mbruhler/claude-orchestration/executing-workflows

Executing Orchestration Workflows

I execute workflows with real-time visualization, progress tracking, and interactive steering at checkpoints.

When I Activate

I automatically activate when you:

• Provide workflow syntax to execute
• Ask to "run a workflow"
• Mention workflow execution
• Want to execute a template
• Ask "how do I run this workflow?"

Quick Start

Just provide workflow syntax and I'll handle the rest:

Explore:"Analyze codebase":analysis ->
implement:"Add feature based on {analysis}":code ->
general-purpose:"Run tests":results

I automatically:

1. Parse and validate syntax
2. Show execution graph visualization
3. Execute agents with progress updates
4. Handle checkpoints and steering
5. Manage errors gracefully
6. Clean up temporary files (with user confirmation)

Execution Process

Phase 1: Parse & Validate

I analyze your workflow:

• Validate syntax correctness
• Check agent references
• Verify variable bindings
• Identify checkpoints
• Map execution graph

Phase 2: Visualize

I show you the execution plan using ASCII art:

Execution Graph:
+-----------------+
| Explore         |
| (Analyze code)  |
+--------+--------+
         |
         v
+-----------------+
| implement       |
| (Add feature)   |
+--------+--------+
         |
         v
+-----------------+
| general-purpose |
| (Run tests)     |
+-----------------+

Phase 3: Execute & Save State (Snapshots)

I run agents sequentially or in parallel. Crucially, I maintain a local execution state file to ensure reliability and enable "Time-Travel" recovery.

1. Before starting, I create/update .orchestration/state.json in the current working directory.
2. After EVERY successful node execution, I update the state.json with captured variables and the node's completed status.
3. If the workflow is interrupted, crashes, or fails, I retain this state file.

Sequential (->):

Running: Explore...  [In Progress]
Result: Analysis complete -> State Saved.
Running: implement...  [In Progress]
Result: Feature added -> State Saved.

⏪ Time-Travel Recovery (Resuming)

If you ask me to "resume" a workflow:

1. I read ./.orchestration/state.json.
2. I inject all previously captured variables back into the context.
3. I skip all nodes marked as completed and instantly start executing from the exact node where it left off.

Parallel (||):

Running: task1...  [In Progress]
Running: task2...  [In Progress]
Running: task3...  [In Progress]
All complete! Merging results...

Phase 4: Steering

At checkpoints (@review), you control flow:

@review-point reached
 
Options:
  [C]ontinue - Proceed with workflow
  [R]etry - Re-run previous step
  [M]odify - Adjust and continue
  [A]bort - Stop workflow
 
Your choice?

Phase 5: Error Recovery

If agent fails, I offer options:

Agent 'implement' failed: Tests not passing
 
Options:
  - Retry with same instruction
  - Modify instruction and retry
  - Skip this step (continue workflow)
  - Abort workflow
 
What would you like to do?

Phase 6: Cleanup (CONDITIONAL)

IMPORTANT: After workflow execution, check for temporary files and ask user before deleting.

Step 1: Detect Temporary Files

Check for existence of temporary files in the current working directory:

# Check temp-agents
TEMP_AGENTS=$(ls ./temp-agents/*.md 2>/dev/null | wc -l)
 
# Check temp-scripts
TEMP_SCRIPTS=$(ls ./temp-scripts/* 2>/dev/null | wc -l)
 
# Check temporary JSON (NOT .flow files!)
TEMP_JSON=$(ls ./examples/*.json 2>/dev/null | wc -l)
 
TOTAL=$((TEMP_AGENTS + TEMP_SCRIPTS + TEMP_JSON))

Step 2: If Files Exist, Ask User

Only if `TOTAL > 0`, use AskUserQuestion:

AskUserQuestion({
  questions: [{
    question: "Found ${TOTAL} temporary files. Do you want to delete them?",
    header: "Cleanup",
    multiSelect: false,
    options: [
      {label: "Yes, delete all", description: "Remove all temporary files (temp-agents, temp-scripts, temp JSON)"},
      {label: "Show me first", description: "List files before deciding"},
      {label: "No, keep them", description: "Leave files for manual review"}
    ]
  }]
})

Step 3: Execute Cleanup if Confirmed

If user chose "Yes, delete all":

# Delete temp-agents (in current working directory)
rm -f ./temp-agents/*.md
 
# Delete temp-scripts (in current working directory)
rm -rf ./temp-scripts/*
 
# Delete temporary JSON only (keep .flow files!)
rm -f ./examples/*.json

Report what was deleted:

Cleaned up ${TOTAL} temporary files:
- ${TEMP_AGENTS} temp agents removed
- ${TEMP_SCRIPTS} temp scripts removed
- ${TEMP_JSON} temporary JSON files removed

If user chose "Show me first":

1. List all files with paths
2. Ask again: "Delete these files?" with Yes/No options

Step 4: Skip if No Files

If TOTAL == 0, skip cleanup silently (don't bother user).

Note for Scheduled/Headless runs: If the workflow is running automatically via @schedule, the interactive cleanup prompt is suppressed. Temporary files will be left intact so they can be reused on the next run, saving computation time.

Syntax Reference

See syntax-reference.md for complete syntax documentation.

