The AI copilot

The chat pane is a thin front over one library loop. crates/vike-studio/src/chat.rs's ChatPane owns the transcript, the provider selector and the last result; the work happens in vike-ai's develop_strategy loop (crates/vike-ai/src/agent.rs) — generate, compile-validate, backtest out-of-sample, repair. The pane runs it the way the Run toolbar runs a backtest: ChatPane::send spawns a worker thread and hands back a receiver, and StudioState::poll drains it beside the run, sweep and walk-forward receivers.

What Send needs

Send is enabled only when four things hold at once: no chat round trip is already running, the selected provider has a key, the input is non-empty, and a data slice is selected in the picker. The last is not decoration — StudioState::start_chat_send says the model needs a (venue, symbol, interval) to backtest against, the same slice the Run toolbar uses (see Data and slices).

Providers and keys

crates/vike-ai/src/provider.rs's Provider offers two: Anthropic, the default, and Cerebras. Each needs one key — ANTHROPIC_API_KEY or CEREBRAS_API_KEY, the names chat.rs's key_env_var fixes for both the lookup and the "no key" error, so the failure message names the variable to set.

ChatApiKeys says where they come from: <project>/settings/secrets.env, the same store every venue credential lives in, resolved by the same walk. The pane itself opens no store and reads no environment — the desktop app resolves the keys from the credential map it already loads and hands them in as a parameter. And they are not venue credentials: absent, the pane offers no provider and Send stays disabled, and nothing about trading changes.

What the model is told

Two things precede every message. The system prompt, agent.rs's STRATEGY_SYSTEM_PROMPT, is the Rhai contract (the scripting tree's host surface): define fn on_bar(); read position(), price() and the host-bound indicators (NaN until warmed up); order with market(dir, qty) or buy/sell; declare sweepable knobs with param("x", default); submit through a submit_strategy tool carrying code and a one-line explanation. An unbound name raises on the first call rather than returning a plausible number.

Then chat.rs's grounding_preamble: the target slice — venue, symbol, interval — a note that the strategy is backtested on that instrument's out-of-sample data, and one valid template as a worked example. It deliberately does not enumerate the indicator registry. The doc records that the reason inverted: the host used to bind only three indicators; now it binds most of the registry, and what remains is a token budget — a fixed roster in every message crowds out the context specific to this chat.

The loop, and how it scores

develop_strategy scores only on the out-of-sample tail the model never saw (holdout_frac). A rejected submission goes back with the compile error or the reason it did not trade, and the model resubmits. max_repairs caps the number of submission attempts actually compiled and backtested — not the number of tool calls — and once the cap is reached every further submit_strategy is refused with a terminal message.

Reviewing the result

The transcript line is summary_of's: accepted, with the OOS Sharpe and trade count; not accepted after N attempts, with the problems; or an error. Below it the pane shows a result chip and a Review changes section — diff_rows, a line-level diff of the current editor buffer against the generated script. apply_result's doc states the contract: the diff comes first so Apply is a reviewed action rather than a blind clobber, and the write itself is still a whole-buffer replace once you accept. Until you press Apply to editor, the editor is untouched.

Memory across sessions

The loop has a ledger. crates/vike-ai/src/ledger.rs keeps two best-effort JSON files under the store root — ai_trials.json and ai_learnings.json — colocated with studio_strategies.json (see Saved strategies and workspace); the Studio resolves the path with LedgerPaths::under because the library reads no environment. Per agent.rs's ledger section it buys two things: prompt grounding, so the loop stops rediscovering duds it already rejected on this exact slice, and cross-session deflation — the deflated Sharpe ratio (Bailey and López de Prado, 2014) counts every prior accepted trial on the same (venue, symbol, interval), so a single-candidate chat that is accepted on a slice already carrying an accepted trial gets a real deflated_sharpe instead of the 0.0 a ledger-less call reports. Fewer than two trials in total — the first chat on a fresh slice included — leaves deflation undefined and the field at 0.0. With a ledger present the model is also offered a record_learning tool for notes future sessions read; without one it is not advertised.

The ledger is bounded — MAX_TRIALS 500 rows, oldest dropped first; equity curves kept on the newest MAX_CURVE_ROWS 50 accepted rows; MAX_LEARNINGS 200 notes of at most MAX_LEARNING_TEXT 1024 bytes — and best-effort: a missing file is empty, a corrupt one is empty plus a warning, a failed write is a warning. The loop never fails because the ledger is unavailable.

ChatOutcome's error arm covers the no-key case and a worker that panicked; both land in the transcript as a readable line rather than a silent stop. The Connect to Claude button at the bottom of the pane uses none of these keys and has its own page.