Method tryReplayCachedResponse

pkg/runtime/cache.go:53–92

tryReplayCachedResponse looks up the latest user message in the agent's response cache. On a hit it replays the cached answer as the assistant message, fires stop hooks (which lets user-defined stop hooks run as they would for a real response), and returns true so the caller can short-circuit the ru

(
	ctx context.Context,
	sess *session.Session,
	a *agent.Agent,
	events EventSink,
)

// because no hook event currently supports short-circuiting the run with
// a synthetic response.
func (r *LocalRuntime) tryReplayCachedResponse(
	ctx context.Context,
	sess *session.Session,
	a *agent.Agent,
	events EventSink,
) bool {
	c := a.Cache()
	if c == nil {
		return false
	}
	question := sess.GetLastUserMessageContent()
	if question == "" {
		return false
	}
	_, cacheSpan := genai.RecordCacheLookup(ctx, "")
	cached, ok := c.Lookup(question)
	cacheSpan.SetHit(ok && cached != "")
	cacheSpan.End()
	// Treat empty stored values as misses: cache_response only stores
	// non-empty responses, so an empty entry only surfaces if the JSON
	// file was hand-edited or downgraded from a future version. Replaying
	// nothing would leave the user staring at a blank assistant message,
	// so we fall through to the model instead.
	if !ok || cached == "" {
		return false
	}

	slog.DebugContext(ctx, "Response cache hit; replaying cached answer",
		"agent", a.Name(), "session_id", sess.ID)
	modelID := a.Model(ctx).ID().String()
	events.Emit(AgentInfo(a.Name(), modelID, a.Description(), a.WelcomeMessage()))
	addAgentMessage(sess, a, &chat.Message{
		Role:      chat.MessageRoleAssistant,
		Content:   cached,
		CreatedAt: time.Now().Format(time.RFC3339),
		Model:     modelID,
	}, events)
	r.executeStopHooks(ctx, sess, a, cached, events)
	return true
}

// cacheResponseBuiltin is the stop-hook builtin that stores the
// assistant's response in the agent's response cache. It is registered