An hourly issue-triage bot for any GitHub repo, running on ORB Cloud.

Run it yourself, or tell your agent to.

# required
export ORB_API_KEY=orb_...
export GITHUB_TOKEN=ghp_...
export TARGET_REPO=myorg/myrepo

# pick ONE auth mode
export OPENAI_API_KEY=sk-...
# export CODEX_AUTH_JSON='{"token":"..."}'

# deploy, posts to /v1/swarms, prints swarm ID + 5 replica IDs
bash <(curl -fsSL https://orbcloud.dev/templates/clawsweeper)

Deploy the clawsweeper template from ORB Cloud.

Read the manifest at https://orbcloud.dev/templates/clawsweeper/template.json, it lists the required secrets, optional inputs, and the deploy command.

Ask me for anything I haven't already given you: ORB_API_KEY, GITHUB_TOKEN, TARGET_REPO, and EITHER OPENAI_API_KEY or CODEX_AUTH_JSON (not both). Then export them and run the deploy command from the manifest.

When it returns, print the swarm ID and link me to https://orbcloud.dev/use-cases/clawsweeper/ so I can watch the replicas idle until the next cron tick.

A maintainer bot for issue triage, not a code reviewer.

For every open issue and PR in your target repo, clawsweeper runs Codex through a structured "should this stay open?" prompt. It can close items only for one of these reasons:

• already implemented on main
• cannot reproduce on current main
• duplicate or superseded by a canonical issue
• concrete but not actionable in this repo
• too incoherent to be actionable
• stale issue older than 60 days with insufficient data

Anything else stays open. Every decision lands in items/<n>.md in your report repo with file paths, line numbers, and SHAs Codex cited, full audit trail. Confidence + protected-label gates mean only high-confidence non-critical items auto-close.

Five workers. Zero coordination.

Each replica reviews items where item.number % 5 == replica_index. No master, no queue, no IPC, pure modulo partition. ORB injects ORB_REPLICA_INDEX and ORB_REPLICA_COUNT into each replica's environment automatically.

Each replica runs in its own ORB computer, own network namespace, own cgroup, own LLM proxy, own checkpoint blob. They sleep independently between hourly cycles. Bursts handled by per-replica wake-on-cron.

Each replica is awake about 1% of the time.

A clawsweeper review takes ~5–15 seconds of Codex thinking. The remaining ~3585 seconds of every hour, the replica is sleeping on NVMe at zero cost. Wake on the next cron, do another ~10 seconds of work, sleep again.

On every other runtime, VM-per-agent, container-per-agent, serverless-per-call, those 99% of idle seconds cost full rate. On ORB they cost zero.

OpenAI API key or ChatGPT plan tokens.

Pick exactly one. The deploy script branches the LLM upstream automatically.

Works on private repos with the same command.

The wrapper clones your target with the GitHub token in the URL, private repos behave identically to public ones. Your token needs read on the target repo and (if you write back review proposals) write on the report repo.

ORB stores the token AES-256-GCM encrypted in its secret store. Never visible after deploy. Rotate by running deploy again with the new value.