feat(ai): session harness contract — incomplete, run_id, confirm bind

Add docs/16 for session state machine, incomplete/resume, queue semantics,
and agent-loop/harness/MCP gap review. Wire RunOutcome on max-rounds,
run_id through audit, TrimMessages/Redact hooks, ConfirmBindHash on write
confirm path; session file skeleton without HTTP yet. No push.

docs/15-AI-MCP-Agent.md

@@ -212,6 +212,20 @@ go build -o ../bin/opsxdb-mcp ./cmd/opsxdb-mcp
 MySQL 3306 demo + Redis 6379：`go run ./cmd/opsxdb-mcp-accept/` 全过。实践文档：`~/Downloads/AI-Agent接入GUI工具开发实践.md`。
 
 
+## 会话产品层（阶段 5，2026-10-15）
+
+工具面（MCP / Ask|Agent / SSE / P1 确认）已齐；下一层是 **会话生命周期**。权威契约见 **[16-AI-Session-Harness.md](./16-AI-Session-Harness.md)**：
+
+| 项 | 说明 |
+|----|------|
+| RunOutcome | `completed` / `incomplete` / `failed` / `cancelled` / `awaiting_confirm` |
+| 真·续跑 | 落盘 tool 轨迹 + 注入 prior；仅气泡回放不算 |
+| Harness | MaxRounds、trim、循环拒、检查点（桌面裁剪，无 Evolver） |
+| 排队 | 忙时单槽 pending（对齐 aiops 14） |
+| 安全 | 脱敏落盘；写确认绑定 connection+语句哈希 |
+
+本轮代码：outcome / run_id 审计 / TrimMessages / Redact / ConfirmBind；会话 HTTP 与排队 UI 后挂。
+
 ## 阶段 3–4 完成记录（2026-10-15）
 
