ridges-bench

         _                                       _
    _  _ | |                                     | | _  _
   | || || |                                     | || || |
 =H| || || |=====================================| || || |H=
   |_||_|| |                                     | ||_||_|
         |_|                                     |_|
 

Synthetic SWE task samples for Ridges — a small set of synthetic coding tasks miners can run their agent against locally to iterate, debug, and finetuning.

Please do not infer the difficulty of what you will in the actual competition from these tasks.

These are samples, they will not be used for evaluations nor their source will be used.

Samples are grouped by the date they were released (e.g. 30_06_2026/).

Each directory is one Harbor task - same format that the subnet works on.

<task-name>/
├── instruction.md     # the problem statement handed to your agent
├── task.toml          # task metadata (difficulty, timeouts, environment)
├── environment/       # the container/env the task runs in
├── tests/             # verifier tests that decide pass/fail (your score)
└── solution/          # the gold patch — reference only, not used when your agent runs

Run a task with the Ridges CLI

You need the ridges CLI installed (from the ridges repo https://github.com/ridgesai/ridges) and an OpenRouter API key.

1. One-time setup:

   ridges miner setup

   This configures your workspace, agent, and provider. Put your OpenRouter key in .env.miner:

   RIDGES_OPENROUTER_API_KEY=sk-or-...
   

   The provider must be configured, or run-local will error.

2. Run your agent against a single task:

   ridges miner run-local --task-path ./30_06_2026/<task-name>

   You'll see the reward and per-test results:

   SUCCEEDED
   reward: 1.0
   tests: 12 total (12 passed, 0 failed, 0 skipped)
   

Local mode runs your agent + verifier but does not enforce the platform's sandbox/budget restrictions — it's for iteration, not official scoring.

Your agent

A miner agent is a single agent.py exposing:

def agent_main(input: dict) -> str:
    # return a unified-diff patch string
    ...