Pipeline

The WorkflowPipeline class in src/autoflow/pipeline.py is the central orchestrator. It manages two execution modes, each running a different sequence of stages.

Pipeline Modes

Synthesize Mode (SITREP Generation)

Produces structured briefing documents from free-form input:

python -m autoflow "Summarize this week's cyber threat landscape"

Generate Mode (Workflow JSON)

Produces Ask Sage Agent Builder workflow JSON:

python -m autoflow "Route support tickets by category" --mode generate

Stage Details

Stage	Agent	Purpose
Navigate	NavigatorAgent	Routes to the right Ask Sage tool (only Agent Builder continues)
Normalize	NormalizerAgent	Extracts objective, deliverable type, constraints from messy input
Parse	BaseAgent (parser)	Splits input into structured sections (synthesize only)
Extract	BaseAgent (extractor)	Extracts events from parsed sections (synthesize only)
Synthesize	BaseAgent (synthesizer)	Produces SITREP output (synthesize only)
Generate	GeneratorAgent (WorkflowCompiler)	Produces v2 workflow JSON with pattern + model hints, then post-processes to ensure valid identity model (generate only)
Evaluate	EvaluatorAgent	Scores on 5 quality dimensions
Critique	CritiqueAgent	Identifies issues and suggests fixes
Envelope	—	Wraps result in Universal Output Envelope

Critique Loop

The evaluate-critique cycle runs up to MAX_CRITIQUE_ITERATIONS (default: 2):

Evaluate — Scores the current output on correctness, completeness, format validity, efficiency, and reusability
Check pass criteria — If correctness >= 4, format_validity == 5, and estimated_human_edit_minutes <= 15, the loop exits with PASS
Critique — If thresholds aren't met, the critique agent identifies specific issues
Regenerate — The generator incorporates critique feedback and produces a new version

If all iterations are exhausted without passing, the best version is returned with a FAIL verdict.

The GeneratorAgent (persona: WorkflowCompiler) includes a post-processing safety net (_ensure_identity_model()) that runs before validation. If the LLM produces v1-style output (snake_case IDs, missing slugs), the post-processor upgrades it to v2 format:

Generates UUID IDs for any non-UUID node IDs
Derives slugs from labels (or preserves old snake_case IDs)
Resolves slug collisions with numeric suffixes
Adds missing key fields
Updates edge source/target references to UUIDs

This ensures generated workflows are always valid v2 format, even if the underlying LLM doesn't follow the compiler-spec prompt perfectly.

Universal Output Envelope

Every pipeline run produces a standardized envelope validated against schemas/universal_output_envelop/schema.json:

{
  "objective": "What the user asked for",
  "deliverable_artifact": { ... },
  "quality_scores": {
    "correctness": 4,
    "completeness": 4,
    "format_validity": 5,
    "efficiency": 4,
    "reusability": 3
  },
  "confidence": 0.85,
  "assumptions": ["..."],
  "constraints": ["..."],
  "known_gaps": ["..."],
  "estimated_human_edit_minutes": 10,
  "recommended_next_actions": ["..."],
  "metadata": {
    "provider": "openai",
    "mode": "generate",
    "iterations": 1,
    "pattern": "classification"
  }
}

If you know Agent Builder is the right tool, skip the Navigator stage:

python -m autoflow "Classify emails" --mode generate --skip-navigator