task-dependency-patterns

Task Dependency Patterns

Overview

Claude Code 2.1.16 introduces a native Task Management System with four tools:

CC 2.1.268 caveat: TaskCreate, TaskGet, TaskUpdate, TaskList, and TodoWrite are offered only on Claude 3.x, Opus 4.0 to 4.7, Sonnet 4.0 to 4.6, and Haiku 4.5. On models outside this allowlist, these tools are absent unless the operator sets CLAUDE_CODE_ENABLE_TODO_TOOLS=1. CC honours the flag only from user or managed settings (or the shell). ork cannot ship it (CC reads only permissions from plugin settings), so treat every Task-tool call in this skill as conditional on the operator's environment.

• TaskCreate: Create new tasks with subject, description, and activeForm
• TaskUpdate: Update status (pending → in_progress → completed), set dependencies
• TaskGet: Retrieve full task details including blockers
• TaskList: View all tasks with status and dependency summary

Tasks enable structured work tracking, parallel coordination, and clear progress visibility.

When to Use

• Breaking down complex multi-step implementations
• Coordinating parallel work across multiple files
• Tracking progress on large features
• Managing dependencies between related changes
• Providing visibility into work status

Key Patterns

1. Task Decomposition

Break complex work into atomic, trackable units:

Feature: Add user authentication

Tasks:
#1. [pending] Create User model
#2. [pending] Add auth endpoints (blockedBy: #1)
#3. [pending] Implement JWT tokens (blockedBy: #2)
#4. [pending] Add auth middleware (blockedBy: #3)
#5. [pending] Write integration tests (blockedBy: #4)

2. Dependency Chains

Use addBlockedBy to create execution order:

// Task #3 cannot start until #1 and #2 complete
{"taskId": "3", "addBlockedBy": ["1", "2"]}

3. Status Workflow

pending → in_progress → completed
   ↓           ↓
(unblocked)  (active)

pending/in_progress → deleted

• pending: Task created but not started
• in_progress: Actively being worked on
• completed: Work finished and verified
• deleted: Task removed — permanently removes the task

Task Deletion

Use status: "deleted" to permanently remove tasks:

// Delete a task
{"taskId": "3", "status": "deleted"}

When to delete:

• Orphaned tasks whose blockers have all failed
• Tasks superseded by a different approach
• Duplicate tasks created in error
• Tasks from a cancelled pipeline

When NOT to delete:

• Tasks that might be retried later (keep as pending)
• Tasks with useful history (mark completed instead)
• Tasks blocked by in_progress work (wait for resolution)

4. activeForm Pattern

Provide present-continuous form for spinner display:

subject (imperative)	activeForm (continuous)
Run tests	Running tests
Update schema	Updating schema
Fix authentication	Fixing authentication

Agent Teams

Agent Teams provides multi-agent coordination with shared task lists and peer-to-peer messaging.

CC 2.1.161 — independent parallel-tool failure: A failed tool call in a parallel batch no longer cancels siblings; each returns its own result. Teammates must check task status independently and handle failures explicitly rather than assuming a batch-wide abort.

Team Workflow

1. (implicit team — CC 2.1.178+)       → one team per session; no TeamCreate
2. TaskCreate(subject, description)    → Add tasks to shared list
3. Agent(name, team_name, prompt)      → Spawn teammates into the implicit team
4. TaskUpdate(owner: "teammate-name")  → Assign tasks
5. SendMessage(to, message, summary)   → Direct teammate communication
6. (turn / background ends)            → teammates wind down; Ctrl+F x2 for bg

Cross-session replies land in the parent (CC 2.1.248): when a subagent sends SendMessage to another session, the reply is delivered to the parent session's conversation, never to the subagent; a subagent sends and moves on, the parent reads the answer. Cross-session SendMessage / ListAgents also work on Bedrock, Vertex and Foundry and with telemetry disabled (CC 2.1.248).

When to Use Teams vs Agent Tool

Criteria	Agent Tool (subagents)	Agent Teams
Independent tasks	Yes	Overkill
Cross-cutting changes	Limited	Yes
Agents need to talk	No (star topology)	Yes (mesh)
Cost sensitivity	Lower (~1x)	Higher (~2.5x)
Complexity < 3.0	Yes	No
Complexity > 3.5	Possible	Recommended

Team Task Patterns

# Spawn teammate into shared task list. A teammate that writes files works in
# a manual worktree the lead creates first (#4557); the lead merges its branch
# before dependent tasks start. No isolation on the call: a call with name and
# isolation launches a plain subagent, not a teammate (Claude Code docs).
Bash("git worktree add .worktrees/backend -b feat/my-feature-backend feat/my-feature")
Agent(
  prompt="You are the backend architect... Work only in .worktrees/backend/ "
         "(cd there, check pwd and branch first). Commit your work there and "
         "message the lead your branch name and commit sha.",
  team_name="my-feature",
  name="backend-architect",
  subagent_type="ork:backend-system-architect"
)

# Teammate claims and works tasks
TaskList → find unblocked, unowned tasks
TaskUpdate(taskId, owner: "backend-architect", status: "in_progress")
# ... do work ...
TaskUpdate(taskId, status: "completed")
TaskList → find next task

Peer Messaging

# Direct message between teammates (params: to, message, summary)
SendMessage(to: "frontend-dev",
  message: "API contract ready: GET /users/:id returns {...}",
  summary: "API contract shared")

# No broadcast primitive — send to each teammate, or post to the shared
# task list (TaskCreate/TaskUpdate) so every teammate sees it
SendMessage(to: "backend-dev",
  message: "Breaking change: auth header format changed",
  summary: "Breaking auth change")

Context Exhaustion Handling

When using Agent Teams, if context limit is reached mid-workflow:

• Collect partial results from completed teammates via TaskList
• Synthesize available outputs — prefer partial results over silent failure
• Log skipped tasks with TaskUpdate(taskId: task_id, status: "completed", metadata: {"skipped": "context limit"})

Anti-Patterns

• Creating tasks for trivial single-step work
• Circular dependencies (A blocks B, B blocks A)
• Leaving tasks in_progress when blocked
• Not marking tasks completed after finishing
• Using broadcast for messages that only concern one teammate
• Spawning teams for simple sequential work (use Agent tool instead)

Related Skills

• ork:implement - Implementation workflow with task tracking and progress updates
• ork:verify - Verification tasks and completion checklists
• ork:fix-issue - Issue resolution with hypothesis-based RCA tracking
• ork:brainstorm - Design exploration with parallel agent tasks

References

Load on demand with Read("references/<file>"):

File	Content
dependency-tracking.md	Dependency tracking patterns
status-workflow.md	Status workflow details
multi-agent-coordination.md	Multi-agent coordination