Recipe Runner Integration with UltraThink Orchestrator¶

Document Type: How-To Guide Audience: Developers using amplihack Status: [IMPLEMENTED] - Production Ready Last Updated: 2026-02-15

Overview¶

The UltraThink orchestrator now uses Recipe Runner as its primary execution path, providing code-enforced workflow execution with automatic fallback to prompt-based methods. This integration transforms workflows from suggestion-based prompts into deterministic, fail-fast Python code.

What This Means:

Workflow steps execute via Python SDK adapters (not just LLM prompts)
Failures stop execution immediately with clear error messages
Context accumulates between steps automatically
Conditional logic works reliably (skip steps based on runtime conditions)
Full backward compatibility maintained (existing workflows unchanged)

Quick Start¶

Using Recipe Runner (Default Behavior)¶

# Recipe Runner is enabled by default
/ultrathink implement JWT authentication

# Recipe Runner automatically:
# 1. Detects this is a development task
# 2. Loads default-workflow recipe
# 3. Executes all 23 steps with code enforcement
# 4. Accumulates context between steps
# 5. Stops on first failure (fail-fast)

Forcing Prompt-Based Execution (When Needed)¶

# Disable Recipe Runner for this session
export AMPLIHACK_USE_RECIPES=0

# Now uses workflow skills or markdown workflows
/ultrathink implement JWT authentication

# Use cases for disabling:
# - Debugging workflow issues
# - Testing prompt-based execution
# - Recipe Runner has bugs
# - Developing new workflow features

Three-Tier Fallback System¶

The ultrathink command follows this execution hierarchy:

┌─────────────────────────────────────────────────────────────┐
│ TIER 1: Recipe Runner (Code-Enforced)                      │
│ ✓ Rust CLI executes each step                              │
│ ✓ Fail-fast on errors                                      │
│ ✓ Context accumulation automatic                           │
│ ✓ Conditional execution reliable                           │
│ ✗ Requires amplihack CLI binary                            │
└─────────────────────────────────────────────────────────────┘
                           ↓ Falls back to
┌─────────────────────────────────────────────────────────────┐
│ TIER 2: Workflow Skills (Prompt-Based + TodoWrite)         │
│ ✓ Structured prompts guide execution                       │
│ ✓ Agent orchestration defined                              │
│ ✓ TodoWrite tracks progress                                │
│ ✗ No code enforcement (relies on Claude following)         │
└─────────────────────────────────────────────────────────────┘
                           ↓ Falls back to
┌─────────────────────────────────────────────────────────────┐
│ TIER 3: Markdown Workflows (Baseline Prompt-Based)         │
│ ✓ Always available (.claude/workflow/*.md)                 │
│ ✓ Baseline workflow instructions                           │
│ ✗ No code enforcement (relies on Claude following)         │
│ ✗ No structured tracking                                   │
└─────────────────────────────────────────────────────────────┘

When Each Tier Activates¶

Recipe Runner (Tier 1):

AMPLIHACK_USE_RECIPES is unset or set to 1 (default)
amplihack CLI binary is installed and available on PATH
Recipe for the workflow exists (default-workflow, investigation-workflow, qa-workflow)

Workflow Skills (Tier 2):

Recipe Runner unavailable (amplihack binary not found on PATH)
OR AMPLIHACK_USE_RECIPES=0 is set
Skill definition exists in .claude/skills/ directory

Markdown Workflows (Tier 3):

Recipe Runner unavailable
Workflow skill unavailable or fails to load
Always works (last resort - workflows always exist in .claude/workflow/)

How Recipe Runner Works¶

Architecture¶

# Recipe Runner execution flow for ultrathink

# 1. Load recipe YAML (default-workflow.yaml)
# 2. Execute steps sequentially via agent adapters
# 3. Accumulate context between steps
# 4. Stop on first error (fail-fast)

amplihack recipe run default-workflow \
  -c task="implement JWT authentication" \
  -c user_requirements="Use RS256 algorithm, store keys in vault" \
  --verbose

# Result output includes:
# - success: whether all steps passed
# - context: accumulated context from all steps
# - errors: any errors that occurred
# - steps_executed: int (number of steps completed)

Step Execution¶

Each workflow step executes through a Python adapter:

# Example: Step 5 (Research and Design) in default-workflow

# Recipe YAML defines:
# - step: research_and_design
#   action: invoke_agent
#   agent: architect
#   inputs:
#     task: "{{ user_context.task }}"
#     requirements: "{{ user_context.user_requirements }}"
#   outputs:
#     design: context.design
#     architecture: context.architecture

