Context sent to the model: live snapshot, prompt and window
Every question to the Senzoukria agent carries a system prompt (role, teaching style, drawing vocabulary, image guide), a live trading snapshot, a reply-language directive, and the last 24 turns of the thread, trimmed to the engine's token budget.
Senzoukria · Documentation · Updated September 2026
Where to find it
- Where
- AI Agent chat → "View live context" / "Hide context" button
- Turns sent
- MAX_CONTEXT_TURNS = 24 (twelve exchanges)
- Token estimate
- 3.5 characters per token, 400 tokens per image block
- Language
- Reply in the interface language; code, symbols and ids untouched
What it does
The system prompt is assembled in a fixed order: the base role (an embedded co-pilot inside a professional orderflow desktop app for US futures and options, read-only access, no buy/sell signals, never invent values missing from the snapshot), the teaching style, the senz-viz drawing vocabulary, the screenshot guide, then the live context, and finally the language directive. Static parts come first so the prefix stays identical and cacheable across turns, which lowers the cost on your own key.
The live context is a Markdown snapshot with a timestamp. The chart is always described first, including the fact that no chart is open, because without that section the model once took the GEX symbol for the chart symbol. Then, only when the module has data: the Gamma Exposure module (its own symbol, spot, zero gamma, call and put walls, total GEX, total DEX, put/call OI ratio, 25Δ skew, ATM IV), Option Flow (last N trades, call and put premium, buy and sell premium, mega trades ≥ $500K, biggest trade), the Account (balance, day P&L, day trades, win rate, up to 8 open positions), the 6 most recent headlines and the next 6 economic events.
The "View live context" button in the chat shows exactly this snapshot so you can check what the model was told before trusting a figure it quotes.
Settings
| Setting | Default | What it changes |
|---|---|---|
| MAX_CONTEXT_TURNS | 24 | Only the last 24 turns are sent; the window always starts on a user turn |
| Token budget (local engines) | engine context window, e.g. 8k in the model picker | Oldest turns are dropped until the rest fits; the last question is always kept |
| estimateTokens | 3.5 characters per token | Sizing only, never billing |
| Image block estimate | 400 tokens | Flat cost per image when fitting the budget |
| Reply language | interface language (English, French, Spanish, Italian, Simplified Chinese, Russian, Arabic) | Directive appended last; follows the user if they write in another language |
| Scripting knowledge | output contract, sandbox limits, conventions, indicator inventory | Added to the scripting assistant prompt; size is bounded by a test |
How to use it
- Open the module you want the model to reason about before asking: a module never opened this session has an empty store and no section in the snapshot.
- Press "View live context" when an answer looks off; if the section is missing, the model was asked to say so and ask you rather than guess.
- Start a new conversation for a new topic. Twelve exchanges are sent each time; an all-day thread pays for its tail on every question.
- For a local model, raise the context window in the model picker if requests are refused; a local engine rejects an oversized prompt instead of truncating it.
Limits or pitfalls
- Trimming applies at send time only. The thread stays whole on screen and on disk; what you pay for is the end.
- The snapshot is a photograph at the moment of sending. It does not update during a long reply.
- GEX values appear only when an options source has returned a chain; the section labels the GEX symbol as options data, not the chart symbol.
- The scripting prompt includes the full indicator inventory read from the registry (about 7,300 characters for 204 entries); it is sent with every scripting question.
Related pages
This page in other languages
Frequently asked questions
- Does the model see my whole conversation?
- It receives the last 24 turns, further reduced for local engines to what fits the token budget. The window always starts on one of your messages, because the API refuses a conversation that opens with an assistant turn.
- Why does the assistant answer in French when my interface is English?
- It should not: a language directive is appended last to every prompt, naming the interface language in English. If you write in another language, the model follows you.
- Why is there no GEX section in the live context?
- The GEX module has no snapshot yet. Open it and wait for an option chain; sections other than the chart are only added when their module holds data.