A hook-based framework for Claude Code that enforces task delegation to specialized agents, enabling structured workflows and expert-level task handling through intelligent orchestration.

See the delegation system in action:

⚡ Soft Enforcement & Lean Startup — Replaces hard-blocking enforcement with adaptive per-turn nudges (silent → hint → warning → strong reminder). SessionStart injection trimmed ~6.6K tokens: stub orchestrator (~1.1KB) injected on startup, full orchestrator loaded only on first delegation. Output style loaded natively from plugin.json (no injection). Net result: lean sessions with automatic escalation when delegation is needed.

🤝 Agent Teams Integration — Native dual-mode execution: workflows automatically select between isolated subagents and collaborative Agent Teams (via TeamCreate + Agent(team_name=...) + SendMessage) based on tool availability. Teammates communicate in real-time, share task lists, and self-coordinate. Enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

This system uses Claude Code's hook mechanism to create a delegation-enforced workflow architecture that routes tasks to specialized agents for expert-level execution.

• Soft Enforcement via Adaptive Nudges - No hard blocks. PreToolUse hook emits per-turn escalating stderr reminders (silent → hint → warning → strong) when main agent bypasses /workflow-orchestrator:delegate. Subagents immune. Nudge counter resets each turn and zeros when delegation runs.
• Lean SessionStart Injection - Consolidated 3 hooks into 1 Python script. Stub orchestrator (~1.1KB) on startup + optional token-efficient CLI guide, full orchestrator (~7.5KB) loaded on-demand. Output style loaded natively from plugin.json (no injection). Saves ~6.6K tokens off session start.
• 8 Specialized Agents - Each agent has domain expertise (code cleanup, testing, architecture, DevOps, etc.)
• Native Plan Mode - Built-in plan mode (EnterPlanMode/ExitPlanMode) handles planning, agent selection, and execution orchestration
• Intelligent Multi-Step Workflows - Sequential execution for dependent phases, parallel for independent phases
• Dual-Mode Execution - Isolated subagent sessions (default) or collaborative Agent Teams with real-time inter-agent communication (experimental)
• Agent Teams Integration - Native TeamCreate + Agent(team_name=...) + SendMessage for peer-to-peer collaboration, shared task lists, and coordinated multi-agent workflows
• Tasks API Integration - Native task tracking via TaskCreate, TaskUpdate, TaskList, TaskGet with structured metadata
• Structured Task Metadata - Wave assignments, phase IDs, agent assignments, and dependencies encoded in task metadata
• Async Hook Support - Non-blocking background tasks for reminders and cleanup operations
• Stateful Session Management - Fresh delegation enforcement per user message with per-turn nudge counter
• Smart Dependency Analysis - Automatically analyzes phase dependencies to determine optimal execution mode
• Parallel Execution Support - Executes independent phases concurrently with automatic wave synchronization
• Visualization & Debugging - Comprehensive logging and debug tools for understanding delegation decisions

The system uses a two-stage execution pipeline:

Stage 0: Planning & Analysis (native plan mode)

• Analyzes task complexity (single-step vs multi-step)
• Decomposes complex tasks into atomic phases
• Performs dependency analysis to determine execution mode
• Assigns specialized agents via keyword matching
• Creates wave assignments for parallel/sequential execution
• Creates tasks via TaskCreate with structured metadata

Stage 1: Execution

• Single-Step Tasks: Hook blocks tools → Delegates to specialized agent → Agent executes → Results returned
• Multi-Step Workflows:
  • Subagent mode (default): Isolated parallel Agent instances per wave. Agents return DONE|{path}. Context-efficient, optimal for most workflows.
  • Team mode (experimental): Native Agent Teams via TeamCreate + Agent(team_name=...). Teammates share context, communicate via SendMessage, and self-coordinate through shared task lists. Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
• Results consolidated and summary provided

Execution Mode Selection: If TeamCreate is in available tools → execution_mode: "team". Otherwise → "subagent". Tool availability is the only signal (set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 to enable). For subagent mode, plan mode further selects sequential (context preservation, dependencies) or parallel (time savings, independence) based on phase dependency analysis.

• uv: https://docs.astral.sh/uv/getting-started/installation/

curl -LsSf https://astral.sh/uv/install.sh | sh

• bun: https://bun.com/docs/installation

curl -fsSL https://bun.com/install | bash

• jq: JSON processor for parallel workflow state tracking.

