Cadre

Measure twice, code once.

A human-governed, Git-aware delivery harness for OpenAI Codex, Claude Code, and Zed Agent.

Cadre turns approved project context into resumable feature and bug tracks. It combines specification, dependency-aware planning, parallel implementation, manual verification, review, revision, Quick changes, Refresh with archival, revert, and Git provenance without asking an agent to invent workflow state.

Documentation: cadre.dhivon.ai · Start the quickstart · Review release notes

Cadre is a product of Dhivon AI.

Cadre 4.2.0 adds optional operation telemetry to help diagnose slow MCP operations and failures. Telemetry-capable builds default on outside CI, with clear installation/startup disclosure, persistent opt-out, and support for your own Sentry DSN or a Sentry-free build. Ten workflows, 21 MCP tools, and immutable template v8 support Codex, Claude Code, and Zed Agent. See cadre-ai on npm.

What changes in 4.2.0

• Understand bottlenecks: operation and nested-phase timing, approval wait, sanitized error codes, inspection counters, and client/release/template versions.
• Observe usage: timestamped project, track, and Quick inventories with random IDs. Counts describe observed project installations; repeated snapshots are not totals.
• Control collection: disable with cadre-ai telemetry disable, cadre-ai install --telemetry off, or DO_NOT_TRACK=1. Saved preferences survive upgrades.
• Choose the destination: replace the default with your own DSN, or build with CADRE_BUILD_TELEMETRY=none to exclude Sentry and the default DSN. Builds without CADRE_BUILD_SENTRY_DSN also omit Sentry entirely; release builds receive it from the CADRE_SENTRY_DSN GitHub Actions secret.
• Keep project state compatible: existing 3.9.x, 4.0.0, and 4.1.0 schema-2 projects remain supported. No migration is needed for telemetry.

See telemetry configuration for collected/excluded data, precedence, and best-effort delivery limits.

What changes in 4.1.0

• Quick changes: deliver one bounded fix, refactor, documentation change, test, or small feature with one final approval, one durable outcome record, and one combined commit.
• Unified Refresh: assess repository changes, project context, existing patterns, selected track learning, and up to 20 quick outcomes together. Update guidance, reassess active-track seeds and dependencies, and archive completed work under one approval, one receipt, and one commit.
• Ten workflows, 21 tools: remove the standalone archive skill and archive_batch_candidate. New archival uses Refresh; approved historical archive journals remain recoverable through candidate_apply.
• Compatible upgrades: schema-2 projects from 3.9.0, 3.9.1, 3.9.2, and 4.0.0 remain readable and usable. Adopt new workflow contracts through an approved Refresh that preserves custom policy, completed history, and recovery authority.
• Template v8 and client cleanup: retain immutable v1–v7 source history; safely remove obsolete Cadre-owned Zed archive links and exact tool permissions.

Quick and Refresh verification includes 660 dummy-project protocol cases across Codex, Claude Code, and Zed identities, plus live Codex delivery and recovery. Claude Code and Zed coverage is protocol/installer coverage, not native model execution. See the testing and release documentation.

Unified Refresh

Refresh now combines context maintenance, learning consolidation and completed-track archival under one approval, receipt and commit. Cadre 4.2.0 exposes ten workflows and 21 MCP tools with template v8. The standalone archive skill and MCP tool are removed. Existing approved archive journals remain recoverable through candidate_apply. Older supported projects adopt the new workflow contract in an approved Refresh.

Create → Track → Implement → Review → completed tracks ─┐
   └──→ Quick → verified outcomes ─────────────────────┤
Context + patterns + repository changes ────────────────┤
                                                       ↓
                                                    Refresh
                          context + learning + active seeds + archival
                                  one approval · one receipt · one commit

Omitting track IDs selects all eligible completed tracks in dependency order. Explicit IDs limit the batch; an explicit empty selection skips track archival. Refresh assesses up to 20 quick outcomes with the selected tracks, with no track batch limit. It also works without completed tracks or new guidance to add.

Quick changes

Use quick for one bounded fix, refactor, documentation change, test, or small feature. It reads relevant Cadre context, implements and verifies the change, then asks for one final approval of the diff and outcome record. Material scope choices are discussed before editing. One local commit contains the change and its durable record. A later approved refresh batch consolidates useful lessons and moves unchanged records to .cadre/archive/.quick/. Routine track context loads no quick history; focused search retrieves relevant outcomes when needed.

Invoke $cadre:quick in Codex, /cadre:quick in Claude Code, or /cadre-quick in Zed. Cadre 4.2.0 includes Quick and unified Refresh with template v8. Existing v5/v6 projects retain their workflows; enabling quick requires an approved refresh adopting its contract. Wisp remains stateless exploration.

Upcoming roadmap

Three optional capabilities are planned beyond 4.2.0:

1. Deeper GitHub / GitLab integration for coordinating work across multiple people, including shared work status, ownership and review handoffs.
2. A knowledge base built as work happens, capturing decisions and learning during delivery and using relevant knowledge as context for subsequent work.
3. Polyrepo support for coordinating dependencies and delivery across multiple repositories while retaining per-repository history and approvals.

Each capability will be independently opt-in and disabled by default. The core single-repository workflow will remain usable without enabling any of them. These are upcoming roadmap items, not features included in 4.2.0; no delivery dates are committed. See the roadmap.

