Measure twice, code once.
Cadre is a context-driven development harness for AI coding agents. It gives Claude Code and OpenAI Codex the same packet-owned workflow: structured setup, spec-first tracks, Beads-backed task memory, review gates, team boards, parallel worker orchestration, and mono/polyrepo delivery.
Cadre is not a prompt collection that asks agents to edit state by hand. The installed plugin bundles a Cadre MCP runtime, workflow protocols, templates, and helper scripts. Agents call Cadre packets, and those packets own Cadre state, Beads writes, review records, provider evidence, and derived indexes.
Why Cadre Exists
AI coding sessions often lose context, skip handoff details, or compete over the same files when a team works in parallel. Cadre makes the workflow durable:
| Need | Cadre answer |
|---|---|
| Project context | cadre/product.md, cadre/tech-stack.json, workflow policy, patterns, and style guides. |
| Work planning | Track specs and plans with testable acceptance criteria and file annotations. |
| Durable memory | Beads stores the task graph, dependencies, notes, blockers, and handoffs. |
| Team safety | Owners, advisory leases, collision scans, review queues, and shared sync. |
| Code intelligence | Repo maps, dependency graph, test impact, diagnostics, and optional LSP review. |
| Delivery | Review gates, hosted provider evidence, monorepo ship, and polyrepo land. |
Quick Start
Install Beads:
npm install -g @beads/bd
bd --version
Install the Cadre plugin from this repository.
Claude Code:
/plugin marketplace add vishal-kr-barnwal/Cadre
/plugin install cadre@cadre
OpenAI Codex:
codex plugin marketplace add vishal-kr-barnwal/Cadre --sparse .agents/plugins --sparse harness/plugins/cadre
codex plugin add cadre@cadre
In a target project, activate Cadre and run setup:
$cadre
cadre-setup
Then use the normal lifecycle:
cadre-newtrack "Add OAuth login"
cadre-implement
cadre-review
cadre-ship
cadre-archive
Use cadre-land instead of cadre-ship when the project is a polyrepo control
repo.
Workflow Lifecycle
flowchart LR
A["cadre-setup"] --> B["cadre-newtrack"]
B --> C["cadre-implement"]
C --> D["cadre-review"]
D --> E{"Topology"}
E -->|Monorepo| F["cadre-ship"]
E -->|Polyrepo| G["cadre-land"]
F --> H["cadre-archive"]
G --> H
H --> I["cadre-release"]
Every project-scoped packet call carries a root argument. Cadre MCP resolves
that root, reads the relevant bounded context, performs the requested operation,
and returns structured next actions. Agents summarize packet results; they do
not manually reconstruct Cadre state.
Documentation Map
- Getting Started: install Beads, install the plugin, run first setup, and verify the runtime.
- How Cadre Works: packet-owned workflows, MCP, Beads, tracks, review gates, provider evidence, and code intelligence.
- Workflows: detailed guide for every
cadre-*workflow. - Architecture: harness package layout, generated plugin bundles, source-of-truth files, and development commands.
- Team And Polyrepo: shared sync, ownership, leases, fleet boards, cross-repo PR groups, and merge train behavior.
- Parallel Execution: plan annotations, worker waves, file claims, merge-back, and recovery.
- Troubleshooting: common install, MCP, Beads, provider, LSP, and generated-bundle failures.
Repository Roles
This repository is the Cadre harness/package repository. The implementation
lives under harness/; the public documentation website lives in the root
docs/ Next.js app, with these Markdown pages under docs/content/.
Generated plugin bundles under harness/.agents/, harness/.claude/, and
harness/plugins/ are rebuilt from master sources. Do not edit generated
bundles by hand.