orc

Orchestrates subagent-driven development to pick, implement, review, and commit GitHub issues autonomously.

Updated Aug 21, 2026
One-click install
npx skills add https://github.com/rogadev/orc-pack --skill orc-rogadev
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: orc
Source: https://github.com/rogadev/orc-pack/tree/main/skills/orc
Command: npx skills add https://github.com/rogadev/orc-pack --skill orc-rogadev

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Orc eliminates the manual babysitting of coding tasks by autonomously carrying a unit of work — a picked issue, a named issue, or a free-text task — all the way to committed, reviewed, green code without mid-run check-ins. ## Core Features & Use Cases - Autonomous issue selection: A scout subagent scans your GitHub issue board with a four-signal ruleset (unblocking value, severity, readiness, age) and picks the highest-value work. - Subagent-driven workflow: Dispatches implementer subagents to write code and tests, a panel of security/architecture/quality/test-coverage reviewers, and a verifier that filters out hallucinated review findings. - Bounded review-fix loop and landing: Loops review and fix up to three rounds, runs the repo's ready check, commits per task, pushes, and closes resolved issues with a structured FINISHED / NOT FINISHED report. - Use Case: Run /orc before leaving for the day and return to a finished issue with commits on the working branch, a green ready check, and a written summary of what changed and why. ## Quick Start Invoke the orc skill with no argument to have it pick the highest-value open GitHub issue and autonomously implement, review, and commit the fix.

Frequently Asked Questions about orc

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run an autonomous coding agent on my GitHub issues?

Invoke /orc with no argument and it dispatches a scout subagent that scans your issue board, picks the highest-value issue using a four-signal ruleset, then implements, reviews, and commits the fix. You can also pass an issue number or a free-text task.

How does subagent-driven code review work in orc?

Each task's diff is written to a file and reviewed in parallel by security, architecture, quality, and test-coverage reviewer subagents. A verifier subagent then filters the findings to discard hallucinated issues before any fix round begins.

Which Claude models does orc support?

Orc runs on Opus 4.8 and Sonnet 5, with Haiku for lightweight scouting and lint/test tasks. It explicitly refuses to run on Opus 5 due to a measured 50.1% hallucination rate, and the preflight check halts the run on that model.

When does orc report NOT FINISHED?

NOT FINISHED is reserved for a closed list: a dirty working tree, being stuck on a protected branch, missing credentials, a blocking review finding after three fix rounds, a ready check that won't go green, a rejected push, or a genuine human-only decision.

Does orc create pull requests or push to main?

No. Orc commits and pushes only to the repo's working branch, never to main, master, or production, and it never opens PRs, force-pushes, rebases, or skips hooks. Promoting to the release branch remains the user's call.