Cadre

Measure twice, code once.

Cadre is a context-driven development harness for AI coding agents. It combines spec-first tracks, native durable task memory, review gates, team boards, parallel worker orchestration, and mono/polyrepo delivery into one packet-owned workflow.

Public docs: https://cadre-docs.pages.dev/

What Cadre Provides

• Structured work: setup, new track, implementation, review, ship/land, archive, release, handoff, refresh, revise, validate, flag, and formula flows.
• Persistent memory: Cadre stores task graph state, dependencies, notes, handoffs, events, and resume evidence through native packet-owned files.
• Team safety: ownership, advisory leases, collision scans, review queues, shared sync, and compact MCP dashboard resources.
• Polyglot intelligence: repo maps, dependency graphs, test impact, workspace diagnostics, LSP setup, warm LSP review, and async job artifacts.
• Two agent surfaces: Claude Code and OpenAI Codex plugins are thin MCP entrypoints. The global cadre-mcp runtime owns the skill contract, protocols, references, templates, and packet tools.

Install

Install Cadre from npm and let the CLI wire detected clients:

npm install -g cadre-ai
cadre install
cadre doctor

Use

In a target project, activate the Cadre skill and ask for the workflow you need:

$cadre
cadre-setup
cadre-newtrack "Add OAuth login"
cadre-implement
cadre-review
cadre-ship

Cadre workflows are packet-owned. The agent verifies Cadre MCP, passes a per-call root, and lets the runtime perform state reads/writes, formula work, parallel worker state, provider evidence write-back, and shared sync. Do not maintain Cadre state by hand.

Setup Outputs

cadre-setup writes the project control plane:

• cadre/product.json plus generated cadre/product.md
• cadre/product_guidelines.json plus generated cadre/product_guidelines.md
• cadre/tech-stack.json
• cadre/workflow.json plus generated cadre/workflow.md
• cadre/tracks.json as the generated track index
• cadre/patterns.jsonl plus generated cadre/patterns.md
• cadre/config.json
• cadre/events.jsonl
• cadre/messages/outbox.jsonl and cadre/messages/inbox.jsonl
• cadre/formulas/*.json when formulas are added
• git-ignored cadre/local/wisps/*.json for local ephemeral formula runs
• cadre/styleguides/*.json plus generated cadre/code_styleguides/*.md
• optional cadre/repos.json for polyrepo topology
• optional cadre/lsp.json for LSP recommendations

Setup also initializes native Cadre state, can configure shared-sync merge attributes, and can scaffold hosted CI checks when requested.

Team And Repo Modes

Cadre supports monorepos and polyrepo control repos. For teams, use shared sync so ownership, leases, review state, blockers, and available work are visible to everyone. Product code publication still happens through ship/land workflows; shared sync is for the Cadre control plane.

Compact MCP resources provide bounded views for larger teams:

• team board and next actions
• review queue and handoff inbox
• quality gate and parallel worker state
• repo topology, repo map, workspace diagnostics, test impact, and LSP status
• provider action plans and async job results

Harness Development

This repository is the Cadre harness/package repo. Runtime sources live in src/, master skill/protocol sources live in skills/cadre/, references live in scripts/agent-refs/, and templates live in templates/.

Run package commands from the repository root:

pnpm --filter cadre-ai generate
pnpm --filter cadre-ai check

Generated plugin bundles under .agents/, .claude/, and plugins/ are rebuilt from master sources. They contain only platform manifests, MCP config, and SKILL.md; the embedded MCP runtime is built under scripts/.

Public documentation lives in the repo-root docs/ Next.js app. Markdown page source is in docs/content/:

• Documentation Home
• Getting Started
• How Cadre Works
• Workflows
• Architecture
• Team And Polyrepo
• Parallel Execution
• Troubleshooting