| name | orcaid |
| description | Daily OrCAID multi-agent delegation — run commit0/self_improve/paperbench tasks, evaluate output, deliver results. Schedule: 09:00 AM. |
| tags | ["orcaid","self-improve","commit0","paperbench","paper2code","daily"] |
| triggers | [{"cron":"0 9 * * *"},{"manual":"delegate_task"}] |
| updated | "2026-05-25T00:00:00.000Z" |
| created_by | agent |
orcaid — OrCAID Multi-Agent Delegation
Run the OrCAID multi-agent delegation framework. Three task types: commit0 (stub implementation), self_improve (self-refactor), paperbench (paper reproduction). Choose the right mechanism for the job.
See Also
references/task-types.md — commit0 / self_improve / paperbench detailsreferences/execution-mechanisms.md — ACTIVE vs PASSIVE mechanismstemplates/run-report.md — run report format
Execution Mechanisms
Mechanism 1 — ACTIVE: orcaid.cli → produces new outcomes
Mechanism 2 — PASSIVE: run_evolution.py --domain orcaid → computes metrics from memory (no new outcomes)
Mechanism 3 — PASSIVE: run_evolution.py --domain paper2code → evaluates Paper2Code outputs
Three Task Types
| Task | Use When | Docker Image |
|---|
commit0 | Fix missing stubs in repo | wentingzhao/minitorch:v0 |
self_improve | Refactor/improve OrCAID itself | python:3.12-slim |
paperbench | Reproduce scientific paper | ghcr.io/openhands/agent-server:latest-python |
Critical Constraints
- Always use
uv run python — NOT bare python
- Orchestrator-memory path:
~/.hermes/orchestrator-memory/ (NOT ~/.orcaid/)
- Retry policy: Always
export ORCAID_RETRY_POLICY=kl
- Report delivery: origin (Discord) — needs human attention on failures
Quality Standards
- Distinguish "task ran" vs "task produced correct output"
- Document Docker image pull time on first run
- Always update carryover — OrCAID debugging depends on historical pattern recognition
