (usually ., or a subdirectory like ./src if you extracted one) — otherwise the profile's frame paths will not resolve onto the graph and most samples come back "unattributed".
  5. Read the ranked views: codespine hotspots --by self-time (the leaves where CPU burns) and codespine cost (which symbols the time is spent under). Report the top frames, tie them to source, and — if asked — to optimization tasks.

Mode B — loadtest: how much load can it take?

  1. Learn the server: find the entrypoint (how it boots, what port), the readiness endpoint, and the routes + their request contracts (read the route handlers). For a write path, find the durability config (does it fsync?).
  2. Scaffold + author: copy templates/loadtest_driver.template.ts to ./.codespine/workload/loadtest_driver.ts and edit the marked block: how to start the server, the readiness path, and buildRequests() (your endpoint mix). A filled example for a typical REST API is in the template's comments.
  3. Run — host baseline first, then the cap:
    # host (uncapped upper bound):
    SERVER_ENTRY=src/main.ts PORT=3000 DB_PATH=/tmp/app.loadtest.db \
      LOADTEST_PROFILE=mixed LOADTEST_SLO_P99_MS=200 \
      node --import tsx .codespine/workload/loadtest_driver.ts
    
    # docker (one constrained "server" box — realism):
    docker build -t codespine-workload-runner .codespine/workload
    docker run --rm --cpus 0.5 --memory 512m --memory-swap 512m \
      -v "$PWD:/work:ro" -v "$PWD/.codespine/workload:/opt/runner/wl" \
      -e SERVER_ENTRY=/work/src/main.ts -e SERVER_CWD=/opt/runner \
      -e PORT=3000 -e DB_PATH=/tmp/app.loadtest.db \
      -e LOADTEST_PROFILE=mixed -e LOADTEST_SLO_P99_MS=200 \
      codespine-workload-runner \
      node --import tsx /opt/runner/wl/loadtest_driver.ts
    
  4. Read the ramp: the driver prints a table (offered rps → p50/p99) and the knee. Capacity = the highest sustained rps with p99 under the SLO. "Time to add a server" = steady-state load approaching that number.
  5. Critical: use a file-backed datastore for any write-inclusive profile (set DB_PATH/equivalent to a file). With an in-memory store, fsync is a no-op and write capacity is fictitious.

Choosing host vs. docker

  • Host — uncapped, uses the project's own node_modules, no container friction. The easy starting point and the upper-bound baseline. A single-threaded server effectively uses ~one core regardless of host core count.
  • Docker — an enforced cgroup cap is the only way to get a faithful "one server" box (macOS has no native per-process hard cap). Use it for the realism answer. Co-locate the load client in the container against 127.0.0.1 (avoids the host↔VM network-latency floor; the trade-off is the client shares the CPU cap). If the project has runtime dependencies with native addons, the container needs a Linux build of them (a macOS-arm64 node_modules will not run in the Linux VM) — provision them into a volume or install in the image. When that is painful, prefer host mode.

Run both and compare: the gap tells you how much of the slowness is the cap (cgroup throttling + a squeezed page cache) versus the program itself.

Interpreting and reporting

  • Always attach the realism caveat to capped numbers (noise; estimate not gate).
  • At low offered rates, p99 is sample-thin (few requests) — lengthen the step duration for a stable tail, and say so.
  • A sharp knee (throughput plateaus while latency explodes) usually means a serial bottleneck (e.g. a synchronous DB on one event loop) — more cores will not help a single-threaded server; "more servers" means more processes/replicas.
  • Tie findings to the graph: cpu-profile → name the hot symbols from hotspots; for a slow endpoint, read its handler → service → datastore path.

Limitations / when to escalate

This MVP assumes the workload is runnable via node/tsx. Projects needing a build step, a non-npm package manager, non-HTTP load, custom readiness, or auth need the driver/run-command adapted by hand — note that to the user rather than reporting a bogus number. For cpu-profile, codespine workload run (host and --docker) and codespine workload scaffold already wrap this orchestration; the loadtest kind is run from the scaffolded driver directly for now. The cpu-profile --docker path auto-provisions a Linux build of the project's dependencies into a cached volume (keyed by the lockfile); the loadtest kind under --docker is not wired yet (use host).

04 同仓库更多 Skills