# macOS
brew install jq
# Linux
sudo apt install jq

Python 3.12+ is required. All hooks use cross-platform Python scripts.

• Python: https://www.python.org/downloads/

  • During installation, ensure "Add Python to PATH" is checked
  • Verify installation: python --version

• uv: https://docs.astral.sh/uv/getting-started/installation/

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

• bun: https://bun.sh/docs/installation

powershell -c "irm bun.sh/install.ps1 | iex"

• jq (optional, for advanced parallel workflow features):

# Using Chocolatey
choco install jq
# Or download from https://jqlang.github.io/jq/download/

Note: The hook system uses Python scripts for cross-platform compatibility. Ensure python is available in your PATH.

This project provides a comprehensive delegation system for Claude Code with multi-agent orchestration. Choose your preferred installation method:

The easiest way to install is via Claude Code's plugin system:

# Add the marketplace
claude plugin marketplace add barkain/claude-code-workflow-orchestration

# Install the plugin
claude plugin install workflow-orchestrator@barkain-plugins  # user-level
# or
claude plugin install workflow-orchestrator@barkain-plugins --scope project  # project-level

Benefits:

• Automatic setup and configuration
• Easy updates via plugin manager
• No manual file copying required

Optional Settings:

• Run /workflow-orchestrator:add-statusline after installation to enable workflow status display
• The plugin includes a technical-adaptive output style optimized for workflow orchestration. To select it, configure outputStyle in your Claude Code settings (project or user level)

Note: Changing the output style requires restarting your Claude Code session for the change to take effect.

For development or custom configurations:

# Clone the repository
git clone https://github.com/barkain/claude-code-workflow-orchestration.git

For project-isolated configurations or version-controlled delegation setups:

cd path/to/project
path/to/repo/install.sh  # follow the installation instructions

Windows users: The install.sh script requires bash (Git Bash or WSL). For Windows, we recommend using the Plugin Installation method above, which works natively on all platforms.

Once installed, the delegation hook is automatically active. Simply use Claude Code normally (Opus 4.5 is preferred):

# Multi-step workflow - enable orchestration for context passing
claude

and then prompt claude with:

> create a simple calculator app with basic math operations.
  add a nice UI and use NextJS/Tailwind to build this out.
  the backend should be implemented in python as a modern uv project.
  add verification steps after each phase.

What happens:

1. The main agent enters native plan mode (EnterPlanMode) to decompose the request, assign specialized agents via keyword matching, schedule waves, and select an execution mode. Mode selection is one rule: if TeamCreate is available (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), it picks team mode; otherwise parallel subagents.

   

2. After ExitPlanMode approval, a dependency graph is rendered showing the parallel waves and the agents assigned to each phase:

   

3. In team mode, the lead calls TeamCreate once and then spawns each phase as a persistent teammate via Agent(team_name=...). Teammates run concurrently in the swarm session and self-coordinate via SendMessage. Open the swarm view with tmux -L claude-swarm-<id> a to watch them work in real time:

   

4. As each wave completes, downstream phases unblock automatically. A final integration/verification phase confirms end-to-end correctness, then the team shuts down cleanly:

   

The delegation system uses adaptive nudges instead of hard blocks:

# 1st direct tool call (imperative STOP)
Read test.py
# stderr: "STOP. This tool call bypasses delegation. Abandon this step and run: /workflow-orchestrator:delegate <your task>"

# 2nd direct tool call (imperative STOP, 2nd-call phrasing)
Read other.py
# stderr: "STOP. 2nd direct tool call this turn. The main agent does not execute work tools. Run: /workflow-orchestrator:delegate <your task>"

# 3rd direct tool call (strong reminder — explains what's being lost)
Edit file.py
# stderr: "STOP. 3 direct tool calls bypassing delegation. You are losing planning, parallelization, and context isolation. Abandon the current plan and run: /workflow-orchestrator:delegate <your task>"

# 4th+ direct tool calls keep the same strong reminder with the updated count
Bash command.sh
# stderr: "STOP. 4 direct tool calls bypassing delegation. ..."

# Delegation resets the counter (state clean)
/workflow-orchestrator:delegate "Create feature"
# Counter zeros. Next turn starts fresh.

