---
license: "mit"
language:
- en
pretty_name: "Tool Errors"
tags:
- benchmark
- llm-evaluation
- ai-benchmark
- model-evaluation
- agent-benchmark
- tool-use
- api-errors
- retries
- idempotency
- fault-injection
- openrouter
- openai
- mcp
- typescript
task_categories:
- text-generation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/tasks.jsonl
---

# Tool Errors

Retry safely when APIs fail.

## What it measures

Safe, bounded recovery from API faults without duplicate side effects or false success reports.

Category: `tools`. Status: `alpha`.

## How to run

```console
$ npx benchmark-tool-errors list
$ npx benchmark-tool-errors run --model <model-id> --provider openrouter
$ npx benchmark-tool-errors score results.json
```

`--provider mock` runs the whole benchmark offline with no keys. Real models run
through OpenRouter (`OPENROUTER_API_KEY`) or any OpenAI-compatible base URL
(`OPENAI_BASE_URL`, `OPENAI_API_KEY`).

## Data fields

`data/tasks.jsonl` holds 40 tasks, one JSON object per line
(kinds: 40 episode).

| field | type |
|---|---|
| `id` | string |
| `kind` | string |
| `prompt` | string |
| `expected` | object |
| `metadata` | object |

## Provenance

# Tool Errors provenance

This file is generated from the benchmark definition.

## Seeded API fault scenarios

- Kind: `synthetic`
- URL: https://github.com/pooriaarab/benchmark-tool-errors/blob/main/scripts/gen-tasks.mjs
- Licence: `MIT`
- Retrieved: 2026-10-15T00:00:00.000Z
- How to get it: Run `node scripts/gen-tasks.mjs --seed 20261015 --split public` from the repository root. Run `node scripts/gen-tasks.mjs --split hidden --seed <private> [--out <path>]` (or set BTER_HIDDEN_SEED) to create the untracked hidden split.
- Contamination notes: The 40 public tasks use seed 20261015 and are visible in the npm package. Final comparisons should use the separately generated, untracked hidden split; its ids, prompts, fault order, and expected answers differ from the public set.
- Refresh: Regenerate and review public fixtures for each minor release; rotate the private hidden seed for every evaluation campaign.


## Grader

Deterministic replay scores end state (45%), side-effect safety (25%), bounded and timed retries (20%), and truthful reporting (10%); duplicate side effects score zero.

## Limitations

This dataset contains only the public task split (40 tasks).
Hidden evaluation splits use a private seed and are deliberately not part of
this dataset. Source kinds: synthetic. These are evaluation prompts
for the `benchmark-tool-errors` benchmark, not training data; scoring requires the package's
deterministic grader. See PROVENANCE.md for contamination notes.

## Citation

```bibtex
@misc{benchmark_tool_errors,
  author = {Pooria Arab},
  title = {Tool Errors task data},
  year = {2026},
  url = {https://github.com/pooriaarab/benchmark-tool-errors}
}
```

## Links

- npm: https://www.npmjs.com/package/benchmark-tool-errors
- GitHub: https://github.com/pooriaarab/benchmark-tool-errors
- Leaderboard: https://aibenchmarkindex.com/benchmarks/benchmark-tool-errors
- Dataset: https://huggingface.co/datasets/pooria/benchmark-tool-errors