# Recipe Runner translates this to:
result = sdk_adapter.invoke_agent(
    agent_type="architect",
    prompt=f"Design: {user_context['task']}\nRequirements: {user_context['user_requirements']}"
)

# Output stored in context for next step:
context['design'] = result['design']
context['architecture'] = result['architecture']

# Next step (Step 6) can access context['design']

Context Accumulation¶

Context flows automatically between steps:

# Step 5: architect agent produces design
context['design'] = {
    'modules': ['auth', 'tokens', 'validation'],
    'contracts': {...}
}

# Step 6: builder agent receives design automatically
result = sdk_adapter.invoke_agent(
    agent_type="builder",
    prompt=f"Implement: {context['design']}"  # Design passed automatically
)

# Step 7: tester agent receives implementation automatically
result = sdk_adapter.invoke_agent(
    agent_type="tester",
    prompt=f"Test: {context['implementation']}"  # Implementation passed automatically
)

Fail-Fast Behavior¶

Recipe Runner stops on first error:

# Step 5: architect agent succeeds
# Step 6: builder agent fails (syntax error)

# Recipe Runner behavior:
# 1. Immediately stops execution
# 2. Returns error details
# 3. Provides context up to failure point
# 4. Does NOT continue to Step 7

# Error handling in ultrathink:
# If recipe runner fails, fall back to workflow skills
amplihack recipe run default-workflow \
  -c task="..." \
  --verbose || echo "Recipe execution failed, falling back to workflow skills..."

Environment Variable Control¶

AMPLIHACK_USE_RECIPES¶

Controls Recipe Runner activation:

# Enable Recipe Runner (DEFAULT)
export AMPLIHACK_USE_RECIPES=1
/ultrathink implement feature

# Or leave unset (also enables)
unset AMPLIHACK_USE_RECIPES
/ultrathink implement feature

# Disable Recipe Runner (force prompt-based)
export AMPLIHACK_USE_RECIPES=0
/ultrathink implement feature

When to Set AMPLIHACK_USE_RECIPES=0¶

Debugging Scenarios:

# Recipe Runner is failing and you need to debug
export AMPLIHACK_USE_RECIPES=0
/ultrathink implement feature
# Now uses workflow skills - easier to see what's happening

Testing Scenarios:

# Testing changes to workflow markdown files
export AMPLIHACK_USE_RECIPES=0
/ultrathink test workflow changes
# Ensures markdown changes work before updating recipes

Development Scenarios:

# Developing new workflow features not yet in recipes
export AMPLIHACK_USE_RECIPES=0
/ultrathink experiment with new feature
# Uses markdown workflows where new features are prototyped

Workaround Scenarios:

# Recipe Runner has a bug in the current version
export AMPLIHACK_USE_RECIPES=0
/ultrathink implement feature
# Temporary workaround until Recipe Runner is fixed

Usage Examples¶

Example 1: Development Task (Recipe Runner Succeeds)¶

# User invokes ultrathink
/ultrathink implement user registration with email verification

# UltraThink execution:
1. Detects task type: Development (keyword "implement")
2. Checks environment: AMPLIHACK_USE_RECIPES not set to 0
3. Tries Recipe Runner:
   amplihack recipe run --help  # SUCCESS — binary found
4. Executes via Recipe Runner:
   amplihack recipe run default-workflow \
     -c task="implement user registration with email verification" \
     --verbose
5. Recipe Runner automatically:
   - Step 1: Task clarification (code-enforced)
   - Step 2: Git branch creation (code-enforced)
   - Step 3: Problem decomposition (code-enforced via prompt-writer agent)
   - Step 4: Research check (code-enforced)
   - Step 5: Architecture design (code-enforced via architect agent)
   - Step 6-22: Implementation, testing, PR creation (all code-enforced)
6. Context accumulates between steps automatically
7. Fail-fast if any step errors
8. Success: Full workflow executed with code enforcement

Example 2: Investigation Task (Recipe Runner Succeeds)¶

# User invokes ultrathink
/ultrathink investigate how the caching layer works

# UltraThink execution:
1. Detects task type: Investigation (keyword "investigate")
2. Checks environment: AMPLIHACK_USE_RECIPES not set to 0
3. Tries Recipe Runner:
   amplihack recipe run --help  # SUCCESS — binary found
4. Executes via Recipe Runner:
   amplihack recipe run investigation-workflow \
     -c task="investigate caching layer" \
     --verbose
5. Recipe Runner automatically:
   - Phase 1: Scope Definition (code-enforced)
   - Phase 2: Exploration Strategy (code-enforced)
   - Phase 3: Parallel Deep Dives (code-enforced via knowledge-archaeologist)
   - Phase 4: Verification & Testing (code-enforced)
   - Phase 5: Synthesis (code-enforced)
   - Phase 6: Knowledge Capture (code-enforced - updates DISCOVERIES.md)
6. Investigation findings stored in context
7. Success: Full investigation with code enforcement

