Chorus
Your AI team, working together. Parallel multi-agent orchestration powered by GitHub Copilot.
Chorus lets you start a coordinated team of AI agents inside VS Code. Give them a goal, and they decompose it into tasks, work in parallel, resolve conflicts, and apply changes to your workspace — with minimal intervention from you.
Features
- Multi-agent orchestration — A Lead agent plans the work, Worker agents execute tasks in parallel, a Change Guardian arbitrates conflicts, and a QA Verifier validates results.
- Dependency-aware task graph — Tasks are organized as a DAG with automatic unblocking when dependencies complete.
- Self-claiming workers — Workers pick up ready tasks from the queue and execute them autonomously.
- Conflict resolution — The Change Guardian detects file conflicts and classifies them into three lanes: auto-pass, deferred, or hard-stop.
- QA validation gate — The QA Verifier blocks premature completion by running targeted checks on changed files.
- Git worktree isolation — Each worker operates in its own git worktree, preventing file conflicts between parallel agents.
- Runtime and budget controls — Set a countdown timer and token budget before starting a run. Limits are enforced automatically.
- In-panel settings — Tune models, Guardian strictness, QA, worktrees, and limits from a built-in settings panel without leaving the sidebar.
- Live header controls — Pause, resume, stop, and reset the team directly from the panel header.
- Agent archiving — Completed workers are automatically archived so you can review their full activity log after they finish.
- Crash recovery — State snapshots allow the team to resume after a reload.
Requirements
- VS Code 1.96.0 or later
- GitHub Copilot Chat extension installed and signed in
Getting Started
- Install the extension from the VS Code Marketplace.
- Open the Chorus panel in the sidebar.
- Optionally adjust the runtime timer and token budget.
- Type your goal into the input box and press Enter.
- Use the panel title bar buttons to control the team:
- Pause — freeze all agent activity
- Resume — continue a paused team
- Stop — halt the team and save state
- Clear — stop and reset everything
All controls live directly in the Chorus panel — no Command Palette needed.
Settings
Most settings are available directly in the Settings tab inside the Chorus panel. You can also configure the extension under Chorus in VS Code Settings (Ctrl+, / Cmd+,):
| Setting |
Type |
Default |
Description |
chorus.modelSelection |
string |
claude-opus-4-6-1m |
Default model family for the Lead agent |
chorus.maxWorkers |
number |
10 |
Maximum concurrent worker agents (1–10) |
chorus.autoRunChecks |
boolean |
true |
Automatically run lint/test/build checks |
chorus.guardianStrictness |
enum |
Balanced |
Change Guardian strictness (Conservative, Balanced, Fast) |
chorus.retryLimit |
number |
10 |
Maximum retries for failed tasks (0–50) |
chorus.verboseLogs |
boolean |
false |
Enable verbose logging output |
In-panel settings
The panel's built-in settings tab provides additional runtime controls:
- Model tiers — Assign different models to fast, standard, and strong tiers.
- Token budget — Cap total token spend across all agents.
- Run timer — Set a countdown (30 min – 24 hrs) that auto-stops the team.
- Max workers — Limit concurrent worker agents.
- Guardian strictness — Adjust conflict detection sensitivity.
- Auto QA — Toggle automatic quality checks on changed files.
- Git worktrees — Toggle per-worker worktree isolation.
Agent Roles
Lead Agent
Plans work by decomposing goals into a dependency-aware task graph, monitors progress, resolves escalations, and synthesizes status. Only asks for help when genuinely blocked.
Worker Agents
Dynamically created by the Lead (1–10 based on goal complexity). Workers self-claim ready tasks, execute them autonomously, and submit change intents that flow through the Guardian. Each worker can run in its own git worktree to prevent file collisions. Workers are automatically archived when their task completes so you can revisit their logs later.
Change Guardian
Inspects every proposed change and classifies it:
- Auto-pass — Uncontested, low-risk changes are approved immediately.
- Deferred — Contested but manageable changes are held briefly, then re-evaluated.
- Hard-stop — High-risk conflicts are blocked and escalated to the Lead.
QA Verifier
Runs targeted validation on changed files. If issues are found, it creates follow-up fix tasks and blocks the team from declaring completion prematurely.
Known Limitations
- No git commit/push automation — all changes stay local.
- State recovery on restart is best-effort, not transactionally durable.
- Single workspace only.
- Timer and budget are client-side estimates; actual API usage may vary.
For the complete list, see docs/known-limitations.md.
License
MIT