Counter resets each user turn and zeros when /workflow-orchestrator:delegate runs. Subagents are immune. Use /workflow-orchestrator:bypass for emergency access without restarting.

The system supports several environment variables for configuration and debugging:

Tasks API Configuration:

CLAUDE_CODE_ENABLE_TASKS=true              # Enable Tasks API (default: true)
CLAUDE_CODE_TASK_LIST_ID=list_id           # Share task list across sessions
CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1     # Disable async background tasks

Agent Teams (Experimental):

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1     # Enable Agent Teams dual-mode execution

Token Efficiency:

CLAUDE_TOKEN_EFFICIENCY=1                  # Enable token-efficient CLI output (default: 1)

Debug & Control:

DEBUG_DELEGATION_HOOK=1                    # Enable hook debug logging
DELEGATION_HOOK_DISABLE=1                  # Emergency bypass (disable enforcement)
CLAUDE_MAX_CONCURRENT=8                    # Max parallel agents per batch (default 8)
CHECK_RUFF=0                               # Skip Ruff validation in PostToolUse
CHECK_PYRIGHT=0                            # Skip Pyright validation in PostToolUse
CLAUDE_SKIP_PYTHON_VALIDATION=1            # Skip all Python validation

See Environment Variables for detailed configuration.

The plugin-hooks.json configures the delegation enforcement hooks using cross-platform Python scripts:

Note: All hooks use uv run --no-project --script for cross-platform compatibility (Windows, macOS, Linux). The --no-project flag allows execution without requiring a pyproject.toml, and --script directly runs Python scripts using uv's managed interpreter.

Hook Events (6 lifecycle points, 14 hooks):

The inject_all.py hook consolidates 3 SessionStart hooks into 1 Python script:

On startup/resume (all sessions):

• Injects orchestrator routing stub (orchestrator_stub.md, ~1.1KB): points the main agent at /workflow-orchestrator:delegate for multi-step work. Minimal overhead (~200 tokens).
• Optionally injects token-efficient CLI guide (token_efficient_cli.md, ~1.9KB, gated by CLAUDE_TOKEN_EFFICIENCY=1 env var). Teaches compact flags and command patterns.
• Output style loaded natively from plugin.json outputStyles field (no injection required). Saves ~1.5K tokens.

On first delegation (lazy load):

• Full orchestrator logic (planning instructions, agent catalog, wave scheduling, team-mode execution) lives inline in commands/delegate.md and is loaded only when /workflow-orchestrator:delegate runs.

Net savings: ~6.6K tokens off session startup. Sessions pay the orchestration tax only when delegation is actually used.

What this enables:

• Multi-step task detection via pattern matching
• Dependency analysis for execution mode selection
• Context passing between workflow phases
• Tasks API integration for progress tracking
• Wave synchronization for parallel execution

Soft enforcement: nudges (never blocks) when main agent uses work-doing tools directly. Tracks 8 stable primitives: Bash, Edit, Write, Read, Glob, Grep, MultiEdit, NotebookEdit. New Claude Code tools never trigger nudges.

Allowed tools (no nudge):

• AskUserQuestion, Skill, SlashCommand - Explicit queries and commands
• TaskCreate, TaskUpdate, TaskList, TaskGet - Task tracking (structured metadata)
• Agent, SubagentTask, AgentTask - Delegation mechanism
• TeamCreate, SendMessage - Agent Teams (when CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
• ToolSearch - Tool discovery
• New Claude Code tools (by definition, never tracked)

Work tools (tracked for nudges, not blocked):

• Bash, Edit, Write, Read, Glob, Grep, MultiEdit, NotebookEdit

Nudge escalation by violation count (per turn):

• 0 violations: silent
• 1 violation: delegate? (~2 tokens)
• 2 violations: nudge: use /workflow-orchestrator:delegate for multi-step work (~12 tokens)
• 3-4 violations: WARNING: N direct tool calls bypassing delegation... (~25 tokens)
• 5+ violations: Strong reminder (~55 tokens)

Note: TaskOutput is prohibited (context exhaustion). Agents write to $CLAUDE_SCRATCHPAD_DIR and return DONE|{path} only. Counter resets each turn; subagents (via CLAUDE_PARENT_SESSION_ID) are exempt.