Example 3: Recipe Runner Unavailable (Falls Back to Skills)¶

# User invokes ultrathink
/ultrathink implement feature X

# UltraThink execution:
1. Detects task type: Development
2. Checks environment: AMPLIHACK_USE_RECIPES not set to 0
3. Tries Recipe Runner:
   amplihack recipe run --help  # FAILS — binary not found on PATH
4. Falls back to workflow skills:
   Skill(skill="default-workflow")
5. Skill loads workflow instructions
6. Claude follows workflow steps via prompts (not code-enforced)
7. TodoWrite tracks progress
8. Success: Workflow completed (prompt-based, not code-enforced)

Example 4: Force Prompt-Based Execution (AMPLIHACK_USE_RECIPES=0)¶

# User sets environment variable
export AMPLIHACK_USE_RECIPES=0

# User invokes ultrathink
/ultrathink implement feature Y

# UltraThink execution:
1. Detects task type: Development
2. Checks environment: AMPLIHACK_USE_RECIPES=0 (skip Recipe Runner)
3. Skips Recipe Runner (forced by environment variable)
4. Uses workflow skills directly:
   Skill(skill="default-workflow")
5. Skill loads workflow instructions
6. Claude follows workflow steps via prompts (not code-enforced)
7. TodoWrite tracks progress
8. Success: Workflow completed (prompt-based, not code-enforced)

Example 5: Recipe Runner Fails Mid-Execution (Falls Back to Skills)¶

# User invokes ultrathink
/ultrathink implement complex feature Z

# UltraThink execution:
1. Detects task type: Development
2. Checks environment: AMPLIHACK_USE_RECIPES not set to 0
3. Tries Recipe Runner:
   amplihack recipe run --help  # SUCCESS — binary found
4. Executes via Recipe Runner:
   amplihack recipe run default-workflow -c task="..." --verbose
5. Recipe Runner executes:
   - Step 1-10: SUCCESS (code-enforced)
   - Step 11: FAILS (adapter error - SDK tool unavailable)
6. Recipe Runner stops immediately (fail-fast)
7. Error message:
   "Recipe execution failed at Step 11 (Mandatory Local Testing): SDK adapter error - Bash tool unavailable in this context"
8. UltraThink falls back to workflow skills:
   Skill(skill="default-workflow")
9. Skill resumes from Step 11 (prompt-based)
10. Success: Workflow completed (hybrid: code-enforced for Steps 1-10, prompt-based for 11-22)

Context Passing to Recipe Runner¶

Recipe Runner receives context from ultrathink via -c flags:

Development Tasks¶

amplihack recipe run default-workflow \
  -c task="implement JWT authentication" \
  -c user_requirements="Use RS256 algorithm, store keys in vault" \
  -c user_constraints="Must pass CI within 30 minutes" \
  -c language="Python" \
  -c framework="FastAPI" \
  --verbose

Investigation Tasks¶

amplihack recipe run investigation-workflow \
  -c task="investigate caching layer" \
  -c focus_areas="Redis integration, cache invalidation, performance" \
  -c depth="deep" \
  -c output_format="architecture-doc" \
  --verbose

Q&A Tasks¶

amplihack recipe run qa-workflow \
  -c question="what is the purpose of the workflow system?" \
  -c context_needed="workflow types, execution hierarchy" \
  -c detail_level="concise" \
  --verbose

Context Available to Recipe Steps¶

Each recipe step can access context variables passed via -c flags, plus accumulated context from previous steps. The recipe runner merges user-provided context with step outputs automatically.

Troubleshooting¶

Problem: Recipe Runner Not Activating¶

Symptoms:

/ultrathink uses workflow skills instead of Recipe Runner
No "Using Recipe Runner for code-enforced execution" message

Diagnosis:

# Check if Recipe Runner binary is installed
amplihack recipe run --help && echo 'Recipe Runner available' || echo 'Recipe Runner not found'

# Check environment variable
echo $AMPLIHACK_USE_RECIPES
# Should be: (empty) or "1" for Recipe Runner
# Should NOT be: "0" (this disables Recipe Runner)

Solution:

# If binary not installed:
cargo install amplihack-rs

# If environment variable is 0:
unset AMPLIHACK_USE_RECIPES
# or
export AMPLIHACK_USE_RECIPES=1

Problem: Recipe Runner Fails with Binary Not Found¶

Symptoms:

