Pbench runner

Skill Yaphet2015/ya-skills/skills/pbench-runner

AI agent skill catalog and yk CLI for installing reusable Claude Code/Codex-ready skills into repos and running workflow tools like PBench, transcripts, and A-share data.

Install

npx -y skills add Yaphet2015/ya-skills --skill pbench-runner

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when a pbench benchmark worktree contains .pbench/run.json and the agent must complete the public replay task before triggering final validation.

SKILL.md

1.9 KB, 387 tokens by cl100k_base, as published. Nobody here has run it

PBench Runner

Use this skill only inside a worktree prepared by yk pbench start.

Workflow

Read .pbench/public/prompt.md, .pbench/public/replay.md, and .pbench/case.public.json.
Complete the benchmark task in the current repository worktree.
When finished, read .pbench/run.json ONLY to obtain its finishCommand, then run that command once.
Report only the finish result and any public work you changed.

Integrity boundaries

The harness (the operator) prepared this worktree and runs the private validators outside it; your only job is the public task plus the one-shot finish. You are being measured, so do not seek information a normal task-solving agent would not have.

Do NOT read, search, grep, or open any of the following — they are out of scope and reading them invalidates the measurement:

Evaluator evidence, validator scripts, and raw session transcripts (the non-public side of the case).
The private-evidence and case-directory environment variables and anything they point at.
The original captured case directory or any captured session material.
The benchmark run-artifacts directory and any prior or current run's outputs (summaries, metrics, validator outcomes, agent logs).
The pbench harness implementation and these skill definitions.

Single attempt: the benchmark is one-shot. Run the finishCommand at most once. If it fails, the run is terminal — report the failure; do not retry, and do not start another run for the same case to improve on it.

Access audit: for every file you open, append one JSON line to .pbench/access-audit.jsonl of the form {"path":"<path>","at":"<iso-8601>"}. This voluntary log is reviewed afterward for sensitive reads; skipping it does not hide reads.