Teamwork Skill

This skill enables dynamic team creation and management for executing complex engineering tasks through coordinated AI agents with intelligent model selection, cost optimization, and continuous performance evaluation.

When to Invoke

Invoke this skill when:

• - User requests execution of complex projects requiring multiple specialized roles
• Tasks need to be broken down into coordinated steps (analysis, design, implementation, testing, review)
• User wants to leverage multiple AI models/providers for optimal cost-performance balance
• Projects require structured workflow with quality assurance and iteration

Initialization & Configuration Management

Automatic Initialization

IMPORTANT: This skill includes an autonomous initialization system. When invoked for the first time or when configuration is missing, it will automatically:

1. 1. Check Configuration Status

- Verify if .trae/config/providers.json exists - Verify if .trae/config/team-roles.json exists - Verify if .trae/data/model_scores.json exists

1. 2. Interactive Setup Process

If configuration files are missing or incomplete, the skill will proactively ask the user:

Step 1: Provider Setup
- Ask: "Which AI providers would you like to configure? (e.g., OpenAI, Anthropic, Google, Azure, etc.)"
- For each provider, collect:
- Provider name
- API key (or environment variable name)
- Base URL (if custom endpoint)

Step 2: Model Configuration
For each provider, ask:
- "Which models from [provider] would you like to use?"
- For each model, collect:
- Model name/identifier
- Pricing model type (subscription/tieredusage/payper_use)
- Pricing details based on type:
- Subscription: cost, start date, end date
- Tiered Usage: daily quota, monthly quota, overage rate
- Pay-Per-Use: input cost per 1k tokens, output cost per 1k tokens
- Capabilities (e.g., reasoning, coding, fast-response)
- Maximum concurrent tasks

Step 3: Host Model Selection
- Ask: "Which model should serve as the primary interface (host model)?"
- Present list of configured models
- User selects one as the main interaction point

Step 4: Budget Configuration
- Ask: "What is your monthly budget limit? (optional)"
- Set alert thresholds

1. 3. Configuration Persistence

- Save all configurations to .trae/config/providers.json - Create default role definitions in .trae/config/team-roles.json - Initialize empty scores database in .trae/data/model_scores.json - Confirm successful setup with user

Configuration Management Commands

Users can manage their configuration at any time using these commands:

View Configuration

User: "Show me my current provider and model configuration"

Response: Display complete configuration from INLINECODE6

Add Provider

User: "Add a new provider: [provider name]"

Action: Interactive prompts for provider details, then append to configuration

Add Model

User: "Add model [model name] to provider [provider name]"

Action: Interactive prompts for model details, then add to provider's model list

Update Model Pricing

User: "Update pricing for [model name]"

Action: Ask for new pricing details and update configuration

Remove Model

User: "Remove model [model name] from provider [provider name]"

Action: Confirm and remove from configuration

Change Host Model

User: "Change the host model to [model name]"

Action: Update host_model configuration

View Model Scores

User: "Show me the performance scores for all models"

Response: Display current model capability scores from INLINECODE7

Reset Configuration

User: "Reset all configurations to default"

Action: Confirm with user, then reinitialize

Configuration File Structure

Provider Configuration (.trae/config/providers.json)
CODEBLOCK8

Team Roles Configuration (.trae/config/team-roles.json)
CODEBLOCK9

Model Scores Database (.trae/data/model_scores.json)
CODEBLOCK10

Initialization Checklist

Before executing any team task, verify:

• - [ ] .trae/config/providers.json exists and contains at least one provider
• [ ] At least one model is configured
• [ ] Host model is designated
• [ ] .trae/config/team-roles.json exists with role definitions
• [ ] .trae/data/model_scores.json exists (can be empty initially)

If any checklist item fails, trigger interactive initialization.

Core System Components

1. Model Performance Evaluation System

Multi-Dimensional Scoring

All models are periodically evaluated by peer models across multiple dimensions:

Evaluation Dimensions:

• - Response Speed: How quickly the model responds to requests
• Response Frequency: Rate of successful responses within time windows
• Thinking Depth: Quality of reasoning and problem-solving approach
• Multi-threading Capability: Ability to handle parallel tasks
• Code Quality: Quality of generated code (for coding tasks)
• Creativity: Novelty and innovation in solutions
• Reliability: Consistency in performance across sessions
• Context Understanding: Ability to maintain context over long conversations

Scoring Mechanism:

• - Each model scores other models on a scale (e.g., 1-10) for each dimension
• Scores are aggregated using weighted average
• Evaluations occur after each task completion
• Historical scores are maintained with decay factor for recent performance
• Final capability score = weighted sum of all dimension scores

Score Storage:
CODEBLOCK11

2. Cost Calculation System

Pricing Models

Subscription-Based (订阅制)

• - Fixed cost for unlimited usage during subscription period
• Lowest effective cost per request when fully utilized
• Marked as expired after subscription ends → excluded from team
• Configuration:

CODEBLOCK12

Tiered Usage (阶段用量制)

• - Lower cost with daily/monthly quotas
• Medium cost effectiveness
• Must monitor quota usage daily
• Configuration:

CODEBLOCK13

Pay-Per-Use (用量计费制)

• - Highest cost per request
• No quota limits
• Best for sporadic or overflow usage
• Configuration:

CODEBLOCK14

Cost Score Calculation:
CODEBLOCK15

3. Model Availability Management

Status Tracking:

• - available: Ready to accept tasks
• INLINECODE15: Currently processing tasks
• INLINECODE16: Subscription expired
• INLINECODE17: Daily/monthly quota reached
• INLINECODE18: Temporarily unavailable due to rate limits
• INLINECODE19: Provider API unavailable

Busy State Management:
CODEBLOCK16

Task Execution Workflow

Phase 1: User Request & Requirement Analysis

Step 1.1: User submits request to Host Model (主模型)

• - User interacts with designated host model (primary interface)
• Host model receives and acknowledges the request

Step 1.2: Host model decomposes requirements

• - Break down request into phases and subtasks
• Identify dependencies between tasks
• Estimate complexity and required capabilities
• Create initial task tree

Step 1.3: User confirmation

• - Present task breakdown to user
• Clarify ambiguities and refine requirements
• Get explicit approval to proceed

Output:
CODEBLOCK17

Phase 2: Team Assembly Meeting

Step 2.1: Host model convenes all available models

• - Filter models by status (exclude expired, offline, quota_exceeded)
• Check busy models for potential availability
• Send meeting invitation to all eligible models

Step 2.2: Task briefing

• - Host model presents all task content to all models
• Share task breakdown, requirements, and constraints
• Distribute context and background information

Step 2.3: Collaborative role definition

• - All models discuss and agree on required roles
• Define capability requirements for each role
• Estimate workload for each role
• Identify potential bottlenecks

Meeting Output:
CODEBLOCK18

Phase 3: Role Assignment

Step 3.1: Self-nomination

• - Each model evaluates own suitability based on:

- Current cost score
- Current capability score
- Current workload (busy status)
- Role requirements match

• - Models submit role preferences

Step 3.2: Conflict resolution

• - If multiple models want same role: democratic voting
• If role has no candidates: negotiate with best-fit models
• Balance workload distribution across models
• Avoid concentrating all tasks on single model

Step 3.3: Final assignment

• - Confirm role assignments with all models
• Document assignment rationale
• Update model busy status

Assignment Algorithm:
CODEBLOCK19

Combined Score Calculation:
CODEBLOCK20

Phase 4: Herald Selection & Communication Setup

Step 4.1: Select Herald (传令官)

• - Choose fastest responding model (not necessarily most capable)
• Herald acts as central communication hub
• All models communicate through herald

Herald Responsibilities:

• - Relay messages between all team members
• Distribute progress updates
• Broadcast requirements and instructions
• Collect and aggregate results
• Monitor task completion status
• Report status to host model
• Handle timeout and failure notifications

Step 4.2: Communication channels
CODEBLOCK21

Herald Configuration:
CODEBLOCK22

Phase 5: Task Execution

Step 5.1: Parallel execution

• - Assigned models work on their respective tasks
• Regular progress updates to herald
• Herald broadcasts relevant updates to team

Step 5.2: Coordination

• - Herald checks task status periodically
• Identifies blockers and delays
• Facilitates inter-model communication
• Escalates issues to host model

Step 5.3: Progress tracking
CODEBLOCK23

Phase 6: Task Completion & Review Meeting

Step 6.1: Completion notification

• - Herald confirms all tasks completed
• Collects final outputs from all models
• Aggregates results

Step 6.2: Summary meeting

• - Host model convenes all participating models
• Each model presents their contribution
• Discuss challenges and solutions
• Evaluate collaboration effectiveness

Step 6.3: Performance re-evaluation

• - Models rate each other's performance
• Update capability scores based on task execution
• Record role-model fit assessments
• Update model scores database

