Method Probe

internal/llm/llm.go:264–278

Probe sends a minimal hello chat just to harvest response headers in one round trip: status validates the URL/model/key combo, X-Context-Window gives the live size, X-Budget-Remaining the live fraction. The body is closed unread; on the cloud proxy that may already charge one token, the cost of a si

(parent context.Context)

	// atomic.Bool: Probe and Chat race on the same Client (startup probe still
	// in flight when the first turn fires) and both read it via postChat; Chat
	// may also write it. A plain bool would be a data race.
	noReasoningEffort atomic.Bool
}

// New builds a Client governed by the caller's context, not http.Client.Timeout.
// That timeout is end-to-end (it covers body reads) and would kill a slow but
// legitimate SSE stream mid-flight on slow local backends. tui.Model's turnCtx
// is the single cancellation source; connect-level safety (DNS/TCP) is already
// bounded by Go's default Dialer (30s).
func New(base, model, token string) *Client {
	return &Client{
		BaseURL:      strings.TrimRight(base, "/"),
		Model:        model,
		Token:        token,
		HTTP:         &http.Client{},
		IdleTimeout:  idleTimeoutFromEnv(),
		RetryBackoff: defaultRetryBackoff,
	}