| # AGENTS.md — writing InfraBench tasks |
|
|
| This file is for coding agents and humans who author **benchmark tasks**. It is intentionally scoped to task packages. The evaluation harness (Syscraft) remains closed-source for now. |
|
|
| ## Non-negotiables |
|
|
| 1. Every task is a directory with `task.toml`, `instruction.md`, `environment/`, and `tests/`. |
| 2. The verifier is the source of truth: write `reward.txt` or `reward.json` under `/logs/verifier/`. |
| 3. Do not bake secrets, API keys, or private CloudLab credentials into the task tree. |
| 4. Prefer structured checks with `scored: true|false` over a single opaque score. |
|
|
| ## Layout |
|
|
| ``` |
| task/ |
| ├── task.toml |
| ├── instruction.md |
| ├── environment/ # Dockerfile OR setup.sh [+ bootstrap.sh] |
| ├── solution/ # optional solve.sh |
| └── tests/ # test.sh (+ helpers) |
| ``` |
|
|
| Canonical examples in this repo: |
|
|
| | Path | Role | |
| |------|------| |
| | `tasks/hello-world` | Minimal Docker task — start here | |
| | `tasks/ipmi-node-power-recovery` | L1 / Easy / BM Cluster | |
| | `tasks/cassandra-nic-split-brain` | L2 / Medium / BM Cluster | |
| | `tasks/cassandra-dead-node-removal` | L3 / Medium / BM Cluster | |
| | `tasks/vm-ceph-bootstrap` | L3 / Hard / VM cluster | |
| | `tasks/db-wal-recovery` | L4 / Hard / Container | |
|
|
| ## Environment types |
|
|
| | `environment.type` | When to use | |
| |--------------------|-------------| |
| | `docker` | Userspace / container-local work | |
| | `vm` | Kernel modules, block devices, single-node systems work | |
| | `vm-cluster` | Multi-node distributed systems | |
| | `cloudlab` | Bare-metal / IPMI / real NIC scenarios | |
|
|
| VM hosts need Linux + KVM. CloudLab tasks need a profile and site access — document that in the PR, do not commit credentials. |
|
|
| ## Verifier patterns |
|
|
| Scalar: |
|
|
| ```bash |
| #!/bin/bash |
| set -euo pipefail |
| # ... assertions ... |
| echo 1.0 > /logs/verifier/reward.txt |
| ``` |
|
|
| Structured (preferred for InfraBench paper-style tasks): |
|
|
| ```bash |
| #!/bin/bash |
| set -euo pipefail |
| mkdir -p /logs/verifier |
| python3 - <<'PY' |
| import json, pathlib |
| checks = [ |
| {"name": "service up", "passed": True, "scored": False}, |
| {"name": "data consistent", "passed": False, "scored": True}, |
| ] |
| reward = 1.0 if all(c["passed"] for c in checks) else 0.0 |
| pathlib.Path("/logs/verifier/reward.json").write_text( |
| json.dumps({"reward": reward, "checks": checks}) |
| ) |
| PY |
| ``` |
|
|
| ## Writing `instruction.md` |
|
|
| - State the **goal** and **success criteria** the agent can observe. |
| - Include operational constraints (do not destroy X, preserve Y). |
| - Do not paste the root-cause spoilers unless the task is explicitly a known-fault drill. |
| - Keep it short enough that an agent can act; put long background in comments inside `environment/` scripts if needed. |
|
|
| ## PR expectations |
|
|
| See [CONTRIBUTING.md](CONTRIBUTING.md) for the human checklist. Agents should open a PR that: |
|
|
| 1. Adds `tasks/<name>/` with a complete package |
| 2. Mentions layer (L1–L4), difficulty, and backend in the PR body |
| 3. Links related paper sections when the task is a HotInfra / InfraBench scenario |
|
|
| ## Out of scope here |
|
|
| - Implementing new agent adapters |
| - Changing the Syscraft CLI or environment factory |
| - Publishing private evaluation traces |
|
|
| Those stay in the closed Syscraft tree until the runtime is released. |
|
|