 | 项 | 状态 |

docs/16-AI-Session-Harness.md

@@ -0,0 +1,255 @@
+# 16 - AI 会话产品层：状态机 · incomplete · Harness · 排队
+
+> 更新：2026-10-15 · 承接 `docs/15-AI-MCP-Agent.md`（工具面已齐）  
+> 借鉴：`/Users/logan/devops/aiops/docs/ai-agent` 的 03 / 05 / 06 / 13 / 14  
+> 产品锁：MCP-first；默认只读；写需 `mode=read_write` + 确认；桌面单用户；**不做** Evolver / 审批流 / 跨产品共用执行器
+
+本文定义 OpsxDB **会话产品层**契约：工具层（aitools / MCP）之上，侧栏 Agent 如何持久化、触顶、续跑、排队与安全确认。
+
+---
+
+## 1. 与 MCP / Agent 的关系
+
+```
+外部 Cursor/Claude ──stdio──▶ opsxdb-mcp ──┐
+                                           ▼
+侧栏 Ask（不调工具）                  aitools.Runtime（同源）
+侧栏 Agent ──▶ ai.runAgent / Harness ──▶ Runtime.Call
+                     │                      │
+                     ▼                      ▼
+              会话落盘（目标）          mcp-audit.jsonl（含 run_id）
+```
+
+| 层 | 职责 | 不负责 |
+|----|------|--------|
+| **aitools.Runtime** | 白名单、只读/可写闸、危险 SQL、确认令牌、审计 | 多轮对话、预算、续跑 |
+| **MCP stdio** | 把 Runtime 暴露给外部 Agent | 会话 UI、incomplete |
+| **ai.RunAgent** | ReAct 循环、步骤事件、SSE | 连接增删、公网后端 |
+| **会话产品层（本文）** | 落盘轨迹、RunOutcome、Harness 预算、排队、脱敏、确认绑定 | Evolver、审批流 |
+
+**真·续跑**（对齐 aiops 13）：必须同时满足 (1) 持久化含 `tool_calls` + `tool` observation（含 `tool_call_id`）；(2) 下次 Run 把历史注入引擎。仅恢复气泡不算 resume。
+
+---
+
+## 2. 会话状态机
+
+### 2.1 Session（会话）
+
+| 状态 | 含义 |
+|------|------|
+| `idle` | 有历史或空会话，无进行中的 run |
+| `running` | 正在执行一轮 Agent |
+| `awaiting_confirm` | 写操作待用户确认（可与 running 同条气泡内嵌） |
+| `incomplete` | 上一趟因预算/超时停住，轨迹可续 |
+| `completed` | 上一趟正常结束 |
+| `failed` | 不可恢复错误 |
+| `cancelled` | 用户停止 |
+
+桌面单用户：会话默认落 `~/.opsxdb/ai-sessions/<session_id>.json`（目标态；首版可仅内存 + 可选落盘）。
+
+### 2.2 RunOutcome（单次运行结束，必填）
+
+对齐 aiops `13-运行结果状态与可续跑契约.md`：
+
+| outcome | 含义 | UI |
+|---------|------|-----|
+| `completed` | 预算内给出最终答案 | 正常 |
+| `incomplete` | 未完成但有部分轨迹 | 「加预算继续」/「继续」 |
+| `failed` | 协议/鉴权/致命工具错 | 失败 |
+| `cancelled` | 用户 Abort | 已取消 |
+| `awaiting_confirm` | 安全闸等待确认 | 确认条 |
+
+### 2.3 IncompleteReason
+
+| reason | 触发 |
+|--------|------|
+| `max_steps` | 触达 MaxRounds / MaxIterations |
+| `context_budget` | trim 后无法安全继续 |
+| `tool_timeout` | 单工具超时 |
+| `llm_timeout` | LLM 超时 |
+| `safety_abort` | 安全中止（非等待确认） |
+
+**规则**：触顶步数 → `incomplete` + `max_steps`，**不得**伪装成 `completed`（现状 `agent.go` 触顶仍发 `done`，见 §6 审查）。
+
+### 2.4 对外字段（API / SSE 目标态）
+
+```text
+run_id                 # 一次 Run 全链路 ID（审计 / 步骤共用）
+session_id             # 可选；有落盘时必填
+run_outcome            # completed | incomplete | failed | cancelled | awaiting_confirm
+incomplete_reason?     # 见上表
+resumable              # 是否满足真·续跑条件
+steps_used / steps_budget
+partial_summary?       # 触顶时进展摘要
+```
+
+SSE：在现有 `step_start` / `step` / `delta` / `done` / `error` 上，目标增加显式 `incomplete` 事件（或 `done` 携带 `run_outcome=incomplete`）。首版代码已在 `AgentRes` / `AgentEvent` 挂上 outcome 字段。
+
+---
+
+## 3. Harness 预算与反馈
+
+参考 aiops 03；OpsxDB 裁剪为桌面 DB Agent：
+
+| 旋钮 | 默认 | 说明 |
+|------|------|------|
+| `MaxRounds` | 8（上限 16） | ReAct 轮数预算，非质量承诺 |
+| `MaxToolResultChars` | 12000 | 回灌 LLM 前截断 |
+| `MaxContextTokens` | 8000（目标） | trimMessages 预算 |
+| `ToolTimeout` | 与 db 层对齐（目标显式化） | 单工具 |
+| `SameCallLimit` | 3（目标） | 同工具同参循环拒 |
+| `Checkpoints` | 第 4 / 6 轮（目标） | 提示换思路 / 收尾 |
+
+### 3.1 trimMessages（目标行为）
+
+1. 永远保留 system  
+2. 从最近向前保留，约 `MaxContextTokens * 90%`  
+3. 至少保留最近 4 条  
+4. 裁剪后按 `tool_call_id` 整组保留 assistant.tool_calls + tool 消息  
+
+代码：`ai.TrimMessages` 钩子已落地；Agent 主循环接入时机在「多轮续跑 / 长会话」打开时启用（单次短问可跳过）。
+
+### 3.2 检查点与循环检测
+
+目标：同参 ≥3 次拒绝并写入 tool observation；第 4/6 轮向 msgs 注入短 system/user 提醒。不做 Evolver。
+
+---
+
+## 4. 会话落盘与真·续跑
+
+### 4.1 存储形状（目标）
+
+```json
+{
+  "id": "sess_…",
+  "title": "列出 demo 表",
+  "createdAt": "…",
+  "updatedAt": "…",
+  "status": "incomplete",
+  "messages": [
+    {"role":"system","content":"…"},
+    {"role":"user","content":"…"},
+    {"role":"assistant","content":"…","toolCalls":[{"id":"c1","function":{…}}]},
+    {"role":"tool","toolCallId":"c1","name":"opsxdb_list_tables","content":"…"}
+  ],
+  "runs": [{"runId":"run_…","outcome":"incomplete","reason":"max_steps","stepsUsed":8}]
+}
+```
+
+脱敏后落盘（见 §5）。密码、连接串明文、大结果全文不进会话文件。
+
+### 4.2 API 契约（目标，大 UI 可后做）
+
+| 方法 | 路径 | 说明 |
+|------|------|------|
+| GET | `/api/ai/sessions` | 列表 |
+| GET | `/api/ai/sessions/:id` | 含 messages |
+| POST | `/api/ai/agent` | 现有；增加可选 `sessionId` / `continue:true` / `prior` |
+| POST | `/api/ai/sessions/:id/continue` | 注入 prior + 可选提高 MaxRounds |
+
+骨架：`server/internal/ai/session.go`（文件存储接口 + 类型；HTTP 路由可后挂）。
+
+---
+
+## 5. 脱敏与确认绑定
+
+### 5.1 脱敏（落盘 / 审计 / 展示）
+
+对齐 aiops 05，桌面范围：
+
+| 类型 | 示例 |
+|------|------|
+| Bearer / API Key | `Bearer …`、`sk-…` |
+| 密码 | `password=`、`pwd=` |
+| 连接串 | `mysql://user:pass@host` |
+| PEM 私钥块 | `BEGIN … PRIVATE KEY` |
+
+`ai.RedactSensitive` 已提供；会话落盘与审计 summary 应过脱敏。
+
+### 5.2 确认绑定（写操作）
+
+写确认不得只靠 UI 气泡里的明文 SQL 再次提交而无绑定：
+
+```text
+confirm_bind = sha256(connection_id + "\n" + normalize(sql_or_command))
+```
+
+规则：
+
+1. `needConfirm` 响应带上 `confirmBind`（或服务端会话内缓存）  
+2. `/api/ai/agent/confirm` 校验：connection_id + 语句 与绑定一致，否则拒绝并要求重确认  
+3. 用户改了 SQL → 哈希变 → 必须重新确认（防 TOCTOU）
+
+首版：`aitools.ConfirmBindHash` + Runtime needConfirm 返回字段；confirm 端点校验（见代码）。
+
+---
+
+## 6. 忙时单槽排队
+
+对齐 aiops 14；桌面侧栏：
+
+| 动作 | 语义 |
+|------|------|
+| 排队 | running 时新输入进 `pending`（单槽；再提交覆盖） |
+| 自动续发 | 当前 run 结束（任一 outcome）后若有 pending → 自动 continue |
+| 立即 | 停止当前 → 立刻跑 pending |
+| 编辑 / 撤销 | 拉回输入框 / 丢弃 |
+
+与 incomplete 续跑可叠加：排队是同会话并发输入；incomplete 是上一趟未完成再继续。
+
+**首版**：文档 + 前端状态契约；服务端仍单请求单 run（Abort 用 ctx cancel）。不做跨会话队列、不做长 FIFO。
+
+---
+
+## 7. 细节审查：现状 → 缺口 → 建议
+
+### 7.1 Agent loop（`server/internal/ai/agent.go`）
+
+| | |
+|--|--|
+| **现状** | ReAct：`ChatWithTools` → `execOneTool` → 回灌；默认 MaxRounds=8（帽 16）；触顶文案后仍 `AgentEventDone`；无 session_id / prior；取消走 ctx → error 事件；NeedConfirm 回灌模型但不暂停 loop 为 `awaiting_confirm` outcome |
+| **缺口** | 无结构化 `run_outcome`；触顶伪装 completed；无真·续跑；无同参循环拒；无中段检查点；确认不打断为等待态 |
+| **建议** | 触顶 → `incomplete`/`max_steps` + `resumable`；`AgentRes` 带 outcome/run_id；可选 `Prior []Message`；NeedConfirm 可提前结束本 run 为 `awaiting_confirm`（UI 已有确认条，语义对齐即可） |
+
+### 7.2 Harness 架构
+
+| | |
+|--|--|
+| **现状** | 预算散落常量（`defaultAgentMaxRounds`、`maxToolResultFeedChars`）；无独立 Harness 配置对象；无 trim / 循环检测 / checkpoint；SSE 事件无 outcome |
+| **缺口** | 与 aiops 03 的「工程 Harness」不对齐；上下文窗口未 own；长会话必炸或静默丢配对 |
+| **建议** | `HarnessConfig` + `TrimMessages` + `DetectRepeatedToolCall`；Run 开头生成 `run_id` 写入 Runtime.Audit；检查点用轻量 user 消息注入，勿上 Evolver |
+
+### 7.3 MCP 同源 Runtime（`server/internal/aitools`）
+
+| | |
+|--|--|
+| **现状** | MCP 与 Agent 共用 `Runtime.Call`；Policy 白名单 + 只读默认；P1 `read_write` + needConfirm；审计 JSONL 有 source；无 run_id；确认端点不校验语句哈希绑定；Agent 剥离模型 confirm_danger |
+| **缺口** | 审计无法按一次 Agent run 聚合；确认存在「改 SQL 仍用旧气泡点执行」窗口；MCP 外部客户端若带 confirm_danger 可在 read_write 下写（设计允许外部 Agent 自带确认语义，但桌面应对齐绑定） |
+| **建议** | `AuditEntry.RunID`；`ConfirmBindHash`；confirm API 校验绑定；MCP 文档写明：外部客户端的 confirm 等同操作者授权，桌面 GUI 必须绑定哈希 |
+
+---
+
+## 8. 落地节奏（本轮 / 下一刀）
+
+| 优先级 | 项 | 本轮 |
+|--------|-----|------|
+| P0 | 文档契约（本文）+ 15 交叉引用 | ✅ |
+| P0 | RunOutcome 类型、触顶 incomplete、run_id 进审计 | ✅ 代码 |
+| P0 | TrimMessages / Redact / ConfirmBind 钩子 | ✅ 代码 |
+| P1 | 会话落盘 + continue API + UI 恢复 | 文档契约 + session 骨架 |
+| P1 | 单槽排队 UI | 文档；前端后做 |
+| P2 | 同参循环拒、检查点、Ask 真 token 流 | 后置 |
+
+---
+
+## 9. 验收（会话层）
+
+- [ ] `go test ./internal/ai/ ./internal/aitools/` 通过（含 outcome / trim / redact / bind）  
+- [ ] 触顶轮次：响应 `run_outcome=incomplete`，`incomplete_reason=max_steps`  
+- [ ] 审计行含 `run_id`（Agent 路径）  
+- [ ] confirm：改 SQL 后旧 bind 拒绝  
+- [ ] （P1）关窗再开：同 session 注入 prior 能继续调工具  
+
+---
+
+*文档版本: v1.0 | 2026-10-15 CST | 先契约与小改，大 UI 后挂*

