claude-copilot

Claude Copilot

An instruction layer for Claude Code — 15 framework agents with strict points of view, a design-led process enforced by mechanical hooks, persistent memory across sessions, and real task/worker orchestration. It makes Claude Code's process repeatable, inspectable, and stateful.

---

What Is Claude Copilot?

Claude Copilot is a set of instructions that sit on top of Claude Code. This is an independent, community-driven framework for Claude Code, unaffiliated with Microsoft Copilot or GitHub Copilot.

It's not separate software—it's markdown files (agents, commands, project instructions) and two CLI tools (cc and tc) that give Claude Code new capabilities:

You Get	What It Does
Persistent Memory	Decisions, lessons, and progress survive across sessions (FTS5 keyword search)
Memory Drift Detection	cc memory check — token-free deterministic checkers (path-exists, command-resolves, version-conflict, staleness); 0–100 score; exits 1 on any fail-severity finding
Usage Observability	cc usage — idle-gated quota probe via Keychain OAuth + anthropic-ratelimit-unified-* headers; producer/consumer split so consumers never corrupt the quota window
15 Framework Agents	Lean agents with on-demand skill loading; methodology-embedded from IDEO to Kent Beck; kc is setup-only (not in the build chain). Authoritative roster: .claude/agents/manifest.json
Auto-Firing Skills	Skills surface automatically from trigger-rich descriptions; code-bearing skills run executable scripts
Parallel Orchestration	Headless workers execute streams concurrently with /orchestrate (works; unproven at large scale — no proven >5-stream run; tests are mock-only)
Pause & Resume	Context switch mid-task with /pause, return with /continue
Task Management	PRDs, tasks, and work products (WP) via tc CLI with minimal context usage
Stream Management	Parallel work streams with conflict detection and dependencies
Known References	Configured paths and refs surface into every session via the Known References registry (cc config set refs.*); cc memory search is available manually — agents do not yet auto-search a company knowledge repo (roadmap: agent auto-pull)
Extensions System	Override or extend agents with your company methodologies
Code-Execution Path	tc.api / cc.api facades for multi-step ops without CLI round-trip token cost
Live Docs	cc docs get <pkg> — version-exact package documentation; agents code against the real installed API, not stale training memory; local-first, offline-safe
Context Engineering	Auto-compaction, continuation enforcement, activation modes

When Claude Code reads these instructions, it gains persistent memory, 15 framework agents (plus kc, setup-only), and a structured process — the design goal being more disciplined, resumable work built from the practices that tend to produce better software. We measure process and context efficiency, not output quality; there is no defect/rework data yet.

→ Why we built this

---

The Problem

Solo developers are expected to be experts in everything. AI assistants help, but they:

• Forget everything between sessions
• Give generic advice without context
• Lack proven processes for complex work