orchestra

Orchestra

Orchestrated multi-agent code review for Claude Code, as an installable plugin.

A /review coordinator asks a bundled MCP server for a filtered diff and a risk tier, then calls a custom spawn_reviewers tool that fans out by spawning concurrent claude -p subprocesses — one per specialist (security, performance, code quality, documentation). Each specialist runs read-only with a tightly-scoped prompt (including explicit "what NOT to flag" rules) and returns structured findings. The coordinator dedupes, judges severity, applies an approval rubric biased toward approval, and emits one consolidated review.

Inspired by Cloudflare's orchestrated AI code review, mapped onto Claude Code primitives with claude -p standing in for OpenCode sessions. Zero runtime dependencies — no build, no npm install.

How it maps to Claude Code

Cloudflare (OpenCode)	Orchestra (Claude Code)
Coordinator OpenCode process	/review skill running in your session
spawn_reviewers SDK tool	spawn_reviewers MCP tool spawning claude -p
Specialist OpenCode sessions	Read-only claude -p subprocesses
Risk tiers (trivial/lite/full)	lib/risk.mjs
Noise filtering	lib/diff.mjs
Telemetry Worker	Local .orchestra/telemetry.jsonl

Reviewers and risk tiers

Tier	Trigger	Roster
trivial	≤10 lines, ≤20 files, no security-sensitive path	code-quality
lite	≤100 lines, ≤20 files	code-quality, documentation
full	>100 lines, >50 files, or any sensitive path	security, performance, code-quality, documentation

Security-sensitive paths (auth/, crypto/, …) always force a full review.

MCP tools

The bundled orchestra MCP server exposes the surface the /review coordinator drives:

Tool	Responsibility
get_diff	Run git diff, filter noise, assess risk tier, materialize per-file patches + shared context.
spawn_reviewers	Fan out to concurrent read-only claude -p specialists; aggregate findings, tokens, and cost.
emit_review	Dedupe, apply the approval rubric, render the consolidated review, write it to .orchestra/reviews/.
record_telemetry	Append a fire-and-forget run record to .orchestra/telemetry.jsonl.
orchestra_health	Liveness probe.

Install

/plugin marketplace add alexanderop/clanker-review
/plugin install orchestra@orchestra

For local development, point Claude Code at a checkout:

claude --plugin-dir /path/to/orchestra

Use

In a repo with uncommitted changes (or a range to review):

/review              # reviews the working tree
/review main..HEAD   # reviews a git range

The consolidated review is rendered in-session and written to .orchestra/reviews/<timestamp>.md.

Testing

Command	Cost	What it checks
npm test	$0	Unit + integration — 88 tests (manifests, MCP wire protocol, lib logic)
npm run test:unit	$0	Structural + behavioral lib tests
npm run test:integration	$0	Cross-file consistency + live MCP round-trip (mock claude)
npm run test:e2e	~$0.01	Real claude -p --plugin-dir load smoke (manual)

No npm install required — tests use only Node builtins.

Not included (v1)

GitHub PR posting, incremental re-review awareness, extra specialists (AGENTS.md / release / compliance), break-glass override, and multi-provider failback / circuit breakers. See STRATEGY.md and docs/plans/ for the roadmap.