What Cadre Provides

• Human-governed delivery: approved scope and approval policy govern execution; clean track completion always requires the author’s explicit approval.
• Resumable state: setup, track, execution, review, revision, refresh, revert, and archive checkpoints survive interrupted sessions.
• Spec-first tracks: feature and bug specifications lead to validated phase/task dependency graphs with derived manual-verification barriers.
• Safe parallel execution: bounded workers operate in isolated Git worktrees while the main agent alone schedules, integrates, resolves conflicts, updates Cadre state, and records approval.
• Incremental learning: phase learning and durable patterns flow forward through dependency-aware seeds and Refresh consolidation.
• Typed runtime: a bundled MCP server provides immutable templates, validation, digest-gated state transitions, and constrained worktree operations.

Install

Cadre supports OpenAI Codex, Claude Code, and native Zed Agent as stable integrations at user scope:

npm install -g cadre-ai
cadre-ai doctor
cadre-ai install

cadre-ai install auto-detects installed clients. Use --target codex, --target claude, --target zed, or --target all to choose explicitly. Start a new client conversation after installation; in Claude Code, run /reload-plugins first. In Zed, run workspace: reload from the command palette, then open a new Agent thread.

Use Cadre

Cadre commands are agent skills rather than shell subcommands:

# Codex
$cadre:create
$cadre:track Add passwordless login as a feature
$cadre:implement passwordless-login
$cadre:review passwordless-login
$cadre:refresh passwordless-login

# Claude Code
/cadre:create
/cadre:track Add passwordless login as a feature
/cadre:implement passwordless-login
/cadre:review passwordless-login
/cadre:refresh passwordless-login

# Zed Agent
/cadre-create
/cadre-track Add passwordless login as a feature
/cadre-implement passwordless-login
/cadre-review passwordless-login
/cadre-refresh passwordless-login

The complete workflow set is create, track, implement, review, revise, refresh, revert, status, wisp, and quick.

An initialized target project keeps approved mutable delivery state under .cadre/. The installed payload retains the runtime, skills, worker definitions, and immutable templates.

Approval modes and continuous Autonomous delivery

Approval mode	Human decisions
Governed	Task, integration, phase/track verification, and review gates
Phase	Phase verification, track verification, and review
Track — default	Track verification and review
Autonomous	One final completion decision after implementation, verification, review, and remediation are clean

Scheduling is separate: Parallel is the default; request Sequential with any approval mode. To use the continuous loop, ask your client’s implement skill to implement the track with Autonomous approval mode. Cadre invokes the review skill through MCP next steps and repeats fixes, verification, and cumulative review without another command.

ready_for_review is an internal handoff during that loop, not an approval pause. Once clean, Cadre returns verification results, remediation history, and the exact completion proposal. Approve it to mark the track completed without running review again, or report additional bugs to continue remediation. Changed implementation or verification inputs require renewed verification/review. Autonomous pauses for scope decisions, unavailable checks, external blockers, or the same unresolved finding after two remediation attempts.

Repository Layout

This is the Cadre harness/package repository, not an initialized target project:

• harness/ contains the TypeScript runtime, installer, workflow skills, worker definitions, versioned templates, plugin manifests, and tests.
• docs/ contains the canonical Next.js/shadcn documentation site; Markdown source lives in docs/content/.
• AGENTS.md and CLAUDE.md define current harness development conventions.

The tracked plugin manifests and marketplace catalogs are source files. harness/dist/ and the installed marketplace under ~/.cadre/marketplaces/cadre are generated.

Develop

From the repository root:

pnpm install
pnpm check

Focused harness validation:

pnpm --filter cadre-ai check
pnpm --filter cadre-ai test
pnpm --filter cadre-ai validate

Cadre is licensed under Apache-2.0. Copyright 2026 Dhivon AI. See NOTICE.

Release preparation includes regression tests, package inspection, and documentation checks. Publishing still requires native installer/discovery/activation checks for Codex, Claude Code, and Zed. Local model tests and byte benchmarks do not establish universal token savings or semantic correctness of generated learning.

Implementation approval modes

Track is the default: phase checks run automatically, followed by track verification and ordinary review approval. Phase retains phase verification gates; Governed retains task and integration gates. Explicit Autonomous implements, verifies, reviews, and remediates in-scope findings until clean, then asks for one final completion approval. Parallel scheduling remains the default and Sequential is available independently. Legacy Autonomous requires an explicit Track-or-Autonomous choice through approved refresh. Policy v2 persists loop authority, cumulative review baseline, and stable findings; a finding unresolved after two remediation attempts pauses for guidance.

Optional Telemetry

Telemetry-capable Cadre 4.2.0 builds default on outside CI and announce it before installation and on MCP startup. Disable before installing with cadre-ai install --target all --telemetry off, or use cadre-ai telemetry disable / DO_NOT_TRACK=1. Saved choices survive upgrades and --yes. Timings, sanitized errors, versions, counts, and random IDs help identify bottlenecks; prompts, files, names, paths, and raw MCP payloads are excluded. Use cadre-ai telemetry configure --dsn <DSN> for your own Sentry destination. See configuration for precedence, privacy, delivery limitations, and Sentry-free builds.