Error message: "amplihack: command not found"
Falls back to workflow skills

Diagnosis:

# Check if amplihack is on PATH
which amplihack

# Check installation
cargo install --list | grep amplihack

Solution:

# Reinstall the amplihack CLI binary
cargo install amplihack-rs --force

# Or build from source
cargo install --git https://github.com/rysweet/amplihack-rs-recipe-runner

Problem: Recipe Execution Fails Mid-Workflow¶

Symptoms:

Recipe starts executing
Fails at specific step with error message
Falls back to workflow skills for remaining steps

Diagnosis:

# Check Recipe Runner logs
cat ~/.amplihack/.claude/runtime/logs/<session_id>/recipe_runner.log

# Look for error at specific step:
# "Recipe execution failed at Step 11 (Mandatory Local Testing): [error details]"

Solution:

# If adapter error (SDK tool unavailable):
# - This is expected in some contexts (e.g., restricted environments)
# - Recipe Runner will fall back to workflow skills
# - No action needed

# If recipe definition error:
# - Report bug to amplihack maintainers
# - Temporary workaround: export AMPLIHACK_USE_RECIPES=0

# If step logic error:
# - Check if workflow step has changed (recipe may be outdated)
# - Update recipe YAML to match current workflow

Problem: Context Not Passing Between Steps¶

Symptoms:

Recipe executes but later steps missing data from earlier steps
Agents report "no design provided" when design was created in Step 5

Diagnosis:

# Check Recipe Runner context logs
cat ~/.amplihack/.claude/runtime/logs/<session_id>/recipe_context.json

# Look for context at each step:
# {
#   "step_1": {"clarified_requirements": {...}},
#   "step_5": {"design": {...}},
#   "step_6": {}  # MISSING implementation
# }

Solution:

# Context accumulation is automatic in Recipe Runner
# If context is missing:
# 1. Check recipe YAML outputs definition for the failing step
# 2. Check if adapter properly extracts outputs from agent responses
# 3. Report bug to amplihack maintainers if both are correct

# Temporary workaround:
export AMPLIHACK_USE_RECIPES=0  # Use prompt-based execution

Problem: Force Prompt-Based Not Working¶

Symptoms:

Set AMPLIHACK_USE_RECIPES=0
Recipe Runner still activates

Diagnosis:

# Check environment variable in current shell
echo $AMPLIHACK_USE_RECIPES

# Check if variable is exported (available to subprocesses)
env | grep AMPLIHACK_USE_RECIPES

Solution:

# Ensure variable is exported
export AMPLIHACK_USE_RECIPES=0

# Verify:
echo $AMPLIHACK_USE_RECIPES  # Should print: 0

# Try ultrathink again:
/ultrathink implement feature

Problem: Recipe Runner Too Slow¶

Symptoms:

Recipe execution takes significantly longer than prompt-based
Steps seem to pause between execution

Diagnosis:

# Check Recipe Runner timing logs
cat ~/.amplihack/.claude/runtime/logs/<session_id>/recipe_timing.log

# Look for slow steps:
# Step 5 (Research and Design): 45.2s  # Expected
# Step 6 (Implementation): 120.5s      # Expected
# Step 11 (Mandatory Local Testing): 180.3s  # TOO SLOW

Solution:

# Recipe Runner may be slower than prompt-based due to code enforcement overhead
# This is expected trade-off for reliability

# If unacceptably slow:
# Option 1: Use prompt-based for time-sensitive work
export AMPLIHACK_USE_RECIPES=0

# Option 2: Optimize recipe steps (advanced)
# - Check if steps can run in parallel
# - Check if conditional skips can reduce work
# - Report performance issues to amplihack maintainers

Migration Guide¶

No migration needed! The Recipe Runner integration is fully backward compatible:

Existing workflows continue to work unchanged
No code changes required
No configuration changes required
Workflows automatically use Recipe Runner when available
Automatic fallback to prompt-based if Recipe Runner unavailable

What Changed Under the Hood¶

Aspect	Before (Prompt-Based)	After (Recipe Runner)
Execution	Claude follows markdown prompts	Python SDK adapters execute steps
Context	Claude remembers context	Context accumulated in Python dict
Errors	Claude continues on errors	Recipe Runner stops immediately (fail-fast)
Conditionals	Claude interprets conditions	Python evaluates conditions
Agents	Invoked via Task tool	Invoked via SDK adapter
Fallback	None (always prompt-based)	Falls back to prompt-based if Recipe Runner unavailable

