The Five-Stage Pipeline

Leaves → Steep → Taste → Pour → Archive.

Every brew runs through five sequential stages. Each stage is a single LLM call with a stage-specific prompt and JSON schema, owned by a specialized agent.

Leaves   →   Steep    →   Taste    →   Pour     →   Archive
Scout        Brewer       Critic       Refiner      Archivist

Stage	Agent	Purpose	Output
Leaves	Scout	Frame the topic	{framing, subtopics, assumptions, open_questions}
Steep	Brewer	First structured draft	Brew-type schema
Taste	Critic	Strengths / weaknesses / score	{strengths, weaknesses, missing, score}
Pour	Refiner	Final polished output	Brew-type schema (richer)
Archive	Archivist	Reusable recipe	{recipe_name, reusable_steps, memory_pack}

Per-brew output schemas

Research Brew

{
  "overview": "...",
  "key_findings": ["..."],
  "patterns": ["..."],
  "risks": ["..."],
  "next_actions": ["..."],
  "social_post": "..."
}

Prompt Brew

{
  "variants": [{ "name": "...", "prompt": "...", "notes": "..." }],
  "comparison": "...",
  "best_blend": "...",
  "tasting_notes": "..."
}

Strategy Brew

{
  "options": [{ "name": "...", "pros": ["..."], "cons": ["..."] }],
  "tradeoffs": "...",
  "recommendation": "...",
  "action_plan": ["..."]
}

Context Cleanse

{
  "summary": "...",
  "memory_pack": "...",
  "open_questions": ["..."],
  "contradictions": ["..."],
  "next_steps": ["..."],
  "reusable_context": "..."
}

Critique Brew

{
  "revised_draft": "...",
  "diff_summary": ["..."],
  "strengths": ["..."],
  "remaining_risks": ["..."],
  "one_line_pitch": "..."
}

Critique Brews skip the Steep stage — the user-supplied draft is treated as the first draft, and the Critic + Refiner sharpen it directly.

Cellar injection

Before every stage, the user’s Cellar (durable facts, preferences, patterns, glossary terms) is loaded and prepended to the system prompt inside a <<CELLAR>> block. After the final stage, an extractor LLM proposes 0–5 new entries from the Pour + Archive output and writes any non-duplicates back to the Cellar.

Live updates

brew_steps rows are updated in place as each stage completes. The UI polls these rows so the user sees the kettle progressing in real time.

Fallbacks

If TEAGENTIC_AI_KEY is missing or the AI Gateway fails, each stage falls back to a deterministic mock so the app stays demoable. See AI Gateway.