fix(provider): stop compaction summaries inheriting reasoning effort / 压缩摘要请求强制无推理，避免思维等级消耗摘要预算 (#9613)

internal/agent/compact.go

@@ -382,6 +382,11 @@ func (a *Agent) summaryRequest(region []provider.Message, instructions string) p
 		Tools:       schemas,
 		MaxTokens:   a.summaryOutputBudget(),
 		Temperature: provider.OptionalTemperature(a.temperature),
+		// The compaction prompt explicitly asks for a direct briefing with no
+		// thinking, and the summary budget must not be consumed by reasoning on
+		// thinking models. Force reasoning off so compaction works even when the
+		// user has a reasoning effort configured (#9613).
+		EffortOverride: "none",
 	}
 }
 

internal/provider/responses/responses.go

@@ -268,7 +268,13 @@ func (c *client) buildRequestBody(req provider.Request) (map[string]any, bool, [
 	messages := provider.SanitizeToolPairing(provider.ModelMessages(req.Messages))
 	body := map[string]any{"model": c.model, "stream": true}
 
-	effort := strings.ToLower(strings.TrimSpace(c.effort))
+	// A per-request EffortOverride (e.g. summary/compaction forcing thinking
+	// off) wins over the client's configured effort; otherwise client effort
+	// reaches every request it builds (#9613).
+	effort := strings.ToLower(strings.TrimSpace(req.EffortOverride))
+	if effort == "" {
+		effort = strings.ToLower(strings.TrimSpace(c.effort))
+	}
 	if c.vendor == "deepseek" && (strings.EqualFold(strings.TrimSpace(c.model), "deepseek-v4-flash") || strings.EqualFold(strings.TrimSpace(c.model), "deepseek-v4-pro")) {
 		if effort == "medium" || effort == "xhigh" {
 			effort = "high"

internal/provider/responses/responses_test.go

@@ -116,6 +116,30 @@ func TestDeepSeekV4ResponsesEffortAliasesNormalizeToHigh(t *testing.T) {
 	}
 }
 
+// TestEffortOverrideWinsOverClientEffort covers #9613: a per-request
+// EffortOverride (used by compaction summary requests to force thinking off)
+// must override the client's configured effort in the serialized body, while a
+// request without an override still inherits the configured effort.
+func TestEffortOverrideWinsOverClientEffort(t *testing.T) {
+	message := []provider.Message{{Role: provider.RoleUser, Content: "hi"}}
+
+	// No override -> client effort (high) reaches the wire unchanged.
+	inherited := New(Config{Name: "deepseek", BaseURL: "https://api.deepseek.com", Model: "deepseek-v4-pro", Effort: "high"}).(*client)
+	body, _, _ := inherited.buildRequestBody(provider.Request{Messages: message})
+	if got, _ := body["reasoning"].(map[string]any)["effort"].(string); got != "high" {
+		t.Fatalf("no override inherited effort = %q, want high", got)
+	}
+
+	// Override "none" wins even when the client effort is high.
+	for _, clientEffort := range []string{"high", ""} {
+		overridden := New(Config{Name: "deepseek", BaseURL: "https://api.deepseek.com", Model: "deepseek-v4-pro", Effort: clientEffort}).(*client)
+		b, _, _ := overridden.buildRequestBody(provider.Request{Messages: message, EffortOverride: "none"})
+		if got, _ := b["reasoning"].(map[string]any)["effort"].(string); got != "none" {
+			t.Fatalf("override none with client effort %q serialized as %q, want none", clientEffort, got)
+		}
+	}
+}
+
 func TestRequestSerializesExplicitMaxOutputTokens(t *testing.T) {
 	client := New(Config{Name: "responses", BaseURL: "https://example.com", Model: "model"}).(*client)
 	body, _, _ := client.buildRequestBody(provider.Request{