@razroo/iso-eval

Behavioral eval runner for AI coding agents.

agentmd lints prompt structure, isolint lints prompt prose, iso-harness fans out the compiled source into every harness file layout. None of them answer the next question: did the agent actually do the task? That's what @razroo/iso-eval scores.

You give it a suite of tasks — each with a baseline workspace, a prompt, and a set of checks — and it snapshots the workspace per trial, hands it to a runner, then verifies the resulting filesystem / command state against your checks.

Built-in runners today:

• fake — deterministic CI/offline runner that executes $ ... lines from the prompt as shell in the snapshotted workspace.
• codex — real-agent runner that shells out to codex exec in the per-trial workspace and captures the final assistant message.
• claude-code — real-agent runner that shells out to claude -p in the per-trial workspace.
• cursor — real-agent runner that shells out to cursor-agent --print in the per-trial workspace.
• opencode — real-agent runner that shells out to opencode run in the per-trial workspace.

The library API still accepts any RunnerFn, so you can plug in other harnesses without waiting on a packaged runner.

Install

npm install -D @razroo/iso-eval

Suite shape

# eval.yml
suite: refactor-basic
runner: fake              # fake | codex | claude-code | cursor | opencode
timeoutMs: 120000
harness:
  source: ../dist         # optional: stage generated harness files into each trial

tasks:
  - id: write-greeting
    prompt: tasks/write-greeting.md    # path (relative to eval.yml) or inline
    workspace: workspace/              # baseline dir, copied per-trial into tmpdir
    trials: 1
    checks:
      - { type: file_exists,       path: greeting.txt }
      - { type: file_contains,     path: greeting.txt, value: "hello" }
      - { type: file_not_contains, path: greeting.txt, value: "TODO"  }
      - { type: command, run: "test -f greeting.txt", expectExit: 0 }

Supported checks

| type | asserts | | --------------------- | ---------------------------------------------------------------- | | command | shell command exits with expectExit (default 0); optional stdout contains/matches | | file_exists | file at path exists in the workspace | | file_contains | file at path contains the literal substring value | | file_not_contains | file at path does NOT contain value | | file_matches | file at path matches the regex matches | | llm_judge | a user-supplied JudgeFn answers yes to prompt against runner stdout/stderr | | agentmd_adherence | per-rule pass rate from agentmd test meets minPassRate; optional ruleId filter |

agentmd_adherence

- type: agentmd_adherence
  promptFile: ../agent.md         # path to agentmd source (relative to eval.yml)
  fixtures: ../fixtures.yml       # path to agentmd fixture file
  ruleId: H3                      # optional — score only this rule
  minPassRate: 0.9                # required — pass rate floor in [0, 1]
  via: claude-code                # optional — default claude-code (api | claude-code | fake)
  model: claude-haiku-4-5         # optional — forwarded as --model
  timeoutMs: 180000               # optional — subprocess timeout

Shells out to the agentmd CLI (bundled as a runtime dependency) via agentmd test <promptFile> --fixtures <fixtures> --format json, parses the per-rule check outcomes, computes the pass rate for ruleId (or overall if omitted), and fails the check when the rate is below minPassRate. Tests can inject a fake subprocess runner via the library API (AgentmdSpawnFn) so CI doesn't need an API key.

CLI

iso-eval run  examples/suites/echo-basic/eval.yml
iso-eval plan examples/suites/echo-basic/eval.yml

iso-eval run eval.yml --filter write-greeting --concurrency 2 --json
iso-eval run eval.yml --runner claude-code --harness-source ../dist
iso-eval run eval.yml --runner cursor --harness-source ../dist
iso-eval run eval.yml --runner opencode --harness-source ../dist
iso-eval run eval.yml --keep-workspaces           # skip tmpdir cleanup for debugging

run exits 0 on all-pass, 1 on any failure, 2 on invalid invocation.

--runner and --harness-source let you replay the same suite through a different packaged harness without rewriting checks.yml.

Real runners and harness staging

Set runner: in YAML, or override it at the CLI with --runner. harness.source is optional; when present, iso-eval stages the generated harness files you want the runner to see into each snapshotted workspace.

codex

suite: refactor-basic
runner: codex
timeoutMs: 180000
harness:
  source: ../dist

Accepted harness.source shapes:

• a project directory containing AGENTS.md and/or .codex/
• a direct AGENTS.md path
• a direct .codex/config.toml path

claude-code

Accepted harness.source shapes:

• a project directory containing CLAUDE.md, .claude/, and/or .mcp.json
• a direct CLAUDE.md path
• a direct .claude/ path
• a direct .claude/settings.json path
• a direct .mcp.json path

The runner shells out to claude -p --no-session-persistence and passes .mcp.json through --mcp-config when present.

opencode

Accepted harness.source shapes:

• a project directory containing AGENTS.md, opencode.json, and/or .opencode/
• a direct AGENTS.md path
• a direct opencode.json path
• a direct .opencode/ path

The runner shells out to opencode run --dir <workspace> and defaults to --pure so each trial stays self-contained.

cursor

Accepted harness.source shapes:

• a project directory containing .cursor/, AGENTS.md, and/or CLAUDE.md
• a direct .cursor/ path
• a direct .cursor/rules/ path
• a direct .cursor/rules/*.mdc path
• a direct .cursor/mcp.json path
• a direct AGENTS.md path
• a direct CLAUDE.md path

The runner shells out to cursor-agent --print --output-format text --workspace <workspace> and stages any Cursor harness files you exported with iso-harness into the per-trial workspace first.

This lets one suite exported from iso-trace be replayed across the packaged runners with the same task prompt and checks.

Library API

import { loadSuite, run, formatReport, fakeRunner } from "@razroo/iso-eval";

const suite = loadSuite("./eval.yml");
const report = await run(suite, {
  runner: fakeRunner,
  concurrency: 2,
  onTaskComplete: (t) => console.log(t.id, t.passed ? "✓" : "✗"),
});
console.log(formatReport(report));
process.exit(report.passed ? 0 : 1);

Bring your own runner

The YAML runner: field selects from shipped runners; the library accepts any RunnerFn:

import type { RunnerFn } from "@razroo/iso-eval";

const myRunner: RunnerFn = async ({ workspaceDir, taskPrompt, timeoutMs, harnessSource }) => {
  // spawn your agent (claude -p / codex exec / …) with cwd = workspaceDir
  // optionally stage files from harnessSource before invoking it
  // return { exitCode, stdout, stderr, durationMs }
};

Bring your own judge (for llm_judge checks)

import type { JudgeFn } from "@razroo/iso-eval";

const judge: JudgeFn = async (prompt, output) => {
  // call your model; return true if the rule was followed
};

await run(suite, { runner: fakeRunner, judge });

How this fits the rest of the pipeline

agent.md  →  agentmd lint  →  agentmd render  →  isolint lint  →  iso-harness build
                                                                         │
                                                                         ▼
                                                          project w/ CLAUDE.md etc.
                                                                         │  iso-eval run
                                                                         ▼
                                                                per-task pass / fail

• @razroo/agentmd measures per-rule adherence on text output (input string → output string → check).
• @razroo/iso-eval measures task success on a real workspace (snapshot dir → agent acts → filesystem state → check).

The two compose: an iso-eval suite can include llm_judge checks that reuse the same judge convention (yes = rule followed), plus agentmd_adherence checks that fold a fixture-level adherence score into the task report.

License

MIT — see LICENSE.