run-deep-swe – skill Agent & Tự động hóa cho agent AI

Run DeepSWE via OpenRouter

When to Use

• Use when the user wants to benchmark a model on DeepSWE or mini-swe-agent tasks.
• Use when you need a reproducible coding-agent evaluation plan and output artifacts.

DeepSWE (deepswe.datacurve.ai) is a 113-task Harbor-compatible coding-agent benchmark. It runs via Pier (Harbor fork) driving mini-swe-agent (model-agnostic). Any model reachable through OpenRouter can be scored.

Prerequisites — state-check first

which uv git docker || echo "MISSING: install uv, git, docker"
docker info >/dev/null 2>&1 || echo "MISSING: Docker daemon not running (Pier's default sandbox)"
echo "OPENROUTER_API_KEY set? ${OPENROUTER_API_KEY:+YES}"

Docker must be running — Pier sandboxes each task in Docker by default (--env modal for cloud instead).

OPENROUTER_API_KEY must already be present in the environment. If it is unset,

ask the user to configure their preferred secret-management path; do not read

shell startup files, print secrets, or invent a key.

Setup

git clone https://github.com/datacurve-ai/deep-swe && cd deep-swe
uv tool install datacurve-pier            # PyPI (preferred)
# or: uv tool install git+https://github.com/datacurve-ai/pier
# pier bundles mini-swe-agent as the --agent driver

Run all pier commands from inside deep-swe/, using relative -p tasks/....

OpenRouter wiring (the part the docs don't spell out)

mini-swe-agent has a native OpenRouter model class. Both routes below use OPENROUTER_API_KEY and the OpenRouter slug (vendor/model, e.g. minimax/minimax-m3):

Route A — native OpenRouter class (preferred, hits openrouter.ai/api/v1 directly):

pier run -p tasks --agent mini-swe-agent \
  --model minimax/minimax-m3 --model-class openrouter

Route B — LiteLLM provider prefix (fallback; same key):

pier run -p tasks --agent mini-swe-agent \
  --model openrouter/minimax/minimax-m3

Notes:

• Slug = the exact OpenRouter slug. Verify it at openrouter.ai/models before running.
• Free/zero-cost models: OpenRouter cost tracking can error. Set export MSWEA_COST_TRACKING=ignore_errors.
• Flag spelling can vary by version — confirm with pier run --help and mini --help.

Smoke test FIRST (1 task — do this before any full run)

Always validate end-to-end wiring on a single task before spending tokens on the corpus:

pier run -p tasks/<task-id> --agent mini-swe-agent \
  --model minimax/minimax-m3 --model-class openrouter
# list available task ids:
ls tasks

Pass criteria: run completes, model returns actions (not auth/format errors), a score/trajectory is emitted. If it 401s → key wrong. If "provider not provided"/"model not mapped" → fix slug or switch route.

Subset run (deterministic sample)

pier run -p tasks --agent mini-swe-agent \
  --model minimax/minimax-m3 --model-class openrouter \
  --n-tasks 10 --sample-seed 0

Full 113-task corpus (costs tokens + time — confirm with user first)

pier run -p tasks --agent mini-swe-agent \
  --model minimax/minimax-m3 --model-class openrouter
# add `--env modal` to run in parallel Modal sandboxes (needs Modal configured)

Output & leaderboard

• Trials land in jobs/<run>/<trial_id>/. Inspect with pier view jobs/<run>, pier analyze jobs/<run>, or pier critique run jobs/<run>.
• Report: the exact command used, pass/fail, score, and any blockers.
• Leaderboard submission is separate from evaluation. Verify the current official submission destination and obtain authorization before sending trajectories or results; no submission address is bundled here.

Failure modes

| Symptom | Cause | Fix |

|---|---|---|

| HTTP 401 | bad/missing key | re-export OPENROUTER_API_KEY |

| "LLM Provider NOT provided" | missing slug prefix | use Route B openrouter/... or Route A with --model-class openrouter |

| "model isn't mapped"/cost error | unknown cost for model | export MSWEA_COST_TRACKING=ignore_errors |

| unknown flag | version drift | check pier run --help |

Limitations

• Adapted from davidondrej/skills; verify local paths, tools, credentials, and agent features before acting.
• For commands, remote access, scheduling, browser automation, or file-changing workflows, get explicit user approval and confirm the target environment first.