orcaid

type	skill-manual
name	orcaid
description	OrCAID multi-agent execution engine — correct entry (orcaid.cli), local path support, auto-build Docker, three task types (commit0/self_improve/paperbench)
triggers	["run OrCAID","commit0","self_improve","paperbench","add OrCAID task"]
tags	["orcaid","multi-agent","delegation"]
updated	"2026-05-24T00:00:00.000Z"

OrCAID Skill Manual

Cron 297092f3b347 is PAUSED. This skill is manual-only — run on-demand via CLI.

Repo: /home/ty/Repositories/ai_workspace/OrCAID/

Correct Entry Point

Always use uv run python -m orcaid.cli — NOT run_infer.py, NOT bare python.

cd /home/ty/Repositories/ai_workspace/OrCAID
uv run python -m orcaid.cli [flags...]

.env file at OrCAID root sets LLM_MODEL and LLM_BASE_URL.

Three Task Types

commit0 — Implement stubs, run pytest (primary task)

cd /home/ty/Repositories/ai_workspace/OrCAID

uv run python -m orcaid.cli \
    --task=commit0 \
    --repo /home/ty/Repositories/ai_workspace/Paper2Code-Enhanced \
    --model minimax/MiniMax-M2.7 \
    --multi_agent=false \
    --max_iterations=5

self_improve — OrCAID self-improvement (ast.parse only, weaker)

cd /home/ty/Repositories/ai_workspace/OrCAID

uv run python -m orcaid.cli \
    --task=self_improve \
    --repo_path /home/ty/Repositories/ai_workspace/OrCAID \
    --task_description "Add memory-of-failures pattern to Manager._verify_and_return" \
    --model minimax/MiniMax-M2.7 \
    --multi_agent=false \
    --max_iterations=3

paperbench — Reproduce ML papers (LLM judge evaluated)

uv run python -m orcaid.cli \
    --task=paperbench \
    --paper_id 2605.18703 \
    --model minimax/MiniMax-M2.7 \
    --multi_agent=false \
    --max_iterations=5

Three-System Pipeline

arXiv cron (arxiv agent) → saves papers to arxiv-papers/
    ↓
Paper2Code-Enhanced (your CLI) → generates code repo
    ↓
OrCAID commit0 → validates + fixes generated repo

This pipeline runs on-demand, not on a cron schedule. Paper2Code generates; commit0 validates.

Local Path Support

--repo accepts absolute local paths (starting with /). When given a local path:

• The local directory is copied into the container after it starts (not git clone)
• DockerDevWorkspace auto-builds an image from a language-appropriate base image (Python → python:3.12-slim, Node → node:20-slim, etc.)
• No pre-built benchmark Docker image needed
• No git authentication required

# Local path — container builds from python:3.12-slim automatically
uv run python -m orcaid.cli \
    --task=commit0 \
    --repo /home/ty/Repositories/ai_workspace/Paper2Code-Enhanced \
    --model minimax/MiniMax-M2.7

Docker Image Requirement

Docker must be running. OrCAID uses openhands.workspace.DockerDevWorkspace:

• openhands.workspace.DockerDevWorkspace builds images from base_image (not from the target repo)
• base_image is set by task_module.get_workspace_config()["base_image"]
• For commit0 on Python repos: python:3.12-slim
• OpenHands' build() function builds from the base image
• Local repo is copied into the running container (not used as build context)

Use --docker_image=docker.io/wentingzhao/minitorch:v0 only when you specifically need the minitorch image.

Key Parameters

Flag	Task	Purpose
--task	all	commit0, self_improve, paperbench
--repo	commit0	GitHub URL (owner/repo) or local absolute path
--repo_path	self_improve	Local absolute path to the repo
--model	all	LiteLLM model (minimax/MiniMax-M2.7)
--multi_agent	all	true (4 engineers) or false (single agent)
--max_iterations	all	Max Manager LLM turns
--max_rounds_chat	all	Subagent chat rounds (default: 2)
--paper_id	paperbench	arXiv paper ID

Key Paths

Path	Role
/home/ty/Repositories/ai_workspace/OrCAID/	OrCAID repo root
/home/ty/Repositories/ai_workspace/Paper2Code-Enhanced/	Target repo for commit0
~/.hermes/orchestrator-memory/verified/	Verified SubAgentResult outcomes
~/.hermes/orchestrator-memory/drift_logs/	Drift/failure records
~/.hermes/orchestrator-memory/escalations/	Human review items
~/.hermes/orcaid-bridge/	Bridge storage

Key Files

File	Role
orcaid/cli.py	Entry point — DockerDevWorkspace setup + workflow
orcaid/core/utils.py	build_task_module() (line 1309), build_llm_kwargs()
orcaid/core/manager.py	Manager + _verify_and_return() (lines 678, 710, 731)
orcaid/core/subagent.py	SubAgentRunner — git worktree per engineer
orcaid/tasks/commit0.py	commit0: pytest, stub implementations
orcaid/tasks/self_improve.py	self_improve: ast.parse syntax check
orcaid/tasks/paperbench.py	paperbench: LLM judge evaluation
orcaid/tasks/paper2code.py	PaperCoder task (stub, wired but not implemented)
orcaid_verification_bridge.py	Self-healing: verify_subagent_completion()

build_task_module Tasks (utils.py line 1309)

Task	Kwarg mapping
commit0	repo → repo_name, also base_branch, docker_image_prefix, docker_image, dataset_path
self_improve	repo_path, task_description
paperbench	paper_id, docker_image, paperbench_dir, ...
paper2code	repo_path, paper_url, output_dir (wired but stub)

MiniMax Configuration

# In OrCAID .env
LLM_MODEL=minimax/MiniMax-M2.7
LLM_BASE_URL=https://api.minimax.io/v1
LLM_API_KEY=<key>

Must be /v1, not /anthropic/v1/messages.

Architecture

User → orcaid.cli → build_task_module() → TaskModule
    ↓
DockerDevWorkspace (base_image=python:3.12-slim, target=source-minimal)
    ↓ (OpenHands builds image + starts container)
container: copy local repo → run agent (Manager + Engineers)
    ↓
Manager: scan_and_analyze() → delegate_tasks() → run_subagents_parallel()
    ↓
collect_and_merge() → _verify_and_return()
    → bridge.verify_subagent_completion()
        → PASS: orchestrator-memory/verified/
        → FAIL: orchestrator-memory/drift_logs/ + correction_context
    ↓
final_review_all() → outputs/

Key line references:

• orcaid/core/manager.py lines 678, 710, 731: _verify_and_return()
• orcaid/core/utils.py line 1309: build_task_module()

Common Issues

Symptom	Cause	Fix
Exit 1, no output	Missing .env or Docker not running	Check LLM_BASE_URL, docker ps
"Unknown model provider"	Wrong LLM_BASE_URL	Must be https://api.minimax.io/v1
Local path not copied	Path must be absolute (start with /)	Use /home/ty/... not ~/...
No container starts	Docker not running	docker ps to verify
paper2code task fails	evaluate() is still a stub	Implement evaluation or use commit0 on the generated repo instead

Add a New Task

1. tasks/my_task.py: MyTaskConfig dataclass + MyTask(TaskModule) implementing:
   • get_docker_image(), get_work_dir(), get_workspace_config()
   • load_task_data(), setup_workspace(), evaluate()
2. tasks/__init__.py: add from .my_task import MyTaskConfig, MyTask
3. orcaid/core/utils.py build_task_module(): add elif task == "my_task": return MyTask(MyTaskConfig(**init))