2. The 5-Layer Multi-Agent AI Pipeline

Requirements generation is an intensive process requiring 60–180 seconds of multi-agent reasoning. SRA orchestrates this using an asynchronous event-driven model:

1. Job Dispatch: queueService.js computes an MD5 hash of the input for idempotency, creates an Analysis record with status PENDING, and publishes the job to Upstash QStash.
2. Worker Ingestion: QStash delivers the webhook to POST /api/worker/process. workerController.js atomically transitions the record to IN_PROGRESS via updateMany to guarantee that duplicate webhook deliveries are safely ignored.
3. Live SSE Streaming: The client connects to GET /api/analyze/:id/stream. Progress events (stage transitions, agent thoughts, section completions) are published to Redis Pub/Sub and pushed to the browser via Server-Sent Events (SSE).
4. Serverless Budget Checkpointing (`pipelineBudget.js`): On serverless environments (e.g. Vercel 300s limit), the pipeline checks execution budgets before heavy stages. If time runs low, it checkpoints intermediate state to PostgreSQL and re-enqueues itself to continue seamlessly without data loss.

Role-Based Agent Choreography

Rather than asking a single LLM to write an entire specification, SRA utilizes specialized agents with distinct domain responsibilities:

1. Product Owner Agent

• Decomposes business goals, identifies primary/secondary user classes, defines operating environments, and establishes boundary constraints.
• Formulates feature candidate lists and stimulus/response sequences.

2. Systems Architect Agent

• Determines external interface requirements (User Interfaces, Hardware Interfaces, Software APIs, and Communication Protocols).
• Generates Mermaid architecture diagrams (Sequence Diagrams, Flowcharts, Entity-Relationship Models).
• Injects relevant historical design patterns retrieved from the vector knowledge base.

3. Developer Agent

• Synthesizes the formal normative requirement statements ("The system shall...").
• Formats each section according to the target standard's exact schema descriptor with intentional rate-limit cooldown throttles.

The 6Cs Quality Standard

Every generated specification is audited by the CriticAgent against six core software engineering metrics:

1. Clarity: Unambiguous phrasing free of vague qualifiers ("fast", "user-friendly", "robust").
2. Completeness: Compulsory coverage of all required sections, external interfaces, and non-functional constraints.
3. Conciseness: High information density without fluff or duplicate requirement definitions.
4. Consistency: Zero internal contradictions between functional rules, database entities, and user roles.
5. Correctness: Valid Mermaid diagram syntax, proper RFC 2119 normative verbs ("SHALL", "SHOULD"), and accurate schema types.
6. Context: Alignment with user-provided domain knowledge and historical organizational standards.

Reflection & Surgical Repair Loop:

• If the aggregate quality score falls below 85/100, the ReviewerAgent flags specific deficient sections.
• The pipeline enters a reflection pass (maximum 2 passes) where the Developer Agent performs surgical refinement only on the flagged sections, preserving the rest of the document intact.