8 specialized agents for different task types:

• tech-lead-architect - Solution design, architecture, research
• codebase-context-analyzer - Code exploration, architecture analysis
• task-completion-verifier - Validation, testing, quality assurance
• code-cleanup-optimizer - Refactoring, technical debt reduction
• code-reviewer - Code review for best practices
• devops-experience-architect - Infrastructure, deployment, CI/CD
• documentation-expert - Documentation creation and maintenance
• dependency-manager - Dependency management and updates

Note: The delegation-orchestrator agent has been deprecated. Its orchestration and routing functionality is now provided by native plan mode (EnterPlanMode/ExitPlanMode), which handles both planning and execution orchestration directly within the main agent.

The /workflow-orchestrator:delegate command provides intelligent task delegation with integrated planning:

/workflow-orchestrator:delegate <task description>

How it works:

1. Enters native plan mode (EnterPlanMode) for unified planning and orchestration
2. Plan mode analyzes task complexity and decomposes into phases
3. Performs dependency analysis to determine execution mode (sequential or parallel)
4. Assigns specialized agents via keyword matching (>=2 match threshold)
5. Creates wave assignments and execution plan
6. Creates task list via TaskCreate
7. Exits plan mode (ExitPlanMode) and executes phases as directed by the plan

The /workflow-orchestrator:delegate slash command loads the full orchestrator logic on demand: multi-step detection, plan-mode instructions, agent catalog, wave scheduling, and team-mode execution. Works in conjunction with native plan mode (EnterPlanMode/ExitPlanMode).

Activate via: Simply start a Claude code session

claude

Multi-step detection patterns:

• Sequential connectors: "and then", "after that", "next"
• Compound indicators: "with [noun]", "including [noun]"
• Multiple verbs: "create X and test Y"

Unified Planning & Execution: Native plan mode (EnterPlanMode/ExitPlanMode) handles both planning and execution orchestration:

1. Task Decomposition - Breaks complex tasks into atomic phases
2. Dependency Analysis - Analyzes phase dependencies to determine optimal execution mode
3. Intelligent Execution Mode Selection - Chooses between sequential and parallel execution
4. Sequential Execution - Dependent phases execute one at a time with context passing
5. Parallel Execution - Independent phases execute concurrently in waves
6. Progress Tracking - Tasks API maintains visible task list throughout
7. State Management - Wave synchronization ensures proper completion order

Execution mode decision logic:

• Sequential: When phases have data dependencies, file conflicts, or state conflicts requiring ordered execution
• Parallel: When phases are independent with no data dependencies, enabling time-efficient concurrent execution
• Conservative fallback: Sequential is chosen when dependencies are uncertain

Complete workflow process:

1. User submits multi-step task (detected by workflow_orchestrator patterns)
2. Native plan mode entered (EnterPlanMode) to decompose and plan execution
3. Task list created via TaskCreate with all phases
4. Phase dependencies analyzed to determine execution mode
5. Sequential Mode: Phases execute one at a time with context passing
6. Parallel Mode: Independent phases grouped into waves and executed concurrently
7. Wave synchronization ensures proper completion order
8. Results consolidated with absolute file paths and summary provided

The framework supports a second execution mode that uses Claude Code's native Agent Teams feature for real-time inter-agent collaboration. When enabled, agents can communicate with each other via SendMessage, share task lists, and self-coordinate -- rather than running as isolated subagents.

Set the environment variable before starting Claude Code:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

No other configuration is required. Plan mode automatically evaluates whether a given task benefits from team-based execution.

During planning, plan mode checks tool availability:

ONE RULE: If TeamCreate is in your available tools → execution_mode: "team". Otherwise → "subagent".

This is detected by attempting to use the tool. Setting CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 makes TeamCreate available, enabling team mode. Without it, only subagent mode is available.

The key difference is one parameter: Agent(team_name="x") makes a teammate; Agent() makes an isolated subagent.

Simple team -- a single AGENT TEAM phase with multiple teammates exploring in parallel. Used for multi-perspective exploration tasks.

> explore the authentication system from different angles

This creates one team phase where each teammate explores a different perspective (e.g., security, performance, architecture), then results are synthesized.

Complex team -- multiple individual phases across waves, all executed as teammates with Agent(team_name=...). Used for collaborative implementation tasks.

> implement the payment service. tasks should be collaborative

All phases run as teammates sharing context and messaging, even though each has a distinct assignment.

