task-management-protocol

安装量: 50
排名: #14875

安装

npx skills add https://github.com/oimiragieo/agent-studio --skill task-management-protocol

Task Management Protocol Standardized protocol for task synchronization, progress tracking, and context handoff between agents and sessions using Claude Code's native task tools. Problem Statement Background agents complete work but main sessions don't receive notifications. Agents don't update task descriptions with findings. No protocol exists for structured context handoff between agents or sessions. This skill solves: Lost context when sessions end or agents complete Background agent findings not surfacing to main session Duplicate work due to poor task visibility No structured way to pass information between agents Core Tools Reference Tool Purpose When to Use TaskList() List all tasks with status Start of work, after completion TaskGet(id) Get full task details Before starting assigned task TaskCreate(...) Create new task Planning phase, discovered subtasks TaskUpdate(...) Update status/metadata Progress, discoveries, completion The Protocol Phase 1: Session Start (MANDATORY) Before doing ANY work, execute this sequence: // Step 1: Check existing tasks TaskList ( ) ; // Step 2: If assigned task exists, read full details TaskGet ( { taskId : '' } ) ; // Step 3: Claim the task TaskUpdate ( { taskId : '' , status : 'in_progress' , activeForm : 'Working on ' , } ) ; Why this matters: Prevents duplicate work (see what's already in progress) Gets full context from task description Signals to other agents/sessions that work has started Phase 2: During Work (Progress Updates) Update tasks when you: Discover important information Find blockers Identify subtasks Make significant progress Discovery Update Pattern // When you discover something important TaskUpdate ( { taskId : 'X' , description : ` ORIGINAL:

Discoveries (

${ new Date ( ) . toISOString ( ) . split ( 'T' ) [ 0 ] } ) - Found: - Files: - Impact: , metadata : { discoveredFiles : [ 'path/to/file1.ts' , 'path/to/file2.ts' ] , discoveries : [ 'Pattern X found' , 'Dependency Y required' ] , lastUpdated : new Date ( ) . toISOString ( ) , } , } ) ; Blocker Update Pattern // When you hit a blocker TaskUpdate ( { taskId : 'X' , description :

BLOCKED (

${ new Date ( ) . toISOString ( ) . split ( 'T' ) [ 0 ] } ) - Blocker: - Needs: - Workaround: , metadata : { status : 'blocked' , blocker : 'Description of blocker' , blockerType : 'dependency|permission|information|external' , needsFrom : 'user|other-agent|external-system' , } , } ) ; Subtask Creation Pattern // When you discover subtasks TaskCreate ( { subject : 'Subtask: <specific-task>' , description : Parent: Task #X

Context

Scope

Acceptance Criteria

  • [ ]
  • [ ] , activeForm : 'Working on <subtask>' , } ) ; // Link to parent TaskUpdate ( { taskId : '<new-subtask-id>' , addBlockedBy : [ 'X' ] , // This subtask blocks parent completion } ) ; Phase 3: Completion (MANDATORY) Never mark a task complete without structured metadata: TaskUpdate ( { taskId : 'X' , status : 'completed' , description :

Completed (

${ new Date ( ) . toISOString ( ) . split ( 'T' ) [ 0 ] } ) - Summary: - Files modified: - Tests: , metadata : { summary : 'Concise summary of completed work' , filesModified : [ 'path/to/file1.ts' , 'path/to/file2.ts' ] , filesCreated : [ 'path/to/new.ts' ] , testsAdded : true , testsPassing : true , outputArtifacts : [ '.claude/context/reports/backend/my-report.md' ] , nextSteps : [ 'Optional follow-up' , 'Another consideration' ] , completedAt : new Date ( ) . toISOString ( ) , } , } ) ; // Check for newly unblocked tasks TaskList ( ) ; Phase 4: Session End / Handoff Before ending a session with incomplete work: // Update all in-progress tasks with current state TaskUpdate ( { taskId : 'X' , description :

