fix: bound summarization max tokens for claude models

agentic 路径(单代理模式)缺少 toolPairReconciler 和 orphanToolPruner，
当 summarization 截断历史破坏配对后，序列化到 OpenAI API 触发 400
"insufficient tool messages following tool_calls message"。

新增 AgenticMessage 版本的 reconciler 和 pruner，与 classic 路径对齐。

Co-authored-by: temp <temp@tempdeMacBook-Pro.local>
Co-authored-by: Cursor <cursoragent@cursor.com>

.gitignore

@@ -1,30 +1,48 @@
 # Runtime data
 data/
+chat_uploads/
 *.db
 *.db-shm
 *.db-wal
+*.db-journal
 *.sqlite
 *.sqlite3
 
 # Local configuration and secrets
 config.yaml
 config.local.yaml
+config.yaml.backup
+*.backup
 .env
 .env.*
+!.env.example
 *.pem
 *.key
 *.crt
 
+# Upgrade / settings backups
+.upgrade-backup/
+
+# Eino runtime task board (under skills_dir)
+skills/.eino/
+
 # Build outputs
 cyberstrike-ai
+cyberstrike-ai.exe
+*.test
 bin/
 dist/
 build/
+.build/
+target/
+.gradle/
 coverage.out
 coverage.html
 
 # Logs and temporary files
 *.log
+*.bak
+*~
 tmp/
 temp/
 
@@ -38,7 +56,12 @@ __pycache__/
 # Go
 vendor/
 
-# macOS / editors
+# Node / browser extension
+node_modules/
+
+# macOS / Windows / editors
 .DS_Store
+Thumbs.db
+desktop.ini
 .idea/
 .vscode/

config.example.yaml

@@ -10,7 +10,7 @@
 # ============================================
 
 # 前端显示的版本号（可选，不填则显示默认版本）
-version: "v1.7.14"
+version: "v1.7.17"
 # 服务器配置
 server:
   host: 0.0.0.0 # 监听地址，0.0.0.0 表示监听所有网络接口
@@ -57,7 +57,7 @@ monitor:
 #   - DeepSeek: https://api.deepseek.com/v1
 #   - 其他兼容 OpenAI 协议的 API
 # 常用模型: gpt-4, gpt-3.5-turbo, deepseek-chat, claude-3-opus 等
-# provider: 可选值 openai_compatible(默认) | claude(自动桥接到 Anthropic Claude Messages API)
+# provider: 可选值 openai_compatible(默认) | claude(Eino 原生 Anthropic Messages API)
 ai:
   default_channel: qwen-max
   channels:
@@ -107,10 +107,10 @@ fofa:
   base_url: https://fofa.info/api/v1/search/all # 可选，留空则使用默认
   api_key: "" # FOFA API Key（可选，建议在系统设置中填写）
 zoomeye:
-  base_url: https://api.zoomeye.org/v2/search # 可选，留空则使用默认
+  base_url: https://api.zoomeye.ai/v2/search # 可选，留空则使用默认
   api_key: "" # ZoomEye API Key；也可通过 ZOOMEYE_API_KEY 环境变量配置
 quake:
-  base_url: https://quake.360.cn/api/v3/search/quake_service # 可选，留空则使用默认
+  base_url: https://quake.360.net/api/v3/search/quake_service # 可选，留空则使用默认
   api_key: "" # Quake API Token；也可通过 QUAKE_API_KEY 环境变量配置
 shodan:
   base_url: https://api.shodan.io # 可选，留空则使用默认
@@ -135,8 +135,12 @@ agent:
 #   approval → audit_agent_prompt
 #   review_edit → audit_agent_prompt_review_edit（可改参后放行）
 hitl:
-  # 全局默认审批方：human=人工审批，audit_agent=审计 Agent；未选会话时切换会写入本项，重启后仍生效
+  # 全局默认人机协同模式：off=关闭，approval=审批模式，review_edit=审查编辑；新建会话无独立配置时沿用
+  default_mode: off
+  # 全局默认审批方：human=人工审批，audit_agent=审计 Agent；新建会话无独立配置时沿用
   default_reviewer: human
+  # 全局默认审批等待时限（秒）：300=5分钟，0=不限时；新建会话无独立配置时沿用
+  default_timeout_seconds: 300
   # 审计 Agent 专用模型；字段留空则复用上方 openai 配置。建议 model 填小模型，用于降低审批成本。
   audit_model:
     provider: "" # openai / claude；留空跟随 openai.provider
@@ -146,7 +150,7 @@ hitl:
   # 已决策审计日志保留天数（与 MCP 监控一致；省略默认 90；0 表示不自动清理）
   retention_days: 90
   # 按你环境里的真实工具名增删（与侧栏一致、小写不敏感）；不需要全局免审批可改为 []
-  tool_whitelist: [read_file, list_dir, glob, grep, tool_search, upsert_project_fact, get_project_fact]
+  tool_whitelist: [read_file, ls, list_dir, glob, grep, tool_search, upsert_project_fact, get_project_fact, list_project_facts, search_project_facts, list_vulnerabilities, get_vulnerability, get_asset, query_assets, list_knowledge_risk_types, get_tool_execution, wait_tool_execution, batch_task_list, batch_task_get, manage_webshell_list, c2_event, c2_file]
   # audit_agent_prompt: |  # 审批模式；留空使用内置默认，可在「人机协同」页编辑
   # audit_agent_prompt_review_edit: |  # 审查编辑模式；留空使用内置默认
 
@@ -304,7 +308,7 @@ multi_agent:
     plan_execute_executed_steps_budget_ratio: 0.2 # plan_execute 中 executed_steps 预算比例
     plan_execute_max_step_result_runes: 4000 # plan_execute 每步结果最大字符数（超出截断）
     plan_execute_keep_last_steps: 8 # plan_execute 仅保留最近 N 步正文，早期步骤折叠为标题
-    checkpoint_dir: data/eino-checkpoints # P0：进程崩溃/OOM 后同会话自动 ADK Resume；正常结束会删 .ckpt；与「中断并继续」(last_react_*) 是两套机制
+    checkpoint_dir: "" # 聊天链路不再使用 ADK checkpoint；跨轮模型态统一走 conversations.last_react_*，便于排查 stale context
     model_retry_max_retries: 0 # Eino 原生 ChatModel retry；408/409/425/429/5xx/网络抖动/空流式输出会重试；0=默认 4（永久性 4xx 不重试）
     model_retry_max_backoff_sec: 0 # Eino 原生 ChatModel retry 单次退避上限秒数；0=默认 30
     model_failover_channels: [] # Eino 原生 ChatModel failover；填写 ai.channels ID，例如 [qwen-plus]；retry 耗尽后按顺序切换

docs/en-US/configuration.md

@@ -68,7 +68,7 @@ ai:
 | `ai.default_channel` | Default channel ID for new conversations and requests without an explicit channel. |
 | `ai.channels.<id>` | Channel config. IDs are normalized to lowercase letters, digits, and hyphens. |
 | `name` | Display name in the Web UI; falls back to the ID. |
-| `provider` | `openai_compatible` or `claude`. OpenAI-compatible channels map to runtime `openai`; Claude channels bridge to Anthropic Messages API. |
+| `provider` | `openai_compatible` or `claude`. OpenAI-compatible channels map to runtime `openai`; Claude channels use Eino's native Anthropic Messages API component. |
 | `base_url/api_key/model` | Required. Base URL usually includes a version path such as `/v1`. |
 | `max_total_tokens` | Shared context budget for compression, attack-chain generation, multi-agent summaries, and similar paths. |
 | `max_completion_tokens` | Per-response output cap; default is used when empty. |

docs/en-US/deployment.md

@@ -10,7 +10,7 @@ CyberStrikeAI can run as a local testing tool, an internal team service, or a pr
 - Python for some MCP servers and tool scripts.
 - SQLite files under `data/`; no external DB is required by default.
 - Actual security tools installed in PATH. YAML files under `tools/` only describe commands.
-- At least one `ai.channels` entry. Use `provider: openai_compatible` for OpenAI-compatible endpoints, or `provider: claude` for the Claude bridge.
+- At least one `ai.channels` entry. Use `provider: openai_compatible` for OpenAI-compatible endpoints, or `provider: claude` for Eino's native Claude component.
 
 Important persistent paths:
 

docs/en-US/hitl-best-practices.md

@@ -25,7 +25,7 @@ hitl:
     api_key: ""
     model: "" # set a small model here; blank reuses the default AI channel model
   retention_days: 90
-  tool_whitelist: [read_file, list_dir, glob, grep, tool_search]
+  tool_whitelist: [read_file, ls, glob, grep, tool_search, get_project_fact, list_project_facts, search_project_facts, list_vulnerabilities, get_vulnerability, get_asset, query_assets, list_knowledge_risk_types, get_tool_execution, wait_tool_execution, batch_task_list, batch_task_get, manage_webshell_list, c2_event, c2_file]
 ```
 
 `audit_model` supports partial configuration. Empty fields inherit from the resolved default AI channel, so the common setup is to fill only `model` and run approvals on a cheaper small model.
@@ -91,10 +91,16 @@ In review-edit mode, you may narrow paths, targets, or command arguments before
 Allowlisted tools skip approval, so keep the list stable and low-risk. Recommended examples:
 
 - `read_file`
-- `list_dir`
+- `ls`
 - `glob`
 - `grep`
 - `tool_search`
+- Project and vulnerability reads: `get_project_fact`, `list_project_facts`, `search_project_facts`, `list_vulnerabilities`, `get_vulnerability`
+- Asset and knowledge-metadata reads: `get_asset`, `query_assets`, `list_knowledge_risk_types`
+- Execution and task-state reads: `get_tool_execution`, `wait_tool_execution`, `batch_task_list`, `batch_task_get`
+- Local management-metadata reads: `manage_webshell_list`, `c2_event`, `c2_file`
+
+These built-in MCP reads remain constrained by RBAC and resource scope; the allowlist only bypasses HITL approval and does not grant additional access. `list_dir` is effective only when that is the actual tool name; the Eino filesystem directory-listing tool is named `ls`.
 
 Avoid globally allowlisting:
 
@@ -102,6 +108,8 @@ Avoid globally allowlisting:
 - File write/delete tools
 - C2 task tools
 - WebShell command execution tools
+- Nominally read-only tools that send requests to a target or external service, such as `webshell_file_read`, `webshell_file_list`, and `search_knowledge_base`
+- Multiplexed tools whose actions include both reads and writes, such as `c2_session`, `c2_listener`, `c2_profile`, and `c2_task_manage`
 
 ## Mode Selection
 

docs/zh-CN/configuration.md

@@ -65,7 +65,7 @@ ai:
 | `ai.default_channel` | 默认通道 ID。新对话、机器人、批量任务和未显式选择通道的请求使用它。 |
 | `ai.channels.<id>` | 通道配置。ID 会归一化为小写、数字和短横线，例如 `Qwen_Max` 会变成 `qwen-max`。 |
 | `name` | Web UI 展示名。留空时使用通道 ID。 |
-| `provider` | `openai_compatible` 或 `claude`。`openai_compatible` 会在运行时映射为 `openai`；`claude` 会桥接到 Anthropic Messages API。 |
+| `provider` | `openai_compatible` 或 `claude`。`openai_compatible` 会在运行时映射为 `openai`；`claude` 使用 Eino 原生 Anthropic Messages API。 |
 | `base_url/api_key/model` | 必填。Base URL 通常需要包含版本路径，如 OpenAI/兼容网关的 `/v1`。 |
 | `max_total_tokens` | 上下文压缩、攻击链构建、多代理摘要等共用的总预算。 |
 | `max_completion_tokens` | 单次模型输出上限；未填时使用默认值。 |

docs/zh-CN/deployment.md

@@ -10,7 +10,7 @@
 - Python：部分 MCP 服务或工具脚本需要 Python 运行环境。
 - SQLite：默认使用文件型数据库，无需单独服务。
 - 安全工具：`tools/` 中的 YAML 只是工具定义，实际命令如 `nmap`、`sqlmap`、`nuclei` 仍需安装到系统 PATH。
-- 模型服务：至少配置一个 `ai.channels` 通道；`provider: openai_compatible` 适用于 OpenAI 兼容 API，`provider: claude` 会走 Claude 桥接。
+- 模型服务：至少配置一个 `ai.channels` 通道；`provider: openai_compatible` 适用于 OpenAI 兼容 API，`provider: claude` 使用 Eino 原生 Claude 组件。
 
 建议目录：
 

docs/zh-CN/hitl-best-practices.md

@@ -25,7 +25,7 @@ hitl:
     api_key: ""
     model: "" # 可填小模型；留空复用默认 AI 通道的模型
   retention_days: 90
-  tool_whitelist: [read_file, list_dir, glob, grep, tool_search]
+  tool_whitelist: [read_file, ls, glob, grep, tool_search, get_project_fact, list_project_facts, search_project_facts, list_vulnerabilities, get_vulnerability, get_asset, query_assets, list_knowledge_risk_types, get_tool_execution, wait_tool_execution, batch_task_list, batch_task_get, manage_webshell_list, c2_event, c2_file]
 ```
 
 `audit_model` 的字段可以只填一部分。空字段会自动继承默认 AI 通道解析后的模型配置，因此常见做法是只填 `model`，让审计 Agent 使用更便宜的小模型。
@@ -91,10 +91,16 @@ hitl:
 白名单工具会跳过审批，因此要保守维护。推荐放：
 
 - `read_file`
-- `list_dir`
+- `ls`
 - `glob`
 - `grep`
 - `tool_search`
+- 项目与漏洞查询：`get_project_fact`、`list_project_facts`、`search_project_facts`、`list_vulnerabilities`、`get_vulnerability`
+- 资产与知识元数据查询：`get_asset`、`query_assets`、`list_knowledge_risk_types`
+- 执行与任务状态查询：`get_tool_execution`、`wait_tool_execution`、`batch_task_list`、`batch_task_get`
+- 本地管理元数据查询：`manage_webshell_list`、`c2_event`、`c2_file`
+
+上述内置 MCP 查询仍受 RBAC 和资源范围约束；白名单只跳过 HITL 审批，不扩大访问权限。`list_dir` 仅在实际工具名为该值时有效；Eino 文件系统的目录列表工具名是 `ls`。
 
 不建议直接全局白名单：
 
@@ -102,6 +108,8 @@ hitl:
 - 文件写入/删除工具
 - C2 任务工具
 - WebShell 命令执行工具
+- 会向目标或外部服务发起请求的“只读”工具，例如 `webshell_file_read`、`webshell_file_list` 和 `search_knowledge_base`
+- 同一工具名同时包含读写 action 的复合工具，例如 `c2_session`、`c2_listener`、`c2_profile`、`c2_task_manage`
 
 ## 模式选择
 

go.mod

@@ -14,6 +14,7 @@ require (
 	github.com/cloudwego/eino-ext/components/document/transformer/splitter/markdown v0.0.0-20260427010451-749e3706378b
 	github.com/cloudwego/eino-ext/components/document/transformer/splitter/recursive v0.0.0-20260427010451-749e3706378b
 	github.com/cloudwego/eino-ext/components/embedding/openai v0.0.0-20260427010451-749e3706378b
+	github.com/cloudwego/eino-ext/components/model/agenticclaude v0.1.3
 	github.com/cloudwego/eino-ext/components/model/agenticopenai v0.2.1
 	github.com/cloudwego/eino-ext/components/model/openai v0.1.13
 	github.com/creack/pty v1.1.24
@@ -24,7 +25,7 @@ require (
 	github.com/gorilla/websocket v1.5.3
 	github.com/larksuite/oapi-sdk-go/v3 v3.4.22
 	github.com/mattn/go-sqlite3 v1.14.18
-	github.com/modelcontextprotocol/go-sdk v1.2.0
+	github.com/modelcontextprotocol/go-sdk v1.3.1
 	github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.1
 	github.com/pkoukk/tiktoken-go v0.1.8
 	github.com/robfig/cron/v3 v3.0.1
@@ -46,11 +47,29 @@ require (
 )
 