server/internal/ai/agent.go

@@ -19,13 +19,17 @@ const (
 
 // AgentReq 侧栏 Agent 请求（与 Ask 的 /api/ai/chat 分路径）。
 type AgentReq struct {
-	Prompt       string `json:"prompt"`
-	ConnectionID string `json:"connectionId,omitempty"`
-	Driver       string `json:"driver,omitempty"`
-	Database     string `json:"database,omitempty"`
-	Model        string `json:"model,omitempty"`
-	MaxRounds    int    `json:"maxRounds,omitempty"`
-	Stream       bool   `json:"stream,omitempty"`
+	Prompt       string    `json:"prompt"`
+	ConnectionID string    `json:"connectionId,omitempty"`
+	Driver       string    `json:"driver,omitempty"`
+	Database     string    `json:"database,omitempty"`
+	Model        string    `json:"model,omitempty"`
+	MaxRounds    int       `json:"maxRounds,omitempty"`
+	Stream       bool      `json:"stream,omitempty"`
+	SessionID    string    `json:"sessionId,omitempty"`
+	Continue     bool      `json:"continue,omitempty"` // 真·续跑：注入 Prior
+	Prior        []Message `json:"prior,omitempty"`    // 含 tool 轨迹的历史（不含本轮 user）
+	RunID        string    `json:"runId,omitempty"`    // 可选；空则自动生成
 }
 
 // AgentStep 挂在回复下的步骤时间线一项。
@@ -38,16 +42,23 @@ type AgentStep struct {
 	Ms           int64             `json:"ms"`
 	NeedConfirm  bool              `json:"needConfirm,omitempty"`
 	ConfirmSQL   string            `json:"confirmSql,omitempty"`
+	ConfirmBind  string            `json:"confirmBind,omitempty"`
 	Pending      bool              `json:"pending,omitempty"`
 	Result       *AgentQueryResult `json:"result,omitempty"`
 }
 
-// AgentRes Agent 最终回答 + 步骤。
+// AgentRes Agent 最终回答 + 步骤 + 运行结果契约。
 type AgentRes struct {
-	Content string      `json:"content"`
-	SQL     string      `json:"sql,omitempty"`
-	Steps   []AgentStep `json:"steps"`
-	Rounds  int         `json:"rounds"`
+	Content          string      `json:"content"`
+	SQL              string      `json:"sql,omitempty"`
+	Steps            []AgentStep `json:"steps"`
+	Rounds           int         `json:"rounds"`
+	RunID            string      `json:"runId,omitempty"`
+	SessionID        string      `json:"sessionId,omitempty"`
+	Outcome          string      `json:"runOutcome,omitempty"`
+	IncompleteReason string      `json:"incompleteReason,omitempty"`
+	Resumable        bool        `json:"resumable,omitempty"`
+	StepsBudget      int         `json:"stepsBudget,omitempty"`
 }
 
 // RunAgent 多轮：LLM tool_call → aitools.Runtime.Call → 回灌。
@@ -69,39 +80,74 @@ func runAgent(ctx context.Context, llm ToolChat, rt *aitools.Runtime, req AgentR
 		return AgentRes{}, fmt.Errorf("tool chat LLM required")
 	}
 	prompt := strings.TrimSpace(req.Prompt)
-	if prompt == "" {
+	if prompt == "" && !(req.Continue && len(req.Prior) > 0) {
 		return AgentRes{}, fmt.Errorf("prompt required")
 	}
-	maxRounds := req.MaxRounds
-	if maxRounds <= 0 {
-		maxRounds = defaultAgentMaxRounds
-	}
-	if maxRounds > 16 {
-		maxRounds = 16
-	}
+	h := DefaultHarness()
+	maxRounds := h.ClampRounds(req.MaxRounds)
 
+	runID := strings.TrimSpace(req.RunID)
+	if runID == "" {
+		runID = NewRunID()
+	}
+	rt.RunID = runID
 	// Agent 路径禁止模型自行带确认令牌
 	rt.RejectCallerConfirm = true
 
 	tools := aitools.OpenAITools()
-	msgs := []Message{
-		{Role: "system", Content: buildAgentSystem(req)},
-		{Role: "user", Content: prompt},
+	var msgs []Message
+	if req.Continue && len(req.Prior) > 0 {
+		msgs = append([]Message{}, req.Prior...)
+		hasSystem := false
+		for _, m := range msgs {
+			if m.Role == "system" {
+				hasSystem = true
+				break
+			}
+		}
+		if !hasSystem {
+			msgs = append([]Message{{Role: "system", Content: buildAgentSystem(req)}}, msgs...)
+		}
+		if prompt != "" {
+			msgs = append(msgs, Message{Role: "user", Content: prompt})
+		}
+		msgs = TrimMessages(msgs, h)
+	} else {
+		msgs = []Message{
+			{Role: "system", Content: buildAgentSystem(req)},
+			{Role: "user", Content: prompt},
+		}
+	}
+
+	baseMeta := func(steps []AgentStep, rounds int) AgentRes {
+		return AgentRes{
+			Steps:       steps,
+			Rounds:      rounds,
+			RunID:       runID,
+			SessionID:   req.SessionID,
+			StepsBudget: maxRounds,
+		}
 	}
 
 	var steps []AgentStep
+	var callHistory []string
 	rounds := 0
 	for round := 0; round < maxRounds; round++ {
 		if err := ctx.Err(); err != nil {
-			res := AgentRes{Steps: steps, Rounds: rounds, Content: "已取消"}
-			emitEvent(emit, AgentEvent{Type: AgentEventError, Error: "cancelled", Steps: steps, Rounds: rounds})
+			res := baseMeta(steps, rounds)
+			res.Content = "已取消"
+			res.Outcome = OutcomeCancelled
+			res.Resumable = len(steps) > 0
+			emitEvent(emit, AgentEvent{Type: AgentEventError, Error: "cancelled", Steps: steps, Rounds: rounds, RunID: runID, Outcome: OutcomeCancelled})
 			return res, err
 		}
 		rounds = round + 1
 		turn, err := llm.ChatWithTools(ctx, req.Model, msgs, tools)
 		if err != nil {
-			emitEvent(emit, AgentEvent{Type: AgentEventError, Error: err.Error(), Steps: steps, Rounds: rounds})
-			return AgentRes{Steps: steps, Rounds: rounds}, err
+			res := baseMeta(steps, rounds)
+			res.Outcome = OutcomeFailed
+			emitEvent(emit, AgentEvent{Type: AgentEventError, Error: err.Error(), Steps: steps, Rounds: rounds, RunID: runID, Outcome: OutcomeFailed})
+			return res, err
 		}
 		if len(turn.ToolCalls) == 0 {
 			content := turn.Content
@@ -111,13 +157,14 @@ func runAgent(ctx context.Context, llm ToolChat, rt *aitools.Runtime, req AgentR
 			for _, part := range chunkDelta(content, 36) {
 				emitEvent(emit, AgentEvent{Type: AgentEventDelta, Delta: part})
 			}
-			res := AgentRes{
-				Content: content,
-				SQL:     extractSQL(content),
-				Steps:   steps,
-				Rounds:  rounds,
-			}
-			emitEvent(emit, AgentEvent{Type: AgentEventDone, Content: res.Content, SQL: res.SQL, Steps: steps, Rounds: rounds})
+			res := baseMeta(steps, rounds)
+			res.Content = content
+			res.SQL = extractSQL(content)
+			res.Outcome = OutcomeCompleted
+			emitEvent(emit, AgentEvent{
+				Type: AgentEventDone, Content: res.Content, SQL: res.SQL, Steps: steps, Rounds: rounds,
+				RunID: runID, Outcome: OutcomeCompleted, StepsBudget: maxRounds,
+			})
 			return res, nil
 		}
 
