The assistant & deep-dive

Forven has a built-in chat assistant: a page-aware guide that can answer questions about the app and your lab, take you to the right page, and — when you let it — create strategies and run backtests for you. Deep-dive is the same assistant focused on a single strategy: open it from a strategy's page and "this strategy" means that one.

Like the rest of the agent layer, the assistant is free and runs on your own LLM key. It can never place or close trades.

Forven is a research tool. Anything the assistant produces — code, parameters, backtest numbers, explanations — is illustrative, not predictive, and nothing here is financial advice.

Opening the assistant

Click the chat button in the bottom-right corner of any page. The assistant docks beside the page instead of covering it, so you can keep working while it answers. It needs a connected AI provider.

It knows where you are: the current page and anything you are looking at (a strategy, for example) are passed to it as context, so you can ask "why is this strategy stuck?" on a strategy page or "what needs my attention?" on the dashboard without explaining further.

What it can do

The assistant's tools come in three tiers.

Always available (read-only). It can read your portfolio, pipeline, and market regime; a strategy's detail and gate report; recent trades; your settings, operations state, ideas and what came of them, bots, routines, and agent runs; and a built-in guide to the app itself. It can also open an app page for you — navigation only, no server-side change.

With actions on (run directly). When Allow actions is ticked, it can also create a strategy, run a backtest, and enter a quick-screen strategy into the gauntlet after a quick pre-screen, without asking. These only create drafts, simulations and evaluation runs — nothing touches money, and nothing is promoted to paper or live.

With actions on (ask first). A few actions always show a confirmation card that you must approve before anything runs: promoting a strategy, assigning a task to an agent, and registering a strategy file (which imports custom code). Approve, and it runs exactly the call you saw; decline, and nothing happens.

Some things are never offered to the assistant at all: placing or closing orders, editing trades, writing arbitrary files, running shell commands, and factory reset.

Untick Allow actions and the assistant answers and advises but takes no actions — useful when you just want to think out loud.

Deep-dive: one strategy at a time

On a strategy's page in The Forge, Deepdive opens the assistant focused on that strategy. Ask it to explain why the strategy diverged out of sample, suggest a parameter change and backtest it, or trace why a signal fires where it does. Every message, tool call, and result is kept in the conversation.

Deep-dive is for refinement, not promotion. A strategy still advances only through the gauntlet and the gates; the assistant can propose a promotion, but only with your confirmation, and the gate still decides.

Limits: tool rounds and cost

• Tool rounds. Within one reply the assistant can loop through tool calls — read, run, read the result — up to 12 rounds by default (Settings → System → Assistant → Chat tool rounds, 2 to 40). Two rounds before the limit it is told to wrap up; at the limit it gives you a summary of what it did and asks whether to continue, rather than failing.
• Cost per conversation. Each conversation has a spend cap — $5 by default. When a conversation reaches it, further messages are refused with a cost-cap error; start a fresh conversation, or change the cap with PUT /api/assistant/cost-cap.
• History. Long conversations replay only their most recent messages to the model, which bounds the cost of each turn. New in the panel header archives the conversation and starts fresh.

Costs are estimates from token counts and list prices. See cost controls.

Steps: work on a strategy with deep-dive

1. Open the strategy in The Forge and click Deepdive.
2. Describe the problem plainly — for example "Walk-forward fails out of sample. What in the entry logic might be overfit?"
3. Tick Allow actions if you want it to run things. Ask it to try a change and backtest it.
4. Read the returned metrics, out-of-sample first.
5. If it proposes a promotion or registering a file, read the confirmation card before you approve.

What you'll see: the reply streams in beside the page, tool calls appear inline as it works, and any action that needs you appears as a Confirm action card with Approve and Reject.

The API (for developers)

The assistant streams over Server-Sent Events (not a WebSocket):

POST /api/assistant/threads                                  create or get a thread
POST /api/assistant/threads/{thread_id}/send                 send a message (SSE stream)
GET  /api/assistant/threads/{thread_id}/messages             the transcript
POST /api/assistant/threads/{thread_id}/archive              archive and start fresh
POST /api/assistant/threads/{thread_id}/actions/{id}/confirm approve or reject a proposed action
GET  /api/assistant/cost-cap · PUT /api/assistant/cost-cap   read or set the per-thread cap

A strategy-scoped deep-dive thread API also remains for scripts — POST /api/deepdive/threads (with a strategy_id), /send, /messages, and /archive — whose tools read and write the strategy's code, update its default parameters, and run its backtest. Its turns share the same tool-round cap.

These routes are operator-gated: open on a keyless localhost, and they require x-operator-key once you set an operator key.

Caveats (beta)

• Edits are real. Code and parameter changes the assistant makes are persisted against the strategy — review them as you would your own.
• It advises; the gates decide. Nothing the assistant does bypasses the gauntlet, the promotion gates, or approvals.
• It spends your tokens. Every round re-sends the conversation; long, tool-heavy turns cost more.

Related

• Agents — the broader AI co-researcher layer
• Cost controls — throughput, caps, and where to watch spend
• The Forge — where strategies are built, optimized, and gauntleted
• Strategy Creator — build a strategy without the chat