 require (
+	cloud.google.com/go/auth v0.7.2 // indirect
+	cloud.google.com/go/auth/oauth2adapt v0.2.3 // indirect
+	cloud.google.com/go/compute/metadata v0.9.0 // indirect
 	github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 // indirect
 	github.com/Azure/azure-sdk-for-go/sdk/internal v1.10.0 // indirect
+	github.com/anthropics/anthropic-sdk-go v1.56.0 // indirect
+	github.com/aws/aws-sdk-go-v2 v1.30.3 // indirect
+	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3 // indirect
+	github.com/aws/aws-sdk-go-v2/config v1.27.27 // indirect
+	github.com/aws/aws-sdk-go-v2/credentials v1.17.27 // indirect
+	github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.16.11 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/configsources v1.3.15 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.6.15 // indirect
+	github.com/aws/aws-sdk-go-v2/internal/ini v1.8.0 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.11.3 // indirect
+	github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.11.17 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sso v1.22.4 // indirect
+	github.com/aws/aws-sdk-go-v2/service/ssooidc v1.26.4 // indirect
+	github.com/aws/aws-sdk-go-v2/service/sts v1.30.3 // indirect
+	github.com/aws/smithy-go v1.20.3 // indirect
 	github.com/bahlo/generic-list-go v0.2.0 // indirect
 	github.com/bmatcuk/doublestar/v4 v4.10.0 // indirect
-	github.com/buger/jsonparser v1.1.1 // indirect
+	github.com/buger/jsonparser v1.1.2 // indirect
 	github.com/bytedance/gopkg v0.1.3 // indirect
 	github.com/bytedance/sonic/loader v0.5.0 // indirect
 	github.com/cenkalti/backoff/v4 v4.3.0 // indirect
@@ -60,6 +79,7 @@ require (
 	github.com/dlclark/regexp2 v1.10.0 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/evanphx/json-patch v0.5.2 // indirect
+	github.com/felixge/httpsnoop v1.0.4 // indirect
 	github.com/gabriel-vasile/mimetype v1.4.2 // indirect
 	github.com/gin-contrib/sse v0.1.0 // indirect
 	github.com/go-logr/logr v1.4.3 // indirect
@@ -71,9 +91,14 @@ require (
 	github.com/go-viper/mapstructure/v2 v2.5.0 // indirect
 	github.com/goccy/go-json v0.10.2 // indirect
 	github.com/gogo/protobuf v1.3.2 // indirect
-	github.com/google/jsonschema-go v0.3.0 // indirect
+	github.com/golang/groupcache v0.0.0-20210331224755-41bb18bfe9da // indirect
+	github.com/golang/protobuf v1.5.4 // indirect
+	github.com/google/jsonschema-go v0.4.2 // indirect
+	github.com/google/s2a-go v0.1.7 // indirect
+	github.com/googleapis/enterprise-certificate-proxy v0.3.2 // indirect
 	github.com/goph/emperror v0.17.2 // indirect
 	github.com/grpc-ecosystem/grpc-gateway/v2 v2.25.1 // indirect
+	github.com/invopop/jsonschema v0.14.0 // indirect
 	github.com/jolestar/go-commons-pool/v2 v2.1.2 // indirect
 	github.com/json-iterator/go v1.1.12 // indirect
 	github.com/klauspost/cpuid/v2 v2.2.10 // indirect
@@ -86,10 +111,14 @@ require (
 	github.com/modern-go/reflect2 v1.0.2 // indirect
 	github.com/nikolalohinski/gonja v1.5.3 // indirect
 	github.com/openai/openai-go/v3 v3.35.0 // indirect
+	github.com/pb33f/ordered-map/v2 v2.3.1 // indirect
 	github.com/pelletier/go-toml/v2 v2.2.3 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
+	github.com/segmentio/asm v1.1.3 // indirect
+	github.com/segmentio/encoding v0.5.4 // indirect
 	github.com/sirupsen/logrus v1.9.3 // indirect
 	github.com/slongfield/pyfmt v0.0.0-20220222012616-ea85ff4c361f // indirect
+	github.com/standard-webhooks/standard-webhooks/libraries v0.0.1 // indirect
 	github.com/tetratelabs/wazero v1.11.0 // indirect
 	github.com/tidwall/gjson v1.18.0 // indirect
 	github.com/tidwall/match v1.1.1 // indirect
@@ -100,17 +129,22 @@ require (
 	github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect
 	github.com/yargevad/filepathx v1.0.0 // indirect
 	github.com/yosida95/uritemplate/v3 v3.0.2 // indirect
+	go.opencensus.io v0.24.0 // indirect
 	go.opentelemetry.io/auto/sdk v1.2.1 // indirect
+	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.49.0 // indirect
+	go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.49.0 // indirect
 	go.opentelemetry.io/otel/exporters/otlp/otlptrace v1.34.0 // indirect
 	go.opentelemetry.io/otel/metric v1.39.0 // indirect
 	go.opentelemetry.io/proto/otlp v1.5.0 // indirect
 	go.uber.org/multierr v1.11.0 // indirect
+	go.yaml.in/yaml/v4 v4.0.0-rc.2 // indirect
 	golang.org/x/arch v0.15.0 // indirect
 	golang.org/x/exp v0.0.0-20250305212735-054e65f0b394 // indirect
 	golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8 // indirect
 	golang.org/x/oauth2 v0.34.0 // indirect
 	golang.org/x/sync v0.20.0 // indirect
 	golang.org/x/sys v0.43.0 // indirect
+	google.golang.org/api v0.189.0 // indirect
 	google.golang.org/genproto/googleapis/api v0.0.0-20251202230838-ff82c1b0f217 // indirect
 	google.golang.org/genproto/googleapis/rpc v0.0.0-20251202230838-ff82c1b0f217 // indirect
 	google.golang.org/grpc v1.79.3 // indirect