What Stayed the Same¶

Workflow markdown files (still authoritative source of truth)
Agent definitions (unchanged)
User commands (/ultrathink syntax unchanged)
Output format (same results, different execution path)
TodoWrite tracking (still used for progress visibility)

Verifying Your Setup¶

# Check if Recipe Runner is active
/ultrathink implement small test feature

# Look for this message in output:
# "Using Recipe Runner for code-enforced execution"

# If you see this instead:
# "Recipe Runner unavailable, falling back to workflow skills"
# Then Recipe Runner is not installed or disabled

# To verify Recipe Runner installation:
amplihack recipe run --help && echo 'Recipe Runner installed' || echo 'Recipe Runner not found'

Benefits of Recipe Runner Integration¶

For Developers¶

Reliability:

Fail-fast prevents cascading errors
Code enforcement ensures steps actually execute (not just suggested)
Deterministic behavior (same inputs → same outputs)

Debugging:

Clear error messages at exact step that failed
Context preserved up to failure point
Can fallback to prompt-based for comparison

Performance:

Context accumulation more efficient (Python dict vs. LLM memory)
Conditional skips work reliably (no ambiguity)
Parallel execution where defined in recipe

For Workflow Authors¶

Maintainability:

Single source of truth (markdown workflows drive recipes)
Recipe YAML is generated from markdown (not manual maintenance)
Changes to markdown automatically update recipes

Evolution:

Can prototype new features in markdown (prompt-based)
Graduate to recipes when proven (code-enforced)
Both modes work simultaneously (gradual migration)

Testing:

Can test recipes independently (unit tests for workflow steps)
Can test prompt-based mode independently (AMPLIHACK_USE_RECIPES=0)
Can compare both modes for regression testing

Advanced Topics¶

Creating Custom Recipes¶

See docs/recipes/README.md for complete guide on creating custom recipes that integrate with ultrathink.

SDK Adapter Development¶

See docs/recipes/SDK_ADAPTER_GUIDE.md for guide on developing SDK adapters that bridge Claude Code tools to Recipe Runner.

Recipe Testing¶

See docs/recipes/TESTING_RECIPES.md for guide on testing recipes independently before integrating with ultrathink.

Recipe Runner Overview - Core Recipe Runner concepts
Workflow System - How workflows are structured
UltraThink Command - Full ultrathink reference
Agent Orchestration - How agents are invoked

Feedback and Issues¶

If you encounter issues with Recipe Runner integration:

Check troubleshooting section above
Try forcing prompt-based mode: export AMPLIHACK_USE_RECIPES=0
Report bugs at: https://github.com/amplihack/amplihack/issues
Include Recipe Runner logs from: ~/.amplihack/.claude/runtime/logs/<session_id>/

Remember: Recipe Runner is an enhancement, not a requirement. If it's not working, the system automatically falls back to proven prompt-based execution. You always have a working workflow system.

Recipe Runner Integration with UltraThink Orchestrator¶

Overview¶

Quick Start¶

Using Recipe Runner (Default Behavior)¶

Forcing Prompt-Based Execution (When Needed)¶

Three-Tier Fallback System¶

When Each Tier Activates¶

How Recipe Runner Works¶

Architecture¶

Step Execution¶

Context Accumulation¶

Fail-Fast Behavior¶

Environment Variable Control¶

AMPLIHACK_USE_RECIPES¶

When to Set AMPLIHACK_USE_RECIPES=0¶

Usage Examples¶

Example 1: Development Task (Recipe Runner Succeeds)¶

Example 2: Investigation Task (Recipe Runner Succeeds)¶

Example 3: Recipe Runner Unavailable (Falls Back to Skills)¶

Example 4: Force Prompt-Based Execution (AMPLIHACK_USE_RECIPES=0)¶

Example 5: Recipe Runner Fails Mid-Execution (Falls Back to Skills)¶

Context Passing to Recipe Runner¶

Development Tasks¶

Investigation Tasks¶

Q&A Tasks¶

Context Available to Recipe Steps¶

Troubleshooting¶

Problem: Recipe Runner Not Activating¶

Problem: Recipe Runner Fails with Binary Not Found¶

Problem: Recipe Execution Fails Mid-Workflow¶

Problem: Context Not Passing Between Steps¶

Problem: Force Prompt-Based Not Working¶

Problem: Recipe Runner Too Slow¶

Migration Guide¶

What Changed Under the Hood¶

What Stayed the Same¶

Verifying Your Setup¶

Benefits of Recipe Runner Integration¶

For Developers¶

For Workflow Authors¶

Advanced Topics¶

Creating Custom Recipes¶

SDK Adapter Development¶

Recipe Testing¶

Related Documentation¶

Feedback and Issues¶