@@ -127,29 +174,76 @@ func runAgent(ctx context.Context, llm ToolChat, rt *aitools.Runtime, req AgentR
 			ToolCalls: turn.ToolCalls,
 		})
 
+		awaiting := false
 		for _, tc := range turn.ToolCalls {
 			name := strings.TrimSpace(tc.Function.Name)
-			pending := AgentStep{Tool: name, Args: truncateRunes(strings.TrimSpace(tc.Function.Arguments), maxStepArgsChars), Pending: true}
-			emitEvent(emit, AgentEvent{Type: AgentEventStepStart, Step: &pending})
+			argsRaw := strings.TrimSpace(tc.Function.Arguments)
+			fp := ToolCallFingerprint(name, argsRaw)
+			if DetectRepeatedToolCall(callHistory, fp, h.SameCallLimit) {
+				step := AgentStep{
+					Tool:    name,
+					Args:    truncateRunes(argsRaw, maxStepArgsChars),
+					OK:      false,
+					Summary: "repeated tool call rejected (same args)",
+					Error:   "repeated tool call rejected (same args)",
+				}
+				steps = append(steps, step)
+				cp := step
+				emitEvent(emit, AgentEvent{Type: AgentEventStep, Step: &cp, RunID: runID})
+				msgs = append(msgs, Message{
+					Role: "tool", ToolCallID: tc.ID, Name: name,
+					Content: step.Error,
+				})
+				callHistory = append(callHistory, fp)
+				continue
+			}
+			callHistory = append(callHistory, fp)
+
+			pending := AgentStep{Tool: name, Args: truncateRunes(argsRaw, maxStepArgsChars), Pending: true}
+			emitEvent(emit, AgentEvent{Type: AgentEventStepStart, Step: &pending, RunID: runID})
 
 			step, toolMsg := execOneTool(ctx, rt, tc)
 			steps = append(steps, step)
 			cp := step
-			ev := AgentEvent{Type: AgentEventStep, Step: &cp}
+			ev := AgentEvent{Type: AgentEventStep, Step: &cp, RunID: runID}
 			if step.Result != nil {
 				ev.Result = step.Result
 			}
 			emitEvent(emit, ev)
 			msgs = append(msgs, toolMsg)
+			if step.NeedConfirm {
+				awaiting = true
+			}
 		}
+		if awaiting {
+			content := "需要确认写操作后才能继续。"
+			res := baseMeta(steps, rounds)
+			res.Content = content
+			res.Outcome = OutcomeAwaitingConfirm
+			res.Resumable = true
+			emitEvent(emit, AgentEvent{
+				Type: AgentEventDone, Content: content, Steps: steps, Rounds: rounds,
+				RunID: runID, Outcome: OutcomeAwaitingConfirm, Resumable: true, StepsBudget: maxRounds,
+			})
+			return res, nil
+		}
+		msgs = TrimMessages(msgs, h)
 	}
 
-	content := fmt.Sprintf("已达最大工具轮次（%d）。请根据已执行步骤自行判断，或缩小问题后重试。", maxRounds)
+	content := fmt.Sprintf("已达最大工具轮次（%d）。可加预算继续，或缩小问题后重试。", maxRounds)
 	for _, part := range chunkDelta(content, 36) {
 		emitEvent(emit, AgentEvent{Type: AgentEventDelta, Delta: part})
 	}
-	res := AgentRes{Content: content, Steps: steps, Rounds: rounds}
-	emitEvent(emit, AgentEvent{Type: AgentEventDone, Content: content, Steps: steps, Rounds: rounds})
+	res := baseMeta(steps, rounds)
+	res.Content = content
+	res.Outcome = OutcomeIncomplete
+	res.IncompleteReason = ReasonMaxSteps
+	res.Resumable = true
+	emitEvent(emit, AgentEvent{
+		Type: AgentEventIncomplete, Content: content, Steps: steps, Rounds: rounds,
+		RunID: runID, Outcome: OutcomeIncomplete, IncompleteReason: ReasonMaxSteps,
+		Resumable: true, StepsBudget: maxRounds,
+	})
 	return res, nil
 }
 
@@ -228,6 +322,9 @@ func execOneTool(ctx context.Context, rt *aitools.Runtime, tc ToolCall) (AgentSt
 			} else if cmd, _ := m["command"].(string); cmd != "" {
 				step.ConfirmSQL = cmd
 			}
+			if bind, _ := m["confirmBind"].(string); bind != "" {
+				step.ConfirmBind = bind
+			}
 			msg, _ := m["message"].(string)
 			if msg == "" {
 				msg = "requires user confirmation (confirm_danger)"

server/internal/ai/agent_stream.go

@@ -12,7 +12,8 @@ const (
 	AgentEventStep      = "step"
 	AgentEventDelta     = "delta"
 	AgentEventDone      = "done"
-	AgentEventError     = "error"
+	AgentEventError      = "error"
+	AgentEventIncomplete = "incomplete"
 )
 
 // AgentQueryResult 可供「在结果区打开」的行列结果。
@@ -26,15 +27,20 @@ type AgentQueryResult struct {
 
 // AgentEvent 流式推送的一项。
 type AgentEvent struct {
-	Type    string            `json:"type"`
-	Step    *AgentStep        `json:"step,omitempty"`
-	Delta   string            `json:"delta,omitempty"`
-	Content string            `json:"content,omitempty"`
-	SQL     string            `json:"sql,omitempty"`
-	Steps   []AgentStep       `json:"steps,omitempty"`
-	Rounds  int               `json:"rounds,omitempty"`
-	Error   string            `json:"error,omitempty"`
-	Result  *AgentQueryResult `json:"result,omitempty"`
+	Type             string            `json:"type"`
+	Step             *AgentStep        `json:"step,omitempty"`
+	Delta            string            `json:"delta,omitempty"`
+	Content          string            `json:"content,omitempty"`
+	SQL              string            `json:"sql,omitempty"`
+	Steps            []AgentStep       `json:"steps,omitempty"`
+	Rounds           int               `json:"rounds,omitempty"`
+	Error            string            `json:"error,omitempty"`
+	Result           *AgentQueryResult `json:"result,omitempty"`
+	RunID            string            `json:"runId,omitempty"`
+	Outcome          string            `json:"runOutcome,omitempty"`
+	IncompleteReason string            `json:"incompleteReason,omitempty"`
+	Resumable        bool              `json:"resumable,omitempty"`
+	StepsBudget      int               `json:"stepsBudget,omitempty"`
 }
 