Evaluation Form:
CODEBLOCK24

Phase 7: Failure Handling & Iteration

Step 7.1: Failure detection

• - Herald detects task failure or timeout
• Collects failure information from relevant models
• Reports to host model with detailed context

Step 7.2: Failure analysis meeting

• - Convene all participating models
• Analyze root cause of failure
• Identify contributing factors
• Propose solutions

Step 7.3: User consultation

• - Host model presents failure analysis to user
• Discuss potential solutions:

- Requirement changes
- Approach modifications
- Team reconfiguration
- Additional resources

• - Get user decision on next steps

Step 7.4: Iteration or termination

• - If user approves changes: restart from appropriate phase
• If user terminates: document lessons learned
• Update model scores based on partial performance

Configuration Files

Provider Configuration (.trae/config/providers.json)

CODEBLOCK25

Team Roles Configuration (.trae/config/team-roles.json)

CODEBLOCK26

Model Scores Database (.trae/data/model_scores.json)

CODEBLOCK27

Best Practices

Model Selection

• - Match model capabilities to task requirements
• Consider cost-effectiveness for routine tasks
• Reserve high-capability models for complex reasoning
• Distribute workload to prevent bottlenecks

Communication

• - Keep messages concise and clear
• Use structured formats for inter-model communication
• Herald should batch non-urgent updates
• Escalate critical issues immediately

Performance Optimization

• - Cache frequently used context
• Batch similar requests when possible
• Monitor quota usage proactively
• Maintain backup models for critical roles

Quality Assurance

• - Always conduct review meetings
• Update scores after each task
• Learn from failures systematically
• Continuously refine role definitions

Error Handling

Provider Failures

• - Retry with exponential backoff
• Switch to backup provider
• Notify team of delays
• Update model availability status

Task Failures

• - Capture detailed error context
• Analyze root cause with team
• Propose remediation strategies
• Consult user for major changes

Communication Failures

• - Herald implements heartbeat checks
• Fallback to direct model-to-model communication
• Reassign herald if unresponsive
• Log all communication issues

Output Format

Final deliverables include:

• - Complete task execution report
• Team composition and role assignments
• Individual model performance metrics
• Cost breakdown and usage statistics
• Updated model capability scores
• Lessons learned and recommendations

Skill Structure & Components

Directory Structure

CODEBLOCK28

Scripts Reference

init.js - Initialization Module

Purpose: Handles skill initialization and configuration loading.

Key Functions:

• - ensureDirectories() - Create required directories
• INLINECODE24 - Verify configuration status
• INLINECODE25 - Create default role definitions
• INLINECODE26 - Initialize empty scores database
• INLINECODE27 - Initialize empty providers config
• INLINECODE28 - Check if initialization is required
• INLINECODE29 - Read JSON configuration file
• INLINECODE30 - Write JSON configuration file

Usage:
CODEBLOCK29

config-manager.js - Configuration Management

Purpose: Manage provider and model configurations.

Key Functions:

• - addProvider(config, providerInfo) - Add new provider
• INLINECODE32 - Add model to provider
• INLINECODE33 - Remove model
• INLINECODE34 - Remove provider
• INLINECODE35 - Update pricing
• INLINECODE36 - Set host model
• INLINECODE37 - Set budget limits
• INLINECODE38 - Get list of available models
• INLINECODE39 - Get model availability status
• INLINECODE40 - Display current configuration

Usage:
CODEBLOCK30

score-manager.js - Performance Score Management

Purpose: Manage model performance evaluation scores.

Key Functions:

• - initializeModelScore(scores, modelName, provider) - Initialize model scores
• INLINECODE42 - Update dimension score
• INLINECODE43 - Calculate weighted overall score
• INLINECODE44 - Update role fit score
• INLINECODE45 - Get top models for a role
• INLINECODE46 - Get models by capability
• INLINECODE47 - Record complete evaluation
• INLINECODE48 - Display all model scores

Usage:
CODEBLOCK31

team-coordinator.js - Team Coordination

Purpose: Coordinate team assembly and task execution.

Key Class: INLINECODE49

Methods:

• - load() - Load configurations
• INLINECODE51 - Get available models list
• INLINECODE52 - Select fastest model as herald
• INLINECODE53 - Assign roles to models
• INLINECODE54 - Calculate selection score
• INLINECODE55 - Create task execution plan
• INLINECODE56 - Generate meeting agenda
• INLINECODE57 - Generate peer evaluation forms
• INLINECODE58 - Generate task report

Usage:
CODEBLOCK32

herald.js - Communication System

Purpose: Manage inter-model communication and coordination.

