pi-evals

name: pi-evals description: Write and run evals for pi extensions and agent behavior using @aliou/pi-evals. Use when creating eval files, writing custom scorers, configuring eval runs, or testing that pi extensions work correctly.

pi-evals

Eval framework for testing pi coding agent behavior. Runs prompts against pi via createAgentSession, then scores the results.

Quick Start

Install:

pnpm add -D @aliou/pi-evals

Create pi-evals.config.ts at the project root:

import { defineConfig } from "@aliou/pi-evals";

export default defineConfig({
  defaults: {
    model: "claude-haiku-4-5",
    provider: "anthropic",
  },
  evalsDir: "./evals",
  timeout: 60_000,
});

Create an eval file in evals/:

// evals/hello.eval.ts
import { evaluate, Scorers } from "@aliou/pi-evals";

evaluate("Create hello file", {
  config: {
    model: "claude-haiku-4-5",
    provider: "anthropic",
  },
  data: [
    {
      input: 'Create a file called hello.txt containing "Hello World"',
      expected: { files: { "hello.txt": "Hello World" } },
    },
  ],
  scorers: [Scorers.files()],
  timeout: 30_000,
});

Run:

pnpm pi-evals              # all evals
pnpm pi-evals --filter "hello"  # by name substring

Eval File Structure

Eval files are *.eval.ts files in the configured evalsDir. Each calls evaluate() to register one eval.

evaluate("Eval name", {
  config: { model, provider, extensions?, env? },
  data: [{ input, expected?, setup?, timeout? }],
  scorers: [...],
  timeout?: number,
});

Test Cases (data)

Each test case runs in an isolated temp directory.

• input: prompt sent to the agent
• expected: optional expected outcome (used by scorers)
• setup.files: files to pre-create in the workspace ({ "path": "content" })
• setup.commands: shell commands to run before the eval
• timeout: override timeout for this case

Config (config)

• model: model name (e.g. "claude-haiku-4-5")
• provider: provider name (e.g. "anthropic", "github-models")
• extensions: array of extension paths, resolved relative to process.cwd()
• env: environment variables to set

For GitHub Models, use:

config: {
  provider: "github-models",
  model: "gpt-4o",
}

github-models is bundled in pi-evals and auto-loaded.

Built-in Scorers

All scorers are accessed via Scorers.*:

Custom Scorers

A scorer is an object with name and score(ctx) => ScoreResult:

import type { Scorer } from "@aliou/pi-evals";

const myScorer: Scorer = {
  name: "my_scorer",
  async score(ctx) {
    // ctx.input      - the prompt
    // ctx.output     - agent's final text response
    // ctx.cwd        - workspace directory
    // ctx.toolCalls  - array of { name, args }
    // ctx.messages   - full conversation
    // ctx.expected   - the expected object from the test case
    // ctx.stats      - { tokens: { input, output, total }, cost }
    return {
      name: "my_scorer",
      score: 1,  // 0 to 1, >= 0.5 passes
      reason: "Looks good",
    };
  },
};

Testing Extensions

Pass extension paths in config.extensions. Paths resolve relative to process.cwd() (the project root), not the temp workspace.

evaluate("My extension eval", {
  config: {
    model: "claude-haiku-4-5",
    provider: "anthropic",
    extensions: ["./extensions/my-ext/index.ts"],
  },
  data: [
    { input: "Use the custom tool provided by my extension." },
  ],
  scorers: [Scorers.toolCalled("my_custom_tool")],
});

CLI Options

-f, --filter <pattern>   Filter evals by name substring
-t, --threshold <pct>    Minimum pass percentage to exit 0
-c, --config <path>      Config file path (default: pi-evals.config.ts)
-m, --model <model>      Override model (env: PI_EVAL_MODEL)
-p, --provider <name>    Override provider (env: PI_EVAL_PROVIDER)
-v, --verbose            Detailed output
    --json               Output results as JSON

Reusable GitHub Workflow

Other repos can run evals via reusable workflow:

jobs:
  evals:
    uses: aliou/pi-evals/.github/workflows/pi-evals.yml@main
    permissions:
      contents: read
      models: read
    secrets: inherit
    with:
      package-manager: npm
      install-command: npm ci
      eval-command: npx pi-evals --json

For pnpm repos, switch package-manager and install-command.

GitHub Action

Other repos can also call the composite action directly:

jobs:
  evals:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      models: read
    steps:
      - uses: actions/checkout@v4
      - uses: aliou/pi-evals@vX.Y.Z
        with:
          package-manager: npm
          install-command: npm ci
          eval-command: npx pi-evals --json
          github-token: ${{ secrets.GITHUB_TOKEN }}

Replace vX.Y.Z with the package version you want to pin.

Session Behavior

Each eval test case runs in an isolated temp directory. Sessions use in-memory storage and are not persisted to the user's session directory.