 // AgentEmitter 可选；nil 时仅返回最终 AgentRes（非流式 fallback）。

server/internal/ai/harness.go

@@ -0,0 +1,165 @@
+package ai
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"encoding/json"
+	"strings"
+)
+
+// HarnessConfig Agent 运行时预算（桌面裁剪版；不做 Evolver）。
+type HarnessConfig struct {
+	MaxRounds           int // 默认 8，硬帽 16
+	MaxToolResultChars  int // 回灌截断
+	MaxContextTokens    int // trim 预算（粗估字符/4）
+	SameCallLimit       int // 同工具同参上限；0=关闭
+	MinRetainMessages   int // trim 至少保留条数（不含 system）
+}
+
+// DefaultHarness 与历史 agent.go 常量对齐。
+func DefaultHarness() HarnessConfig {
+	return HarnessConfig{
+		MaxRounds:          defaultAgentMaxRounds,
+		MaxToolResultChars: maxToolResultFeedChars,
+		MaxContextTokens:   8000,
+		SameCallLimit:      3,
+		MinRetainMessages:  4,
+	}
+}
+
+func (h HarnessConfig) ClampRounds(n int) int {
+	if n <= 0 {
+		n = h.MaxRounds
+		if n <= 0 {
+			n = defaultAgentMaxRounds
+		}
+	}
+	if n > 16 {
+		n = 16
+	}
+	return n
+}
+
+// NewRunID 生成短 run_id，贯穿审计与步骤。
+func NewRunID() string {
+	var b [8]byte
+	_, _ = rand.Read(b[:])
+	return "run_" + hex.EncodeToString(b[:])
+}
+
+// EstimateTokens 粗估（字符/4）；含 tool_calls JSON。
+func EstimateTokens(msgs []Message) int {
+	n := 0
+	for _, m := range msgs {
+		n += len(m.Content)
+		if len(m.ToolCalls) > 0 {
+			if b, err := json.Marshal(m.ToolCalls); err == nil {
+				n += len(b)
+			}
+		}
+		n += len(m.Name) + len(m.ToolCallID) + 8
+	}
+	if n < 0 {
+		return 0
+	}
+	return n / 4
+}
+
+// TrimMessages 保留 system + 最近消息，控制在 MaxContextTokens*90%；
+// 并按 tool_call_id 整组保留 assistant.tool_calls 与对应 tool。
+func TrimMessages(msgs []Message, h HarnessConfig) []Message {
+	if len(msgs) == 0 {
+		return msgs
+	}
+	budget := h.MaxContextTokens
+	if budget <= 0 {
+		budget = 8000
+	}
+	limit := budget * 90 / 100
+	minRetain := h.MinRetainMessages
+	if minRetain <= 0 {
+		minRetain = 4
+	}
+
+	var system []Message
+	var rest []Message
+	for _, m := range msgs {
+		if m.Role == "system" {
+			system = append(system, m)
+		} else {
+			rest = append(rest, m)
+		}
+	}
+	if EstimateTokens(msgs) <= limit {
+		return ensureToolPairing(msgs)
+	}
+
+	// 从尾部向前取，直到接近预算
+	var kept []Message
+	for i := len(rest) - 1; i >= 0; i-- {
+		candidate := append([]Message{}, rest[i:]...)
+		candidate = append(append([]Message{}, system...), candidate...)
+		if EstimateTokens(candidate) > limit && len(kept) >= minRetain {
+			break
+		}
+		kept = rest[i:]
+		if len(kept) >= minRetain && EstimateTokens(candidate) <= limit {
+			break
+		}
+	}
+	if len(kept) == 0 && len(rest) > 0 {
+		// 至少留最后一条
+		kept = rest[len(rest)-1:]
+	}
+	out := append(append([]Message{}, system...), kept...)
+	return ensureToolPairing(out)
+}
+
+// ensureToolPairing 丢弃没有对应 assistant tool_calls 的孤立 tool，以及裁切后半截的 tool_calls。
+func ensureToolPairing(msgs []Message) []Message {
+	toolIDs := map[string]bool{}
+	for _, m := range msgs {
+		if m.Role == "assistant" {
+			for _, tc := range m.ToolCalls {
+				if tc.ID != "" {
+					toolIDs[tc.ID] = true
+				}
+			}
+		}
+	}
+	out := make([]Message, 0, len(msgs))
+	for _, m := range msgs {
+		if m.Role == "tool" {
+			if m.ToolCallID == "" || !toolIDs[m.ToolCallID] {
+				continue
+			}
+		}
+		if m.Role == "assistant" && len(m.ToolCalls) > 0 {
+			// 若后续缺少任一 tool 结果，整组 assistant 仍保留（由调用方补 incomplete）；
+			// 这里只滤 tool 侧孤儿。
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+// ToolCallFingerprint 同参循环检测键。
+func ToolCallFingerprint(name, argsJSON string) string {
+	name = strings.TrimSpace(name)
+	argsJSON = strings.Join(strings.Fields(argsJSON), " ")
+	return name + "\n" + argsJSON
+}
+
+// DetectRepeatedToolCall 若同一指纹已出现 limit 次（不含本次），返回 true。
+func DetectRepeatedToolCall(history []string, fp string, limit int) bool {
+	if limit <= 0 || fp == "" {
+		return false
+	}
+	n := 0
+	for _, h := range history {
+		if h == fp {
+			n++
+		}
+	}
+	return n >= limit
+}