go.sum

@@ -1,39 +1,82 @@
+cloud.google.com/go v0.26.0/go.mod h1:aQUYkXzVsufM+DwF1aE+0xfcU+56JwCaLick0ClmMTw=
+cloud.google.com/go/auth v0.7.2 h1:uiha352VrCDMXg+yoBtaD0tUF4Kv9vrtrWPYXwutnDE=
+cloud.google.com/go/auth v0.7.2/go.mod h1:VEc4p5NNxycWQTMQEDQF0bd6aTMb6VgYDXEwiJJQAbs=
+cloud.google.com/go/auth/oauth2adapt v0.2.3 h1:MlxF+Pd3OmSudg/b1yZ5lJwoXCEaeedAguodky1PcKI=
+cloud.google.com/go/auth/oauth2adapt v0.2.3/go.mod h1:tMQXOfZzFuNuUxOypHlQEXgdfX5cuhwU+ffUuXRJE8I=
 cloud.google.com/go/compute/metadata v0.3.0/go.mod h1:zFmK7XCadkQkj6TtorcaGlCW1hT1fIilQDwofLpJ20k=
+cloud.google.com/go/compute/metadata v0.9.0 h1:pDUj4QMoPejqq20dK0Pg2N4yG9zIkYGdBtwLoEkH9Zs=
+cloud.google.com/go/compute/metadata v0.9.0/go.mod h1:E0bWwX5wTnLPedCKqk3pJmVgCBSM6qQI1yTBdEb3C10=
 github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 h1:g0EZJwz7xkXQiZAI5xi9f3WWFYBlX1CPTrR+NDToRkQ=
 github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0/go.mod h1:XCW7KnZet0Opnr7HccfUw1PLc4CjHqpcaxW8DHklNkQ=
+github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.7.0 h1:tfLQ34V6F7tVSwoTf/4lH5sE0o6eCJuNDTmH09nDpbc=
+github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.7.0/go.mod h1:9kIvujWAA58nmPmWB1m23fyWic1kYZMxD9CxaWn4Qpg=
 github.com/Azure/azure-sdk-for-go/sdk/internal v1.10.0 h1:ywEEhmNahHBihViHepv3xPBn1663uRv2t2q/ESv9seY=
 github.com/Azure/azure-sdk-for-go/sdk/internal v1.10.0/go.mod h1:iZDifYGJTIgIIkYRNWPENUnqx6bJ2xnSDFI2tjwZNuY=
+github.com/AzureAD/microsoft-authentication-library-for-go v1.2.2 h1:XHOnouVk1mxXfQidrMEnLlPk9UMeRtyBTnEFtxkV0kU=
+github.com/AzureAD/microsoft-authentication-library-for-go v1.2.2/go.mod h1:wP83P5OoQ5p6ip3ScPr0BAq0BvuPAvacpEuSzyouqAI=
+github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU=
 github.com/Masterminds/semver/v3 v3.4.0 h1:Zog+i5UMtVoCU8oKka5P7i9q9HgrJeGzI9SA1Xbatp0=
 github.com/Masterminds/semver/v3 v3.4.0/go.mod h1:4V+yj/TJE1HU9XfppCwVMZq3I84lprf4nC11bSS5beM=
 github.com/airbrake/gobrake v3.6.1+incompatible/go.mod h1:wM4gu3Cn0W0K7GUuVWnlXZU11AGBXMILnrdOU8Kn00o=
+github.com/anthropics/anthropic-sdk-go v1.56.0 h1:idVU14wOZ06D0GBNEvuhn927xXmBVEquo0469iDwLsc=
+github.com/anthropics/anthropic-sdk-go v1.56.0/go.mod h1:3EfIfmFqxH6rbiLcIP4tPFyXL/IHakx2wDG4OU+TIEI=
+github.com/aws/aws-sdk-go-v2 v1.30.3 h1:jUeBtG0Ih+ZIFH0F4UkmL9w3cSpaMv9tYYDbzILP8dY=
+github.com/aws/aws-sdk-go-v2 v1.30.3/go.mod h1:nIQjQVp5sfpQcTc9mPSr1B0PaWK5ByX9MOoDadSN4lc=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3 h1:tW1/Rkad38LA15X4UQtjXZXNKsCgkshC3EbmcUmghTg=
+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3/go.mod h1:UbnqO+zjqk3uIt9yCACHJ9IVNhyhOCnYk8yA19SAWrM=
+github.com/aws/aws-sdk-go-v2/config v1.27.27 h1:HdqgGt1OAP0HkEDDShEl0oSYa9ZZBSOmKpdpsDMdO90=
+github.com/aws/aws-sdk-go-v2/config v1.27.27/go.mod h1:MVYamCg76dFNINkZFu4n4RjDixhVr51HLj4ErWzrVwg=
+github.com/aws/aws-sdk-go-v2/credentials v1.17.27 h1:2raNba6gr2IfA0eqqiP2XiQ0UVOpGPgDSi0I9iAP+UI=
+github.com/aws/aws-sdk-go-v2/credentials v1.17.27/go.mod h1:gniiwbGahQByxan6YjQUMcW4Aov6bLC3m+evgcoN4r4=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.16.11 h1:KreluoV8FZDEtI6Co2xuNk/UqI9iwMrOx/87PBNIKqw=
+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.16.11/go.mod h1:SeSUYBLsMYFoRvHE0Tjvn7kbxaUhl75CJi1sbfhMxkU=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.3.15 h1:SoNJ4RlFEQEbtDcCEt+QG56MY4fm4W8rYirAmq+/DdU=
+github.com/aws/aws-sdk-go-v2/internal/configsources v1.3.15/go.mod h1:U9ke74k1n2bf+RIgoX1SXFed1HLs51OgUSs+Ph0KJP8=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.6.15 h1:C6WHdGnTDIYETAm5iErQUiVNsclNx9qbJVPIt03B6bI=
+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.6.15/go.mod h1:ZQLZqhcu+JhSrA9/NXRm8SkDvsycE+JkV3WGY41e+IM=
+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.0 h1:hT8rVHwugYE2lEfdFE0QWVo81lF7jMrYJVDWI+f+VxU=
+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.0/go.mod h1:8tu/lYfQfFe6IGnaOdrpVgEL2IrrDOf6/m9RQum4NkY=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.11.3 h1:dT3MqvGhSoaIhRseqw2I0yH81l7wiR2vjs57O51EAm8=
+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.11.3/go.mod h1:GlAeCkHwugxdHaueRr4nhPuY+WW+gR8UjlcqzPr1SPI=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.11.17 h1:HGErhhrxZlQ044RiM+WdoZxp0p+EGM62y3L6pwA4olE=
+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.11.17/go.mod h1:RkZEx4l0EHYDJpWppMJ3nD9wZJAa8/0lq9aVC+r2UII=
+github.com/aws/aws-sdk-go-v2/service/sso v1.22.4 h1:BXx0ZIxvrJdSgSvKTZ+yRBeSqqgPM89VPlulEcl37tM=
+github.com/aws/aws-sdk-go-v2/service/sso v1.22.4/go.mod h1:ooyCOXjvJEsUw7x+ZDHeISPMhtwI3ZCB7ggFMcFfWLU=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.26.4 h1:yiwVzJW2ZxZTurVbYWA7QOrAaCYQR72t0wrSBfoesUE=
+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.26.4/go.mod h1:0oxfLkpz3rQ/CHlx5hB7H69YUpFiI1tql6Q6Ne+1bCw=
+github.com/aws/aws-sdk-go-v2/service/sts v1.30.3 h1:ZsDKRLXGWHk8WdtyYMoGNO7bTudrvuKpDKgMVRlepGE=
+github.com/aws/aws-sdk-go-v2/service/sts v1.30.3/go.mod h1:zwySh8fpFyXp9yOr/KVzxOl8SRqgf/IDw5aUt9UKFcQ=
+github.com/aws/smithy-go v1.20.3 h1:ryHwveWzPV5BIof6fyDvor6V3iUL7nTfiTKXHiW05nE=
+github.com/aws/smithy-go v1.20.3/go.mod h1:krry+ya/rV9RDcV/Q16kpu6ypI4K2czasz0NC3qS14E=
 github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
 github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
 github.com/bitly/go-simplejson v0.5.0/go.mod h1:cXHtHw4XUPsvGaxgjIAn8PhEWG9NfngEKAMDJEczWVA=
 github.com/bmatcuk/doublestar/v4 v4.10.0 h1:zU9WiOla1YA122oLM6i4EXvGW62DvKZVxIe6TYWexEs=
 github.com/bmatcuk/doublestar/v4 v4.10.0/go.mod h1:xBQ8jztBU6kakFMg+8WGxn0c6z1fTSPVIjEY1Wr7jzc=
 github.com/bmizerany/assert v0.0.0-20160611221934-b7ed37b82869/go.mod h1:Ekp36dRnpXw/yCqJaO+ZrUyxD+3VXMFFr56k5XYrpB4=
-github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=
-github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
+github.com/buger/jsonparser v1.1.2 h1:frqHqw7otoVbk5M8LlE/L7HTnIq2v9RX6EJ48i9AxJk=
+github.com/buger/jsonparser v1.1.2/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
 github.com/bugsnag/bugsnag-go v1.4.0/go.mod h1:2oa8nejYd4cQ/b0hMIopN0lCRxU0bueqREvZLWFrtK8=
 github.com/bugsnag/panicwrap v1.2.0/go.mod h1:D/8v3kj0zr8ZAKg1AQ6crr+5VwKN5eIywRkfhyM/+dE=
 github.com/bwmarrin/discordgo v0.29.0 h1:FmWeXFaKUwrcL3Cx65c20bTRW+vOb6k8AnaP+EgjDno=
 github.com/bwmarrin/discordgo v0.29.0/go.mod h1:NJZpH+1AfhIcyQsPeuBKsUtYrRnjkyu0kIVMCHkZtRY=
 github.com/bytedance/gopkg v0.1.3 h1:TPBSwH8RsouGCBcMBktLt1AymVo2TVsBVCY4b6TnZ/M=
 github.com/bytedance/gopkg v0.1.3/go.mod h1:576VvJ+eJgyCzdjS+c4+77QF3p7ubbtiKARP3TxducM=
-github.com/bytedance/mockey v1.3.0 h1:ONLRdvhqmCfr9rTasUB8ZKCfvbdD2tohOg4u+4Q/ed0=
-github.com/bytedance/mockey v1.3.0/go.mod h1:1BPHF9sol5R1ud/+0VEHGQq/+i2lN+GTsr3O2Q9IENY=
 github.com/bytedance/mockey v1.4.6 h1:pPkAFB6yiaaybvgp7DP1Rj4Ztiew3nsaMizoNkzsvNA=
+github.com/bytedance/mockey v1.4.6/go.mod h1:1BPHF9sol5R1ud/+0VEHGQq/+i2lN+GTsr3O2Q9IENY=
 github.com/bytedance/sonic v1.15.0 h1:/PXeWFaR5ElNcVE84U0dOHjiMHQOwNIx3K4ymzh/uSE=
 github.com/bytedance/sonic v1.15.0/go.mod h1:tFkWrPz0/CUCLEF4ri4UkHekCIcdnkqXw9VduqpJh0k=
 github.com/bytedance/sonic/loader v0.5.0 h1:gXH3KVnatgY7loH5/TkeVyXPfESoqSBSBEiDd5VjlgE=
 github.com/bytedance/sonic/loader v0.5.0/go.mod h1:AR4NYCk5DdzZizZ5djGqQ92eEhCCcdf5x77udYiSJRo=
 github.com/cenkalti/backoff/v4 v4.3.0 h1:MyRJ/UdXutAwSAT+s3wNd7MfTIcy71VQueUuFK343L8=
 github.com/cenkalti/backoff/v4 v4.3.0/go.mod h1:Y3VNntkOUPxTVeUxJ/G5vcM//AlwfmyYozVcomhLiZE=
+github.com/census-instrumentation/opencensus-proto v0.2.1/go.mod h1:f6KPmirojxKA12rnyqOA5BBL4O983OfeGPqjHWSTneU=
 github.com/certifi/gocertifi v0.0.0-20190105021004-abcd57078448/go.mod h1:GJKEexRPVJrBSOjoqN5VNOIKJ5Q3RViH6eu3puDRwx4=
 github.com/cespare/xxhash/v2 v2.1.2/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
 github.com/cespare/xxhash/v2 v2.2.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
 github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=
 github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
+github.com/client9/misspell v0.3.4/go.mod h1:qj6jICC3Q7zFZvVWo7KLAzC3yx5G7kyvSDkc90ppPyw=
 github.com/cloudwego/base64x v0.1.6 h1:t11wG9AECkCDk5fMSoxmufanudBtJ+/HemLstXDLI2M=
 github.com/cloudwego/base64x v0.1.6/go.mod h1:OFcloc187FXDaYHvrNIjxSe8ncn0OOM8gEHfghB2IPU=
 github.com/cloudwego/eino v0.9.14 h1:suNVibjtkPMiW7csFBdBqN3FRG0nRlCqoGKE7t0UDwY=
@@ -48,14 +91,17 @@ github.com/cloudwego/eino-ext/components/document/transformer/splitter/recursive
 github.com/cloudwego/eino-ext/components/document/transformer/splitter/recursive v0.0.0-20260427010451-749e3706378b/go.mod h1:9R0RQrQSpg1JaNnRtw7+RfRAAv0HgdE348YnrlZ6coo=
 github.com/cloudwego/eino-ext/components/embedding/openai v0.0.0-20260427010451-749e3706378b h1:pOqupZQyc46rw2Z0HeybtTmSMTwqfTrbRuGDuDsNf2A=
 github.com/cloudwego/eino-ext/components/embedding/openai v0.0.0-20260427010451-749e3706378b/go.mod h1:zyPrZT2bO6LyRJgVksQowR18jVgyLSvqK93hnO53/Lc=
+github.com/cloudwego/eino-ext/components/model/agenticclaude v0.1.3 h1:nfO9jUPhIKTHDPyL7rso6y7Mx8cAmrIgkeJC6KqTuXI=
+github.com/cloudwego/eino-ext/components/model/agenticclaude v0.1.3/go.mod h1:/dX1P7HNC7D3/nYLDoXBvdKAFnGS4tBbtJ8NrkaXuUk=
 github.com/cloudwego/eino-ext/components/model/agenticopenai v0.2.1 h1:bIMyUxEZdBEofWbXQCozfGxXCsbyNjpZcRMfW+Kt5u0=
 github.com/cloudwego/eino-ext/components/model/agenticopenai v0.2.1/go.mod h1:84jUcVAcmlGDRVX0ExABVJEfrIaplaAeRbjFYW9mzkI=
 github.com/cloudwego/eino-ext/components/model/openai v0.1.13 h1:5XHRTiTD5bt9KQrMHcfvuWNklEC3tpm3XHejdozt9vM=
 github.com/cloudwego/eino-ext/components/model/openai v0.1.13/go.mod h1:mgIoqYYOc0eECCqvLbEYpOJrQNTNxkwXzSJzFU+v5sQ=
-github.com/cloudwego/eino-ext/libs/acl/openai v0.1.17 h1:EeVcR1TslRA2IdNW1h/2LaGbPlffwGhQm99jM3zWZiI=
-github.com/cloudwego/eino-ext/libs/acl/openai v0.1.17/go.mod h1:Zkcx6DPTR2NfWmtSXbhItswGw6hqUezNPhNcke0pOG8=
 github.com/cloudwego/eino-ext/libs/acl/openai v0.1.18-0.20260527084435-846f52bd97c6 h1:ES/xufN5eqJ3h+9tw/tq6F8kkgnAxBAHVUB6nqKsIDU=
 github.com/cloudwego/eino-ext/libs/acl/openai v0.1.18-0.20260527084435-846f52bd97c6/go.mod h1:5Xj74dGrfHo1z7I07Fzp3SlTF7Bt4tss3A2FSt8SqQ4=
+github.com/cncf/udpa/go v0.0.0-20191209042840-269d4d468f6f/go.mod h1:M8M6+tZqaGXZJjfX53e64911xZQV5JYwmTeXPW+k8Sc=
+github.com/cncf/xds/go v0.0.0-20251210132809-ee656c7534f5 h1:6xNmx7iTtyBRev0+D/Tv1FZd4SCg8axKApyNyRsAt/w=
+github.com/cncf/xds/go v0.0.0-20251210132809-ee656c7534f5/go.mod h1:KdCmV+x/BuvyMxRnYBlmVaq4OLiKW6iRQfvC62cvdkI=
 github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/creack/pty v1.1.24 h1:bJrF4RRfyJnbTJqzRLHzcGaZK1NeM5kTC9jGgovnR1s=
 github.com/creack/pty v1.1.24/go.mod h1:08sCNb52WyoAwi2QDyzUCTgcvVFhUzewun7wtTfvcwE=
@@ -67,12 +113,25 @@ github.com/disintegration/imaging v1.6.2 h1:w1LecBlG2Lnp8B3jk5zSuNqd7b4DXhcjwek1
 github.com/disintegration/imaging v1.6.2/go.mod h1:44/5580QXChDfwIclfc/PCwrr44amcmDAg8hxG0Ewe4=
 github.com/dlclark/regexp2 v1.10.0 h1:+/GIL799phkJqYW+3YbOd8LCcbHzT0Pbo8zl70MHsq0=
 github.com/dlclark/regexp2 v1.10.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
+github.com/dnaeon/go-vcr v1.2.0 h1:zHCHvJYTMh1N7xnV7zf1m1GPBF9Ad0Jk/whtQ1663qI=
+github.com/dnaeon/go-vcr v1.2.0/go.mod h1:R4UdLID7HZT3taECzJs4YgbbH6PIGXB6W/sc5OLb6RQ=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/eino-contrib/jsonschema v1.0.3 h1:2Kfsm1xlMV0ssY2nuxshS4AwbLFuqmPmzIjLVJ1Fsp0=
 github.com/eino-contrib/jsonschema v1.0.3/go.mod h1:cpnX4SyKjWjGC7iN2EbhxaTdLqGjCi0e9DxpLYxddD4=
+github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
+github.com/envoyproxy/go-control-plane v0.9.1-0.20191026205805-5f8ba28d4473/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4=
+github.com/envoyproxy/go-control-plane v0.9.4/go.mod h1:6rpuAdCZL397s3pYoYcLgu1mIlRU8Am5FuJP05cCM98=
+github.com/envoyproxy/go-control-plane v0.14.0 h1:hbG2kr4RuFj222B6+7T83thSPqLjwBIfQawTkC++2HA=
+github.com/envoyproxy/go-control-plane/envoy v1.36.0 h1:yg/JjO5E7ubRyKX3m07GF3reDNEnfOboJ0QySbH736g=
+github.com/envoyproxy/go-control-plane/envoy v1.36.0/go.mod h1:ty89S1YCCVruQAm9OtKeEkQLTb+Lkz0k8v9W0Oxsv98=
+github.com/envoyproxy/protoc-gen-validate v0.1.0/go.mod h1:iSmxcyjqTsJpI2R4NaDN7+kN2VEUnK/pcBlmesArF7c=
+github.com/envoyproxy/protoc-gen-validate v1.3.0 h1:TvGH1wof4H33rezVKWSpqKz5NXWg5VPuZ0uONDT6eb4=
+github.com/envoyproxy/protoc-gen-validate v1.3.0/go.mod h1:HvYl7zwPa5mffgyeTUHA9zHIH36nmrm7oCbo4YKoSWA=
 github.com/evanphx/json-patch v0.5.2 h1:xVCHIVMUu1wtM/VkR9jVZ45N3FhZfYMMYGorLCR8P3k=
 github.com/evanphx/json-patch v0.5.2/go.mod h1:ZWS5hhDbVDyob71nXKNL0+PWn6ToqBHMikGIFbs31qQ=
+github.com/felixge/httpsnoop v1.0.4 h1:NFTV2Zj1bL4mc9sqWACXbQFVBBg2W3GPvqp8/ESS2Wg=
+github.com/felixge/httpsnoop v1.0.4/go.mod h1:m8KPJKqk1gH5J9DgRY2ASl2lWCfGKXixSwevea8zH2U=
 github.com/fortytw2/leaktest v1.3.0 h1:u8491cBMTQ8ft8aeV+adlcytMZylmA5nnwwkRZjI8vw=
 github.com/fortytw2/leaktest v1.3.0/go.mod h1:jDsjWgpAGjm2CA7WthBh/CdZYEPF31XHquHwclZch5g=
 github.com/fsnotify/fsnotify v1.4.7/go.mod h1:jwhsz4b93w/PPRr/qN1Yymfu8t87LnFCMoQvtojpjFo=
@@ -117,33 +176,49 @@ github.com/gogo/protobuf v1.3.2 h1:Ov1cvc58UF3b5XjBnZv7+opcTcQFZebYjWzi34vdm4Q=
 github.com/gogo/protobuf v1.3.2/go.mod h1:P1XiOD3dCwIKUDQYPy72D8LYyHL2YPYrpS2s69NZV8Q=
 github.com/golang-jwt/jwt/v5 v5.2.2 h1:Rl4B7itRWVtYIHFrSNd7vhTiz9UpLdi6gZhZ3wEeDy8=
 github.com/golang-jwt/jwt/v5 v5.2.2/go.mod h1:pqrtFR0X4osieyHYxtmOUWsAWrfe1Q5UVIyoH402zdk=
+github.com/golang/glog v0.0.0-20160126235308-23def4e6c14b/go.mod h1:SBH7ygxi8pfUlaOkMMuAQtPIUF8ecWP5IEl/CR7VP2Q=
+github.com/golang/groupcache v0.0.0-20200121045136-8c9f03a8e57e/go.mod h1:cIg4eruTrX1D+g88fzRXU5OdNfaM+9IcxsU14FzY7Hc=
+github.com/golang/groupcache v0.0.0-20210331224755-41bb18bfe9da h1:oI5xCqsCo564l8iNU+DwB5epxmsaqB+rhGL0m5jtYqE=
+github.com/golang/groupcache v0.0.0-20210331224755-41bb18bfe9da/go.mod h1:cIg4eruTrX1D+g88fzRXU5OdNfaM+9IcxsU14FzY7Hc=
+github.com/golang/mock v1.1.1/go.mod h1:oTYuIxOrZwtPieC+H1uAHpcLFnEyAGVDL/k47Jfbm0A=
 github.com/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
+github.com/golang/protobuf v1.3.2/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U=
 github.com/golang/protobuf v1.4.0-rc.1/go.mod h1:ceaxUfeHdC40wWswd/P6IGgMaK3YpKi5j83Wpe3EHw8=
 github.com/golang/protobuf v1.4.0-rc.1.0.20200221234624-67d41d38c208/go.mod h1:xKAWHe0F5eneWXFV3EuXVDTCmh+JuBKY0li0aMyXATA=
 github.com/golang/protobuf v1.4.0-rc.2/go.mod h1:LlEzMj4AhA7rCAGe4KMBDvJI+AwstrUpVNzEA03Pprs=
 github.com/golang/protobuf v1.4.0-rc.4.0.20200313231945-b860323f09d0/go.mod h1:WU3c8KckQ9AFe+yFwt9sWVRKCVIyN9cPHBJSNnbL67w=
 github.com/golang/protobuf v1.4.0/go.mod h1:jodUvKwWbYaEsadDk5Fwe5c77LiNKVO9IDvqG2KuDX0=
+github.com/golang/protobuf v1.4.1/go.mod h1:U8fpvMrcmy5pZrNK1lt4xCsGvpyWQ/VVv6QDs8UjoX8=
 github.com/golang/protobuf v1.4.2/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI=
+github.com/golang/protobuf v1.4.3/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI=
 github.com/golang/protobuf v1.5.0/go.mod h1:FsONVRAS9T7sI+LIUmWTfcYkHO4aIWwzhcaSAoJOfIk=
 github.com/golang/protobuf v1.5.2/go.mod h1:XVQd3VNwM+JqD3oG2Ue2ip4fOMUkwXdXDdiuN0vRsmY=
 github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=
 github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6rSs7xps=
+github.com/google/go-cmp v0.2.0/go.mod h1:oXzfMopK8JAjlY9xF4vHSVASa0yLyX7SntLO5aqRK0M=
 github.com/google/go-cmp v0.3.0/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU=
 github.com/google/go-cmp v0.3.1/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU=
 github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
+github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
+github.com/google/go-cmp v0.5.3/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
 github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
-github.com/google/jsonschema-go v0.3.0 h1:6AH2TxVNtk3IlvkkhjrtbUc4S8AvO0Xii0DxIygDg+Q=
-github.com/google/jsonschema-go v0.3.0/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=
+github.com/google/jsonschema-go v0.4.2 h1:tmrUohrwoLZZS/P3x7ex0WAVknEkBZM46iALbcqoRA8=
+github.com/google/jsonschema-go v0.4.2/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=
 github.com/google/pprof v0.0.0-20260402051712-545e8a4df936 h1:EwtI+Al+DeppwYX2oXJCETMO23COyaKGP6fHVpkpWpg=
 github.com/google/pprof v0.0.0-20260402051712-545e8a4df936/go.mod h1:MxpfABSjhmINe3F1It9d+8exIHFvUqtLIRCdOGNXqiI=
+github.com/google/s2a-go v0.1.7 h1:60BLSyTrOV4/haCDW4zb1guZItoSq8foHCXrAnjBo/o=
+github.com/google/s2a-go v0.1.7/go.mod h1:50CgR4k1jNlWBu4UfS4AcfhVe1r6pdZPygJ3R8F0Qdw=
+github.com/google/uuid v1.1.2/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/google/uuid v1.3.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/googleapis/enterprise-certificate-proxy v0.3.2 h1:Vie5ybvEvT75RniqhfFxPRy3Bf7vr3h0cechB90XaQs=
+github.com/googleapis/enterprise-certificate-proxy v0.3.2/go.mod h1:VLSiSSBs/ksPL8kq3OBOQ6WRI2QnaFynd1DCjZ62+V0=
 github.com/goph/emperror v0.17.2 h1:yLapQcmEsO0ipe9p5TaN22djm3OFV/TfM/fcYP0/J18=
 github.com/goph/emperror v0.17.2/go.mod h1:+ZbQ+fUNO/6FNiUo0ujtMjhgad9Xa6fQL9KhH4LNHic=
 github.com/gopherjs/gopherjs v1.17.2 h1:fQnZVsXk8uxXIStYb0N4bGk7jeyTalG/wsZjQ25dO0g=
