OpenClaw Supervisor Swarm & Automation

DevDiff integrates with OpenClaw environments for local-first, privacy-respecting LLM routing, task delegation, and multi-agent swarm orchestration.

Beginning in DevDiff v1.0.5, the OpenClaw Supervisor manages automated pipelines, error recovery, package verification, and interactive human reviews.

---

📋 Supervisor configuration (supervisor-pipeline.yaml)

Orchestrate your development agents by placing a supervisor-pipeline.yaml automation script inside your repository's .devdiff/automations/ directory. This script configures roles, routing models, thresholds, and delegation rules:

version: "1.0"
pipeline:
  name: "OpenClaw Swarm Supervisor"
  defaultModel: "ollama://qwen2.5-coder:7b"

  agents:
    - name: "writer"
      role: "Changelog synthesis and formatting"
      model: "ollama://qwen2.5-coder:7b"
      confidenceThreshold: 80

    - name: "security_reviewer"
      role: "Security scanner and dependency auditor"
      model: "ollama://llama3.2:3b"
      confidenceThreshold: 90

    - name: "compliance_officer"
      role: "License and framework compliance verification"
      model: "ollama://llama3.2:3b"
      confidenceThreshold: 90

  tasks:
    - id: "diff_analysis"
      description: "Analyze git diff for logic changes"
      assignedTo: "writer"

    - id: "security_scan"
      description: "Check diffs for security vulnerabilities"
      assignedTo: "security_reviewer"
      dependsOn: "diff_analysis"

Once defined, the DevDiff core orchestrator swarms the task analysis across your local agents and synthesizes the finalized changelog.

---

🛠️ Resilient Diagnostics & Error Recovery

The OpenClaw supervisor wraps all agent tasks in a resilient error recovery handler that intercepts failures and auto-retries tasks with fallbacks.

The handler diagnoses and recovers from:

• Connection Refusal: Auto-starts local Ollama endpoints or switches to backup cloud models.
• Context Length Overflow: Intelligently prunes unnecessary diff files, removes test code, and reduces the token payload.
• Output Hallucinations: Detects if an agent returns invalid JSON format or empty outputs, re-prompting with strict schema validations.
• Model Timeouts: Retries the request using faster, lightweight local models (e.g. llama3.2:3b) to ensure completion.

---

📦 Dependency Auto-Installation

To save developer setup time, the supervisor contains an automated Dependency Manager that checks for missing local packages and models on startup:

1. Package Managers: Auto-detects active package tools (npm, pnpm, yarn, bun).
2. Missing Dependencies: Identifies missing CLI binaries or configurations.
3. Auto-Install: Promptlessly runs pnpm add or system installations to configure the environment, ensuring all agent dependencies are linked.

---

👤 Human-in-the-Loop Interruption

If an agent finishes analysis with a low confidence score, or detects a critical concern (like a security flag), the supervisor halts the pipeline and starts the Human Review System:

┌─────────────────────────────────────────────┐
│  👤 HUMAN REVIEW REQUESTED                   │
├─────────────────────────────────────────────┤
│  Type: security_finding                     │
│  Confidence: 45                             │
│  Priority: CRITICAL                         │
├─────────────────────────────────────────────┤
│  AI Output:                                 │
│  Found critical CVE vulnerability in database│
├─────────────────────────────────────────────┤
│  Evidence:                                  │
│  • package.json                             │
└─────────────────────────────────────────────┘

  Options:
  [A] Approve — Accept AI output as-is
  [R] Reject  — Discard AI output, retry with different approach
  [M] Modify  — Edit the AI output
  [D] Delegate — Send to another AI agent

Review Modes:

• Interactive (TTY): Prompts the developer in their terminal with options: [A] Approve, [R] Reject, [M] Modify, or [D] Delegate (routes to another model).
• Non-Interactive (CI/CD): When running in automated environments (like GitHub Actions), the system bypasses interactive prompts, logs warnings, and falls back to a safe confidence resolution.