Assistant Agent

Overview

Layers 1–3 and 5 on the Mental Model. The Assistant owns chat and the plan. hear_prompt has no tools — dish name only, short confirm, or refuse. After the prompt is saved, develop_plan (same harness in agents_run.py, driven by worker.py) seeds catalog and public recipes onto the Plan, fills open slots via the Researcher, then drafts summary, advice, and directions.

Behavior

• Dish names only in chat. Non-dish prompts are answered in chat and not saved.
• Confirm before side effects. Nothing is queued or researched from hear_prompt.
• SQLite capture queue. Confirm enqueues a server job (queue.db); the worker claims it, retries on failure, and clears the row after Neo4j save.
• Seed, then fill. Catalog favorites first, then public Neo4j matches not already offered this session; each is linked onto the plan immediately. Target is three recipes. Researcher scrapes only remaining slots.
• Orchestrator-Worker. Researcher hands scraped pages onto the plan as they are accepted; Assistant compiles the draft. See Building effective agents.
• Keep stored directions. Catalog and public matches keep their stored directions and nutrition. No analyze_recipe tool on this path.

Plan flow

One confirmed dish through the SQLite capture queue, then the worker. Diagram key: ✦ AI agent · second line = Anthropic workflow · purple = actor · white = service.

---
config:
  theme: default
  themeVariables:
    background: "#ffffff"
    fontSize: "14px"
---
flowchart TB
    person[Person]
    chat["✦ Assistant hear prompt
Prompt Chaining"]
    queue[SQLite capture queue]
    worker[Worker]
    plan["✦ Assistant develop plan
Orchestrator-Worker"]
    fav[Catalog favorites]
    pub[Public Neo4j]
    specialist["✦ fill open slots"]
    neo[Neo4j Plan]

    person -->|POST api listen| chat
    chat -->|yes enqueue capture| queue --> worker --> plan
    plan --> fav
    plan --> pub
    plan -->|remaining| specialist
    plan --> neo

    classDef agent fill:#e8f0ee,stroke:#0f3d36,stroke-width:2px,color:#0d0d0d
    classDef service fill:#fafaf8,stroke:#c8c8c2,stroke-width:1px,color:#0d0d0d
    classDef actor fill:#f3eef8,stroke:#6b4f9a,stroke-width:2px,color:#0d0d0d
    class chat,plan,specialist agent
    class queue,worker,fav,pub,neo service
    class person actor

Prompt: assistant_agent.md. Chat: POST /api/listen. Capture queue → worker → develop_plan. Retoast re-enters at the worker and leaves prior plan recipes out.