@@ -155,6 +230,8 @@ github.com/gorilla/websocket v1.5.3/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/ad
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.25.1 h1:VNqngBF40hVlDloBruUehVYC3ArSgIyScOAyMRqBxRg=
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.25.1/go.mod h1:RBRO7fro65R6tjKzYgLAFo0t1QEXY1Dp+i/bvpRiqiQ=
 github.com/hpcloud/tail v1.0.0/go.mod h1:ab1qPbhIpdTxEkNHXyeSf5vhxWSCs/tWer42PpOxQnU=
+github.com/invopop/jsonschema v0.14.0 h1:MHQqLhvpNUZfw+hM3AZDYK7jxO8FZoQeQM77g8iyZjg=
+github.com/invopop/jsonschema v0.14.0/go.mod h1:ygm6C2EaVNMBDPpaPlnOA2pFAxBnxGjFlMZABxm9n2I=
 github.com/jessevdk/go-flags v1.4.0/go.mod h1:4FA24M0QyGHXBuZZK/XkWh8h0e1EYbRYJSGM75WSRxI=
 github.com/jolestar/go-commons-pool/v2 v2.1.2 h1:E+XGo58F23t7HtZiC/W6jzO2Ux2IccSH/yx4nD+J1CM=
 github.com/jolestar/go-commons-pool/v2 v2.1.2/go.mod h1:r4NYccrkS5UqP1YQI1COyTZ9UjPJAAGTUxzcsK1kqhY=
@@ -179,6 +256,8 @@ github.com/kr/pty v1.1.1/go.mod h1:pFQYn66WHrOpPYNljwOMqo10TkYh1fy3cYio2l3bCsQ=
 github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
+github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=
+github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=
 github.com/larksuite/oapi-sdk-go/v3 v3.4.22 h1:57daKuslQPX9X3hC2idc5bu8bl2krfsBGWGJ6b5FlD8=
 github.com/larksuite/oapi-sdk-go/v3 v3.4.22/go.mod h1:ZEplY+kwuIrj/nqw5uSCINNATcH3KdxSN7y+UxYY5fI=
 github.com/leodido/go-urn v1.2.4 h1:XlAE/cm/ms7TE/VMVoduSpNBoyc2dOxHs5MZSwAN63Q=
@@ -195,8 +274,8 @@ github.com/meguminnnnnnnnn/go-openai v0.1.2 h1:iXombGGjqjBrmE9WaSidUhhi3YQhf42QT
 github.com/meguminnnnnnnnn/go-openai v0.1.2/go.mod h1:qs96ysDmxhE4BZoU45I43zcyfnaYxU3X+aRzLko/htY=
 github.com/mgutz/ansi v0.0.0-20170206155736-9520e82c474b h1:j7+1HpAFS1zy5+Q4qx1fWh90gTKwiN4QCGoY9TWyyO4=
 github.com/mgutz/ansi v0.0.0-20170206155736-9520e82c474b/go.mod h1:01TrycV0kFyexm33Z7vhZRXopbI8J3TDReVlkTgMUxE=
-github.com/modelcontextprotocol/go-sdk v1.2.0 h1:Y23co09300CEk8iZ/tMxIX1dVmKZkzoSBZOpJwUnc/s=
-github.com/modelcontextprotocol/go-sdk v1.2.0/go.mod h1:6fM3LCm3yV7pAs8isnKLn07oKtB0MP9LHd3DfAcKw10=
+github.com/modelcontextprotocol/go-sdk v1.3.1 h1:TfqtNKOIWN4Z1oqmPAiWDC2Jq7K9OdJaooe0teoXASI=
+github.com/modelcontextprotocol/go-sdk v1.3.1/go.mod h1:DgVX498dMD8UJlseK1S5i1T4tFz2fkBk4xogC3D15nw=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w8PVh93nsPXa1VrQ6jlwL5oN8l14QlcNfg=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
@@ -221,16 +300,23 @@ github.com/onsi/gomega v1.40.0 h1:Vtol0e1MghCD2ZVIilPDIg44XSL9l2QAn8ZNaljWcJc=
 github.com/onsi/gomega v1.40.0/go.mod h1:M/Uqpu/8qTjtzCLUA2zJHX9Iilrau25x1PdoSRbWh5A=
 github.com/openai/openai-go/v3 v3.35.0 h1:109x3epXMSE423KW2euR506GGFezcEt0s87MoWejpH0=
 github.com/openai/openai-go/v3 v3.35.0/go.mod h1:cdufnVK14cWcT9qA1rRtrXx4FTRsgbDPW7Ia7SS5cZo=
+github.com/pb33f/ordered-map/v2 v2.3.1 h1:5319HDO0aw4DA4gzi+zv4FXU9UlSs3xGZ40wcP1nBjY=
+github.com/pb33f/ordered-map/v2 v2.3.1/go.mod h1:qxFQgd0PkVUtOMCkTapqotNgzRhMPL7VvaHKbd1HnmQ=
 github.com/pelletier/go-toml/v2 v2.2.3 h1:YmeHyLY8mFWbdkNWwpr+qIL2bEqT0o95WSdkNHvL12M=
 github.com/pelletier/go-toml/v2 v2.2.3/go.mod h1:MfCQTFTvCcUyyvvwm1+G6H/jORL20Xlb6rzQu9GuUkc=
+github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c h1:+mdjkGKdHQG3305AYmdv1U2eRNDiU2ErMBj1gwrq8eQ=
+github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c/go.mod h1:7rwL4CYBLnjLxUqIJNnCWiEdr3bn6IUYi15bNlnbCCU=
 github.com/pkg/diff v0.0.0-20210226163009-20ebb0f2a09e/go.mod h1:pJLUxLENpZxwdsKMEsNbx1VGcRFpLqf3715MtcvvzbA=
 github.com/pkg/errors v0.8.0/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkoukk/tiktoken-go v0.1.8 h1:85ENo+3FpWgAACBaEUVp+lctuTcYUO7BtmfhlN/QTRo=
 github.com/pkoukk/tiktoken-go v0.1.8/go.mod h1:9NiV+i9mJKGj1rYOT+njbv+ZwA/zJxYdewGl6qVatpg=
+github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10 h1:GFCKgmp0tecUJ0sJuv4pzYCqS9+RGSn52M3FUwPs+uo=
+github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10/go.mod h1:t/avpk3KcrXxUnYOhZhMXJlSEyie6gQbtLq5NM3loB8=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/prometheus/client_model v0.0.0-20190812154241-14fe0d1b01d4/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA=
 github.com/robfig/cron/v3 v3.0.1 h1:WdRxkvbJztn8LMz/QEvLN5sBU+xKpSqwwUO1Pjr4qDs=
 github.com/robfig/cron/v3 v3.0.1/go.mod h1:eQICP3HwyT7UooqI/z+Ov+PtYAWygg1TEWWzGIFLtro=
 github.com/rogpeppe/go-internal v1.6.1/go.mod h1:xXDCJY+GAPziupqXw64V24skbSoqbTEfhy4qGm1nDQc=
@@ -238,6 +324,10 @@ github.com/rogpeppe/go-internal v1.9.0/go.mod h1:WtVeX8xhTBvf0smdhujwtBcq4Qrzq/f
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
 github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=
 github.com/rollbar/rollbar-go v1.0.2/go.mod h1:AcFs5f0I+c71bpHlXNNDbOWJiKwjFDtISeXco0L5PKQ=
+github.com/segmentio/asm v1.1.3 h1:WM03sfUOENvvKexOLp+pCqgb/WDjsi7EK8gIsICtzhc=
+github.com/segmentio/asm v1.1.3/go.mod h1:Ld3L4ZXGNcSLRg4JBsZ3//1+f/TjYl0Mzen/DQy1EJg=
+github.com/segmentio/encoding v0.5.4 h1:OW1VRern8Nw6ITAtwSZ7Idrl3MXCFwXHPgqESYfvNt0=
+github.com/segmentio/encoding v0.5.4/go.mod h1:HS1ZKa3kSN32ZHVZ7ZLPLXWvOVIiZtyJnO1gPH1sKt0=
 github.com/sirupsen/logrus v1.2.0/go.mod h1:LxeOpSwHxABJmUn/MG1IvRgCAasNZTLOkJPxbbu5VWo=
 github.com/sirupsen/logrus v1.9.3 h1:dueUQJ1C2q9oE3F7wvmSGAaVtTmUizReu6fjN8uqzbQ=
 github.com/sirupsen/logrus v1.9.3/go.mod h1:naHLuLoDiP4jHNo9R0sCBMtWGeIprob74mVsIT4qYEQ=
@@ -251,6 +341,8 @@ github.com/smarty/assertions v1.16.0 h1:EvHNkdRA4QHMrn75NZSoUQ/mAUXAYWfatfB01yTC
 github.com/smarty/assertions v1.16.0/go.mod h1:duaaFdCS0K9dnoM50iyek/eYINOZ64gbh1Xlf6LG7AI=
 github.com/smartystreets/goconvey v1.8.1 h1:qGjIddxOk4grTu9JPOU31tVfq3cNdBlNa5sSznIX1xY=
 github.com/smartystreets/goconvey v1.8.1/go.mod h1:+/u4qLyY6x1jReYOp7GOM2FSt8aP9CzCZL03bI28W60=
+github.com/standard-webhooks/standard-webhooks/libraries v0.0.1 h1:uOfcYT+3QungH6tIGSVCR/Y3KJmgJiHcojJbMTPDZAI=
+github.com/standard-webhooks/standard-webhooks/libraries v0.0.1/go.mod h1:L1MQhA6x4dn9r007T033lsaZMv9EmBAdXyU/+EF40fo=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/objx v0.1.1/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=
@@ -272,14 +364,12 @@ github.com/tencent-connect/botgo v0.2.1 h1:+BrTt9Zh+awL28GWC4g5Na3nQaGRWb0N5IctS
 github.com/tencent-connect/botgo v0.2.1/go.mod h1:oO1sG9ybhXNickvt+CVym5khwQ+uKhTR+IhTqEfOVsI=
 github.com/tetratelabs/wazero v1.11.0 h1:+gKemEuKCTevU4d7ZTzlsvgd1uaToIDtlQlmNbwqYhA=
 github.com/tetratelabs/wazero v1.11.0/go.mod h1:eV28rsN8Q+xwjogd7f4/Pp4xFxO7uOGbLcD/LzB1wiU=
-github.com/tidwall/gjson v1.9.3 h1:hqzS9wAHMO+KVBBkLxYdkEeeFHuqr95GfClRLKlgK0E=
 github.com/tidwall/gjson v1.9.3/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY=
 github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA=
 github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=
-github.com/tidwall/pretty v1.2.0 h1:RWIZEg2iJ8/g6fDDYzMpobmaoGh5OLl4AXtGUGPcqCs=
 github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/pretty v1.2.1 h1:qjsOFOWWQl+N3RsoF5/ssm1pHmJJwhjlSbZ51I6wMl4=
 github.com/tidwall/pretty v1.2.1/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
@@ -302,8 +392,14 @@ github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT0
 github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.2.1/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=
+go.opencensus.io v0.24.0 h1:y73uSU6J157QMP2kn2r30vwW1A2W2WFwSCGnAVxeaD0=
+go.opencensus.io v0.24.0/go.mod h1:vNK8G9p7aAivkbmorf4v+7Hgx+Zs0yY+0fOtgBfjQKo=
 go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=
 go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=
+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.49.0 h1:4Pp6oUg3+e/6M4C0A/3kJ2VYa++dsWVTtGgLVj5xtHg=
+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.49.0/go.mod h1:Mjt1i1INqiaoZOMGR1RIUJN+i3ChKoFRqzrRQhlkbs0=
+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.49.0 h1:jq9TW8u3so/bN+JPT166wjOI6/vQPF6Xe7nMNIltagk=
+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.49.0/go.mod h1:p8pYQP+m5XfbZm9fxtSKAbM6oIllS7s2AfxrChvc7iw=
 go.opentelemetry.io/otel v1.39.0 h1:8yPrr/S0ND9QEfTfdP9V+SiwT4E0G7Y5MO7p85nis48=
 go.opentelemetry.io/otel v1.39.0/go.mod h1:kLlFTywNWrFyEdH0oj2xK0bFYZtHRYUdv1NklR/tgc8=
 go.opentelemetry.io/otel/exporters/otlp/otlptrace v1.34.0 h1:OeNbIYk/2C15ckl7glBlOBp5+WlYsOElzTNmiPW/x60=
@@ -332,6 +428,8 @@ go.uber.org/zap v1.26.0 h1:sI7k6L95XOKS281NhVKOFCUNIvv9e0w4BF8N3u+tCRo=
 go.uber.org/zap v1.26.0/go.mod h1:dtElttAiwGvoJ/vj4IwHBS/gXsEu/pZ50mUIRWuG0so=
 go.yaml.in/yaml/v3 v3.0.4 h1:tfq32ie2Jv2UxXFdLJdh3jXuOzWiL1fo0bu/FbuKpbc=
 go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=
+go.yaml.in/yaml/v4 v4.0.0-rc.2 h1:/FrI8D64VSr4HtGIlUtlFMGsm7H7pWTbj6vOLVZcA6s=
+go.yaml.in/yaml/v4 v4.0.0-rc.2/go.mod h1:aZqd9kCMsGL7AuUv/m/PvWLdg5sjJsZ4oHDEnfPPfY0=
 golang.org/x/arch v0.15.0 h1:QtOrQd0bTUnhNVNndMpLHNWrDmYzZ2KDqSrEymqInZw=
 golang.org/x/arch v0.15.0/go.mod h1:JmwW7aLIoRUKgaTzhkiEFxvcEiQGyOg9BMonBJUS7EE=
 golang.org/x/crypto v0.0.0-20180904163835-0709b304e793/go.mod h1:6SG95UA2DQfeDnfUPMdvaQW0Q7yPrPDi9nlGo2tz2b4=
@@ -343,22 +441,31 @@ golang.org/x/crypto v0.0.0-20210921155107-089bfa567519/go.mod h1:GvvjBRRGRdwPK5y
 golang.org/x/crypto v0.16.0/go.mod h1:gCAAfMLgwOJRpTjQ2zCCt2OcSfYMTeZVSRtQlPC7Nq4=
 golang.org/x/crypto v0.50.0 h1:zO47/JPrL6vsNkINmLoo/PH1gcxpls50DNogFvB5ZGI=
 golang.org/x/crypto v0.50.0/go.mod h1:3muZ7vA7PBCE6xgPX7nkzzjiUq87kRItoJQM1Yo8S+Q=
+golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20250305212735-054e65f0b394 h1:nDVHiLt8aIbd/VzvPWN6kSOPE7+F/fNFDSXLVYkE/Iw=
 golang.org/x/exp v0.0.0-20250305212735-054e65f0b394/go.mod h1:sIifuuw/Yco/y6yb6+bDNfyeQ/MdPUy/hKEMYQV17cM=
 golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8 h1:hVwzHzIUGRjiF7EcUjqNxk3NCfkPxbDKRdnNE1Rpg0U=
 golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
+golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE=
+golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU=
+golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc=
 golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA=
 golang.org/x/mod v0.6.0-dev.0.20220419223038-86c51ed26bb4/go.mod h1:jJ57K6gSWd91VN4djpZkiMVwK6gcyfeH4XE8wZrZaV4=
 golang.org/x/mod v0.8.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs=
 golang.org/x/mod v0.35.0 h1:Ww1D637e6Pg+Zb2KrWfHQUnH2dQRLBQyAtpr/haaJeM=
 golang.org/x/mod v0.35.0/go.mod h1:+GwiRhIInF8wPm+4AoT6L0FA1QWAad3OMdTRx4tFYlU=
+golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
 golang.org/x/net v0.0.0-20180906233101-161cd47e91fd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4=
+golang.org/x/net v0.0.0-20190311183353-d8887717615a/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
 golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg=
 golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
 golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=
 golang.org/x/net v0.0.0-20200520004742-59133d7f0dd7/go.mod h1:qpuaurCH72eLCgpAm/N6yyVIVM9cpaDIP3A8BGJEC5A=
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
+golang.org/x/net v0.0.0-20201110031124-69a78807bb2b/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210428140749-89ef3d95e781/go.mod h1:OJAsFXCWl8Ukc7SiCT/9KSuxbyM7479/AVlXFRxuMCk=
@@ -368,10 +475,12 @@ golang.org/x/net v0.10.0/go.mod h1:0qNGK6F8kojg2nk9dLZ2mShWaEBan6FAoqfSigmmuDg=
 golang.org/x/net v0.19.0/go.mod h1:CfAk/cbD4CthTvqiEl8NpboMuiuOYsAr/7NOjZJtv1U=
 golang.org/x/net v0.53.0 h1:d+qAbo5L0orcWAr0a9JweQpjXF19LMXJE8Ey7hwOdUA=
 golang.org/x/net v0.53.0/go.mod h1:JvMuJH7rrdiCfbeHoo3fCQU24Lf5JJwT9W3sJFulfgs=
+golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.23.0/go.mod h1:XYTD2NtWslqkgxebSiOHnXEap4TF09sJSc7H1sXbhtI=
 golang.org/x/oauth2 v0.34.0 h1:hqK/t4AKgbqWkdkcAeI8XLmbK+4m4G5YeQRrmiotGlw=
 golang.org/x/oauth2 v0.34.0/go.mod h1:lzm5WQJQwKZ3nwavOZ3IS5Aulzxi68dUSgRHujetwEA=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
+golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -379,6 +488,7 @@ golang.org/x/sync v0.0.0-20220722155255-886fb9371eb4/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
 golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20180905080454-ebe1bf3edb33/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20180909124046-d0be0721c37e/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
@@ -421,6 +531,10 @@ golang.org/x/text v0.37.0/go.mod h1:a5sjxXGs9hsn/AJVwuElvCAo9v8QYLzvavO5z2PiM38=
 golang.org/x/time v0.14.0 h1:MRx4UaLrDotUKUdCIqzPC48t1Y9hANFKIRpNx+Te8PI=
 golang.org/x/time v0.14.0/go.mod h1:eL/Oa2bBBK0TkX57Fyni+NgnyQQN4LitPmob2Hjnqw4=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
+golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
+golang.org/x/tools v0.0.0-20190226205152-f727befe758c/go.mod h1:9Yl7xja0Znq3iFh3HoIrodX9oNMXvdceNzlUR8zjMvY=
+golang.org/x/tools v0.0.0-20190311212946-11955173bddd/go.mod h1:LCzVGOaR6xXOjkQ3onu1FJEFr0SW1gC7cKk1uF8kGRs=
+golang.org/x/tools v0.0.0-20190524140312-2c0ae7006135/go.mod h1:RgjU9mgBXZiqYHBnxXauZ1Gv1EHHAz9KjViQ78xBX0Q=
 golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo=
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20201224043029-2b0845dc783e/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
@@ -435,10 +549,22 @@ golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8T
 golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 gonum.org/v1/gonum v0.16.0 h1:5+ul4Swaf3ESvrOnidPp4GZbzf0mxVQpDCYUQE7OJfk=
 gonum.org/v1/gonum v0.16.0/go.mod h1:fef3am4MQ93R2HHpKnLk4/Tbh/s0+wqD5nfa6Pnwy4E=
+google.golang.org/api v0.189.0 h1:equMo30LypAkdkLMBqfeIqtyAnlyig1JSZArl4XPwdI=
+google.golang.org/api v0.189.0/go.mod h1:FLWGJKb0hb+pU2j+rJqwbnsF+ym+fQs73rbJ+KAUgy8=
+google.golang.org/appengine v1.1.0/go.mod h1:EbEs0AVv82hx2wNQdGPgUI5lhzA/G0D9YwlJXL52JkM=
+google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4=
+google.golang.org/genproto v0.0.0-20180817151627-c66870c02cf8/go.mod h1:JiN7NxoALGmiZfu7CAH4rXhgtRTLTxftemlI0sWmxmc=
+google.golang.org/genproto v0.0.0-20190819201941-24fa4b261c55/go.mod h1:DMBHOl98Agz4BDEuKkezgsaosCRResVns1a3J2ZsMNc=
+google.golang.org/genproto v0.0.0-20200526211855-cb27e3aa2013/go.mod h1:NbSheEEYHJ7i3ixzK3sjbqSGDJWnxyFXZblF3eUsNvo=
 google.golang.org/genproto/googleapis/api v0.0.0-20251202230838-ff82c1b0f217 h1:fCvbg86sFXwdrl5LgVcTEvNC+2txB5mgROGmRL5mrls=
 google.golang.org/genproto/googleapis/api v0.0.0-20251202230838-ff82c1b0f217/go.mod h1:+rXWjjaukWZun3mLfjmVnQi18E1AsFbDN9QdJ5YXLto=
 google.golang.org/genproto/googleapis/rpc v0.0.0-20251202230838-ff82c1b0f217 h1:gRkg/vSppuSQoDjxyiGfN4Upv/h/DQmIR10ZU8dh4Ww=
 google.golang.org/genproto/googleapis/rpc v0.0.0-20251202230838-ff82c1b0f217/go.mod h1:7i2o+ce6H/6BluujYR+kqX3GKH+dChPTQU19wjRPiGk=
+google.golang.org/grpc v1.19.0/go.mod h1:mqu4LbDTu4XGKhr4mRzUsmM4RtVoemTSY81AxZiDr8c=
+google.golang.org/grpc v1.23.0/go.mod h1:Y5yQAOtifL1yxbo5wqy6BxZv8vAUGQwXBOALyacEbxg=
+google.golang.org/grpc v1.25.1/go.mod h1:c3i+UQWmh7LiEpx4sFZnkU36qjEYZ0imhYfXVyQciAY=
+google.golang.org/grpc v1.27.0/go.mod h1:qbnxyOmOxrQa7FizSgH+ReBfzJrCY1pSN7KXBS8abTk=
+google.golang.org/grpc v1.33.2/go.mod h1:JMHMWHQWaTccqQQlmk3MJZS+GWXOdAesneDmEnv2fbc=
 google.golang.org/grpc v1.79.3 h1:sybAEdRIEtvcD68Gx7dmnwjZKlyfuc61Dyo9pGXXkKE=
 google.golang.org/grpc v1.79.3/go.mod h1:KmT0Kjez+0dde/v2j9vzwoAScgEPx/Bw1CYChhHLrHQ=
 google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8=
@@ -446,7 +572,10 @@ google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ
 google.golang.org/protobuf v0.0.0-20200228230310-ab0ca4ff8a60/go.mod h1:cfTl7dwQJ+fmap5saPgwCLgHXTUD7jkjRqWcaiX5VyM=
 google.golang.org/protobuf v1.20.1-0.20200309200217-e05f789c0967/go.mod h1:A+miEFZTKqfCUM6K7xSMQL9OKL/b6hQv+e19PK+JZNE=
 google.golang.org/protobuf v1.21.0/go.mod h1:47Nbq4nVaFHyn7ilMalzfO3qCViNmqZ2kzikPIcrTAo=
+google.golang.org/protobuf v1.22.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
 google.golang.org/protobuf v1.23.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
+google.golang.org/protobuf v1.23.1-0.20200526195155-81db48ad09cc/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU=
+google.golang.org/protobuf v1.25.0/go.mod h1:9JNX74DMeImyA3h4bdi1ymwjUzf21/xIlbajtzgsN7c=
 google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw=
 google.golang.org/protobuf v1.26.0/go.mod h1:9q0QmTI4eRPtz6boOQmLYwt+qCgq0jsYwAQnmE0givc=
 google.golang.org/protobuf v1.36.10 h1:AYd7cD/uASjIL6Q9LiTjz8JLcrh/88q5UObnmY3aOOE=
@@ -462,7 +591,10 @@ gopkg.in/yaml.v2 v2.2.1/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v2 v2.2.4/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
 gopkg.in/yaml.v2 v2.3.0/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=
+gopkg.in/yaml.v2 v2.4.0 h1:D8xgwECY7CYvx+Y2n4sBz93Jn9JRvxdiyyo8CTfuKaY=
 gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=
 gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
 gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
+honnef.co/go/tools v0.0.0-20190102054323-c2f93a96b099/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4=
+honnef.co/go/tools v0.0.0-20190523083050-ea95bdfd59fc/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4=