Session Paused (

${ new Date ( ) . toISOString ( ) . split ( 'T' ) [ 0 ] } ) - Progress: - Current state: - Next step: - Files to review: ` , metadata : { sessionPaused : true , progress : '60%' , currentState : 'Description of current state' , immediateNextStep : 'The very next thing to do' , keyFiles : [ 'file1.ts' , 'file2.ts' ] , keyDecisions : [ 'Decision 1' , 'Decision 2' ] , pausedAt : new Date ( ) . toISOString ( ) , } , } ) ; Context Handoff Structure Metadata Schema for Handoff Use this consistent structure for context handoff between agents: interface TaskHandoffMetadata { // Progress tracking status ? : 'not_started' | 'in_progress' | 'blocked' | 'completed' ; progress ? : string ; // e.g., "60%", "3/5 steps" // Discovery context discoveredFiles ? : string [ ] ; discoveries ? : string [ ] ; patterns ? : string [ ] ; // Blocker information blocker ? : string ; blockerType ? : 'dependency' | 'permission' | 'information' | 'external' ; needsFrom ? : string ; // Completion context summary ? : string ; filesModified ? : string [ ] ; filesCreated ? : string [ ] ; outputArtifacts ? : string [ ] ; // Continuation context currentState ? : string ; immediateNextStep ? : string ; keyFiles ? : string [ ] ; keyDecisions ? : string [ ] ; // Timestamps lastUpdated ? : string ; completedAt ? : string ; pausedAt ? : string ; } Reading Handoff Context When starting work on a task that another agent worked on: // Get full task details including metadata const task = TaskGet ( { taskId : 'X' } ) ; // Check metadata for context if ( task . metadata ?. sessionPaused ) { // Previous session paused - read currentState and immediateNextStep } if ( task . metadata ?. discoveries ) { // Previous agent found things - review discoveries array } if ( task . metadata ?. blocker ) { // Task was blocked - check if blocker is resolved } Cross-Session Coordination Environment Variable: CLAUDE_CODE_TASK_LIST_ID Use this environment variable to share task lists across sessions:

Set shared task list for all sessions

export CLAUDE_CODE_TASK_LIST_ID = "my-project-tasks"