server/internal/ai/harness_test.go

@@ -0,0 +1,155 @@
+package ai
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"opsxdb/server/internal/aitools"
+)
+
+func TestTrimMessagesKeepsSystemAndDropsOrphanTool(t *testing.T) {
+	h := DefaultHarness()
+	h.MaxContextTokens = 40
+	msgs := []Message{
+		{Role: "system", Content: "sys"},
+		{Role: "user", Content: strings.Repeat("u", 400)},
+		{Role: "assistant", Content: "a", ToolCalls: []ToolCall{{ID: "c1", Type: "function", Function: ToolCallFunc{Name: "opsxdb_list_tables", Arguments: "{}"}}}},
+		{Role: "tool", ToolCallID: "c1", Name: "opsxdb_list_tables", Content: `{"tables":["users"]}`},
+		{Role: "user", Content: "continue please"},
+	}
+	out := TrimMessages(msgs, h)
+	hasSys := false
+	for _, m := range out {
+		if m.Role == "system" {
+			hasSys = true
+		}
+	}
+	if !hasSys {
+		t.Fatal("system dropped")
+	}
+	orphan := TrimMessages([]Message{
+		{Role: "system", Content: "s"},
+		{Role: "tool", ToolCallID: "x", Content: "nope"},
+		{Role: "user", Content: "hi"},
+	}, h)
+	for _, m := range orphan {
+		if m.Role == "tool" {
+			t.Fatalf("expected orphan tool dropped: %+v", orphan)
+		}
+	}
+}
+
+func TestDetectRepeatedToolCall(t *testing.T) {
+	fp := ToolCallFingerprint("opsxdb_list_tables", `{"connection_id":"c1"}`)
+	hist := []string{fp, fp, fp}
+	if !DetectRepeatedToolCall(hist, fp, 3) {
+		t.Fatal("want reject at 3")
+	}
+	if DetectRepeatedToolCall(hist[:2], fp, 3) {
+		t.Fatal("should allow under limit")
+	}
+}
+
+func TestRedactSensitive(t *testing.T) {
+	out := RedactSensitive("Bearer tokensecret password=secret mysql://user:pass@host/db")
+	if strings.Contains(out, "tokensecret") || strings.Contains(out, "password=secret") || strings.Contains(out, ":pass@") {
+		t.Fatalf("not redacted: %s", out)
+	}
+	if !strings.Contains(RedactSensitive("sk-abcdefghijklmnop"), "[REDACTED]") {
+		t.Fatal("sk key")
+	}
+}
+
+func TestRunAgentMaxRoundsIncomplete(t *testing.T) {
+	rt := aitools.NewRuntime(aitools.Policy{
+		Mode:                 aitools.ModeReadOnly,
+		AllowedConnectionIDs: []string{"c1"},
+		DefaultLimit:         100,
+		MaxLimit:             500,
+	})
+	rt.LoadConns = func() ([]aitools.SavedConnection, error) {
+		return []aitools.SavedConnection{{ID: "c1", Name: "d", Driver: "mysql"}}, nil
+	}
+	sawRunID := false
+	rt.Audit = func(e aitools.AuditEntry) {
+		if e.RunID != "" {
+			sawRunID = true
+		}
+	}
+	var turns []ToolsTurn
+	for i := 0; i < 3; i++ {
+		turns = append(turns, ToolsTurn{ToolCalls: []ToolCall{{
+			ID: "c", Type: "function",
+			Function: ToolCallFunc{Name: aitools.ToolListConnections, Arguments: "{}"},
+		}}})
+	}
+	llm := &scriptedLLM{turns: turns}
+	res, err := RunAgent(context.Background(), llm, rt, AgentReq{Prompt: "loop", MaxRounds: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Outcome != OutcomeIncomplete {
+		t.Fatalf("outcome=%s want incomplete", res.Outcome)
+	}
+	if res.IncompleteReason != ReasonMaxSteps {
+		t.Fatalf("reason=%s", res.IncompleteReason)
+	}
+	if !res.Resumable || res.RunID == "" {
+		t.Fatalf("resumable/runId: %+v", res)
+	}
+	if !sawRunID {
+		t.Fatal("expected audit run_id")
+	}
+}
+
+func TestRunAgentCompletedHasOutcome(t *testing.T) {
+	rt := aitools.NewRuntime(aitools.Policy{Mode: aitools.ModeReadOnly, AllowedConnectionIDs: []string{"c1"}, DefaultLimit: 100, MaxLimit: 500})
+	rt.Audit = func(aitools.AuditEntry) {}
+	llm := &scriptedLLM{turns: []ToolsTurn{{Content: "hello"}}}
+	res, err := RunAgent(context.Background(), llm, rt, AgentReq{Prompt: "hi"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if res.Outcome != OutcomeCompleted {
+		t.Fatalf("outcome=%s", res.Outcome)
+	}
+	if res.RunID == "" || res.StepsBudget != defaultAgentMaxRounds {
+		t.Fatalf("%+v", res)
+	}
+}
+
+func TestRunAgentRepeatCallRejected(t *testing.T) {
+	rt := aitools.NewRuntime(aitools.Policy{Mode: aitools.ModeReadOnly, AllowedConnectionIDs: []string{"c1"}, DefaultLimit: 100, MaxLimit: 500})
+	rt.LoadConns = func() ([]aitools.SavedConnection, error) {
+		return []aitools.SavedConnection{{ID: "c1", Name: "d", Driver: "mysql"}}, nil
+	}
+	rt.Audit = func(aitools.AuditEntry) {}
+	args := `{"connection_id":"c1"}`
+	var turns []ToolsTurn
+	// 同一轮多个相同 tool_call，以及跨轮：SameCallLimit=3 指 history 已有 3 次则拒第 4 次
+	// 每轮一个相同调用，第 4 轮应被拒但仍继续到最终答案
+	for i := 0; i < 4; i++ {
+		turns = append(turns, ToolsTurn{ToolCalls: []ToolCall{{
+			ID: "c" + string(rune('1'+i)), Type: "function",
+			Function: ToolCallFunc{Name: aitools.ToolListConnections, Arguments: args},
+		}}})
+	}
+	turns = append(turns, ToolsTurn{Content: "done"})
+	llm := &scriptedLLM{turns: turns}
+	res, err := RunAgent(context.Background(), llm, rt, AgentReq{Prompt: "x", MaxRounds: 8})
+	if err != nil {
+		t.Fatal(err)
+	}
+	rejected := 0
+	for _, s := range res.Steps {
+		if strings.Contains(s.Summary, "repeated tool call") {
+			rejected++
+		}
+	}
+	if rejected < 1 {
+		t.Fatalf("expected at least one repeated rejection, steps=%+v", res.Steps)
+	}
+	_ = json.Marshal
+}

server/internal/ai/outcome.go

@@ -0,0 +1,19 @@
+package ai
+
+// RunOutcome 单次 Agent 运行结束状态（对齐 aiops docs/ai-agent/13）。
+const (
+	OutcomeCompleted       = "completed"
+	OutcomeIncomplete      = "incomplete"
+	OutcomeFailed          = "failed"
+	OutcomeCancelled       = "cancelled"
+	OutcomeAwaitingConfirm = "awaiting_confirm"
+)
+
+// IncompleteReason 仅当 Outcome==incomplete 时有意义。
+const (
+	ReasonMaxSteps       = "max_steps"
+	ReasonContextBudget  = "context_budget"
+	ReasonToolTimeout    = "tool_timeout"
+	ReasonLLMTimeout     = "llm_timeout"
+	ReasonSafetyAbort    = "safety_abort"
+)

server/internal/ai/redact.go

@@ -0,0 +1,42 @@
+package ai
+
+import (
+	"regexp"
+	"strings"
+)
+
+var (
+	reBearer   = regexp.MustCompile(`(?i)(Bearer\s+)[A-Za-z0-9._\-]+`)
+	reSkKey    = regexp.MustCompile(`(?i)\b(sk-[A-Za-z0-9]{8,})`)
+	rePassword = regexp.MustCompile(`(?i)(password|passwd|pwd)\s*[=:]\s*\S+`)
+	reDSN      = regexp.MustCompile(`(?i)\b((?:mysql|postgres|postgresql|mongodb|redis)://[^:\s/]+):([^@\s]+)@`)
+	rePEM      = regexp.MustCompile(`(?s)-----BEGIN [A-Z0-9 ]*PRIVATE KEY-----.*?-----END [A-Z0-9 ]*PRIVATE KEY-----`)
+)
+
+// RedactSensitive 脱敏：落盘 / 审计 summary / 展示共用。
+func RedactSensitive(s string) string {
+	if s == "" {
+		return s
+	}
+	s = rePEM.ReplaceAllString(s, "[REDACTED_PRIVATE_KEY]")
+	s = reDSN.ReplaceAllString(s, "${1}:[REDACTED]@")
+	s = reBearer.ReplaceAllString(s, "${1}[REDACTED]")
+	s = reSkKey.ReplaceAllString(s, "sk-[REDACTED]")
+	s = rePassword.ReplaceAllString(s, "${1}=[REDACTED]")
+	return s
+}
+
+// RedactMessages 对消息 Content 做脱敏（不改 tool_call 结构）。
+func RedactMessages(msgs []Message) []Message {
+	out := make([]Message, len(msgs))
+	copy(out, msgs)
+	for i := range out {
+		out[i].Content = RedactSensitive(out[i].Content)
+	}
+	return out
+}
+
+// NormalizeStatement 确认绑定前规范化 SQL/命令。
+func NormalizeStatement(s string) string {
+	return strings.Join(strings.Fields(strings.TrimSpace(s)), " ")
+}

