Method streamWithFallback

cmd/wasm/runtime_wasm.go:390–428

streamWithFallback tries the primary model, then fallbacks on failure.

(ctx context.Context, a *agent.Agent, messages []chat.Message, agentTools []tools.Tool)

// streamWithFallback tries the primary model, then fallbacks on failure.
func (rt *wasmRuntime) streamWithFallback(ctx context.Context, a *agent.Agent, messages []chat.Message, agentTools []tools.Tool) (*wasmStreamResult, error) {
	prov := rt.providers[a.Name()]
	fallbacks := rt.fallbacks[a.Name()]

	chain := append([]provider.Provider{prov}, fallbacks...)

	var lastErr error
	for i, p := range chain {
		if ctx.Err() != nil {
			return nil, ctx.Err()
		}

		if i > 0 {
			rt.emitEvent(map[string]any{
				"type":    "fallback",
				"from":    chain[i-1].ID(),
				"to":      p.ID(),
				"attempt": float64(i + 1),
				"reason":  lastErr.Error(),
			})
			// Brief backoff between fallback attempts.
			select {
			case <-ctx.Done():
				return nil, ctx.Err()
			case <-time.After(time.Duration(i) * 500 * time.Millisecond):
			}
		}

		result, err := rt.streamCompletion(ctx, p, messages, agentTools)
		if err != nil {
			lastErr = err
			slog.WarnContext(ctx, "Model attempt failed", "model", p.ID(), "attempt", i+1, "error", err)
			continue
		}
		return result, nil
	}

	return nil, fmt.Errorf("all models failed: %w", lastErr)
}

// streamCompletion runs one streaming completion call and emits deltas.
func (rt *wasmRuntime) streamCompletion(ctx context.Context, prov provider.Provider, messages []chat.Message, agentTools []tools.Tool) (*wasmStreamResult, error) {