feat: AutoZero 通用 Agent 框架（从 AutoGo 五阶段拆分完成）

llm（三协议 Provider）/ session（JSONL 持久化）/ tools（Registry、围栏、
builtin、computer）/ cdp（browser）/ skills / mcp / agent（循环、两级压缩、
编排、审批门控）+ 底座 jsonl/textx/execx/diag/scratch/spill/imagecache。

产品装配代码（config 桥接、schedule 工具、memory/plan/retro/prompt）
留在 AutoGo，宿主接入方式见 MIGRATION.md。

llm/retry_test.go

@@ -0,0 +1,104 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+)
+
+// TestRetryOn500：前两次 500，第三次成功——provider 应自动重试。
+func TestRetryOn500(t *testing.T) {
+	calls := 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls++
+		if calls <= 2 {
+			w.WriteHeader(http.StatusInternalServerError)
+			return
+		}
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.Write([]byte("data: {\"choices\":[{\"delta\":{\"content\":\"ok\"}}]}\n\n"))
+		w.Write([]byte("data: [DONE]\n\n"))
+	}))
+	defer srv.Close()
+
+	p := &OpenAIProvider{BaseURL: srv.URL, APIKey: "k", Model: "m", HTTP: srv.Client()}
+	final, err := p.Stream(context.Background(), "sys", []Message{{Role: RoleUser, Content: "hi"}}, nil, StreamOptions{}, nil)
+	if err != nil {
+		t.Fatalf("stream should succeed after retries: %v", err)
+	}
+	if final.Content != "ok" || calls != 3 {
+		t.Fatalf("content=%q calls=%d", final.Content, calls)
+	}
+}
+
+// TestNoRetryOn400：4xx 属不可重试错误，应立即返回。
+func TestNoRetryOn400(t *testing.T) {
+	calls := 0
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls++
+		w.WriteHeader(http.StatusBadRequest)
+		w.Write([]byte(`{"error":"bad request"}`))
+	}))
+	defer srv.Close()
+
+	p := &OpenAIProvider{BaseURL: srv.URL, APIKey: "k", Model: "m", HTTP: srv.Client()}
+	_, err := p.Stream(context.Background(), "sys", nil, nil, StreamOptions{}, nil)
+	if err == nil || !strings.Contains(err.Error(), "400") {
+		t.Fatalf("expected 400 error, got %v", err)
+	}
+	if calls != 1 {
+		t.Fatalf("4xx must not retry, calls=%d", calls)
+	}
+}
+
+// TestAnthropicCacheControlAndMaxTokens：system/tools 标记 cache_control，
+// max_tokens 来自 provider 配置。
+func TestAnthropicCacheControlAndMaxTokens(t *testing.T) {
+	var body map[string]any
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewDecoder(r.Body).Decode(&body)
+		w.Header().Set("Content-Type", "text/event-stream")
+		// 带一句正文：只发 message_stop 的空流现在会被当成失败（见 ErrEmptyCompletion），
+		// 而这条用例断言的是**请求体**形状。
+		w.Write([]byte("data: {\"type\":\"content_block_delta\",\"index\":0,\"delta\":{\"type\":\"text_delta\",\"text\":\"ok\"}}\n\n"))
+		w.Write([]byte("data: {\"type\":\"message_stop\"}\n\n"))
+	}))
+	defer srv.Close()
+
+	p := &AnthropicProvider{BaseURL: srv.URL, APIKey: "k", Model: "m", MaxTokens: 4096, HTTP: srv.Client()}
+	_, err := p.Stream(context.Background(), "system prompt",
+		[]Message{{Role: RoleUser, Content: "hi"}},
+		[]ToolDef{{Name: "read", Description: "d", Parameters: map[string]any{"type": "object"}}}, StreamOptions{}, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if body["max_tokens"].(float64) != 4096 {
+		t.Fatalf("max_tokens: %v", body["max_tokens"])
+	}
+	sys := body["system"].([]any)
+	s0 := sys[0].(map[string]any)
+	if s0["cache_control"].(map[string]any)["type"] != "ephemeral" {
+		t.Fatalf("system cache_control missing: %v", s0)
+	}
+	tools := body["tools"].([]any)
+	last := tools[len(tools)-1].(map[string]any)
+	if last["cache_control"].(map[string]any)["type"] != "ephemeral" {
+		t.Fatalf("tools cache_control missing: %v", last)
+	}
+}
+
+func TestParseRetryAfter(t *testing.T) {
+	if got := parseRetryAfter("3"); got != 3*time.Second {
+		t.Fatalf("got %v", got)
+	}
+	if got := parseRetryAfter(""); got != 0 {
+		t.Fatalf("got %v", got)
+	}
+	if got := parseRetryAfter("99999"); got != 0 { // 上限保护
+		t.Fatalf("got %v", got)
+	}
+}