Key Class: INLINECODE59

Methods:

• - initializeTeam(team) - Initialize team status tracking
• INLINECODE61 - Broadcast message to all
• INLINECODE62 - Send direct message
• INLINECODE63 - Update task progress
• INLINECODE64 - Get current team status
• INLINECODE65 - Check for timeout conditions
• INLINECODE66 - Request status from all members
• INLINECODE67 - Send status report to host
• INLINECODE68 - Notify failure
• INLINECODE69 - Notify completion
• INLINECODE70 - Get overall task progress

Usage:
CODEBLOCK33

Templates Reference

task-report.md - Task Execution Report

Purpose: Document complete task execution details.

Variables:

• - task_id - Unique task identifier
• INLINECODE72 - Report generation time
• INLINECODE73 - Task status (completed/failed/in_progress)
• INLINECODE74 - Executive summary
• INLINECODE75 - Array of team member details
• INLINECODE76 - Array of execution phases
• INLINECODE77 - Performance metrics per model
• INLINECODE78 - Total execution cost
• INLINECODE79 - Array of deliverables
• INLINECODE80 - Lessons learned
• INLINECODE81 - Recommendations
• INLINECODE82 - Model score updates

Usage:
CODEBLOCK34

meeting-minutes.md - Meeting Documentation

Purpose: Document team meeting discussions and decisions.

Variables:

• - meeting_id - Unique meeting identifier
• INLINECODE84 - Type of meeting
• INLINECODE85 - Meeting date
• INLINECODE86 - Meeting duration
• INLINECODE87 - Array of participants
• INLINECODE88 - Meeting agenda
• INLINECODE89 - Voting results (if applicable)
• INLINECODE90 - Action items from meeting
• INLINECODE91 - Next steps to take

failure-report.md - Failure Analysis

Purpose: Document and analyze task failures.

Variables:

• - task_id - Failed task identifier
• INLINECODE93 - Time of failure
• INLINECODE94 - Type of failure
• INLINECODE95 - Failure severity
• INLINECODE96 - Timeline of events
• INLINECODE97 - Root cause
• INLINECODE98 - Contributing factors
• INLINECODE99 - Actions taken for recovery
• INLINECODE100 - Recommendations to prevent recurrence

evaluation-form.md - Model Evaluation

Purpose: Document peer model evaluations.

Variables:

• - evaluator_model - Evaluating model
• INLINECODE102 - Model being evaluated
• INLINECODE103 - Related task
• INLINECODE104 - Role in task
• INLINECODE105 through context_understanding - Dimension scores
• INLINECODE107 - Overall role fit assessment
• INLINECODE108 - Model strengths
• INLINECODE109 - Areas for improvement

Utilities Reference

helpers.js - General Utilities

Functions:

• - generateId(prefix) - Generate unique identifier
• INLINECODE111 - Format date to ISO string
• INLINECODE112 - Format milliseconds to readable duration
• INLINECODE113 - Calculate API cost
• INLINECODE114 - Deep clone object
• INLINECODE115 - Deep merge objects
• INLINECODE116 - Retry with exponential backoff
• INLINECODE117 - Split array into chunks
• INLINECODE118 - Group array by key
• INLINECODE119 - Sort array by key
• INLINECODE120 - Remove duplicates by key

logger.js - Logging System

Classes: INLINECODE121

Log Levels: DEBUG, INFO, WARN, ERROR

Methods:

• - debug(message, data) - Log debug message
• INLINECODE123 - Log info message
• INLINECODE124 - Log warning message
• INLINECODE125 - Log error message
• INLINECODE126 - Set log level
• INLINECODE127 - Set log file path

Usage:
CODEBLOCK35

errors.js - Custom Errors

Error Classes:

• - ValidationError - Input validation errors
• INLINECODE129 - Configuration errors
• INLINECODE130 - Model not found errors
• INLINECODE131 - Provider not found errors
• INLINECODE132 - Task execution errors
• INLINECODE133 - Timeout errors
• INLINECODE134 - Budget exceeded errors
• INLINECODE135 - Quota exceeded errors
• INLINECODE136 - Herald communication errors

Functions:

• - handleError(error, logger) - Standardized error handling
• INLINECODE138 - Check if error is recoverable

API Reference

Quick Start

CODEBLOCK36

Version History

• - v1.0.0 (2026-02-12): Initial release with full feature set

- Multi-provider support - Three pricing models - 8-dimension performance evaluation - Herald communication system - Complete workflow management - Template-based reporting