benchmark — Deterministic Agent Evaluation

This package exposes a consistent MCP workflow for deterministic benchmark-driven agent evaluation.

The rebuilt core is organized around challenge catalogs, typed task assets, and explicit checker kinds so coding, web, and OS-style tasks can share one MCP surface without relying on LLM judges.

Tools

| Tool | Purpose | |------|---------| | list_challenges | List deterministic benchmark suites with family, runner, and checker metadata | | get_catalog_status | Inspect catalog source configuration, cache state, and availability | | sync_catalog | Fetch and cache the remote benchmark catalog when a URL source is configured | | start_challenge | Start an attempt and return the first task | | submit_solution | Grade one task and return checker evidence plus the next task or final score | | get_asset | Read an attached benchmark asset by asset_id | | get_attempt | Inspect attempt status, current task, and paginated evaluation history | | cancel_attempt | Cancel an active attempt |

Workflow

1. Call list_challenges
2. Pick a challenge_id
3. Call start_challenge
4. If the task has assets, call get_asset
5. Call submit_solution
6. Repeat until done: true

Each response includes machine-readable guidance for the most likely next tool call.

Fallback benchmark families

• Code — deterministic JSON/text answers backed by explicit checkers
• Web — DOM snapshot extraction tasks with JSON field assertions
• OS — filesystem/log review tasks with deterministic text grading

Zero-config run

Run with the bundled fallback catalog and no extra configuration:

npx @fre4x/benchmark

Or from this repo:

cd /home/fritzprix/my_works/b1te
npm run inspector -w @fre4x/benchmark

Mock Mode

Run with the same bundled fallback catalog in mock mode:

MOCK=true npx @fre4x/benchmark

Optional environment

BENCHMARK_CATALOG_FILE=/absolute/path/to/benchmark-catalog.json
BENCHMARK_CATALOG_URL=https://example.com/benchmark-catalog.json
BENCHMARK_CACHE_DIR=/absolute/path/to/catalog-cache
BENCHMARK_CACHE_TTL_SECONDS=3600
BENCHMARK_STATE_DIR=/absolute/path/to/store-attempt-json
BENCHMARK_MOCK=true

• BENCHMARK_CATALOG_FILE: Optional JSON file with challenge definitions in the rebuilt deterministic catalog format
• BENCHMARK_CATALOG_URL: Optional remote JSON catalog URL for fetch/cache based ingestion
• BENCHMARK_CACHE_DIR: Optional cache directory for remote catalog snapshots
• BENCHMARK_CACHE_TTL_SECONDS: Freshness window for remote catalog cache reuse
• BENCHMARK_STATE_DIR: Where attempt files and lock directories are persisted
• BENCHMARK_MOCK: Alternate mock-mode flag

BENCHMARK_GAIA_DATA_FILE is still accepted as a backward-compatible alias, but the rebuilt package is no longer GAIA-first.

When BENCHMARK_CATALOG_URL is set, the package will reuse a fresh cached copy when available and can be explicitly refreshed with sync_catalog.

Catalog shape

External catalogs must be a JSON array of challenge definitions shaped like:

[
  {
    "challenge_id": "custom_suite",
    "benchmark_id": "custom",
    "family": "code",
    "runner_kind": "code_runner",
    "title": "Custom Challenge",
    "description": "Deterministic single-task suite",
    "version": "v1",
    "source": "external",
    "tasks": [
      {
        "task_id": "custom-1",
        "title": "Return yes",
        "prompt": "Return only yes.",
        "response_format": "text",
        "difficulty": 1,
        "assets": [],
        "checkers": [
          {
            "checker_id": "custom-yes",
            "kind": "exact_text",
            "expected": "yes"
          }
        ]
      }
    ]
  }
]

Supported checker kinds today:

• exact_text
• normalized_text
• regex_match
• contains_all_text
• json_field_equals
• runner_fact_equals
• runner_log_contains_text

Each submission now also records runner execution metadata:

• workspace directory
• materialized asset and submission artifacts
• runner facts
• runner logs

Example external catalog

The repo includes the bundled fallback catalog source at:

benchmark/catalogs/expanded-catalog.json

Use it like this:

cd /home/fritzprix/my_works/b1te
BENCHMARK_CATALOG_FILE=/home/fritzprix/my_works/b1te/benchmark/catalogs/expanded-catalog.json npm run inspector -w @fre4x/benchmark

If no catalog env is provided at runtime, the published package falls back to the bundled copy of this catalog automatically.

Claude Desktop

{
  "mcpServers": {
    "benchmark": {
        "command": "npx",
        "args": ["-y", "@fre4x/benchmark"],
        "env": {
          "BENCHMARK_CATALOG_URL": "https://example.com/benchmark-catalog.json",
          "BENCHMARK_CACHE_DIR": "/absolute/path/to/benchmark-cache"
        }
      }
    }
}

Development

npm install
npm run build -w @fre4x/benchmark
npm run typecheck -w @fre4x/benchmark
npm test -w @fre4x/benchmark
MOCK=true npm run inspector -w @fre4x/benchmark