Quick reference:

Agent Types

Built-in Claude Code agents (no prefix):

• Explore - Fast codebase exploration and search
• Plan - Planning and breaking down tasks
• general-purpose - Versatile agent for complex multi-step tasks

Plugin agents (orchestration: prefix):

• orchestration:workflow-socratic-designer - Workflow creation via Socratic method
• orchestration:workflow-syntax-designer - Custom syntax design

External agents (registered via /orchestration:init):

• Agents from ~/.claude/agents/ can be registered and used directly
• Example: expert-code-implementer, code-optimizer (if registered)

Temp agents ($name):

• Created during workflow execution in ./temp-agents/
• Automatically cleaned up after workflow (with user confirmation)
• Can be promoted to permanent agents if useful

Variable Passing

See variables.md for advanced variable usage.

Capture output:

Explore:"Find routes":routes ->
analyze:"Check {routes}":findings

Conditional on variables:

test:"Run tests":results ->
(if results.passed)~> deploy ->
(if results.failed)~> debug

Error Handling

Common error patterns:

Retry on failure:

@attempt ->
operation:"Try task" ->
(if failed)~> wait:"Wait 5s" -> @attempt ~>
(if passed)~> continue

Fallback path:

primary:"Try primary" ->
(if failed)~> backup:"Use backup" ~>
(if passed)~> process

Stop on critical error:

security-scan:"Scan" ->
(if critical-issues)~> @emergency-stop -> abort ~>
(if clean)~> deploy

Checkpoints

See checkpoints.md for checkpoint details.

Basic checkpoint:

implement -> @review -> deploy

Labeled checkpoint:

@quality-gate:"Review code quality. Approve?"

Conditional checkpoint:

(if security-critical)~> @security-review

Parallel Execution

See parallel.md for parallel execution patterns.

Basic parallel:

[task1 || task2 || task3] -> merge

Parallel with individual variables:

[
  task1:"First":result1 ||
  task2:"Second":result2 ||
  task3:"Third":result3
] ->
general-purpose:"Process {result1}, {result2}, {result3}"

Conditional parallel:

(if needs-full-scan)~> [security || performance || style] ~>
(if needs-quick-check)~> basic-lint

Examples

See examples/ for categorized workflow examples:

• sequential.md - Sequential workflows
• parallel.md - Parallel execution
• conditional.md - Conditional logic
• error-handling.md - Error recovery
• checkpoints.md - Manual gates

Execution Modes

Normal mode (default):

• Full execution with all phases
• Interactive checkpoints
• Error recovery prompts

Dry-Run Unit Tests (.flow.test / Mock Mode):

• Parse and validate only.
• If you run a file with .flow.test extension or explicitly ask for a "dry-run with mocks":

1. I will look for a mocks: section in the YAML frontmatter of the workflow (where you define fake outputs for variables).
2. I will traverse the execution graph without invoking any actual agents.
3. Whenever a variable is needed for a conditional (~>) or a semantic router (route() =>), I will use your provided mocked data to decide the path.
4. I will output the exact execution path taken based on the mocks, allowing you to unit-test complex logic before running on production code.

Auto mode / Headless:

• Skip checkpoint prompts
• Automatic error retry (up to 3 times)
• Minimal user interaction

Progress Tracking

During execution, I show:

Workflow: TDD Implementation
Progress: [========..] 80%
 
Phase 1: Done - Requirements analyzed
Phase 2: Done - Tests written
Phase 3: Done - Tests verified failing
Phase 4: Paused - Checkpoint: review-test-coverage
Phase 5: In Progress - Implementing code...
Phase 6: Pending
Phase 7: Pending

Workflow Metadata

Track execution metadata:

Workflow: debug-and-fix.flow
Started: 2025-01-08 14:32:10
Duration: 5m 23s
Agents used: 8
Checkpoints: 2
Status: Complete
 
Agents executed:
- Explore (x1)
- general-purpose (x5)
- expert-code-implementer (x2)
 
Resources:
- Files read: 12
- Files modified: 3
- Tests run: 1

Tips for Successful Execution

1. Start simple - Test with small workflows first
2. Use checkpoints - Add review points for critical steps
3. Capture outputs - Use variables to pass data between agents
4. Handle errors - Add fallback paths for critical operations
5. Monitor progress - Watch execution visualization

Common Issues

Agent not found:

• Check agent name spelling
• Verify temp agent exists in ./temp-agents/
• Ensure namespace prefix for plugin agents

Variable not found:

• Verify variable was captured with :varname
• Check variable name spelling in {varname}
• Ensure variable set before use

Checkpoint skipped:

• Checkpoints only work in normal mode
• Check checkpoint syntax: @checkpoint-name

Parallel execution failed:

• Ensure parallel tasks are independent
• Check bracket syntax: [a || b]
• Verify no shared state between parallel tasks

Related Skills

• creating-workflows: Design and create workflows
• managing-agents: Create and manage custom agents
• debugging-workflows: Debug workflow issues
• using-templates: Execute workflow templates

Commands

• /orchestration:run - Execute workflow from file or inline
• /orchestration:template - Execute saved template
• /orchestration:explain - Explain workflow execution plan

Ready to execute? Provide your workflow syntax or template name!