internal/agent/agent.go

@@ -820,6 +820,19 @@ func (a *Agent) UpdateMCPExecutionDisplayResult(executionID, resultText string)
 	}
 }
 
+// MCPExecutionResultText returns the monitor-facing result text after storage
+// guards such as large-output spilling have been applied.
+func (a *Agent) MCPExecutionResultText(executionID string) string {
+	if a == nil || a.mcpServer == nil || strings.TrimSpace(executionID) == "" {
+		return ""
+	}
+	exec, ok := a.mcpServer.GetExecution(executionID)
+	if !ok || exec == nil || exec.Result == nil {
+		return ""
+	}
+	return mcp.ToolResultPlainText(exec.Result)
+}
+
 // CancelMCPToolExecutionWithNote 取消一次进行中的 MCP 工具（先内部后外部），与监控页「终止工具」一致；note 非空时合并进返回给模型的文本。
 func (a *Agent) CancelMCPToolExecutionWithNote(executionID, note string) bool {
 	executionID = strings.TrimSpace(executionID)

internal/agentfinalizer/decision.go

@@ -65,12 +65,6 @@ func FromRunResult(db *database.DB, result *multiagent.RunResult, in Input) Deci
 		if len(in.MCPExecutionIDs) == 0 {
 			in.MCPExecutionIDs = result.MCPExecutionIDs
 		}
-		if strings.TrimSpace(in.Status) == "" {
-			in.Status = result.Status
-		}
-		if strings.TrimSpace(in.CompletionReason) == "" {
-			in.CompletionReason = result.CompletionReason
-		}
 	}
 	d := Decide(db, in)
 	if result != nil {

internal/agentfinalizer/decision_test.go

@@ -7,6 +7,7 @@ import (
 
 	"cyberstrike-ai/internal/database"
 	"cyberstrike-ai/internal/mcp"
+	"cyberstrike-ai/internal/multiagent"
 
 	"go.uber.org/zap"
 )
@@ -130,3 +131,23 @@ func TestDecideAllowsInformationalAnswerWhenExecutionEvidenceIsNotRequired(t *te
 		t.Fatalf("informational response should finalize when execution evidence is not required: %+v", d)
 	}
 }
+
+func TestFromRunResultDoesNotReusePreviousFinalizationStatusAsRunStatus(t *testing.T) {
+	db := newDecisionTestDB(t)
+	saveDecisionTestExecution(t, db, "run-slow", mcp.ToolExecutionStatusRunning)
+	result := &multiagent.RunResult{
+		Response:        "工具已触发，按用户要求直接总结。",
+		MCPExecutionIDs: []string{"run-slow"},
+	}
+
+	first := FromRunResult(db, result, Input{})
+	if first.Finalizable || first.CompletionReason != ReasonPendingTools || result.Status != StatusInProgress {
+		t.Fatalf("first decision should mark pending and write metadata: decision=%+v result=%+v", first, result)
+	}
+
+	saveDecisionTestExecution(t, db, "run-slow", mcp.ToolExecutionStatusCancelled)
+	second := FromRunResult(db, result, Input{})
+	if !second.Finalizable || !second.Finalized || second.Status != StatusCompleted {
+		t.Fatalf("second decision should ignore previous result status after pending cleanup: decision=%+v result=%+v", second, result)
+	}
+}

internal/app/app.go

@@ -972,6 +972,8 @@ func setupRoutes(
 		protected.GET("/hitl/tool-whitelist", agentHandler.GetHITLGlobalToolWhitelist)
 		protected.PUT("/hitl/tool-whitelist", agentHandler.SetHITLGlobalToolWhitelist)
 		protected.POST("/hitl/tool-whitelist", agentHandler.MergeHITLGlobalToolWhitelist)
+		protected.GET("/hitl/default-config", agentHandler.GetHITLDefaultConfig)
+		protected.PUT("/hitl/default-config", agentHandler.UpdateHITLDefaultConfig)
 		protected.GET("/hitl/default-reviewer", agentHandler.GetHITLDefaultReviewer)
 		protected.PUT("/hitl/default-reviewer", agentHandler.UpdateHITLDefaultReviewer)
 		protected.GET("/hitl/audit-strategy", agentHandler.GetHITLAuditStrategy)
@@ -1027,9 +1029,11 @@ func setupRoutes(
 		protected.DELETE("/batch-tasks/:queueId/tasks/:taskId", agentHandler.DeleteBatchTask)
 
 		// 对话历史
+		protected.GET("/usage/tokens", conversationHandler.GetTokenUsageStats)
 		protected.POST("/conversations", conversationHandler.CreateConversation)
 		protected.GET("/conversations", conversationHandler.ListConversations)
 		protected.GET("/conversations/:id", conversationHandler.GetConversation)
+		protected.GET("/conversations/:id/token-usage", conversationHandler.GetConversationTokenUsageStats)
 		protected.GET("/conversations/:id/plan-tasks", conversationHandler.GetConversationPlanTasks)
 		protected.GET("/messages/:id/process-details", conversationHandler.GetMessageProcessDetails)
 		protected.GET("/process-details/:id", conversationHandler.GetProcessDetail)

internal/config/config.go

@@ -298,7 +298,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	PlanExecuteMaxStepResultRunes int `yaml:"plan_execute_max_step_result_runes,omitempty" json:"plan_execute_max_step_result_runes,omitempty"`
 	// PlanExecuteKeepLastSteps keeps only the tail steps in prompt view (default 8).
 	PlanExecuteKeepLastSteps int `yaml:"plan_execute_keep_last_steps,omitempty" json:"plan_execute_keep_last_steps,omitempty"`
-	// CheckpointDir when non-empty enables adk.Runner CheckPointStore (file-backed) for interrupt/resume persistence.
+	// CheckpointDir is retained for config compatibility. Chat agent runs do
+	// not consume it; cross-turn recovery is centralized in conversations.last_react_*.
 	CheckpointDir string `yaml:"checkpoint_dir,omitempty" json:"checkpoint_dir,omitempty"`
 	// DeepOutputKey passed to deep.Config OutputKey (session final text); empty = off.
 	DeepOutputKey string `yaml:"deep_output_key,omitempty" json:"deep_output_key,omitempty"`
@@ -817,7 +818,7 @@ type MCPConfig struct {
 }
 
 type OpenAIConfig struct {
-	Provider            string `yaml:"provider,omitempty" json:"provider,omitempty"` // API 提供商: "openai"(默认) 或 "claude"，claude 时自动桥接为 Anthropic Messages API
+	Provider            string `yaml:"provider,omitempty" json:"provider,omitempty"` // API 提供商: "openai"(默认) 或 "claude"，claude 使用 Eino 原生 Anthropic Messages API
 	APIKey              string `yaml:"api_key" json:"api_key"`
 	BaseURL             string `yaml:"base_url" json:"base_url"`
 	Model               string `yaml:"model" json:"model"`
@@ -959,13 +960,12 @@ func (c OpenAIConfig) MaxCompletionTokensEffective() int {
 }
 
 // IsDeepSeekEndpointOrModel reports whether the channel targets DeepSeek's
-// official-compatible API or a DeepSeek model family. This is separate from the
-// reasoning profile: profile controls field mapping, while DeepSeek has provider
-// constraints such as default thinking mode and no tool_choice in thinking mode.
+// official-compatible API endpoint. The historical name is kept for compatibility;
+// model names alone are not enough to infer DeepSeek wire behavior behind
+// OpenAI-compatible gateways.
 func (c OpenAIConfig) IsDeepSeekEndpointOrModel() bool {
 	baseURL := strings.ToLower(strings.TrimSpace(c.BaseURL))
-	model := strings.ToLower(strings.TrimSpace(c.Model))
-	return strings.Contains(baseURL, "deepseek") || strings.Contains(model, "deepseek")
+	return strings.Contains(baseURL, "deepseek")
 }
 
 // OpenAIReasoningConfig 全局默认与网关 profile（对话页可通过 ChatRequest.reasoning 覆盖，受 AllowClientReasoning 约束）。
@@ -1062,8 +1062,24 @@ type HitlConfig struct {
 	AuditAgentPromptReviewEdit string `yaml:"audit_agent_prompt_review_edit,omitempty" json:"audit_agent_prompt_review_edit,omitempty"`
 	// RetentionDays 已决策审计日志（hitl_interrupts 非 pending）保留天数；省略时默认 90；0 表示不自动清理。
 	RetentionDays *int `yaml:"retention_days,omitempty" json:"retention_days,omitempty"`
-	// DefaultReviewer 全局默认审批方（human | audit_agent）；未选会话时切换会写入 config.yaml；新建会话无独立配置时沿用。
+	// DefaultMode 全局默认人机协同模式（off | approval | review_edit）；新建会话无独立配置时沿用。
+	DefaultMode string `yaml:"default_mode,omitempty" json:"default_mode,omitempty"`
+	// DefaultReviewer 全局默认审批方（human | audit_agent）；新建会话无独立配置时沿用。
 	DefaultReviewer string `yaml:"default_reviewer,omitempty" json:"default_reviewer,omitempty"`
+	// DefaultTimeoutSeconds 全局默认审批等待秒数；nil 表示使用前端历史默认 300 秒，0 表示不限时。
+	DefaultTimeoutSeconds *int `yaml:"default_timeout_seconds,omitempty" json:"default_timeout_seconds,omitempty"`
+}
+
+// EffectiveDefaultMode returns off, approval, or review_edit; omitted or unknown values default to off.
+func (h HitlConfig) EffectiveDefaultMode() string {
+	switch strings.ToLower(strings.TrimSpace(h.DefaultMode)) {
+	case "feedback", "followup":
+		return "approval"
+	case "approval", "review_edit":
+		return strings.ToLower(strings.TrimSpace(h.DefaultMode))
+	default:
+		return "off"
+	}
 }
 
 // EffectiveDefaultReviewer returns human or audit_agent; omitted or unknown values default to human.
@@ -1076,6 +1092,17 @@ func (h HitlConfig) EffectiveDefaultReviewer() string {
 	}
 }
 
+// EffectiveDefaultTimeoutSeconds returns the default HITL approval timeout; nil defaults to 5 minutes.
+func (h HitlConfig) EffectiveDefaultTimeoutSeconds() int {
+	if h.DefaultTimeoutSeconds == nil {
+		return 300
+	}
+	if *h.DefaultTimeoutSeconds < 0 {
+		return 0
+	}
+	return *h.DefaultTimeoutSeconds
+}
+
 // RetentionDaysEffective returns retention; 0 means keep forever; omitted defaults to 90.
 func (h HitlConfig) RetentionDaysEffective() int {
 	if h.RetentionDays == nil {

internal/config/config_test.go

@@ -95,6 +95,29 @@ func TestHitlAuditModelEffectiveFallsBackToMainConfig(t *testing.T) {
 	}
 }
 
+func TestHitlDefaultConfigEffectiveValues(t *testing.T) {
+	if got := (HitlConfig{}).EffectiveDefaultMode(); got != "off" {
+		t.Fatalf("empty default mode = %q, want off", got)
+	}
+	if got := (HitlConfig{DefaultMode: "review-edit"}).EffectiveDefaultMode(); got != "off" {
+		t.Fatalf("unknown default mode = %q, want off", got)
+	}
+	if got := (HitlConfig{DefaultMode: "review_edit"}).EffectiveDefaultMode(); got != "review_edit" {
+		t.Fatalf("review_edit default mode = %q, want review_edit", got)
+	}
+	if got := (HitlConfig{}).EffectiveDefaultTimeoutSeconds(); got != 300 {
+		t.Fatalf("empty default timeout = %d, want 300", got)
+	}
+	zero := 0
+	if got := (HitlConfig{DefaultTimeoutSeconds: &zero}).EffectiveDefaultTimeoutSeconds(); got != 0 {
+		t.Fatalf("zero default timeout = %d, want 0", got)
+	}
+	neg := -1
+	if got := (HitlConfig{DefaultTimeoutSeconds: &neg}).EffectiveDefaultTimeoutSeconds(); got != 0 {
+		t.Fatalf("negative default timeout = %d, want 0", got)
+	}
+}
+
 func TestLoadUsesAIDefaultChannelAsRuntimeOpenAI(t *testing.T) {
 	dir := t.TempDir()
 	path := filepath.Join(dir, "config.yaml")

internal/database/conversation.go

@@ -1350,6 +1350,8 @@ func (db *DB) AddProcessDetailWithID(messageID, conversationID, eventType, messa
 		return "", fmt.Errorf("添加过程详情失败: %w", err)
 	}
 
+	db.maybeRecordModelTokenUsage(messageID, conversationID, id, eventType, data)
+
 	return id, nil
 }
 
@@ -1538,6 +1540,11 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 		return nil, fmt.Errorf("统计工具调用详情失败: %w", err)
 	}
 
