devteam

DevTeam: Multi-Agent Autonomous Development System

A Claude Code plugin providing 127 specialized AI agents with:

• Interview-driven planning - Clarify requirements before work begins
• Codebase research - Investigate patterns and blockers before implementation
• SQLite state management - Reliable session tracking and cost analytics
• Model escalation - Automatic haiku → sonnet → opus progression
• Bug Council - 5-agent diagnostic team for complex issues
• Eco mode - Cost-optimized execution mode for simpler tasks
• Quality gates - Tests, types, lint, security, coverage enforcement

---

How This Works

This is a Claude Code plugin composed of:

• Markdown agent instructions (agents/*.md) — Claude Code reads these and follows them as subagent prompts via the Task tool
• YAML configuration (.devteam/*.yaml) — defines capabilities, thresholds, and agent selection triggers
• Shell scripts (scripts/*.sh, hooks/*.sh) — handle state persistence (SQLite), event logging, hook lifecycle, and database management
• Slash commands (commands/*.md, skills/*/SKILL.md) — user-facing commands that orchestrate agent workflows

There is no separate executable orchestrator. Claude Code itself is the runtime — it reads the agent markdown files, selects appropriate agents based on task characteristics, and executes them as subagents. The shell scripts provide supporting infrastructure (database, hooks, state tracking) but the orchestration logic lives in the agent instructions themselves.

---

Key Features

Autonomous Development with Task Loop

Task Loop is the iterative quality enforcement system that ensures every task is completed to specification:

┌─────────────────────────────────────────────────────────────┐
│                       TASK LOOP                              │
│                                                              │
│   Execute → Quality Gates → Pass? → Complete                 │
│      ↑           │                                           │
│      │          Fail                                         │
│      │           ↓                                           │
│      └─── Fix Tasks ← Model Escalation (if needed)          │
│                                                              │
│   Loop until: ALL QUALITY GATES PASS                        │
└─────────────────────────────────────────────────────────────┘

Features:

• Automatic model escalation (haiku → sonnet → opus) after failures
• Stuck loop detection with Bug Council activation
• Quality gates: tests, types, lint, security, coverage
• Anti-abandonment system prevents agents from giving up
• Maximum 10 iterations with human notification

Intelligent Agent Selection

The /devteam:implement command automatically selects the best agents for your task:

/devteam:implement "Add user authentication with JWT tokens"

The /devteam:implement command analyzes your task description, file types involved, and project context to select appropriate agents. It considers keyword matches, file extensions, task type (feature vs. bug), and detected language/framework.

Bug Council: Multi-Perspective Debugging

For complex bugs, the Bug Council convenes 5 specialized analysts:

┌─────────────────────────────────────────────────────────────┐
│                       BUG COUNCIL                            │
├─────────────────────────────────────────────────────────────┤
│  Root Cause Analyst    │ Error analysis, stack traces       │
│  Code Archaeologist    │ Git history, regression detection  │
│  Pattern Matcher       │ Similar bugs, anti-patterns        │
│  Systems Thinker       │ Dependencies, integration issues   │
│  Adversarial Tester    │ Edge cases, security vectors       │
└─────────────────────────────────────────────────────────────┘
                              ↓
                    Synthesized Solution

Activation Triggers:

• Critical/high severity bugs
• 3+ failed opus attempts
• Complexity score ≥ 10
• Explicit bug_council: true flag

Scope Enforcement

Agents are strictly confined to their assigned scope:

scope:
  allowed_files:
    - "src/auth/*.py"
  forbidden_directories:
    - "src/billing/"
  max_files_changed: 5

6 Enforcement Layers:

1. Task scope definition in YAML
2. Agent prompt constraints
3. Scope validator agent with VETO power
4. Pre-commit hook blocking
5. Runtime file access control
6. Out-of-scope observations logging

Anti-Abandonment System

Agents cannot give up. The persistence system ensures completion:

Abandonment Attempt → Detected → Re-engagement Prompt
         ↓
    Still stuck?
         ↓
    Model Escalation (haiku → sonnet → opus)
         ↓
    Still stuck?
         ↓
    Bug Council Activation
         ↓
    Still stuck?
         ↓
    Human Notification (but keep trying)

---

127 Specialized Agents

Enterprise Roles

Category	Agents	Capabilities
SRE	Site Reliability Engineer	SLOs, incident response, chaos engineering
SRE	Platform Engineer	Internal developer platforms, golden paths
SRE	Observability Engineer	Metrics, logging, tracing, alerting
Security	Penetration Tester	OWASP testing, API security, vuln assessment
Security	Compliance Engineer	SOC2, HIPAA, GDPR, PCI-DSS
Product	Product Manager	PRDs, roadmaps, user research
Quality	Accessibility Specialist	WCAG 2.1, screen readers, inclusive design
DevRel	Developer Advocate	Technical content, community, DX

Orchestration Agents

Agent	Purpose
Autonomous Controller	Execution loop management, state transitions, circuit breaker
Bug Council Orchestrator	Multi-perspective bug analysis
Code Review Coordinator	Cross-agent code review orchestration
Quality Gate Enforcer	Run and aggregate quality gate results
Requirements Validator	Validate acceptance criteria met
Scope Validator	Enforce scope boundaries
Sprint Loop	Sprint-level quality validation after all tasks
Sprint Orchestrator	Sprint execution management, task sequencing
Task Loop	Iterative quality loop for single task execution
Track Merger	Merge parallel worktree tracks
Workflow Compliance	Meta-validator auditing orchestration process