Start claude code - will use shared task list

claude When to use: Multiple terminals working on same project Background agents that should share task state Team collaboration on task lists Shared Task List Pattern // Session A creates task TaskCreate ( { subject : 'Implement feature X' , description : '...' , metadata : { owner : 'session-a' , priority : 'high' , } , } ) ; // Session B (same CLAUDE_CODE_TASK_LIST_ID) picks up task TaskList ( ) ; // Sees task from Session A TaskUpdate ( { taskId : '1' , status : 'in_progress' , metadata : { owner : 'session-b' , // Claims ownership previousOwner : 'session-a' , } , } ) ; Iron Laws (MUST FOLLOW) 1. Never Complete Without Summary // WRONG - No context for future reference TaskUpdate ( { taskId : 'X' , status : 'completed' } ) ; // CORRECT - Full context preserved TaskUpdate ( { taskId : 'X' , status : 'completed' , metadata : { summary : 'Added auth middleware with JWT validation' , filesModified : [ 'src/middleware/auth.ts' ] , completedAt : new Date ( ) . toISOString ( ) , } , } ) ; 2. Always Update on Discovery // WRONG - Discoveries lost // ... agent finds important pattern but doesn't record it ... // CORRECT - Discoveries preserved TaskUpdate ( { taskId : 'X' , metadata : { discoveries : [ ... existingDiscoveries , 'Found circular dependency in module X' ] , } , } ) ; 3. Always TaskList After Completion // WRONG - May have unblocked other tasks TaskUpdate ( { taskId : "X" , status : "completed" } ) ; // ... session ends ... // CORRECT - Check for follow-up work TaskUpdate ( { taskId : "X" , status : "completed" , metadata : { ... } } ) ; TaskList ( ) ; // Find newly unblocked tasks 4. Use Metadata for Structure, Description for Prose // WRONG - Structured data in prose TaskUpdate ( { taskId : 'X' , description : 'Files: a.ts, b.ts. Blocked by: auth issue. Progress: 50%' , } ) ; // CORRECT - Structured metadata + prose description TaskUpdate ( { taskId : 'X' , description : 'Implementing auth flow. Hit a blocker with token refresh.' , metadata : { filesModified : [ 'a.ts' , 'b.ts' ] , blocker : 'auth issue' , progress : '50%' , } , } ) ; Integration with Memory Protocol Task metadata complements but does not replace Memory Protocol: Information Type Task Metadata Memory Files Task-specific discoveries Yes No Project-wide patterns Reference only Yes (learnings.md) Architecture decisions Reference only Yes (decisions.md) Blocking issues Yes Yes (issues.md) Progress state Yes No Completion summary Yes Yes (learnings.md) Pattern: Record task-specific context in task metadata Record project-wide learnings in memory files Cross-reference between them // Task completion with memory reference TaskUpdate ( { taskId : 'X' , status : 'completed' , metadata : { summary : 'Implemented auth flow' , memoryUpdates : [ 'learnings.md: JWT refresh pattern' , 'decisions.md: ADR-005 auth architecture' ] , } , } ) ; Examples Example 1: Developer Agent Completing Feature // Start TaskList ( ) ; TaskGet ( { taskId : '5' } ) ; TaskUpdate ( { taskId : '5' , status : 'in_progress' , activeForm : 'Implementing login flow' } ) ; // Discovery during work TaskUpdate ( { taskId : '5' , metadata : { discoveries : [ 'Existing auth module at src/auth/' , 'Uses JWT not sessions' ] , keyFiles : [ 'src/auth/jwt.ts' , 'src/middleware/auth.ts' ] , } , } ) ; // Completion TaskUpdate ( { taskId : '5' , status : 'completed' , metadata : { summary : 'Added login endpoint with JWT auth' , filesCreated : [ 'src/routes/login.ts' , 'src/routes/login.test.ts' ] , filesModified : [ 'src/routes/index.ts' ] , testsAdded : true , testsPassing : true , completedAt : new Date ( ) . toISOString ( ) , } , } ) ; TaskList ( ) ; // Check for unblocked tasks Example 2: Background Agent with Handoff // Background agent starting long task TaskUpdate ( { taskId : '10' , status : 'in_progress' , metadata : { runningInBackground : true } , } ) ; // Progress update (multiple during execution) TaskUpdate ( { taskId : '10' , metadata : { progress : '40%' , currentState : 'Processed 400/1000 files' , discoveries : [ 'Found 12 security issues' , '3 critical in auth module' ] , } , } ) ; // Completion - main session can check this TaskUpdate ( { taskId : '10' , status : 'completed' , metadata : { summary : 'Security scan complete: 12 issues found (3 critical)' , outputArtifacts : [ '.claude/context/reports/security/security-scan.md' ] , criticalFindings : 3 , highFindings : 5 , mediumFindings : 4 , completedAt : new Date ( ) . toISOString ( ) , } , } ) ; Example 3: Multi-Agent Coordination // Planner creates tasks with dependencies TaskCreate ( { subject : 'Design auth system' , description : '...' , activeForm : 'Designing auth' } ) ; TaskCreate ( { subject : 'Implement auth backend' , description : '...' , activeForm : 'Implementing auth' , } ) ; TaskCreate ( { subject : 'Add auth tests' , description : '...' , activeForm : 'Testing auth' } ) ; TaskUpdate ( { taskId : '2' , addBlockedBy : [ '1' ] } ) ; // Implementation blocked by design TaskUpdate ( { taskId : '3' , addBlockedBy : [ '2' ] } ) ; // Tests blocked by implementation // Architect completes design TaskUpdate ( { taskId : '1' , status : 'completed' , metadata : { summary : 'Auth design complete - JWT with refresh tokens' , outputArtifacts : [ '.claude/context/plans/auth-design.md' ] , keyDecisions : [ 'JWT over sessions' , 'Redis for token store' ] , } , } ) ; // Developer can now start (task 2 unblocked) TaskList ( ) ; // Shows task 2 now available TaskGet ( { taskId : '2' } ) ; // Gets design context from task 1's metadata

返回排行榜