@diegopetrucci/pi-test-runner0.1.0

test-runner

Adds a test_runner subagent tool that adapts The Last Harness test-runner prompt into a standalone pi extension.

Use it to delegate exact validation commands (for example npm ci, then npm test) to an isolated child agent. The subagent runs the steps exactly in order, stops at the first failure, never edits, fixes, installs on its own, or commits, and reports each step as PASS or FAIL plus an overall result. Blockers are reported in the final result; the subagent cannot ask follow-up questions.

Install

Standalone npm package

pi install npm:@diegopetrucci/pi-test-runner

Collection package

pi install npm:@diegopetrucci/pi-extensions

GitHub package

pi install git:github.com/diegopetrucci/pi-extensions

Then reload pi:

/reload

test_runner tool behavior

The tool accepts:

• steps — required, non-empty array of exact shell commands, executed in order
• context — optional extra context such as validation scope or known flaky steps
• cwd — optional working directory for the steps; defaults to the current session directory

When called, it launches an isolated in-memory child agent with:

• no inherited extensions, skills, prompt templates, themes, context files, or agents files;
• the current session model (no automatic model selection and no persisted /config commands);
• low thinking, clamped to the model's supported levels;
• only unrestricted bash plus configured MCP tools (no read, write, edit, grep, find, or ls);
• configured MCP tools (run without confirmation; may have side effects; requires host Pi >=1.0.4; older hosts degrade to no MCP, no error); the global mcp.json always applies, and a project .pi/mcp.json is loaded only when the host session has already trusted the project.

The prompt lists the steps exactly as given, numbered in order, and tells the subagent to stop at the first failure, never edit files, fix failures, install dependencies, update snapshots, or stage/commit/push, and to report the scope, each step with PASS/FAIL, the overall result, and skipped steps. The tool appends a run-details suffix with the model, effective thinking level, turn count, tool calls, duration, working directory, and tools used.

Budgets

Validation can be long, so the budgets are generous: 60 turns, 60 minutes total wall-clock, and a default 30-minute timeout per bash call (applied when the model does not set one). Tool use is disabled on the final turn so the subagent can report.

Guarantees and limitations

• bash is unrestricted and not sandboxed. The steps you pass run with your user's permissions, and the subagent could run other commands. The "never edit/fix/install/commit" rules are prompt guidance, not enforcement; only pass commands you would run yourself. Steps such as npm ci legitimately modify the checkout.
• Only bash and MCP tools are available; other tool calls are blocked. Configured MCP tools run without confirmation and are limited only by prompt guidance.
• Fail-fast and step order are enforced by the prompt, not by a runtime scheduler.
• If the run times out or fails, steps may have been partially executed.

Example

Use test_runner with:

{
  "steps": ["npm ci", "npm run typecheck", "npm test"]
}