Token report shows task approaching 40% threshold (80K tokens)
Subagent has experienced compaction events
PLAN.md analysis reveals task is larger than expected
Partial collection indicates significant remaining work
Pre-emptive decomposition during planning phase

Workflow

1. Analyze Current Task Scope

TASK_DIR=".claude/cat/issues/v${MAJOR}/v${MAJOR}.${MINOR}/${TASK_NAME}"

# Read current PLAN.md
cat "${TASK_DIR}/PLAN.md"

# Read STATE.md for progress
cat "${TASK_DIR}/STATE.md"

# If subagent exists, check its progress
if [ -d ".worktrees/${TASK}-sub-${UUID}" ]; then
  # Review commits made
  cd ".worktrees/${TASK}-sub-${UUID}"
  git log --oneline origin/HEAD..HEAD
fi

2. Identify Logical Split Points

Analyze PLAN.md for natural boundaries:

Good split points:

Between independent features
Between layers (model, service, controller)
Between read and write operations
Between setup and implementation
Between implementation and testing

Poor split points:

Middle of a refactoring
Between tightly coupled components
In the middle of a transaction boundary

3. Create New Task Directories

# Original task: 1.2/implement-parser
# New tasks: parser-lexer, parser-ast, parser-semantic (within same minor)

# Create directories for new tasks
mkdir -p ".claude/cat/issues/v1/v1.2/parser-lexer"
mkdir -p ".claude/cat/issues/v1/v1.2/parser-ast"
mkdir -p ".claude/cat/issues/v1/v1.2/parser-semantic"

4. Create PLAN.md for Each New Task

Each new task gets its own focused PLAN.md:

# 1.2a-parser-lexer/PLAN.md
---
task: 1.2a-parser-lexer
parent: 1.2-implement-parser
sequence: 1 of 3
---

# Implement Parser Lexer

## Objective
Implement the lexical analysis phase of the parser.

## Scope
- Token definitions
- Lexer implementation
- Lexer unit tests

## Dependencies
- None (first in sequence)

## Deliverables
- src/parser/Token.java
- src/parser/Lexer.java
- test/parser/LexerTest.java

5. Define Dependencies Between New Tasks

# Dependency graph
dependencies:
  1.2a-parser-lexer: []  # No dependencies
  1.2b-parser-ast:
    - 1.2a-parser-lexer  # Depends on lexer
  1.2c-parser-semantic:
    - 1.2b-parser-ast    # Depends on AST

6. Update STATE.md Files

Original task STATE.md:

# 1.2-implement-parser/STATE.md

- **Status:** decomposed
- **Progress:** 0%
- **Decomposed At:** 2026-01-10T16:00:00Z
- **Reason:** Task exceeded context threshold (85K tokens used)

## Decomposed Into
- 1.2a-parser-lexer
- 1.2b-parser-ast
- 1.2c-parser-semantic

## Progress Preserved
- Lexer implementation 80% complete in subagent work
- Will be merged to 1.2a branch

New task STATE.md:

# 1.2a-parser-lexer/STATE.md

- **Status:** pending
- **Progress:** 0%
- **Created From:** 1.2-implement-parser
- **Inherits Progress:** true (will receive merge from parent subagent)
- **Dependencies:** []

7. Handle Existing Subagent Work

If decomposing due to subagent context limits:

# Collect partial results from subagent
collect-results "${SUBAGENT_ID}"

# Determine which new task inherits the work
# Usually the first or most complete component

# Merge subagent work to appropriate new task branch
git checkout "1.2a-parser-lexer"
git merge "${SUBAGENT_BRANCH}" -m "Inherit partial progress from decomposed parent"

8. Generate Parallel Execution Plan

MANDATORY: Analyze dependencies and create sub-task-based execution plan.

After decomposition, determine which subtasks can run concurrently:

# Dependency analysis
subtasks:
  - id: 1.2a-parser-lexer
    dependencies: []
    estimated_tokens: 25000
  - id: 1.2b-parser-ast
    dependencies: [1.2a-parser-lexer]
    estimated_tokens: 30000
  - id: 1.2c-parser-tests
    dependencies: []
    estimated_tokens: 20000

# Sub-task-based parallel plan
parallel_execution_plan:
  sub_task_1:
    # Tasks with no dependencies - can run concurrently
    tasks: [1.2a-parser-lexer, 1.2c-parser-tests]
    max_concurrent: 2
    reason: "Both have no dependencies, can execute in parallel"

  sub_task_2:
    # Tasks that depend on sub_task_1 completion
    tasks: [1.2b-parser-ast]
    depends_on: [sub_task_1]
    reason: "Depends on 1.2a-parser-lexer from sub_task_1"

execution_order:
  1. Spawn subagents for sub_task_1 tasks (parallel)
  2. Monitor and collect sub_task_1 results
  3. Merge sub_task_1 branches
  4. Spawn subagents for sub_task_2 tasks (parallel)
  5. Monitor and collect sub_task_2 results
  6. Merge sub_task_2 branches

Output parallel plan to STATE.md:

## Parallel Execution Plan

### Sub-task 1 (Concurrent)
| Task | Est. Tokens | Dependencies |
|------|-------------|--------------|
| 1.2a-parser-lexer | 25K | None |
| 1.2c-parser-tests | 20K | None |

### Sub-task 2 (After Sub-task 1)
| Task | Est. Tokens | Dependencies |
|------|-------------|--------------|
| 1.2b-parser-ast | 30K | 1.2a-parser-lexer |

**Total sub-tasks:** 2
**Max concurrent subagents:** 2

Conflict detection for parallel tasks:

