fix(loop): 原生工具调用从 Response.ToolCalls 解码 + 逐次回填工具结果 — 修复工具消息无回复

- 根因: loop/agent.decodeToolCalls 只读 resp.Message.ToolCallName/resp.Content JSON信封, 从不看 api 原生解析出的 resp.ToolCalls(Anthropic tool_use / OpenAI tool_calls)
  → 模型返回 thinking+tool_use(无text) 被判定'无工具调用' → 空 Content 当最终 text → res.Text="" → no reply produced; 纯聊天消息不调工具直接出 text, 故表现为'部分消息无回复'
- agent/step.go: ToolCall 增 ID; ExtractToolCalls 导出并优先 resp.ToolCalls(保留原生id+归一化nil Args), 回落两旧约定
- loop/request.go: decodeToolCalls 委托 agent.ExtractToolCalls(流式/非流式共用)
- loop/tool_history.go(新): 原生批次回填 assistant tool_use + 带tool_use_id的tool_result; 旧JSON信封批次保持 RoleTool 纯文本(向后兼容)
- openai.go: tool_call parts 编成 tool_calls + tool_call_id
- 新增 native_toolcall_test + native_toolcall_http_test(httptest假Anthropic SSE: round1 tool_use(无text)→round2含tool_use_id/tool_result→res.Text非空)
- build/vet/test绿+交叉编译

internal/runtime/loop/request.go

@@ -6,7 +6,6 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/agent"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/capacity"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
-	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/tool"
 )
 
 // buildRequest assembles a model.Request from the agent's public surface,
@@ -164,26 +163,11 @@ func (l *Loop) capacityRecall(ctx context.Context, messages []model.Message) []c
 	return items
 }
 
-// decodeToolCalls decodes tool calls from a model.Response using the same two
-// conventions as the agent package: the normalized RoleTool placeholder, then a
-// JSON envelope of the form {"tool_calls":[...]} in Content. It reports ok=true
-// (with a non-empty slice) when a call is decoded.
+// decodeToolCalls decodes tool calls from a model.Response via the agent's shared
+// decoder, so the streaming and non-streaming paths agree on the provider-native
+// tool calls the adapters emit (Anthropic tool_use / OpenAI tool_calls) as well
+// as the legacy JSON-envelope convention. It reports ok=true (with a non-empty
+// slice) when a call is decoded.
 func decodeToolCalls(resp model.Response) ([]agent.ToolCall, bool) {
-	if resp.Message.Role == model.RoleAssistant && resp.Message.ToolCallName != "" {
-		return []agent.ToolCall{{Name: resp.Message.ToolCallName, Args: map[string]any{}}}, true
-	}
-	var payload struct {
-		ToolCalls []agent.ToolCall `json:"tool_calls"`
-	}
-	// Lenient decode repairs common model-emitted malformations (code fences,
-	// trailing commas) so a slightly broken tool-call envelope does not turn into
-	// a text answer; an unrecoverable payload still yields ok=false and the round
-	// proceeds as text.
-	if err := tool.LenientJSONUnmarshal([]byte(resp.Content), &payload); err != nil {
-		return nil, false
-	}
-	if len(payload.ToolCalls) == 0 {
-		return nil, false
-	}
-	return payload.ToolCalls, true
+	return agent.ExtractToolCalls(resp)
 }