orc

Invoke the skill with a request like "use orc to build <feature>" or the /orc command. It runs intake questions, drafts an intent spec for sign-off, plans tasks, scores each one to pick a model, and dispatches conflict-free waves of subagents through review, verify, and ship.

The planner emits per-task facets (breadth, novelty, logic, test surface, risk, uncertainty), and the orchestrator computes a 0-100 score with a fixed arithmetic formula. The score maps through a six-band table from Haiku 4.5 up to Opus 5 medium, and the full scoring table is shown before dispatch.

Yes. The orchestrator writes a checkpoint.json and state-of-play.md to disk after every wave, escalation, and phase transition. At any pause it emits a resume block you can paste into a fresh session to continue with near-zero context cost.

This repository is Claude Code-native, but a provider-agnostic port called orc-open exists for other coding agents. An optional extra_enabled config can also route low-score task bands to foreign models, with mandatory disclosure at Phase 1.

Every P0-P2 finding must be anchored with a file:line and verbatim quote, which the orchestrator spot-checks before acting. A mismatched or missing quote demotes the finding to advisory P3, so it never triggers a fix or blocks shipping.

Use orc-mini for smaller jobs that can skip review, verify, and summary with a single Sonnet subagent, and orc-fast for the quickest lane gated on a fresh wiki and pattern cache. The full /orc lane suits multi-task features needing review, verification, and PR shipping.