Ensure no parallel tasks modify the same files:

conflict_check:
  task_1: 1.2a-parser-lexer
    files: [src/parser/Lexer.java, test/parser/LexerTest.java]
  task_2: 1.2c-parser-tests
    files: [test/parser/ParserIntegrationTest.java]

  overlap: []  # No conflicts - safe to parallelize

  # If overlap exists:
  conflict_resolution:
    move_conflicting_task_to_next_sub_task: true

9. Mark Original Task as Decomposed

# Update original PLAN.md
echo "---
status: DECOMPOSED
decomposed_into: [1.2a, 1.2b, 1.2c]
parallel_plan: sub_task_1=[1.2a, 1.2c], sub_task_2=[1.2b]
---" >> "${TASK_DIR}/PLAN.md"

Examples

Pre-Planning Decomposition

When analyzing requirements reveals a task is too large:

# Original task seemed manageable
task: 1.5-implement-authentication

# Analysis reveals scope
components:
  - User model and repository
  - Password hashing service
  - JWT token generation
  - Login/logout endpoints
  - Session management
  - Password reset flow
  - Email verification

# Too many components - decompose before starting
decompose_to:
  - 1.5a-auth-user-model
  - 1.5b-auth-password-service
  - 1.5c-auth-jwt-tokens
  - 1.5d-auth-endpoints
  - 1.5e-auth-sessions
  - 1.5f-auth-password-reset
  - 1.5g-auth-email-verify

Mid-Execution Decomposition

When subagent hits context limits:

decomposition_trigger:
  task: 1.3-implement-formatter
  subagent_tokens: 85000
  compaction_events: 1
  completed_work:
    - Basic formatter structure
    - Indentation handling
  remaining_work:
    - Line wrapping
    - Comment formatting
    - Multi-line string handling

decomposition_result:
  - task: 1.3a-formatter-core
    inherits: subagent work
    status: nearly_complete
  - task: 1.3b-formatter-wrapping
    status: ready
  - task: 1.3c-formatter-comments
    status: ready

Emergency Decomposition

When subagent is stuck or confused:

emergency_decomposition:
  trigger: "Subagent making no progress for 30+ minutes"
  analysis: |
    Task scope unclear, subagent attempting multiple
    approaches without success.

  action:
    - Collect any usable partial work
    - Re-analyze requirements
    - Create smaller, more specific tasks
    - Add explicit acceptance criteria to each

Anti-Patterns

Split at logical boundaries

# ❌ Splitting at arbitrary points
1.2a: "Lines 1-100 of Parser.java"
1.2b: "Lines 101-200 of Parser.java"

# ✅ Split at logical boundaries
1.2a: "Lexer component"
1.2b: "AST builder component"

Model actual dependencies accurately

# ❌ Treating all subtasks as independent
1.2a-parser-lexer: []
1.2b-parser-ast: []    # Actually needs lexer!
1.2c-parser-semantic: []  # Actually needs AST!

# ✅ Model actual dependencies
1.2a-parser-lexer: []
1.2b-parser-ast: [1.2a]
1.2c-parser-semantic: [1.2b]

Preserve partial progress when decomposing

# ❌ Starting fresh after decomposition
decompose_task "1.2-parser"
# Subagent work discarded!

# ✅ Preserve progress
collect_results "${SUBAGENT}"
decompose_task "1.2-parser"
merge_to_appropriate_subtask "${SUBAGENT_WORK}"

Create meaningful chunks (avoid over-decomposition)

# ❌ Too granular
1.2a: "Define Token class"
1.2b: "Define TokenType enum"
1.2c: "Implement nextToken method"
1.2d: "Implement peek method"
# ...20 more tiny tasks

# ✅ Meaningful chunks
1.2a: "Implement Lexer (tokens, types, core methods)"
1.2b: "Implement Parser (AST, expressions, statements)"

Always update orchestration when decomposing

# ❌ Create subtasks, forget to track
mkdir 1.2a 1.2b 1.2c
# Parent doesn't know about them!

# ✅ Full state update
create_subtasks "1.2a" "1.2b" "1.2c"
update_parent_state "decomposed" "1.2a,1.2b,1.2c"
update_orchestration_plan

Distinguish runtime dependencies from extraction dependencies

For code extraction/refactoring tasks, runtime method calls are NOT task dependencies.

# ❌ Confusing runtime calls with task dependencies
# "parseUnary calls parsePostfix, so extract-unary must run before extract-postfix"
subtasks:
  extract-unary: []
  extract-postfix: [extract-unary]  # Wrong! Just copying code, not executing it

# ✅ Extraction tasks that write to different sections can run in parallel
# Methods call each other at RUNTIME, but extraction is just copying text
subtasks:
  extract-unary: [setup-interface]      # Both depend on interface setup
  extract-postfix: [setup-interface]    # Both can run concurrently

# Key insight: "Does method A call method B?" is irrelevant for extraction order.
# Ask instead: "Do both tasks write to the same file section?"
# If writing to different sections of the same file → can parallelize
# Only the final integration task depends on all extractions completing

Dependency analysis questions:

Does task B need OUTPUT from task A? (Real dependency)
Does task B just reference CODE that task A also references? (Not a dependency)
Are both tasks copying different methods to the same target file? (Parallelizable with merge)

Related Skills

cat:token-report - Triggers decomposition decisions
cat:collect-results - Preserves progress before decomposition
cat:spawn-subagent - Launches work on decomposed tasks
cat:parallel-execute - Can run independent subtasks concurrently

cat:decompose-task

Decompose Task

Purpose

When to Use