server/internal/ai/session.go

@@ -0,0 +1,89 @@
+package ai
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"time"
+)
+
+// Session 桌面本地会话（目标态落盘；HTTP 路由可后挂）。
+type Session struct {
+	ID        string         `json:"id"`
+	Title     string         `json:"title,omitempty"`
+	Status    string         `json:"status"` // idle|running|incomplete|completed|failed|cancelled
+	CreatedAt string         `json:"createdAt"`
+	UpdatedAt string         `json:"updatedAt"`
+	Messages  []Message      `json:"messages"`
+	Runs      []SessionRun   `json:"runs,omitempty"`
+}
+
+// SessionRun 一次运行摘要。
+type SessionRun struct {
+	RunID             string `json:"runId"`
+	Outcome           string `json:"outcome"`
+	IncompleteReason  string `json:"incompleteReason,omitempty"`
+	Resumable         bool   `json:"resumable"`
+	StepsUsed         int    `json:"stepsUsed"`
+	StepsBudget       int    `json:"stepsBudget"`
+	FinishedAt        string `json:"finishedAt"`
+}
+
+func sessionsDir() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(home, ".opsxdb", "ai-sessions"), nil
+}
+
+// SessionPath 返回会话文件路径。
+func SessionPath(id string) (string, error) {
+	dir, err := sessionsDir()
+	if err != nil {
+		return "", err
+	}
+	return filepath.Join(dir, id+".json"), nil
+}
+
+// SaveSession 脱敏后写入磁盘。
+func SaveSession(s Session) error {
+	if s.ID == "" {
+		return fmt.Errorf("session id required")
+	}
+	s.Messages = RedactMessages(s.Messages)
+	s.UpdatedAt = time.Now().Format(time.RFC3339)
+	if s.CreatedAt == "" {
+		s.CreatedAt = s.UpdatedAt
+	}
+	path, err := SessionPath(s.ID)
+	if err != nil {
+		return err
+	}
+	if err := os.MkdirAll(filepath.Dir(path), 0700); err != nil {
+		return err
+	}
+	b, err := json.MarshalIndent(s, "", "  ")
+	if err != nil {
+		return err
+	}
+	return os.WriteFile(path, b, 0600)
+}
+
+// LoadSession 读会话；不存在返回 os.ErrNotExist。
+func LoadSession(id string) (Session, error) {
+	path, err := SessionPath(id)
+	if err != nil {
+		return Session{}, err
+	}
+	b, err := os.ReadFile(path)
+	if err != nil {
+		return Session{}, err
+	}
+	var s Session
+	if err := json.Unmarshal(b, &s); err != nil {
+		return Session{}, err
+	}
+	return s, nil
+}

server/internal/aitools/audit.go

@@ -15,6 +15,7 @@ type AuditEntry struct {
 	ConnectionID string `json:"connection_id,omitempty"`
 	Mode         string `json:"mode"`
 	Source       string `json:"source,omitempty"` // mcp | agent
+	RunID        string `json:"run_id,omitempty"`
 	OK           bool   `json:"ok"`
 	Summary      string `json:"summary,omitempty"`
 	Error        string `json:"error,omitempty"`

server/internal/aitools/confirm_bind.go

@@ -0,0 +1,29 @@
+package aitools
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"fmt"
+	"strings"
+)
+
+// ConfirmBindHash 绑定 connection + 规范化语句；改语句则哈希变，须重确认。
+func ConfirmBindHash(connectionID, statement string) string {
+	connectionID = strings.TrimSpace(connectionID)
+	statement = strings.Join(strings.Fields(strings.TrimSpace(statement)), " ")
+	sum := sha256.Sum256([]byte(connectionID + "\n" + statement))
+	return hex.EncodeToString(sum[:])
+}
+
+// VerifyConfirmBind 校验客户端带回的 bind；空 bind 视为不合法（桌面路径必须带）。
+func VerifyConfirmBind(connectionID, statement, bind string) error {
+	bind = strings.TrimSpace(bind)
+	if bind == "" {
+		return fmt.Errorf("confirmBind required")
+	}
+	want := ConfirmBindHash(connectionID, statement)
+	if !strings.EqualFold(bind, want) {
+		return fmt.Errorf("confirmBind mismatch: statement or connection changed; re-confirm required")
+	}
+	return nil
+}

server/internal/aitools/confirm_bind_test.go

@@ -0,0 +1,57 @@
+package aitools
+
+import "testing"
+
+func TestConfirmBindHashStableAndSensitive(t *testing.T) {
+	a := ConfirmBindHash("c1", "INSERT INTO t VALUES (1)")
+	b := ConfirmBindHash("c1", "  INSERT  INTO   t VALUES (1) ")
+	if a != b {
+		t.Fatalf("normalize mismatch %s vs %s", a, b)
+	}
+	c := ConfirmBindHash("c1", "INSERT INTO t VALUES (2)")
+	if a == c {
+		t.Fatal("different sql should differ")
+	}
+	if err := VerifyConfirmBind("c1", "INSERT INTO t VALUES (1)", a); err != nil {
+		t.Fatal(err)
+	}
+	if err := VerifyConfirmBind("c1", "INSERT INTO t VALUES (2)", a); err == nil {
+		t.Fatal("want mismatch")
+	}
+	if err := VerifyConfirmBind("c1", "INSERT INTO t VALUES (1)", ""); err == nil {
+		t.Fatal("want empty bind error")
+	}
+}
+
+func TestNeedConfirmIncludesBind(t *testing.T) {
+	rt := NewRuntime(Policy{
+		Mode:                 ModeReadWrite,
+		AllowedConnectionIDs: []string{"c1"},
+		DefaultLimit:         100,
+		MaxLimit:             500,
+	})
+	rt.RejectCallerConfirm = true
+	rt.FindConn = func(id string) (*SavedConnection, error) {
+		return &SavedConnection{ID: id, Driver: "mysql", Host: "127.0.0.1"}, nil
+	}
+	rt.Audit = func(AuditEntry) {}
+	out, err := rt.Call(nil, ToolExecuteQuery, map[string]any{
+		"connection_id": "c1",
+		"sql":           "DELETE FROM users WHERE id=1",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, ok := out.(map[string]any)
+	if !ok || m["needConfirm"] != true {
+		t.Fatalf("%v", out)
+	}
+	bind, _ := m["confirmBind"].(string)
+	if bind == "" {
+		t.Fatal("missing confirmBind")
+	}
+	want := ConfirmBindHash("c1", "DELETE FROM users WHERE id=1")
+	if bind != want {
+		t.Fatalf("bind %s want %s", bind, want)
+	}
+}

server/internal/aitools/runtime.go

@@ -53,6 +53,8 @@ type Runtime struct {
 	AllowConfirmedWrite bool
 	// Source 写入审计：mcp | agent。
 	Source string
+	// RunID 写入审计，串联一次 Agent run。
+	RunID string
 }
 
 func NewRuntime(policy Policy) *Runtime {
@@ -77,6 +79,9 @@ func (r *Runtime) audit(e AuditEntry) {
 	if e.Source == "" {
 		e.Source = r.Source
 	}
+	if e.RunID == "" && r.RunID != "" {
+		e.RunID = r.RunID
+	}
 	if r.Audit != nil {
 		r.Audit(e)
 		return
@@ -399,11 +404,13 @@ func (r *Runtime) executeQuery(ctx context.Context, args map[string]any) (any, e
 		// P1：可写已开启，但无确认令牌则返回 needConfirm（永不静默写）。
 		if !confirmed {
 			r.audit(AuditEntry{Tool: ToolExecuteQuery, ConnectionID: id, OK: false, Error: "needConfirm", Summary: truncate(sqlText, 80)})
+			bind := ConfirmBindHash(id, sqlText)
 			return map[string]any{
-				"needConfirm": true,
-				"sql":         sqlText,
-				"message":     "write requires explicit user confirmation",
-				"reason":      readReason,
+				"needConfirm":  true,
+				"sql":          sqlText,
+				"confirmBind":  bind,
+				"message":      "write requires explicit user confirmation",
+				"reason":       readReason,
 			}, nil
 		}
 	}
@@ -471,11 +478,13 @@ func (r *Runtime) executeRedis(ctx context.Context, args map[string]any) (any, e
 		}
 		if !confirmed {
 			r.audit(AuditEntry{Tool: ToolExecuteRedisCommand, ConnectionID: id, OK: false, Error: "needConfirm", Summary: truncate(command, 80)})
+			bind := ConfirmBindHash(id, command)
 			return map[string]any{
-				"needConfirm": true,
-				"command":     command,
-				"message":     "redis write requires explicit user confirmation",
-				"reason":      rerr.Error(),
+				"needConfirm":  true,
+				"command":      command,
+				"confirmBind":  bind,
+				"message":      "redis write requires explicit user confirmation",
+				"reason":       rerr.Error(),
 			}, nil
 		}
 	}