+	pendingToolStatus := "result_missing"
+	if summary.Status == "running" {
+		pendingToolStatus = "running"
+	}
+
 	execRows, err := db.Query(
 		"SELECT id, event_type, data FROM process_details WHERE message_id = ? AND event_type IN ('tool_call', 'tool_result') ORDER BY created_at ASC, rowid ASC",
 		messageID,
@@ -1548,12 +1555,12 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 	seenExecIDs := make(map[string]bool)
 	// A provider may reuse a fallback toolCallId across streaming rounds. Keep a
 	// FIFO per ID instead of a single index so every persisted call gets at most
-	// one result. Results without a stable ID are kept separate instead of being
-	// guessed by order; showing no link is safer than linking to the wrong tool.
+	// one result. ID-less results still attach to an unmatched call with the same
+	// tool name (parallel nmap 1/2, 2/2 often lose one ID); different tools stay
+	// unlinked so a leftover preview cannot steal another call's slot.
 	toolIndexesByCallID := make(map[string][]int)
 	lastMatchedToolIndexByCallID := make(map[string]int)
 	matchedToolIndexes := make([]bool, 0)
-	nextUnmatchedToolIdx := 0
 	for execRows.Next() {
 		var detailID string
 		var eventType string
@@ -1569,33 +1576,19 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 		if err := json.Unmarshal([]byte(dataJSON), &payload); err != nil {
 			continue
 		}
-		toolName, _ := payload["toolName"].(string)
-		toolName = strings.TrimSpace(toolName)
-		toolCallID, _ := payload["toolCallId"].(string)
-		toolCallID = strings.TrimSpace(toolCallID)
-		execID, _ := payload["executionId"].(string)
-		execID = strings.TrimSpace(execID)
-		status := ""
-		if eventType == "tool_result" {
-			if success, ok := payload["success"].(bool); ok {
-				if success {
-					status = "completed"
-				} else {
-					status = "failed"
-				}
-			} else if isErr, ok := payload["isError"].(bool); ok && isErr {
-				status = "failed"
-			}
-		}
+		toolName := processDetailString(payload, "toolName")
+		toolCallID := processDetailString(payload, "toolCallId")
+		execID := processDetailString(payload, "executionId")
+		status := toolResultStatusFromPayload(payload, eventType)
 		if eventType == "tool_call" {
 			summary.ToolExecutions = append(summary.ToolExecutions, ProcessDetailsToolExecution{
 				ProcessDetailID: strings.TrimSpace(detailID),
 				ToolName:        toolName,
 				ToolCallID:      toolCallID,
-				// This summary is reconstructed from persisted history, not live
-				// execution state. Until a matching result is found the honest state
-				// is "result_missing", never "running".
-				Status: "result_missing",
+				// This summary is reconstructed from persisted history. For an
+				// active assistant turn, a missing result means the call is still
+				// pending; after the turn is terminal it is genuinely incomplete.
+				Status: pendingToolStatus,
 			})
 			matchedToolIndexes = append(matchedToolIndexes, false)
 			if toolCallID != "" {
@@ -1603,36 +1596,14 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 			}
 		}
 		if eventType == "tool_result" {
-			idx := -1
-			if toolCallID != "" {
-				queue := toolIndexesByCallID[toolCallID]
-				for len(queue) > 0 {
-					candidate := queue[0]
-					queue = queue[1:]
-					if candidate >= 0 && candidate < len(matchedToolIndexes) && !matchedToolIndexes[candidate] {
-						idx = candidate
-						break
-					}
-				}
-				toolIndexesByCallID[toolCallID] = queue
-				if idx < 0 {
-					// Multiple persisted result events for one call (for example an
-					// agent-facing reduced result replacing an earlier preview) update
-					// that call instead of consuming an unrelated FIFO entry.
-					if previous, ok := lastMatchedToolIndexByCallID[toolCallID]; ok {
-						idx = previous
-					}
-				}
-			}
-			if idx < 0 && toolCallID != "" {
-				for nextUnmatchedToolIdx < len(matchedToolIndexes) && matchedToolIndexes[nextUnmatchedToolIdx] {
-					nextUnmatchedToolIdx++
-				}
-				if nextUnmatchedToolIdx < len(matchedToolIndexes) {
-					idx = nextUnmatchedToolIdx
-					nextUnmatchedToolIdx++
-				}
-			}
+			idx := matchToolExecutionIndex(
+				summary.ToolExecutions,
+				matchedToolIndexes,
+				toolCallID,
+				toolName,
+				toolIndexesByCallID,
+				lastMatchedToolIndexByCallID,
+			)
 			if idx >= 0 && idx < len(summary.ToolExecutions) {
 				matchedToolIndexes[idx] = true
 				if toolCallID != "" {
@@ -1648,6 +1619,8 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 				summary.ToolExecutions[idx].ExecutionID = execID
 				if status != "" {
 					summary.ToolExecutions[idx].Status = status
+				} else {
+					summary.ToolExecutions[idx].Status = "completed"
 				}
 			} else {
 				summary.ToolExecutions = append(summary.ToolExecutions, ProcessDetailsToolExecution{
@@ -1670,6 +1643,7 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 		return nil, fmt.Errorf("遍历工具执行摘要失败: %w", err)
 	}
 	execRows.Close()
+	db.applyPersistedToolExecutionStatuses(summary.ToolExecutions)
 
 	rows, err := db.Query(
 		"SELECT data FROM process_details WHERE message_id = ? AND event_type = 'iteration' ORDER BY created_at ASC, rowid ASC",
@@ -1704,6 +1678,100 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 	return summary, nil
 }
 
+func processDetailString(payload map[string]interface{}, key string) string {
+	if payload == nil {
+		return ""
+	}
+	v, ok := payload[key]
+	if !ok || v == nil {
+		return ""
+	}
+	s := strings.TrimSpace(fmt.Sprint(v))
+	if s == "" || s == "<nil>" {
+		return ""
+	}
+	return s
+}
+
+func toolResultStatusFromPayload(payload map[string]interface{}, eventType string) string {
+	if eventType != "tool_result" {
+		return ""
+	}
+	if status := processDetailString(payload, "status"); strings.EqualFold(status, "background_running") {
+		return "background_running"
+	}
+	if success, ok := payload["success"].(bool); ok {
+		if success {
+			return "completed"
+		}
+		return "failed"
+	}
+	if isErr, ok := payload["isError"].(bool); ok && isErr {
+		return "failed"
+	}
+	return "completed"
+}
+
+func (db *DB) applyPersistedToolExecutionStatuses(executions []ProcessDetailsToolExecution) {
+	for i := range executions {
+		execID := strings.TrimSpace(executions[i].ExecutionID)
+		if execID == "" {
+			continue
+		}
+		var status string
+		if err := db.QueryRow(`SELECT status FROM tool_executions WHERE id = ?`, execID).Scan(&status); err != nil {
+			continue
+		}
+		status = strings.ToLower(strings.TrimSpace(status))
+		if status == "" {
+			continue
+		}
+		executions[i].Status = status
+	}
+}
+
+func matchToolExecutionIndex(
+	executions []ProcessDetailsToolExecution,
+	matched []bool,
+	toolCallID, toolName string,
+	toolIndexesByCallID map[string][]int,
+	lastMatchedToolIndexByCallID map[string]int,
+) int {
+	if toolCallID != "" {
+		queue := toolIndexesByCallID[toolCallID]
+		for len(queue) > 0 {
+			candidate := queue[0]
+			queue = queue[1:]
+			if candidate >= 0 && candidate < len(matched) && !matched[candidate] {
+				toolIndexesByCallID[toolCallID] = queue
+				return candidate
+			}
+		}
+		toolIndexesByCallID[toolCallID] = queue
+		if previous, ok := lastMatchedToolIndexByCallID[toolCallID]; ok {
+			return previous
+		}
+	}
+	if toolName != "" {
+		for i := range matched {
+			if matched[i] {
+				continue
+			}
+			if strings.EqualFold(strings.TrimSpace(executions[i].ToolName), toolName) {
+				return i
+			}
+		}
+	}
+	if toolCallID != "" {
+		for i := range matched {
+			if !matched[i] {
+				return i
+			}
+		}
+	}
+	return -1
+}
+
 // GetProcessDetailsPage 分页获取消息的过程详情（按时间升序）。
 func (db *DB) GetProcessDetailsPage(messageID string, limit, offset int) ([]ProcessDetail, int, error) {
 	var total int

internal/database/database.go

@@ -216,6 +216,32 @@ func (db *DB) initTables() error {
 		FOREIGN KEY (conversation_id) REFERENCES conversations(id) ON DELETE CASCADE
 	);`
 
+	// 创建模型 Token 用量表：process_details 负责时间线回放，本表负责结构化聚合统计。
+	createModelTokenUsageTable := `
+	CREATE TABLE IF NOT EXISTS model_token_usage (
+		id TEXT PRIMARY KEY,
+		process_detail_id TEXT NOT NULL UNIQUE,
+		message_id TEXT NOT NULL,
+		conversation_id TEXT NOT NULL,
+		project_id TEXT,
+		source TEXT NOT NULL DEFAULT '',
+		orchestration TEXT NOT NULL DEFAULT '',
+		reason TEXT NOT NULL DEFAULT '',
+		model TEXT NOT NULL DEFAULT '',
+		model_calls INTEGER NOT NULL DEFAULT 0,
+		prompt_tokens INTEGER NOT NULL DEFAULT 0,
+		completion_tokens INTEGER NOT NULL DEFAULT 0,
+		total_tokens INTEGER NOT NULL DEFAULT 0,
+		cached_tokens INTEGER NOT NULL DEFAULT 0,
+		reasoning_tokens INTEGER NOT NULL DEFAULT 0,
+		created_at DATETIME NOT NULL,
+		updated_at DATETIME NOT NULL,
+		FOREIGN KEY (process_detail_id) REFERENCES process_details(id) ON DELETE CASCADE,
+		FOREIGN KEY (message_id) REFERENCES messages(id) ON DELETE CASCADE,
+		FOREIGN KEY (conversation_id) REFERENCES conversations(id) ON DELETE CASCADE,
+		FOREIGN KEY (project_id) REFERENCES projects(id) ON DELETE SET NULL
+	);`
+
 	// 创建工具执行记录表
 	createToolExecutionsTable := `
 	CREATE TABLE IF NOT EXISTS tool_executions (
@@ -719,6 +745,10 @@ func (db *DB) initTables() error {
 	CREATE INDEX IF NOT EXISTS idx_conversations_updated_at ON conversations(updated_at);
 	CREATE INDEX IF NOT EXISTS idx_process_details_message_id ON process_details(message_id);
 	CREATE INDEX IF NOT EXISTS idx_process_details_conversation_id ON process_details(conversation_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_created_at ON model_token_usage(created_at);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_conversation ON model_token_usage(conversation_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_project ON model_token_usage(project_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_model ON model_token_usage(model);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_tool_name ON tool_executions(tool_name);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_start_time ON tool_executions(start_time);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_status ON tool_executions(status);
@@ -806,6 +836,10 @@ func (db *DB) initTables() error {
 		return fmt.Errorf("创建process_details表失败: %w", err)
 	}
 
+	if _, err := db.Exec(createModelTokenUsageTable); err != nil {
+		return fmt.Errorf("创建model_token_usage表失败: %w", err)
+	}
+
 	if _, err := db.Exec(createToolExecutionsTable); err != nil {
 		return fmt.Errorf("创建tool_executions表失败: %w", err)
 	}
@@ -981,6 +1015,10 @@ func (db *DB) initTables() error {
 	if _, err := db.Exec(createIndexes); err != nil {
 		return fmt.Errorf("创建索引失败: %w", err)
 	}
+
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		return fmt.Errorf("回填模型Token用量失败: %w", err)
+	}
 	db.logger.Debug("数据库表初始化完成")
 	return nil
 }

