Programmatic Task Management for Claude Code

Five verified approaches for interacting with Claude Code's task/todo system from code. Each suits different use cases — from lightweight shell scripts to full orchestration systems.

Quick Reference: Which Approach to Use

Goal	Approach	Language
Watch task progress in real time	Agent SDK message stream	Python / TypeScript
React to task changes with side effects	Hooks (PostToolUse on TodoWrite)	Shell / Python
Block task completion until criteria met	Hooks (TaskCompleted)	Shell / Python / LLM
Add custom task-like tools Claude can call	Custom MCP tools	Python / TypeScript
Parse task updates from CLI output	Headless CLI streaming	Any (parses NDJSON)
Delegate work to specialized sub-agents	Subagents via Agent SDK	Python / TypeScript

Approach 1: Agent SDK Message Stream

The most direct way to monitor tasks programmatically. The query() function returns an async generator. Every TodoWrite call appears as a tool_use block in the stream.

Install:

bash

npm install @anthropic-ai/claude-agent-sdk   # TypeScript
pip install claude-agent-sdk                  # Python

TodoWrite Schema

typescript

// Input: complete replacement of the todo list each time
interface TodoWriteInput {
  todos: Array<{
    content: string;                              // Imperative form: "Fix the bug"
    status: 'pending' | 'in_progress' | 'completed';
    activeForm: string;                           // Present participle: "Fixing the bug"
  }>;
}

// Output: confirmation with stats
interface TodoWriteOutput {
  message: string;
  stats: { total: number; pending: number; in_progress: number; completed: number };
}

Only one todo should be in_progress at any time. The list is replaced wholesale on each call (not diffed).

TypeScript: Monitor Todos

typescript

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Refactor the auth module and track your progress",
  options: { maxTurns: 15 }
})) {
  if (message.type === "assistant") {
    for (const block of message.message.content) {
      if (block.type === "tool_use" && block.name === "TodoWrite") {
        const todos = block.input.todos;
        console.log("--- Todo Update ---");
        todos.forEach((todo, i) => {
          const icon = { completed: "done", in_progress: ">>", pending: "  " }[todo.status];
          console.log(`  [${icon}] ${todo.content}`);
        });
      }
    }
  }
}

Python: Monitor Todos

python

from claude_agent_sdk import query, AssistantMessage, ToolUseBlock

async for message in query(
    prompt="Refactor the auth module and track your progress",
    options={"max_turns": 15}
):
    if isinstance(message, AssistantMessage):
        for block in message.content:
            if isinstance(block, ToolUseBlock) and block.name == "TodoWrite":
                todos = block.input["todos"]
                for todo in todos:
                    status_icon = {"completed": "done", "in_progress": ">>", "pending": "  "}
                    print(f"  [{status_icon[todo['status']]}] {todo['content']}")

For a full real-time progress tracker class with change detection, see references/progress-tracker.md.

Approach 2: Hooks

Hooks fire at lifecycle points and are configured in .claude/settings.json (project-scoped) or ~/.claude/settings.json (global). Three hook handler types exist: command (shell script), prompt (single LLM call), and agent (multi-turn subagent with tool access).

Intercept TodoWrite Updates

Use PostToolUse with matcher "TodoWrite" to react after a todo update, or PreToolUse to validate/block before it happens.

Log every todo update (.claude/settings.json):

json

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "TodoWrite",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/log-todos.sh"
          }
        ]
      }
    ]
  }
}

.claude/hooks/log-todos.sh:

bash

#!/bin/bash
INPUT=$(cat)
TODOS=$(echo "$INPUT" | jq '.tool_input.todos')
TIMESTAMP=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
echo "{\"timestamp\": \"$TIMESTAMP\", \"todos\": $TODOS}" >> "$CLAUDE_PROJECT_DIR/.claude/todo-log.jsonl"
exit 0

Block certain todos via PreToolUse:

bash

#!/bin/bash
INPUT=$(cat)
FORBIDDEN=$(echo "$INPUT" | jq -r '.tool_input.todos[].content' | grep -i "deploy to production")
if [ -n "$FORBIDDEN" ]; then
  jq -n '{hookSpecificOutput:{hookEventName:"PreToolUse",permissionDecision:"deny",permissionDecisionReason:"Production deploys require manual approval"}}'
else
  exit 0
fi

TaskCompleted Hook

Fires when any task is being marked as completed. Exit code 2 blocks completion and feeds stderr back to the agent as corrective feedback. No matcher support — fires on every completion.