server/internal/server/router.go

@@ -532,6 +532,8 @@ func NewRouter(token string) *gin.Engine {
 				Command      string `json:"command"`
 				Database     string `json:"database"`
 				Kind         string `json:"kind"` // query | redis
+				ConfirmBind  string `json:"confirmBind"`
+				RunID        string `json:"runId"`
 			}
 			if err := c.ShouldBindJSON(&req); err != nil {
 				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
@@ -546,7 +548,16 @@ func NewRouter(token string) *gin.Engine {
 				c.JSON(http.StatusForbidden, gin.H{"error": "writes disabled: set mode=read_write in ~/.opsxdb/mcp.json"})
 				return
 			}
+			stmt := strings.TrimSpace(req.SQL)
+			if stmt == "" {
+				stmt = strings.TrimSpace(req.Command)
+			}
+			if err := aitools.VerifyConfirmBind(req.ConnectionID, stmt, req.ConfirmBind); err != nil {
+				c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+				return
+			}
 			rt.Source = "agent"
+			rt.RunID = strings.TrimSpace(req.RunID)
 			rt.RejectCallerConfirm = false
 			rt.AllowConfirmedWrite = true
 			kind := strings.ToLower(strings.TrimSpace(req.Kind))

src/AiPanel.tsx

@@ -43,9 +43,17 @@ function splitContent(content: string): ContentSeg[] {
 /** R2-F7:等宽代码块 + 顶部「复制」「插入编辑器」小按钮 */
 function CodeBlock({ code }: { code: string }) {
   const notify = useWorkspaceStore((s) => s.notify);
+  const notifyError = useWorkspaceStore((s) => s.notifyError);
+  const [copied, setCopied] = useState(false);
+  const copiedTimer = useRef<ReturnType<typeof setTimeout> | null>(null);
+  useEffect(() => () => { if (copiedTimer.current) clearTimeout(copiedTimer.current); }, []);
   const copy = async () => {
-    await navigator.clipboard.writeText(code);
-    notify({ kind: "success", title: "已复制", message: "SQL 已复制", mode: "toast", durationMs: 1200 });
+    try {
+      await navigator.clipboard.writeText(code);
+      setCopied(true);
+      if (copiedTimer.current) clearTimeout(copiedTimer.current);
+      copiedTimer.current = setTimeout(() => setCopied(false), 1200);
+    } catch (error) { notifyError(error, "toast"); }
   };
   const insertToEditor = () => {
     const st = useWorkspaceStore.getState();
@@ -61,7 +69,7 @@ function CodeBlock({ code }: { code: string }) {
         style={{ borderColor: "var(--ops-border-soft)" }}
       >
         <button type="button" className="text-[10px] hover:underline" style={{ color: "var(--ops-accent)" }} onClick={() => void copy()}>
-          复制
+          {copied ? "已复制 ✓" : "复制"}
         </button>
         <button type="button" className="text-[10px] hover:underline" style={{ color: "var(--ops-accent)" }} onClick={insertToEditor}>
           插入编辑器
@@ -161,12 +169,16 @@ function DangerConfirmBar({
   connectionId,
   database,
   kind,
+  confirmBind,
+  runId,
   onDone,
 }: {
   sql: string;
   connectionId?: string;
   database?: string;
   kind: "query" | "redis";
+  confirmBind?: string;
+  runId?: string;
   onDone: (ok: boolean, msg?: string) => void;
 }) {
   const [busy, setBusy] = useState(false);
@@ -180,6 +192,8 @@ function DangerConfirmBar({
         command: kind === "redis" ? sql : undefined,
         database,
         kind,
+        confirmBind,
+        runId,
       });
       onDone(true);
     } catch (e) {
@@ -528,6 +542,7 @@ export function AiPanel() {
               content: final.content,
               steps: final.steps || [],
               streaming: false,
+              runId: final.runId,
             });
             if (final.sql && tab) updateTabSqlIfUnchanged(tab.id, sql, tab.connectionId, tab.database, final.sql);
           },
@@ -674,6 +689,8 @@ export function AiPanel() {
                 connectionId={conn?.id}
                 database={database || undefined}
                 kind={pendingConfirm.step.tool?.includes("redis") ? "redis" : "query"}
+                confirmBind={pendingConfirm.step.confirmBind}
+                runId={m.runId}
                 onDone={(ok, msg) => {
                   if (ok) {
                     patchAiMessage(m.id, (cur) => ({

src/api.ts

@@ -540,6 +540,12 @@ export type AiAgentRes = {
   sql?: string;
   steps: AiAgentStep[];
   rounds: number;
+  runId?: string;
+  sessionId?: string;
+  runOutcome?: string;
+  incompleteReason?: string;
+  resumable?: boolean;
+  stepsBudget?: number;
 };
 
 export async function callAiAgent(payload: {
@@ -564,6 +570,7 @@ export type AiAgentStepExt = AiAgentStep & {
   pending?: boolean;
   needConfirm?: boolean;
   confirmSql?: string;
+  confirmBind?: string;
   result?: AiAgentQueryResult;
 };
 
@@ -691,6 +698,8 @@ export async function confirmAiAgentWrite(payload: {
   command?: string;
   database?: string;
   kind?: "query" | "redis";
+  confirmBind?: string;
+  runId?: string;
 }) {
   return post<Record<string, unknown>>("/api/ai/agent/confirm", payload);
 }

src/store.ts

@@ -207,6 +207,7 @@ export type AiAgentStepView = {
   pending?: boolean;
   needConfirm?: boolean;
   confirmSql?: string;
+  confirmBind?: string;
   result?: AiAgentQueryResultView;
 };
 
@@ -215,6 +216,7 @@ export type AiMessage = {
   role: "user" | "assistant" | "system";
   content: string;
   at: string;
+  runId?: string;
   /** Agent 工具步骤时间线（仅 assistant） */
   steps?: AiAgentStepView[];
   streaming?: boolean;
@@ -296,6 +298,8 @@ type State = {
     detail?: string;
     tips?: string[];
     mode?: "modal" | "toast";
+    /** 成功操作默认静默；长任务完成等需要用户注意时显式设为 visible。 */
+    feedback?: "visible";
     durationMs?: number;
     raw?: unknown;
   }) => void;
@@ -534,6 +538,7 @@ export const useWorkspaceStore = create<State>((set, get) => ({
     const parsed = input.raw !== undefined ? formatDbError(input.raw) : null;
     const kind = input.kind ?? parsed?.kind ?? "info";
     const mode = input.mode ?? (kind === "error" ? "modal" : "toast");
+    if (kind === "success" && mode === "toast" && input.feedback !== "visible") return;
     const payload: NotifyPayload = {
       id: notifyId(),
       kind,