> explore the codebase from different angles
> design the API with a team of specialists
> implement the feature collaboratively
> use a team to review and refactor the auth module
> brainstorm together on the CLI design

Before creating a team, the framework presents the team plan and asks for confirmation:

• Team name, execution mode, number of phases, wave structure
• If declined, execution falls back to subagent mode automatically

Team mode creates two additional state files (automatically cleaned up on completion or next user prompt):

The framework minimizes command output to reduce context consumption and preserve tokens for meaningful work. Token efficiency is enabled by default (CLAUDE_TOKEN_EFFICIENCY=1).

1. Behavioral Guidance — The token_efficient_cli.md system prompt (injected via SessionStart) teaches compact flag usage:

   • git status -sb (short branch format)
   • pytest -q --tb=short (quiet mode, short tracebacks)
   • npm test -- -q (quiet test output)
   • Encourages --help parsing and targeted commands

2. Output Compression — The token_rewrite_hook.py PreToolUse hook rewrites matching Bash commands through compact_run.py, which compresses git/test/log output post-execution:

   • Git: push, pull, commit, merge, rebase, status, etc.
   • Test runners: pytest, cargo test, npm/pnpm/yarn/bun test, vitest, jest, mocha, etc.
   • Logs: docker logs, kubectl logs, make output
   • Build tools: eslint, next, tsc
   • Command chaining: cd && command pattern support

3. Conditional System Prompt Injection — The orchestrator is injected conditionally:

   • On session startup: Stub version (~200 tokens) provides minimal direction
   • On first plan mode entry (via /workflow-orchestrator:delegate or detected multi-step): Full version (~11K tokens) for complete planning capability
   • Saves tokens for single-step and read-only tasks

To temporarily disable token-efficient output:

export CLAUDE_TOKEN_EFFICIENCY=0

This disables both the behavioral guidance and output compression layers.

We welcome contributions to the Claude Code Workflow Orchestration System! Whether you're fixing bugs, adding features, or improving documentation, your help is appreciated.

Found a bug or have a feature request? Please open a GitHub Issue with:

• Clear description of the issue or feature request
• Steps to reproduce (for bugs)
• Expected vs. actual behavior
• Your environment (Windows/macOS/Linux, Claude Code version, Python version, etc.)
• The used claude code model
• Relevant logs or screenshots if applicable

1. Fork the repository and create a feature branch:

   git checkout -b feature/your-feature-name

2. Make your changes following our code style guidelines (see below)

3. Run quality checks (if applicable) before submitting:

   # Format code
   uvx ruff format .
   
   # Lint code
   uvx ruff check --no-fix .
   
   # Type checking
   uvx pyright .
   
   # Run tests
   uv run pytest

   All checks must pass before submission.

4. Commit with clear messages:

   git commit -m "feat: description of your changes"

   Use conventional commit format: feat:, fix:, docs:, refactor:, etc.

5. Push to your fork and submit a Pull Request to the main branch with a clear description of changes

• Python 3.12+ with modern syntax (e.g., list[str], str | None)
• Type hints on all functions and variables
• No print statements - use structured logging with logger calls
• Comprehensive docstrings with examples for public APIs
• Clear variable and function names that reflect intent
• Automatic enforcement via Ruff (formatting), Pyright (types), and Pytest (tests)

Always run quality checks locally before submitting to catch issues early.

The project includes a comprehensive test suite covering hooks, token efficiency, and integration:

# Run all tests
uv run pytest

# Run with verbose output and coverage
uv run pytest -v --cov=hooks --cov=scripts --cov=system-prompts

# Run specific test file
uv run pytest tests/test_token_rewrite_hook.py -v

# Run tests matching a pattern
uv run pytest -k "token_efficiency" -v

Test files:

• tests/test_token_rewrite_hook.py - Token rewriting hook tests
• tests/test_inject_token_efficiency.py - Token efficiency injection tests
• tests/test_compact_run.py - Compact output runner tests
• tests/test_integration.py - End-to-end integration tests
• tests/conftest.py - Test fixtures and configuration

• Clear, well-documented code
• Tests for new functionality
• Documentation updates for new features
• Constructive feedback and collaboration
• Diverse perspectives and creative solutions

Thank you for contributing to making Claude Code workflows even better!