reactive-loop

name: reactive-loop description: "Start the Clean Claude reactive multi-agent loop. Agents collaborate with smart routing: test failures → Dev, design flaws → Architect, spec gaps → PO. Full mesh of reactive links." user-invocable: false context: fork agent: orchestrator allowed-tools: Read, Write, Edit, Bash, Glob, Grep, Task

Reactive Loop Skill

You are initiating the Clean Claude Reactive Loop — a self-correcting multi-agent system where agents collaborate automatically to deliver features.

What Happens

When you invoke /reactive-loop:

1. Initialize — Set up .clean-claude/ shared state
2. Define — Product Owner creates user story
3. Design — Architect designs the solution
4. Implement — Frontend Engineer builds the feature
5. Verify — QA Engineer runs tests
6. Fix Loop — If tests fail, Dev fixes and QA re-verifies (up to 3 retries)
7. Complete — Feature delivered with all tests passing

Setup

First, ensure the project has the Clean Claude hooks configured:

# Check if .clean-claude exists
if [ ! -d .clean-claude ]; then
    mkdir -p .clean-claude
    echo '{"workflow":null,"feature":null,"phase":"idle","retryCount":0,"maxRetries":3,"agents":{"lastActive":null,"history":[]},"status":"ready"}' > .clean-claude/state.json
    touch .clean-claude/errors.jsonl
    touch .clean-claude/events.jsonl
    touch .clean-claude/learnings.jsonl
    echo '{}' > .clean-claude/context.json
fi

Starting the Loop

Step 1: Get Feature Description

Ask the user what feature they want to build:

• What is the feature?
• Who is it for?
• What problem does it solve?

Step 2: Initialize State

cat > .clean-claude/state.json << EOF
{
  "workflow": "feature",
  "feature": "<FEATURE_NAME>",
  "phase": "define",
  "retryCount": 0,
  "maxRetries": 3,
  "agents": {"lastActive": "orchestrator", "history": ["orchestrator"]},
  "status": "in_progress"
}
EOF

cat > .clean-claude/context.json << EOF
{
  "feature": "<FEATURE_NAME>",
  "description": "<FEATURE_DESCRIPTION>",
  "startedAt": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "requestedBy": "user"
}
EOF

Step 3: Start Phase 1 — Define

Spawn the product-owner agent to create the user story:

Use the product-owner agent to define user stories for this feature:

Feature: <FEATURE_NAME>
Description: <FEATURE_DESCRIPTION>

Create a complete user story with:
1. As a / I want / So that
2. Acceptance criteria (Given/When/Then)
3. Examples
4. Out of scope
5. Success metrics

Save the user story to: docs/features/<feature-name>/user-story.md

Step 4: Continue the Loop

After each agent completes:

1. Read state: cat .clean-claude/state.json
2. Check for trigger: cat .clean-claude/trigger 2>/dev/null
3. Spawn next agent based on phase

Phase Transitions

Error Type Detection

Step 5: Monitor Progress

Keep the user informed:

[CLEAN CLAUDE] Phase: implement (2/5)
[CLEAN CLAUDE] Agent: frontend-engineer
[CLEAN CLAUDE] Status: Implementing feature...

Step 6: Handle Errors (Smart Routing)

When any agent reports an issue, route based on error type:

6a. Test Failures → Dev

Use the <dev-agent> agent to fix the following test failure:

Error: <ERROR_MESSAGE>
File: <FILE_PATH>
Ownership: <AGENT_WHO_WROTE_IT>

Fix the error and ensure tests pass.

6b. Design Flaws → Architect

Use the architect agent to fix this design issue:

Problem: <DESIGN_ISSUE>
Detection: <HOW_IT_WAS_FOUND>
Files: <AFFECTED_FILES>

Redesign and update technical spec.