Input fields: task_id, task_subject, task_description (optional), teammate_name (optional), team_name (optional).

Require tests to pass:

json

{
  "hooks": {
    "TaskCompleted": [
      {
        "hooks": [{ "type": "command", "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/require-tests.sh" }]
      }
    ]
  }
}

bash

#!/bin/bash
INPUT=$(cat)
TASK=$(echo "$INPUT" | jq -r '.task_subject')
if ! npm test 2>&1; then
  echo "Tests failing. Fix before completing: $TASK" >&2
  exit 2
fi
exit 0

LLM-based completion gate:

json

{
  "hooks": {
    "TaskCompleted": [
      {
        "hooks": [{
          "type": "prompt",
          "prompt": "Evaluate whether this task is truly complete: $ARGUMENTS. Check all subtasks are done and no errors remain.",
          "timeout": 30
        }]
      }
    ]
  }
}

The LLM responds with {"ok": true} or {"ok": false, "reason": "..."}.

Agent-based deep verification:

json

{
  "hooks": {
    "TaskCompleted": [
      {
        "hooks": [{
          "type": "agent",
          "prompt": "Verify task completion: check tests pass, no TODO comments in changed files, docs updated. Context: $ARGUMENTS",
          "timeout": 120
        }]
      }
    ]
  }
}

Agent hooks spawn a subagent with access to Read, Grep, and Glob tools for multi-turn investigation.

Other Useful Hook Events

Event	Matcher	Use Case
`Stop`	none	Prevent Claude from stopping if tasks remain incomplete
`SubagentStart`	agent type	React when a subagent spawns
`SubagentStop`	agent type	Block subagent from finishing; can force more work
`SessionStart`	startup/resume	Inject task context at session start

For the complete hook reference (all events, inputs, outputs, exit codes), see references/hooks-reference.md.

Approach 3: Custom MCP Tools

Build your own task management tools as an in-process MCP server. Claude can call these alongside its built-in TodoWrite.

Important: Custom MCP tools require streaming input mode (async generator for the prompt parameter, not a plain string).

TypeScript: Custom Task Board

typescript

import { query, tool, createSdkMcpServer } from "@anthropic-ai/claude-agent-sdk";
import { z } from "zod";

const taskBoard = createSdkMcpServer({
  name: "task-board",
  version: "1.0.0",
  tools: [
    tool("create_task", "Create a task on the external board",
      { title: z.string(), priority: z.enum(["low", "medium", "high"]) },
      async (args) => {
        // Insert into DB, call API, etc.
        return { content: [{ type: "text", text: `Created: ${args.title} [${args.priority}]` }] };
      }
    ),
    tool("complete_task", "Mark a task as done",
      { task_id: z.string() },
      async (args) => {
        return { content: [{ type: "text", text: `Completed: ${args.task_id}` }] };
      }
    )
  ]
});

// Streaming input is required for MCP servers
async function* messages() {
  yield { type: "user" as const, message: { role: "user" as const, content: "Create a task to fix the login bug" } };
}

for await (const msg of query({
  prompt: messages(),
  options: {
    mcpServers: { "task-board": taskBoard },
    allowedTools: ["mcp__task-board__create_task", "mcp__task-board__complete_task"]
  }
})) {
  if ("result" in msg) console.log(msg.result);
}

Python: Custom Task Board

python

from claude_agent_sdk import tool, create_sdk_mcp_server
from typing import Any

@tool("create_task", "Create a task on the external board", {"title": str, "priority": str})
async def create_task(args: dict[str, Any]) -> dict[str, Any]:
    return {"content": [{"type": "text", "text": f"Created: {args['title']} [{args['priority']}]"}]}

@tool("complete_task", "Mark a task as done", {"task_id": str})
async def complete_task(args: dict[str, Any]) -> dict[str, Any]:
    return {"content": [{"type": "text", "text": f"Completed: {args['task_id']}"}]}

task_server = create_sdk_mcp_server(name="task-board", version="1.0.0", tools=[create_task, complete_task])

Tool names follow the pattern mcp__{server_name}__{tool_name} (e.g., mcp__task-board__create_task).

