PromptWheel

Orchestrate AI coding agents. Plan, execute, learn, repeat.

PromptWheel coordinates autonomous coding agents across your codebase — executing in parallel, learning across runs, and producing draft PRs you review before merging.

Quick Start

Inside Claude Code (recommended)

/plugin install promptwheel-ai/promptwheel

/promptwheel:run

Uses your Claude Code subscription directly. No API key needed.

From the terminal

npm install -g @promptwheel/cli
cd your-project
promptwheel init

# Quick spin while you grab coffee
promptwheel

# Afternoon run while you step away
promptwheel --hours 2

# Overnight with milestones
promptwheel --hours 8 --batch-size 30

# Human-in-the-loop planning mode
promptwheel --plan

Also works with Codex (--codex).

How It Gets Smarter

PromptWheel isn't a one-shot tool. Each session compounds on the last:

Session 1                    Session 5                    Session 20
Scout everywhere             Focus shifts to weak areas   High-yield areas targeted
Execute, learn from failures Learnings prevent repeats    Confidence tuned per category
Track what works             Drill generates trajectories Feedback loop tightening

Learnings — Failures and successes persist across sessions with temporal decay. Relevant learnings are injected into future prompts so the same mistakes aren't repeated.

Drill — In spin mode, PromptWheel auto-generates multi-step trajectories from scout proposals, sequencing related improvements and adapting ambition based on completion history.

Trajectories

For structured multi-step improvements, define a trajectory:

# .promptwheel/trajectories/harden-api.yaml
name: harden-api
description: Security hardening for API layer
steps:
  - id: input-validation
    title: Add input validation to all API endpoints
    scope: "src/api/**"
    categories: [security]
    acceptance_criteria:
      - All endpoints validate input before processing
    verification_commands:
      - npm test -- src/api

  - id: rate-limiting
    title: Add rate limiting to login and reset endpoints
    scope: "src/api/**,src/middleware/**"
    depends_on: [input-validation]
    acceptance_criteria:
      - Login endpoint rate-limited to 5 attempts per minute

/promptwheel:trajectory activate harden-auth
/promptwheel:run

Each session focuses on the current step, advancing through the DAG as acceptance criteria are met. Steps can have dependencies, scoped paths, and verification commands.

Drill Mode (Auto-Trajectories)

In spin mode, PromptWheel automatically generates multi-step trajectories from scout proposals. Each trajectory sequences related improvements into ordered steps.

promptwheel                               # Spin with drill (default)
promptwheel --no-drill                    # Spin without drill
promptwheel solo nudge --drill-pause      # Pause drill during a session

Configure in .promptwheel/config.json:

{ "auto": { "drill": { "enabled": true, "cooldownStalled": 5, "minProposals": 3 } } }

Integrations

Connect external MCP tools that run on a schedule during spin mode:

# .promptwheel/integrations.yaml
providers:
  - name: securitychecks
    command: "npx @securitychecks/mcp-server"
    tool: security_scan
    every: 5
    feed: proposals

  - name: patternstack
    command: "npx @patternstack/mcp-server"
    tool: analyze_patterns
    every: 10
    feed: learnings

Each provider is an MCP server that PromptWheel spawns and queries on cadence. Results feed back as proposals (scout output), learnings (cross-run memory), or nudges (session guidance).

Safety

• Scope enforcement — each ticket sandboxed to specific file paths
• Typecheck + tests — every change verified before merging
• Trust ladder — categories controlled; tests opt-in (--tests), safe mode (--safe)
• Spindle detection — catches oscillation, QA ping-pong, stalling, and token budget overruns
• Adversarial review — devil's advocate scoring challenges every proposal before execution
• Draft PRs by default — nothing merges without your review

Core Loop

Scout ──> Filter ──> Execute (parallel) ──> QA ──> PR
  |                                                  |
  └──── next cycle (learnings grow, drill adapts) ────┘

1. Scout — scans your codebase for improvements
2. Filter — dedup, impact scoring, adversarial review, trust ladder
3. Execute — parallel tickets in isolated git worktrees
4. QA — typecheck, tests, lint (retries with test-fix on failure)
5. PR — draft PR per ticket, or direct commits

Plugin (Claude Code)

/promptwheel:run                         Spin+drill (default)
/promptwheel:run plan                    Planning mode (approve first)
/promptwheel:run hours=4                 Timed spin
/promptwheel:run deep                    Architectural review
/promptwheel:trajectory activate <name>  Start a trajectory
/promptwheel:status                      Check progress
/promptwheel:nudge hint="focus on auth"  Steer mid-run

See Plugin Reference for all slash commands.

Commands

promptwheel                     # Spin+drill (default)
promptwheel --hours 4           # Timed spin
promptwheel --plan              # Planning mode (scout, approve, execute)
promptwheel --deep              # Architectural review
promptwheel --dry-run           # Scout only, no execution
promptwheel nudge "hint"        # Steer a running session
promptwheel status              # View current state

See CLI Reference for the complete list.

Requirements

• Node.js 18+
• Git repository
• Claude Code (plugin) or an Anthropic API key

Configuration

Optional .promptwheel/config.json. See Configuration.

FAQ

How is this different from just running Claude Code? PromptWheel is an orchestration layer. It coordinates autonomous agents in parallel, tracks learnings across runs, enforces scope isolation, deduplicates work, and progresses through structured trajectories — producing draft PRs you review before merging.

Will it break my code? Every change runs through typecheck and tests. Failed tickets are blocked, not merged. Scope enforcement sandboxes each ticket. All PRs are drafts by default.

How much does it cost? PromptWheel is free and open source. It uses your existing Claude Code subscription or Anthropic API key.

Contributing

See CONTRIBUTING.md.

License

Apache 2.0 - See LICENSE