Node Types

# With agent:
claude --agent code-reviewer --print "Review changed files for security issues"

# Without agent:
claude --print "Generate unit tests for src/utils.ts"

# Cost is extracted from Claude CLI stderr:
# Looks for "Total cost: $X.XX" or "total_cost_usd: X.XX"

# Commands run via:
bash -c "{instructions}"

# stdout and stderr are streamed to the live log viewer
# Exit code 0 = success, non-zero = failure

npm run test -- --coverage
cargo build --release
docker compose up -d
eslint src/ --fix

git checkout -b feature/auto-fixes
git add -A && git commit -m "feat: AI-generated changes"
git push origin HEAD
git merge origin/main --no-edit

1. All child nodes are spawned concurrently via tokio::task::JoinSet
2. Results are processed in completion order (not spawn order)
3. Execution blocks until ALL children complete
4. Status = "success" only if ALL children succeed
5. Status = "failed" if ANY child fails
6. Individual child status is tracked separately

{
  "id": "node-parallel",
  "name": "Run All Checks",
  "type": "parallel",
  "instructions": "Run lint, test, and type-check in parallel",
  "children": ["node-lint", "node-test", "node-typecheck"],
  "inputs": [],
  "outputs": [],
  "position": { "x": 250, "y": 150 }
}

{
  "loop_config": {
    "separator": "newline",   // "newline" | "comma" | custom string
    "max_iterations": 100,    // Cap iterations (1-1000)
    "timeout": 600,           // Per-loop timeout in seconds
    "model": "sonnet"         // Model override for AI children
  }
}

$LOOP_ITEM   — The current item from the list
$LOOP_INDEX  — Zero-based index of the current iteration
$LOOP_COUNT  — Total number of items in the list

# Example: if instructions = "file1.ts\nfile2.ts\nfile3.ts"
# Iteration 0: $LOOP_ITEM="file1.ts", $LOOP_INDEX=0, $LOOP_COUNT=3
# Iteration 1: $LOOP_ITEM="file2.ts", $LOOP_INDEX=1, $LOOP_COUNT=3
# Iteration 2: $LOOP_ITEM="file3.ts", $LOOP_INDEX=2, $LOOP_COUNT=3

{
  "id": "node-loop",
  "name": "Review Each File",
  "type": "loop",
  "instructions": "src/auth.ts\nsrc/api.ts\nsrc/db.ts",
  "children": ["node-review"],
  "loop_config": {
    "separator": "newline",
    "max_iterations": 50
  },
  "inputs": [],
  "outputs": [],
  "position": { "x": 250, "y": 150 }
}

1. Node enters "running" state
2. Emits "approval-requested" event: { runId, nodeId, name }
3. UI shows approval dialog with the node instructions
4. Blocks until respond_to_approval(approved: bool) is called
5. If approved → status "success", approval_state "approved"
6. If rejected → status "failed", approval_state "rejected"
7. If timeout → status "failed" (no approval received)

1. Loads {project}/.claude/pipelines/{pipeline_ref}.pipeline.json
2. Executes the full pipeline recursively (run_pipeline_loop)
3. Circular references are detected and blocked:
   - Ancestor pipeline names are tracked through the chain
   - If pipeline_ref appears in ancestors → error
4. Cost is aggregated from all nodes in the sub-pipeline
5. Sub-pipeline success/failure propagates to parent

{
  "id": "node-deploy",
  "name": "Deploy to Staging",
  "type": "sub-pipeline",
  "instructions": "Execute the staging deployment pipeline",
  "pipeline_ref": "deploy-staging",
  "inputs": [],
  "outputs": [],
  "position": { "x": 250, "y": 300 }
}

{
  "id": "node-comment-1",
  "name": "Deployment Notes",
  "type": "comment",
  "instructions": "This section handles the staging deployment. Requires VPN access and valid AWS credentials.",
  "inputs": [],
  "outputs": [],
  "position": { "x": 50, "y": 600 }
}

{
  "retry": {
    "max": 3,    // Maximum number of attempts (1 = no retries)
    "delay": 10  // Seconds to wait between retries
  }
}

{
  "timeout": 300  // Kill process after 300 seconds (5 minutes)
}