Bug Council Agents

Agent	Specialty
Root Cause Analyst	Error analysis, hypothesis generation
Code Archaeologist	Git history, regression detection
Pattern Matcher	Similar bugs, anti-pattern identification
Systems Thinker	Dependencies, architectural issues
Adversarial Tester	Edge cases, security vulnerabilities

Implementation Agents

Backend (by language):

• Python (FastAPI, Django, Flask)
• TypeScript (Express, NestJS, Fastify)
• Go (Gin, Echo, Fiber)
• Java (Spring Boot, Micronaut)
• C# (ASP.NET Core)
• Ruby (Rails, Sinatra)
• PHP (Laravel, Symfony)

Frontend:

• React, Vue, Svelte, Angular specialists
• Accessibility specialist
• Performance auditor

Database:

• Schema designers
• Query optimization specialists
• Migration specialists

DevOps:

• CI/CD Specialist (GitHub Actions, Jenkins, GitLab CI)
• Docker Specialist
• Kubernetes Specialist
• Terraform Specialist

Quality Agents

Agent	Focus
Test Writer	Unit, integration, e2e tests
Security Auditor	OWASP Top 10, vulnerability scanning
Performance Auditor	Profiling, optimization, load testing
Accessibility Specialist	WCAG compliance, inclusive design
E2E Tester	Playwright, Cypress, browser testing

---

Quick Start

Planning + Implementation (Recommended)

# Plan a new feature (interview → research → PRD → tasks → sprints)
/devteam:plan --feature "Add user authentication with OAuth"

# Execute the plan
/devteam:implement

# Or execute specific sprint
/devteam:implement --sprint 1

The system will:

1. Interview - Clarify requirements with targeted questions
2. Research - Analyze codebase, identify patterns and blockers
3. Plan - Generate PRD, tasks, and sprints
4. Execute - Run with Task Loop quality loop and model escalation
5. Verify - Pass all quality gates before completion

Bug Fixing

# Fix a local bug (interview → diagnose → fix → verify)
/devteam:bug "Login fails for guest users"

# Fix a GitHub issue
/devteam:issue 123

# Force Bug Council for complex issues
/devteam:bug "Memory leak in image processor" --council

Cost-Optimized Mode

# Use eco mode (lower-cost models for simpler tasks)
/devteam:implement --eco
/devteam:bug "Minor CSS issue" --eco

Monitoring

# Check status, costs, progress
/devteam:status

# List plans and tasks
/devteam:list

# Reset stuck sessions
/devteam:reset

---

Configuration

Task Loop Configuration (.devteam/task-loop-config.yaml)

loop_settings:
  max_iterations: 10

model_escalation:
  enabled: true
  consecutive_failures:
    haiku_to_sonnet: 2
    sonnet_to_opus: 2
    opus_max_failures: 3  # Then Bug Council

quality_gates:
  required:
    tests_passing: true
    type_check: true
    lint: true
  security:
    on_finding: create_fix_task

Scope Definition (per task)

# In task definition
scope:
  allowed_files:
    - "src/auth/*.py"
  allowed_patterns:
    - "tests/auth/**/*.py"
  forbidden_directories:
    - "src/billing/"
    - "src/admin/"
  max_files_changed: 10

Agent Selection (.devteam/agent-capabilities.yaml)

categories:
  security:
    agents:
      - id: penetration_tester
        triggers:
          keywords: [pentest, security testing, vulnerability]
          task_types: [security_testing]

---

Hooks System

The system uses Claude Code hooks for autonomous execution. All hooks support both Linux/macOS (Bash) and Windows (PowerShell).

Hook	Linux/macOS	Windows	Purpose
Stop Hook	stop-hook.sh	stop-hook.ps1	Blocks exit without EXIT_SIGNAL: true
Persistence Hook	persistence-hook.sh	persistence-hook.ps1	Detects and prevents abandonment
Scope Check	scope-check.sh	scope-check.ps1	Validates commits stay in scope
Pre-Compact	pre-compact.sh	pre-compact.ps1	Preserves state before context compaction
Pre-Tool-Use	pre-tool-use-hook.sh	pre-tool-use-hook.ps1	Pre-execution validation
Post-Tool-Use	post-tool-use-hook.sh	post-tool-use-hook.ps1	Post-execution logging
Session Start	session-start.sh	session-start.ps1	Session initialization
Session End	session-end.sh	session-end.ps1	Session cleanup
Install	install.sh	install.ps1	Hook installation script

When installed via the marketplace or as a plugin, all hooks are configured automatically through hooks/hooks.json. No manual settings.json editing is required.

See hooks/README.md for hook details and troubleshooting.

---

Model Tiers & Cost Optimization

Automatic Model Selection

Complexity	Model	Characteristics	Use Case
1-4	Haiku	Fast, lowest cost	Simple fixes, docs
5-8	Sonnet	Balanced	Standard features
9-14	Opus	Most capable, highest cost	Complex architecture

Escalation Flow