internal/database/model_token_usage.go

@@ -0,0 +1,485 @@
+package database
+
+import (
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"math"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+	"go.uber.org/zap"
+)
+
+const modelTokenUsageEventType = "eino_usage_summary"
+
+// ModelTokenUsage records one model-usage summary emitted by an Agent run.
+type ModelTokenUsage struct {
+	ID               string    `json:"id"`
+	ProcessDetailID  string    `json:"processDetailId"`
+	MessageID        string    `json:"messageId"`
+	ConversationID   string    `json:"conversationId"`
+	ProjectID        string    `json:"projectId,omitempty"`
+	Source           string    `json:"source"`
+	Orchestration    string    `json:"orchestration"`
+	Reason           string    `json:"reason"`
+	Model            string    `json:"model,omitempty"`
+	ModelCalls       int64     `json:"modelCalls"`
+	PromptTokens     int64     `json:"promptTokens"`
+	CompletionTokens int64     `json:"completionTokens"`
+	TotalTokens      int64     `json:"totalTokens"`
+	CachedTokens     int64     `json:"cachedTokens"`
+	ReasoningTokens  int64     `json:"reasoningTokens"`
+	CreatedAt        time.Time `json:"createdAt"`
+	UpdatedAt        time.Time `json:"updatedAt"`
+}
+
+// ModelTokenUsageSummary is the aggregate shape used by dashboard and APIs.
+type ModelTokenUsageSummary struct {
+	Events           int64 `json:"events"`
+	ModelCalls       int64 `json:"modelCalls"`
+	PromptTokens     int64 `json:"promptTokens"`
+	CompletionTokens int64 `json:"completionTokens"`
+	TotalTokens      int64 `json:"totalTokens"`
+	CachedTokens     int64 `json:"cachedTokens"`
+	ReasoningTokens  int64 `json:"reasoningTokens"`
+}
+
+// ModelTokenUsageBreakdown is a grouped aggregate row.
+type ModelTokenUsageBreakdown struct {
+	Key              string `json:"key"`
+	Label            string `json:"label,omitempty"`
+	Events           int64  `json:"events"`
+	ModelCalls       int64  `json:"modelCalls"`
+	PromptTokens     int64  `json:"promptTokens"`
+	CompletionTokens int64  `json:"completionTokens"`
+	TotalTokens      int64  `json:"totalTokens"`
+	CachedTokens     int64  `json:"cachedTokens"`
+	ReasoningTokens  int64  `json:"reasoningTokens"`
+}
+
+// ModelTokenUsageStats is a compact API response for usage dashboards.
+type ModelTokenUsageStats struct {
+	Summary         ModelTokenUsageSummary     `json:"summary"`
+	Today           ModelTokenUsageSummary     `json:"today"`
+	ByDay           []ModelTokenUsageBreakdown `json:"byDay"`
+	ByModel         []ModelTokenUsageBreakdown `json:"byModel"`
+	ByOrchestration []ModelTokenUsageBreakdown `json:"byOrchestration"`
+	Recent          []ModelTokenUsage          `json:"recent"`
+}
+
+// ModelTokenUsageFilter scopes usage queries.
+type ModelTokenUsageFilter struct {
+	ConversationID string
+	ProjectID      string
+	Since          time.Time
+	Until          time.Time
+	Days           int
+	Access         RBACListAccess
+	Limit          int
+}
+
+func modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID string, data interface{}) (ModelTokenUsage, bool) {
+	m := mapFromUsageData(data)
+	if len(m) == 0 {
+		return ModelTokenUsage{}, false
+	}
+	usage := ModelTokenUsage{
+		ID:               uuid.New().String(),
+		ProcessDetailID:  strings.TrimSpace(processDetailID),
+		MessageID:        strings.TrimSpace(messageID),
+		ConversationID:   strings.TrimSpace(conversationID),
+		Source:           strings.TrimSpace(fmt.Sprint(m["source"])),
+		Orchestration:    strings.TrimSpace(fmt.Sprint(m["orchestration"])),
+		Reason:           strings.TrimSpace(fmt.Sprint(m["reason"])),
+		Model:            strings.TrimSpace(fmt.Sprint(m["model"])),
+		ModelCalls:       usageInt64(m["modelCalls"]),
+		PromptTokens:     usageInt64(m["promptTokens"]),
+		CompletionTokens: usageInt64(m["completionTokens"]),
+		TotalTokens:      usageInt64(m["totalTokens"]),
+		CachedTokens:     usageInt64(m["cachedTokens"]),
+		ReasoningTokens:  usageInt64(m["reasoningTokens"]),
+	}
+	if usage.TotalTokens == 0 && (usage.PromptTokens > 0 || usage.CompletionTokens > 0) {
+		usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	}
+	if usage.ProcessDetailID == "" || usage.MessageID == "" || usage.ConversationID == "" {
+		return ModelTokenUsage{}, false
+	}
+	if usage.ModelCalls == 0 && usage.TotalTokens == 0 && usage.PromptTokens == 0 && usage.CompletionTokens == 0 && usage.CachedTokens == 0 && usage.ReasoningTokens == 0 {
+		return ModelTokenUsage{}, false
+	}
+	return usage, true
+}
+
+func mapFromUsageData(data interface{}) map[string]interface{} {
+	switch v := data.(type) {
+	case nil:
+		return nil
+	case map[string]interface{}:
+		return v
+	case string:
+		var m map[string]interface{}
+		if err := json.Unmarshal([]byte(v), &m); err == nil {
+			return m
+		}
+	case []byte:
+		var m map[string]interface{}
+		if err := json.Unmarshal(v, &m); err == nil {
+			return m
+		}
+	default:
+		raw, err := json.Marshal(v)
+		if err == nil {
+			var m map[string]interface{}
+			if err := json.Unmarshal(raw, &m); err == nil {
+				return m
+			}
+		}
+	}
+	return nil
+}
+
+func usageInt64(v interface{}) int64 {
+	switch n := v.(type) {
+	case int:
+		return int64(n)
+	case int8:
+		return int64(n)
+	case int16:
+		return int64(n)
+	case int32:
+		return int64(n)
+	case int64:
+		return n
+	case uint:
+		return int64(n)
+	case uint8:
+		return int64(n)
+	case uint16:
+		return int64(n)
+	case uint32:
+		return int64(n)
+	case uint64:
+		if n > math.MaxInt64 {
+			return math.MaxInt64
+		}
+		return int64(n)
+	case float32:
+		return int64(n)
+	case float64:
+		return int64(n)
+	case json.Number:
+		i, _ := n.Int64()
+		return i
+	case string:
+		i, _ := strconv.ParseInt(strings.TrimSpace(n), 10, 64)
+		return i
+	default:
+		i, _ := strconv.ParseInt(strings.TrimSpace(fmt.Sprint(v)), 10, 64)
+		return i
+	}
+}
+
+func (db *DB) maybeRecordModelTokenUsage(messageID, conversationID, processDetailID, eventType string, data interface{}) {
+	if db == nil || eventType != modelTokenUsageEventType {
+		return
+	}
+	usage, ok := modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID, data)
+	if !ok {
+		return
+	}
+	if err := db.UpsertModelTokenUsage(usage); err != nil && db.logger != nil {
+		db.logger.Warn("保存模型Token用量失败",
+			zap.String("processDetailId", processDetailID),
+			zap.String("conversationId", conversationID),
+			zap.Error(err))
+	}
+}
+
+// UpsertModelTokenUsage persists usage with process_detail_id idempotency.
+func (db *DB) UpsertModelTokenUsage(usage ModelTokenUsage) error {
+	if db == nil {
+		return fmt.Errorf("database is nil")
+	}
+	now := time.Now()
+	createdAt := usage.CreatedAt
+	if createdAt.IsZero() {
+		createdAt = now
+	}
+	if usage.ID == "" {
+		usage.ID = uuid.New().String()
+	}
+	var projectID sql.NullString
+	if err := db.QueryRow(`SELECT project_id FROM conversations WHERE id = ?`, usage.ConversationID).Scan(&projectID); err != nil && err != sql.ErrNoRows {
+		return fmt.Errorf("查询对话项目失败: %w", err)
+	}
+	projectValue := interface{}(nil)
+	if projectID.Valid && strings.TrimSpace(projectID.String) != "" {
+		projectValue = strings.TrimSpace(projectID.String)
+	}
+	_, err := db.Exec(`
+INSERT INTO model_token_usage (
+	id, process_detail_id, message_id, conversation_id, project_id,
+	source, orchestration, reason, model, model_calls,
+	prompt_tokens, completion_tokens, total_tokens, cached_tokens, reasoning_tokens,
+	created_at, updated_at
+) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+ON CONFLICT(process_detail_id) DO UPDATE SET
+	message_id = excluded.message_id,
+	conversation_id = excluded.conversation_id,
+	project_id = excluded.project_id,
+	source = excluded.source,
+	orchestration = excluded.orchestration,
+	reason = excluded.reason,
+	model = excluded.model,
+	model_calls = excluded.model_calls,
+	prompt_tokens = excluded.prompt_tokens,
+	completion_tokens = excluded.completion_tokens,
+	total_tokens = excluded.total_tokens,
+	cached_tokens = excluded.cached_tokens,
+	reasoning_tokens = excluded.reasoning_tokens,
+	created_at = excluded.created_at,
+	updated_at = excluded.updated_at`,
+		usage.ID, usage.ProcessDetailID, usage.MessageID, usage.ConversationID, projectValue,
+		usage.Source, usage.Orchestration, usage.Reason, usage.Model, usage.ModelCalls,
+		usage.PromptTokens, usage.CompletionTokens, usage.TotalTokens, usage.CachedTokens, usage.ReasoningTokens,
+		createdAt, now,
+	)
+	if err != nil {
+		return fmt.Errorf("写入模型Token用量失败: %w", err)
+	}
+	return nil
+}
+
+// BackfillModelTokenUsageFromProcessDetails makes existing timeline usage events queryable.
+func (db *DB) BackfillModelTokenUsageFromProcessDetails() error {
+	if db == nil {
+		return nil
+	}
+	rows, err := db.Query(`
+SELECT pd.id, pd.message_id, pd.conversation_id, pd.data, pd.created_at
+FROM process_details pd
+LEFT JOIN model_token_usage mtu ON mtu.process_detail_id = pd.id
+WHERE pd.event_type = ?
+	AND (mtu.id IS NULL OR mtu.created_at != pd.created_at)`, modelTokenUsageEventType)
+	if err != nil {
+		return fmt.Errorf("查询历史模型Token用量失败: %w", err)
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var processDetailID, messageID, conversationID string
+		var data sql.NullString
+		var createdAt string
+		if err := rows.Scan(&processDetailID, &messageID, &conversationID, &data, &createdAt); err != nil {
+			return fmt.Errorf("扫描历史模型Token用量失败: %w", err)
+		}
+		if !data.Valid {
+			continue
+		}
+		usage, ok := modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID, data.String)
+		if !ok {
+			continue
+		}
+		usage.CreatedAt = parseModelTokenUsageTime(createdAt)
+		if err := db.UpsertModelTokenUsage(usage); err != nil {
+			return err
+		}
+	}
+	if err := rows.Err(); err != nil {
+		return fmt.Errorf("遍历历史模型Token用量失败: %w", err)
+	}
+	return nil
+}
+
+func (db *DB) GetModelTokenUsageStats(filter ModelTokenUsageFilter) (*ModelTokenUsageStats, error) {
+	if db == nil {
+		return nil, fmt.Errorf("database is nil")
+	}
+	if filter.Days <= 0 {
+		filter.Days = 7
+	}
+	if filter.Limit <= 0 {
+		filter.Limit = 10
+	}
+	where, args := buildModelTokenUsageWhere(filter, "mtu", "c")
+	summary, err := db.queryModelTokenUsageSummary("SELECT "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where, args...)
+	if err != nil {
+		return nil, err
+	}
+	todayFilter := filter
+	now := time.Now()
+	todayFilter.Since = time.Date(now.Year(), now.Month(), now.Day(), 0, 0, 0, 0, now.Location())
+	todayWhere, todayArgs := buildModelTokenUsageWhere(todayFilter, "mtu", "c")
+	today, err := db.queryModelTokenUsageSummary("SELECT "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+todayWhere, todayArgs...)
+	if err != nil {
+		return nil, err
+	}
+	byDay, err := db.queryModelTokenUsageBreakdown(
+		"SELECT date(mtu.created_at) AS k, date(mtu.created_at) AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY date(mtu.created_at) ORDER BY k DESC LIMIT ?",
+		append(args, filter.Days)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	byModel, err := db.queryModelTokenUsageBreakdown(
+		"SELECT COALESCE(NULLIF(TRIM(mtu.model), ''), 'unknown') AS k, COALESCE(NULLIF(TRIM(mtu.model), ''), 'Unknown') AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY k ORDER BY SUM(mtu.total_tokens) DESC LIMIT ?",
+		append(args, filter.Limit)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	byOrch, err := db.queryModelTokenUsageBreakdown(
+		"SELECT COALESCE(NULLIF(TRIM(mtu.orchestration), ''), 'unknown') AS k, COALESCE(NULLIF(TRIM(mtu.orchestration), ''), 'Unknown') AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY k ORDER BY SUM(mtu.total_tokens) DESC LIMIT ?",
+		append(args, filter.Limit)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	recent, err := db.ListModelTokenUsage(filter)
+	if err != nil {
+		return nil, err
+	}
+	return &ModelTokenUsageStats{
+		Summary:         summary,
+		Today:           today,
+		ByDay:           byDay,
+		ByModel:         byModel,
+		ByOrchestration: byOrch,
+		Recent:          recent,
+	}, nil
+}
+
+func modelTokenUsageSummarySelect(alias string) string {
+	p := ""
+	if alias != "" {
+		p = alias + "."
+	}
+	return fmt.Sprintf(`COUNT(%sid),
+COALESCE(SUM(%smodel_calls), 0),
+COALESCE(SUM(%sprompt_tokens), 0),
+COALESCE(SUM(%scompletion_tokens), 0),
+COALESCE(SUM(%stotal_tokens), 0),
+COALESCE(SUM(%scached_tokens), 0),
+COALESCE(SUM(%sreasoning_tokens), 0)`, p, p, p, p, p, p, p)
+}
+
+func buildModelTokenUsageWhere(filter ModelTokenUsageFilter, usageAlias, convAlias string) (string, []interface{}) {
+	where := " WHERE 1=1"
+	args := []interface{}{}
+	uPrefix := ""
+	if usageAlias != "" {
+		uPrefix = usageAlias + "."
+	}
+	if cid := strings.TrimSpace(filter.ConversationID); cid != "" {
+		where += " AND " + uPrefix + "conversation_id = ?"
+		args = append(args, cid)
+	}
+	where, args = appendConversationProjectFilter(where, args, filter.ProjectID, usageAlias)
+	if !filter.Since.IsZero() {
+		where += " AND " + uPrefix + "created_at >= ?"
+		args = append(args, filter.Since)
+	}
+	if !filter.Until.IsZero() {
+		where += " AND " + uPrefix + "created_at <= ?"
+		args = append(args, filter.Until)
+	}
+	where, args = appendConversationAccessFilter(where, args, filter.Access.UserID, filter.Access.Scope, convAlias)
+	return where, args
+}
+
+func (db *DB) queryModelTokenUsageSummary(query string, args ...interface{}) (ModelTokenUsageSummary, error) {
+	var s ModelTokenUsageSummary
+	err := db.QueryRow(query, args...).Scan(
+		&s.Events, &s.ModelCalls, &s.PromptTokens, &s.CompletionTokens,
+		&s.TotalTokens, &s.CachedTokens, &s.ReasoningTokens,
+	)
+	if err != nil {
+		return s, fmt.Errorf("查询模型Token用量汇总失败: %w", err)
+	}
+	return s, nil
+}
+
+func (db *DB) queryModelTokenUsageBreakdown(query string, args ...interface{}) ([]ModelTokenUsageBreakdown, error) {
+	rows, err := db.Query(query, args...)
+	if err != nil {
+		return nil, fmt.Errorf("查询模型Token用量分组失败: %w", err)
+	}
+	defer rows.Close()
+	out := []ModelTokenUsageBreakdown{}
+	for rows.Next() {
+		var row ModelTokenUsageBreakdown
+		if err := rows.Scan(
+			&row.Key, &row.Label, &row.Events, &row.ModelCalls, &row.PromptTokens,
+			&row.CompletionTokens, &row.TotalTokens, &row.CachedTokens, &row.ReasoningTokens,
+		); err != nil {
+			return nil, fmt.Errorf("扫描模型Token用量分组失败: %w", err)
+		}
+		out = append(out, row)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("遍历模型Token用量分组失败: %w", err)
+	}
+	return out, nil
+}
+
+func (db *DB) ListModelTokenUsage(filter ModelTokenUsageFilter) ([]ModelTokenUsage, error) {
+	if filter.Limit <= 0 {
+		filter.Limit = 20
+	}
+	if filter.Limit > 500 {
+		filter.Limit = 500
+	}
+	where, args := buildModelTokenUsageWhere(filter, "mtu", "c")
+	args = append(args, filter.Limit)
+	rows, err := db.Query(`
+SELECT mtu.id, mtu.process_detail_id, mtu.message_id, mtu.conversation_id,
+	COALESCE(mtu.project_id, ''), mtu.source, mtu.orchestration, mtu.reason, mtu.model,
+	mtu.model_calls, mtu.prompt_tokens, mtu.completion_tokens, mtu.total_tokens,
+	mtu.cached_tokens, mtu.reasoning_tokens, mtu.created_at, mtu.updated_at
+FROM model_token_usage mtu
+JOIN conversations c ON c.id = mtu.conversation_id`+where+`
+ORDER BY mtu.created_at DESC, mtu.rowid DESC
+LIMIT ?`, args...)
+	if err != nil {
+		return nil, fmt.Errorf("查询模型Token用量明细失败: %w", err)
+	}
+	defer rows.Close()
+	out := []ModelTokenUsage{}
+	for rows.Next() {
+		var u ModelTokenUsage
+		var createdAt, updatedAt string
+		if err := rows.Scan(
+			&u.ID, &u.ProcessDetailID, &u.MessageID, &u.ConversationID, &u.ProjectID,
+			&u.Source, &u.Orchestration, &u.Reason, &u.Model, &u.ModelCalls,
+			&u.PromptTokens, &u.CompletionTokens, &u.TotalTokens, &u.CachedTokens,
+			&u.ReasoningTokens, &createdAt, &updatedAt,
+		); err != nil {
+			return nil, fmt.Errorf("扫描模型Token用量明细失败: %w", err)
+		}
+		u.CreatedAt = parseModelTokenUsageTime(createdAt)
+		u.UpdatedAt = parseModelTokenUsageTime(updatedAt)
+		out = append(out, u)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("遍历模型Token用量明细失败: %w", err)
+	}
+	return out, nil
+}
+
+func parseModelTokenUsageTime(s string) time.Time {
+	for _, layout := range []string{
+		"2006-01-02 15:04:05.999999999-07:00",
+		"2006-01-02 15:04:05.999999-07:00",
+		"2006-01-02 15:04:05",
+		time.RFC3339Nano,
+		time.RFC3339,
+	} {
+		if t, err := time.Parse(layout, strings.TrimSpace(s)); err == nil {
+			return t
+		}
+	}
+	return time.Time{}
+}

internal/database/model_token_usage_test.go

@@ -0,0 +1,85 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+func TestModelTokenUsagePersistsFromUsageProcessDetail(t *testing.T) {
+	db := newModelTokenUsageTestDB(t)
+	conv, err := db.CreateConversation("usage", ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	msg, err := db.AddMessage(conv.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	if err := db.AddProcessDetail(msg.ID, conv.ID, modelTokenUsageEventType, "usage", map[string]interface{}{
+		"source":           "eino",
+		"orchestration":    "deep",
+		"reason":           "final",
+		"model":            "gpt-test",
+		"modelCalls":       2,
+		"promptTokens":     10,
+		"completionTokens": 3,
+		"totalTokens":      13,
+		"cachedTokens":     4,
+		"reasoningTokens":  1,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail: %v", err)
+	}
+
+	stats, err := db.GetModelTokenUsageStats(ModelTokenUsageFilter{})
+	if err != nil {
+		t.Fatalf("GetModelTokenUsageStats: %v", err)
+	}
+	if stats.Summary.Events != 1 || stats.Summary.ModelCalls != 2 || stats.Summary.TotalTokens != 13 || stats.Summary.CachedTokens != 4 || stats.Summary.ReasoningTokens != 1 {
+		t.Fatalf("summary = %#v", stats.Summary)
+	}
+	if len(stats.ByModel) != 1 || stats.ByModel[0].Key != "gpt-test" || stats.ByModel[0].TotalTokens != 13 {
+		t.Fatalf("by model = %#v", stats.ByModel)
+	}
+}
+
+func TestModelTokenUsageBackfillIsIdempotent(t *testing.T) {
+	db := newModelTokenUsageTestDB(t)
+	conv, err := db.CreateConversation("usage", ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	msg, err := db.AddMessage(conv.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	if err := db.AddProcessDetail(msg.ID, conv.ID, modelTokenUsageEventType, "usage", map[string]interface{}{
+		"source": "eino", "modelCalls": 1, "promptTokens": 7, "completionTokens": 5, "totalTokens": 12,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail: %v", err)
+	}
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		t.Fatalf("Backfill 1: %v", err)
+	}
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		t.Fatalf("Backfill 2: %v", err)
+	}
+	stats, err := db.GetModelTokenUsageStats(ModelTokenUsageFilter{})
+	if err != nil {
+		t.Fatalf("GetModelTokenUsageStats: %v", err)
+	}
+	if stats.Summary.Events != 1 || stats.Summary.TotalTokens != 12 {
+		t.Fatalf("summary after backfill = %#v", stats.Summary)
+	}
+}
+
+func newModelTokenUsageTestDB(t *testing.T) *DB {
+	t.Helper()
+	db, err := NewDB(filepath.Join(t.TempDir(), "usage.db"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	return db
+}

internal/database/process_details_summary_test.go

@@ -8,7 +8,7 @@ import (
 	"go.uber.org/zap"
 )
 
-func TestProcessDetailsSummaryDoesNotGuessIDLessResultsByOrder(t *testing.T) {
+func TestProcessDetailsSummaryDoesNotGuessIDLessResultsOntoDifferentTool(t *testing.T) {
 	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
 	for _, id := range []string{"call-1", "call-2", "call-3", "call-4"} {
 		if err := db.AddProcessDetail(messageID, conversationID, "tool_call", "call", map[string]interface{}{
@@ -20,8 +20,8 @@ func TestProcessDetailsSummaryDoesNotGuessIDLessResultsByOrder(t *testing.T) {
 	results := []map[string]interface{}{
 		{"toolName": "http-framework-test", "toolCallId": "call-1", "success": true},
 		{"toolName": "http-framework-test", "toolCallId": "call-2", "success": true},
-		{"toolName": "http-framework-test", "success": true},
-		{"toolName": "http-framework-test", "success": true},
+		{"toolName": "other-tool", "success": true},
+		{"toolName": "other-tool", "success": true},
 	}
 	var resultIDs []string
 	for _, result := range results {
@@ -53,12 +53,71 @@ func TestProcessDetailsSummaryDoesNotGuessIDLessResultsByOrder(t *testing.T) {
 		}
 	}
 	for i, execution := range summary.ToolExecutions[4:] {
-		if execution.Status != "completed" || execution.ToolCallID != "" {
+		if execution.Status != "completed" || execution.ToolCallID != "" || execution.ToolName != "other-tool" {
 			t.Fatalf("idless result %d = %#v, want separate completed result without toolCallId", i, execution)
 		}
 	}
 }
 