Approach 4: CLI Headless Streaming

Parse task updates from claude -p output without any SDK code. Ideal for shell scripts, CI/CD, and language-agnostic integrations.

Stream and Filter TodoWrite Events

bash

claude -p "Refactor this module and track progress" \
  --output-format stream-json \
  --verbose \
  --include-partial-messages | \
  jq -c 'select(.type == "assistant") |
    .message.content[]? |
    select(.type == "tool_use" and .name == "TodoWrite") |
    .input.todos'

Session Management

bash

# Capture session ID
SESSION=$(claude -p "Start analysis" --output-format json | jq -r '.session_id')

# Resume later
claude -p "Continue the analysis" --resume "$SESSION"

# Or just continue most recent
claude -p "What's left to do?" --continue

Key CLI Flags

Flag	Purpose
`-p`	Non-interactive (headless) mode
`--output-format stream-json`	NDJSON streaming for real-time parsing
`--output-format json`	Structured JSON with session_id and metadata
`--verbose`	Full turn-by-turn output
`--include-partial-messages`	Token-level streaming
`--continue` / `--resume <id>`	Session continuity
`--allowedTools "Bash,Read,Edit"`	Auto-approve specific tools
`--append-system-prompt "..."`	Add instructions while keeping defaults

Approach 5: Subagents

Define specialized agents that handle focused subtasks. The main agent delegates via the Task tool.

typescript

for await (const message of query({
  prompt: "Review auth security, then run tests",
  options: {
    allowedTools: ["Read", "Grep", "Glob", "Bash", "Task"],
    agents: {
      "security-reviewer": {
        description: "Security code review specialist.",
        prompt: "Identify vulnerabilities, injection risks, and auth flaws.",
        tools: ["Read", "Grep", "Glob"],   // read-only
        model: "sonnet"
      },
      "test-runner": {
        description: "Test execution and analysis specialist.",
        prompt: "Run test suites and analyze failures.",
        tools: ["Bash", "Read", "Grep"]
      }
    }
  }
})) {
  // Detect delegation
  if (message.type === "assistant") {
    for (const block of message.message.content) {
      if (block.type === "tool_use" && block.name === "Task") {
        console.log(`Delegated to: ${block.input.subagent_type}`);
      }
    }
  }
}

Key rules: subagents cannot spawn their own subagents; include Task in allowedTools to enable delegation; omit tools to inherit all parent tools.

For resuming subagents and advanced patterns, see references/subagent-patterns.md.

Combining Approaches

These compose naturally. Common combinations:

•SDK + Hooks: Monitor via message stream while hooks enforce quality gates.
•CLI + Custom MCP: Use claude -p in CI with project-specific task tools.
•SDK + MCP + Subagents: Full orchestration — custom task tools, specialized agents, real-time monitoring.

For detailed combined examples, see references/combined-patterns.md.

claude-code-tasks

Programmatic Task Management for Claude Code

Quick Reference: Which Approach to Use

Approach 1: Agent SDK Message Stream

TodoWrite Schema

TypeScript: Monitor Todos

Python: Monitor Todos

Approach 2: Hooks

Intercept TodoWrite Updates

TaskCompleted Hook

Other Useful Hook Events

Approach 3: Custom MCP Tools

TypeScript: Custom Task Board

Python: Custom Task Board

Approach 4: CLI Headless Streaming

Stream and Filter TodoWrite Events

Session Management

Key CLI Flags

Approach 5: Subagents

Combining Approaches

Official Documentation