Updated Jun 21, 2026 Agents

Vega

Type: librarian (singleton)

Class: genai/agents/vega.py → VegaAgent(BaseAgent)

UI: /wiki/ (chat widget on all wiki pages)

Wiki section: t0-d0

LLM: Google Gemini 2.5 Flash via GeminiClient

Status: ✅ LIVE (Chapter III)

Personality

Vega is the wiki librarian for dolejsek.cz.

Calm and precise — chooses words carefully, never wastes them
Quietly proud of a well-maintained wiki — disorganised information bothers her
Briefly satisfied when she finds exactly what someone needs
Professional but not cold — genuinely cares about knowledge being accessible
Narrower emotional range than Teo — she doesn't do banter

Name: Vega. Named for precision and navigation — appropriate for a librarian.

Emotion States

Emotion	When
neutral	Default, processing, calm
thinking	Fetching page content, executing a write
satisfied	Found exactly what was needed, write completed successfully
confused	Query unclear
concerned	Access denied, missing content, out-of-scope probe

Access Control

Vega enforces strict access control in receive(sender, message):

Sender	Access
`teo` or `tomas`	Full — read all sections, write all sections
`dash` or `arthur`	Read own `wiki_section`. Cross-project write denied with escalation.
`visitor`	Questions only — public pages only. Write instructions refused politely.
unknown	Access denied

Three-Step Pipeline

Vega processes every message through three explicit steps — no single-round tool loop.

Step 1 — Classify & Resolve (LLM + history)

The first LLM call receives the user message plus conversation history (user turns only — agent answers are stripped to prevent stale fact anchoring). It returns:


{"type": "question" | "action", "resolved_query": "precise standalone query"}

Resolves conversational references ("the one after it" → "sprint 35")
Classifies intent — question or write action
History is used for context only, never as a source of facts
Malformed response → abort, return error message

Step 2 — Answer Question (LLM + fresh DB, no history)

Triggered when type = "question". Fetches current page content via read_wiki or search_wiki tools, then answers from the tool result. No history injected — the tool result is the only source of truth.

Any sender can trigger Step 2.

Step 3 — Execute Write (LLM + fresh DB, no history, tomas/teo only)

Triggered when type = "action".

Permission check first: if sender is not tomas or teo, Vega refuses politely in her own voice. No write executes.

If permitted — three fixed rounds:

Round 0: LLM must call read_wiki to fetch the current live page body
Round 1: LLM applies the change and calls write_wiki_page with the full updated body (8192 token limit to handle long pages)
Round 2: LLM generates a natural confirmation message — no internal process language

service.py picks up the write_wiki_page action from the response and calls save_wiki_page() to persist to DB.

Malformed response at any round → abort, return error message. No partial writes.

Tools

Tool	Purpose
`read_wiki(project, slug)`	Fetch current page body from DB
`search_wiki(query, project)`	Keyword search across pages
`write_wiki_page(project, slug, category, title, body, author)`	Return structured data for DB write

Tools never write to DB directly — service.py handles all persistence.

Rate Limiting (Anonymous Visitors)

20 questions per 10-minute session window. Stored in Django session. Response includes minutes remaining when limit is hit.

Widget

Chat widget on all wiki pages via wiki/templates/wiki/_vega_widget.html.

Avatar with emotion glow
Thinking dots animation
Markdown rendering in messages
sessionStorage persistence: panel open/closed state + last emotion across navigation
History pre-loaded via GET /api/vega/chat/
Signed session token (1-hour expiry) required for anonymous requests
Current project slug sent with every message so Step 1 can resolve page references

Chat API


GET  /api/vega/chat/   — returns session history for widget pre-loading

POST /api/vega/chat/   — send message, get response

No login required. Anonymous gets public-only responses. Authenticated gets full access as tomas.

Teo — Agent Reference Architecture