6c. Spec Gaps → Product Owner

Use the product-owner agent to complete the specification:

Missing: <EDGE_CASE_OR_REQUIREMENT>
Context: <WHERE_FOUND>
Reported by: <AGENT>

Update the spec with the missing criteria.

6d. Unclear Criteria → Product Owner

Use the product-owner agent to clarify this acceptance criterion:

Criterion: <AMBIGUOUS_CRITERION>
Question: <WHAT_NEEDS_CLARIFICATION>
Impact: <WHAT_CANT_BE_TESTED>

Clarify and update the user story.

6e. Contradictions → Product Owner

Use the product-owner agent to resolve this contradiction:

Requirement A: <FIRST_REQUIREMENT>
Requirement B: <CONFLICTING_REQUIREMENT>
Analysis: <WHY_THEY_CONFLICT>

Decide on the tradeoff and update the spec.

6f. Design Blocks → Architect

Use the architect agent to adjust the design:

Blocker: <WHAT_CANT_BE_IMPLEMENTED>
Current Design: <DESIGN_DECISION>
Suggested: <DEV_SUGGESTION>

Adjust the design to unblock implementation.

Routing Logic

┌─────────────────────────────────────────────────────────────────┐
│                    ERROR ROUTING LOGIC                          │
│                                                                 │
│  1. Parse error message                                         │
│  2. Detect error type:                                          │
│     - test_failure    → Check ownership → Route to owner dev    │
│     - type_error      → Route to architect             │
│     - design_flaw     → Route to architect             │
│     - spec_gap        → Route to product-owner                  │
│     - unclear_criteria→ Route to product-owner                  │
│     - contradiction   → Route to product-owner                  │
│     - blocked_by_design → Route to architect           │
│  3. Spawn agent with error context                              │
│  4. After fix, re-verify with qa-engineer                       │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Step 7: Complete

When all tests pass:

1. Update state to complete
2. Record any new learnings
3. Report success to user

[CLEAN CLAUDE] ✓ Feature complete!
[CLEAN CLAUDE]
[CLEAN CLAUDE] Summary:
[CLEAN CLAUDE] - User story: docs/features/<feature>/user-story.md
[CLEAN CLAUDE] - Design: docs/features/<feature>/technical-design.md
[CLEAN CLAUDE] - Tests: All passing
[CLEAN CLAUDE] - Retries: <N>
[CLEAN CLAUDE] - Learnings: <N> new patterns recorded

Error Recovery

If max retries exceeded:

[CLEAN CLAUDE] ✗ Max retries (3) exceeded
[CLEAN CLAUDE]
[CLEAN CLAUDE] Last error:
[CLEAN CLAUDE] <ERROR_DETAILS>
[CLEAN CLAUDE]
[CLEAN CLAUDE] Options:
[CLEAN CLAUDE] 1. Manually fix and run: /reactive-loop continue
[CLEAN CLAUDE] 2. Reset and start over: /reactive-loop reset
[CLEAN CLAUDE] 3. Ask for help with the specific error

Commands

Start new feature

/reactive-loop
> What feature do you want to build?

Continue after manual fix

/reactive-loop continue

Check status

/reactive-loop status

Reset

/reactive-loop reset

File Outputs

The loop creates:

docs/features/<feature-name>/
├── user-story.md        # From product-owner
├── technical-design.md  # From architect
└── qa-report.md         # From qa-engineer

.clean-claude/
├── state.json           # Current state
├── errors.jsonl         # Error history
├── events.jsonl         # Event log
├── learnings.jsonl      # Patterns learned
└── context.json         # Feature context

The Magic

The reactive loop is self-correcting:

1. QA finds bug → Automatically spawns Dev
2. Dev fixes → Automatically spawns QA
3. QA verifies → Loop until success or max retries
4. Learnings accumulate → Future fixes are faster

Each iteration makes the system smarter. Patterns that work get recorded. Errors that repeat get recognized. The agents learn from each other.

Start the loop and watch the agents collaborate.