+func TestProcessDetailsSummaryPairsIDLessResultsWithSameToolName(t *testing.T) {
+	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
+	for i, id := range []string{"call-1", "call-2"} {
+		if err := db.AddProcessDetail(messageID, conversationID, "tool_call", "call", map[string]interface{}{
+			"toolName": "nmap", "toolCallId": id, "index": i + 1, "total": 2,
+		}); err != nil {
+			t.Fatalf("AddProcessDetail(tool_call): %v", err)
+		}
+	}
+	var resultIDs []string
+	for i := 0; i < 2; i++ {
+		resultID, err := db.AddProcessDetailWithID(messageID, conversationID, "tool_result", "result", map[string]interface{}{
+			"toolName": "nmap", "success": true,
+		})
+		if err != nil {
+			t.Fatalf("AddProcessDetail(tool_result): %v", err)
+		}
+		resultIDs = append(resultIDs, resultID)
+	}
+
+	summary, err := db.GetProcessDetailsSummary(messageID)
+	if err != nil {
+		t.Fatalf("GetProcessDetailsSummary: %v", err)
+	}
+	if len(summary.ToolExecutions) != 2 {
+		t.Fatalf("tool executions = %d, want 2", len(summary.ToolExecutions))
+	}
+	for i, execution := range summary.ToolExecutions {
+		if execution.Status != "completed" {
+			t.Fatalf("execution %d status = %q, want completed", i, execution.Status)
+		}
+		if execution.ResultDetailID != resultIDs[i] {
+			t.Fatalf("execution %d result detail id = %q, want %q", i, execution.ResultDetailID, resultIDs[i])
+		}
+	}
+}
+
+func TestProcessDetailsSummaryPairedResultWithoutSuccessIsCompleted(t *testing.T) {
+	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
+	if err := db.AddProcessDetail(messageID, conversationID, "tool_call", "call", map[string]interface{}{
+		"toolName": "nmap", "toolCallId": "call-1",
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+	if err := db.AddProcessDetail(messageID, conversationID, "tool_result", "result", map[string]interface{}{
+		"toolName": "nmap", "toolCallId": "call-1", "resultPreview": "open 22",
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_result): %v", err)
+	}
+
+	summary, err := db.GetProcessDetailsSummary(messageID)
+	if err != nil {
+		t.Fatalf("GetProcessDetailsSummary: %v", err)
+	}
+	if len(summary.ToolExecutions) != 1 || summary.ToolExecutions[0].Status != "completed" {
+		t.Fatalf("tool executions = %#v, want completed", summary.ToolExecutions)
+	}
+}
+
 func TestProcessDetailsSummaryPairsRepeatedToolCallIDsFIFO(t *testing.T) {
 	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
 	for i := 0; i < 2; i++ {
@@ -106,6 +165,32 @@ func TestProcessDetailsSummaryDoesNotReportPersistedOrphanAsRunning(t *testing.T
 	}
 }
 
+func TestProcessDetailsSummaryReportsUnmatchedToolCallAsRunningForActiveTurn(t *testing.T) {
+	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
+	if _, err := db.Exec(
+		"UPDATE messages SET content = ?, updated_at = ? WHERE id = ?",
+		"处理中...", "2026-08-10T08:00:00Z", messageID,
+	); err != nil {
+		t.Fatalf("update running message: %v", err)
+	}
+	if err := db.AddProcessDetail(messageID, conversationID, "tool_call", "call", map[string]interface{}{
+		"toolName": "execute", "toolCallId": "pending",
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+
+	summary, err := db.GetProcessDetailsSummary(messageID)
+	if err != nil {
+		t.Fatalf("GetProcessDetailsSummary: %v", err)
+	}
+	if summary.Status != "running" {
+		t.Fatalf("summary status = %q, want running", summary.Status)
+	}
+	if len(summary.ToolExecutions) != 1 || summary.ToolExecutions[0].Status != "running" {
+		t.Fatalf("tool executions = %#v, want running", summary.ToolExecutions)
+	}
+}
+
 func TestProcessDetailsSummaryIncludesPersistedTurnTiming(t *testing.T) {
 	db, _, messageID := setupProcessDetailsSummaryTest(t)
 	startedAt := "2026-08-10T08:00:00Z"

internal/database/tool_execution_args_lookup.go

@@ -20,16 +20,20 @@ func (db *DB) FindNearestToolExecutionArguments(conversationID, toolName string,
 	if window <= 0 {
 		window = 5 * time.Second
 	}
+	names := []string{toolName}
+	if !strings.Contains(toolName, "::") {
+		names = append(names, "eino_fs::"+toolName)
+	}
 	start := at.Add(-window)
 	end := at.Add(window)
 	rows, err := db.Query(`
 SELECT id, arguments
 FROM tool_executions
 WHERE conversation_id = ?
-  AND tool_name = ?
+  AND tool_name IN (?, ?)
   AND julianday(start_time) BETWEEN julianday(?) AND julianday(?)
 ORDER BY ABS(julianday(start_time) - julianday(?)) ASC, start_time ASC
-LIMIT 1`, conversationID, toolName, start, end, at)
+LIMIT 1`, conversationID, names[0], names[len(names)-1], start, end, at)
 	if err != nil {
 		return "", nil, err
 	}

internal/handler/agent.go

@@ -315,12 +315,13 @@ func (h *AgentHandler) SetHitlToolWhitelistSaver(s HitlToolWhitelistSaver) {
 	h.hitlWhitelistSaver = s
 }
 
-// HitlDefaultReviewerSaver 持久化全局默认审批方到 config.yaml。
+// HitlDefaultReviewerSaver 持久化全局默认人机协同配置到 config.yaml。
 type HitlDefaultReviewerSaver interface {
+	UpdateHitlDefaultConfig(mode, reviewer string, timeoutSeconds int) error
 	UpdateHitlDefaultReviewer(reviewer string) error
 }
 
-// SetHitlDefaultReviewerSaver 设置 HITL 默认审批方落盘。
+// SetHitlDefaultReviewerSaver 设置 HITL 默认配置落盘。
 func (h *AgentHandler) SetHitlDefaultReviewerSaver(s HitlDefaultReviewerSaver) {
 	h.hitlDefaultReviewerSaver = s
 }
@@ -332,6 +333,35 @@ func (h *AgentHandler) hitlEffectiveDefaultReviewer() string {
 	return "human"
 }
 
+func (h *AgentHandler) hitlEffectiveDefaultMode() string {
+	if h != nil && h.config != nil {
+		return normalizeHitlDefaultMode(h.config.Hitl.EffectiveDefaultMode())
+	}
+	return "off"
+}
+
+func (h *AgentHandler) hitlEffectiveDefaultTimeoutSeconds() int {
+	if h != nil && h.config != nil {
+		timeout := h.config.Hitl.EffectiveDefaultTimeoutSeconds()
+		if timeout < 0 {
+			return 0
+		}
+		return timeout
+	}
+	return 300
+}
+
+func (h *AgentHandler) hitlEffectiveDefaultRequest() *HITLRequest {
+	mode := h.hitlEffectiveDefaultMode()
+	return &HITLRequest{
+		Enabled:        mode != "off",
+		Mode:           mode,
+		Reviewer:       h.hitlEffectiveDefaultReviewer(),
+		SensitiveTools: []string{},
+		TimeoutSeconds: h.hitlEffectiveDefaultTimeoutSeconds(),
+	}
+}
+
 // HITLNeedsToolApproval 供 C2 危险任务门控：与会话侧人机协同及免审批白名单判定一致。
 func (h *AgentHandler) HITLNeedsToolApproval(conversationID, toolName string) bool {
 	if h == nil || h.hitlManager == nil {
@@ -698,18 +728,19 @@ func (h *AgentHandler) mergeAssistantMessagePartialOnCancel(messageID, partial s
 
 // ChatResponse 聊天响应
 type ChatResponse struct {
-	Response            string    `json:"response"`
-	MCPExecutionIDs     []string  `json:"mcpExecutionIds,omitempty"` // 本次对话中执行的MCP调用ID列表
-	ConversationID      string    `json:"conversationId"`            // 对话ID
-	Time                time.Time `json:"time"`
-	Finalizable         bool      `json:"finalizable"`
-	Finalized           bool      `json:"finalized"`
-	Status              string    `json:"status,omitempty"`
-	CompletionReason    string    `json:"completionReason,omitempty"`
-	EvidenceVerified    bool      `json:"evidenceVerified"`
-	EvidenceRefs        []string  `json:"evidenceRefs,omitempty"`
-	PendingExecutionIDs []string  `json:"pendingExecutionIds,omitempty"`
-	MissingChecks       []string  `json:"missingChecks,omitempty"`
+	Response                         string    `json:"response"`
+	MCPExecutionIDs                  []string  `json:"mcpExecutionIds,omitempty"` // 本次对话中执行的MCP调用ID列表
+	ConversationID                   string    `json:"conversationId"`            // 对话ID
+	Time                             time.Time `json:"time"`
+	Finalizable                      bool      `json:"finalizable"`
+	Finalized                        bool      `json:"finalized"`
+	Status                           string    `json:"status,omitempty"`
+	CompletionReason                 string    `json:"completionReason,omitempty"`
+	EvidenceVerified                 bool      `json:"evidenceVerified"`
+	EvidenceRefs                     []string  `json:"evidenceRefs,omitempty"`
+	PendingExecutionIDs              []string  `json:"pendingExecutionIds,omitempty"`
+	MissingChecks                    []string  `json:"missingChecks,omitempty"`
+	AutoCancelledPendingExecutionIDs []string  `json:"autoCancelledPendingExecutionIds,omitempty"`
 }
 
 func (h *AgentHandler) finalizeRobotAgentError(ctx context.Context, assistantMessageID, conversationID string, resultMA *multiagent.RunResult, errMA error) (string, string, error) {
@@ -724,8 +755,13 @@ func (h *AgentHandler) finalizeRobotAgentError(ctx context.Context, assistantMes
 	return "", conversationID, errMA
 }
 
-func (h *AgentHandler) finalizeRobotAgentSuccess(assistantMessageID, conversationID string, resultMA *multiagent.RunResult) (string, string, error) {
-	decision := h.finalizeAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(resultMA.LastAgentTraceInput), true)
+func (h *AgentHandler) finalizeRobotAgentSuccess(taskCtx context.Context, assistantMessageID, conversationID string, resultMA *multiagent.RunResult) (string, string, error) {
+	reasoningContent := multiagent.AggregatedReasoningFromTraceJSON(resultMA.LastAgentTraceInput)
+	decision := h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, true)
+	if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, true)
+	}
+	h.persistFinalizationDecision(conversationID, assistantMessageID, "robot", resultMA.MCPExecutionIDs, reasoningContent, decision)
 	responseText := decision.FinalText
 	if !decision.Finalizable {
 		responseText = finalizationBlockedMessage(decision)
@@ -758,7 +794,7 @@ func (h *AgentHandler) runRobotEinoSingleWithRetry(
 		*taskStatus = "failed"
 		return h.finalizeRobotAgentError(taskCtx, assistantMessageID, conversationID, resultMA, errMA)
 	}
-	return h.finalizeRobotAgentSuccess(assistantMessageID, conversationID, resultMA)
+	return h.finalizeRobotAgentSuccess(taskCtx, assistantMessageID, conversationID, resultMA)
 }
 
 func (h *AgentHandler) runRobotMultiAgentWithRetry(
@@ -779,7 +815,7 @@ func (h *AgentHandler) runRobotMultiAgentWithRetry(
 		*taskStatus = "failed"
 		return h.finalizeRobotAgentError(taskCtx, assistantMessageID, conversationID, resultMA, errMA)
 	}
-	return h.finalizeRobotAgentSuccess(assistantMessageID, conversationID, resultMA)
+	return h.finalizeRobotAgentSuccess(taskCtx, assistantMessageID, conversationID, resultMA)
 }
 
 // ProcessMessageForRobot 供机器人（企业微信/钉钉/飞书）调用：Eino 单/多代理执行路径（含 progressCallback、过程详情），仅不发送 SSE，最后返回完整回复

internal/handler/batch_queue_executor.go

@@ -281,7 +281,12 @@ func (h *AgentHandler) executeOneBatchSubTask(queueID string, queue *BatchTaskQu
 	if useBatchMulti {
 		agentMode = "batch_eino_" + batchOrch
 	}
-	decision := h.finalizeAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, reasoningContent, true)
+	decision := h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, true)
+	autoCancelledPendingExecutionIDs := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback)
+	if len(autoCancelledPendingExecutionIDs) > 0 {
+		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, true)
+	}
+	h.persistFinalizationDecision(conversationID, assistantMessageID, agentMode, mcpIDs, reasoningContent, decision)
 	resText := decision.FinalText
 	if !decision.Finalizable {
 		resText = finalizationBlockedMessage(decision)
@@ -289,14 +294,15 @@ func (h *AgentHandler) executeOneBatchSubTask(queueID string, queue *BatchTaskQu
 		sendEvent("finalization_check", resText, decision)
 	}
 	sendEvent("response", resText, finalizationResponsePayload(decision, map[string]interface{}{
-		"conversationId":   conversationID,
-		"messageId":        assistantMessageID,
-		"agentMode":        agentMode,
-		"mcpExecutionIds":  mcpIDs,
-		"batchQueueId":     queueID,
-		"batchTaskId":      task.ID,
-		"batchTaskStatus":  map[bool]string{true: string(BatchTaskStatusCompleted), false: string(BatchTaskStatusFailed)}[decision.Finalizable],
-		"candidatePreview": safeTruncateString(resultMA.Response, 500),
+		"conversationId":                   conversationID,
+		"messageId":                        assistantMessageID,
+		"agentMode":                        agentMode,
+		"mcpExecutionIds":                  mcpIDs,
+		"batchQueueId":                     queueID,
+		"batchTaskId":                      task.ID,
+		"batchTaskStatus":                  map[bool]string{true: string(BatchTaskStatusCompleted), false: string(BatchTaskStatusFailed)}[decision.Finalizable],
+		"candidatePreview":                 safeTruncateString(resultMA.Response, 500),
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 
 	if assistantMessageID == "" {

internal/handler/config.go

@@ -18,11 +18,13 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/database"
 	"cyberstrike-ai/internal/knowledge"
+	"cyberstrike-ai/internal/llm"
 	"cyberstrike-ai/internal/mcp"
 	"cyberstrike-ai/internal/mcp/builtin"
 	"cyberstrike-ai/internal/openai"
 	"cyberstrike-ai/internal/security"
 
+	"github.com/cloudwego/eino/schema"
 	"github.com/gin-gonic/gin"
 	"go.uber.org/zap"
 	"gopkg.in/yaml.v3"
@@ -889,7 +891,14 @@ func (h *ConfigHandler) UpdateConfig(c *gin.Context) {
 	if req.Hitl != nil {
 		h.config.Hitl.AuditModel = req.Hitl.AuditModel
 		h.config.Hitl.ToolWhitelist = mergeHitlToolWhitelistSlice(nil, req.Hitl.ToolWhitelist)
+		if strings.TrimSpace(req.Hitl.DefaultMode) != "" {
+			h.config.Hitl.DefaultMode = req.Hitl.EffectiveDefaultMode()
+		}
 		h.config.Hitl.DefaultReviewer = req.Hitl.EffectiveDefaultReviewer()
+		if req.Hitl.DefaultTimeoutSeconds != nil {
+			v := req.Hitl.EffectiveDefaultTimeoutSeconds()
+			h.config.Hitl.DefaultTimeoutSeconds = &v
+		}
 		h.config.Hitl.AuditAgentPrompt = strings.TrimSpace(req.Hitl.AuditAgentPrompt)
 		h.config.Hitl.AuditAgentPromptReviewEdit = strings.TrimSpace(req.Hitl.AuditAgentPromptReviewEdit)
 		if req.Hitl.RetentionDays != nil {
@@ -1216,7 +1225,7 @@ func (h *ConfigHandler) TestOpenAI(c *gin.Context) {
 		"max_completion_tokens": 5,
 	}
 
-	// 使用内部 openai Client 进行测试，若 provider 为 claude 会自动走桥接层
+	// OpenAI-compatible 通道使用内部客户端；Claude 通道在下方直接使用 Eino agenticclaude。
 	tmpCfg := &config.OpenAIConfig{
 		Provider: req.Provider,
 		BaseURL:  baseURL,
@@ -1229,6 +1238,28 @@ func (h *ConfigHandler) TestOpenAI(c *gin.Context) {
 	defer cancel()
 
 	start := time.Now()
+	if llm.IsClaudeProvider(req.Provider) {
+		nativeModel, err := llm.NewClaudeAgenticModel(ctx, *tmpCfg, nil, 5, nil)
+		if err == nil {
+			_, err = nativeModel.Generate(ctx, []*schema.AgenticMessage{
+				schema.UserAgenticMessage("Hi"),
+			})
+		}
+		if err != nil {
+			c.JSON(http.StatusOK, gin.H{
+				"success": false,
+				"error":   "连接失败: " + err.Error(),
+			})
+			return
+		}
+		c.JSON(http.StatusOK, gin.H{
+			"success":    true,
+			"model":      tmpCfg.Model,
+			"latency_ms": time.Since(start).Milliseconds(),
+		})
+		return
+	}
+
 	var chatResp struct {
 		ID      string `json:"id"`
 		Object  string `json:"object"`
@@ -2117,12 +2148,35 @@ func updateHitlConfig(doc *yaml.Node, cfg config.HitlConfig) {
 	setStringInMap(auditModelNode, "model", cfg.AuditModel.Model)
 	// flow 样式 [a, b, c] 单行展示，工具多时比块序列省行数
 	setFlowStringSliceInMap(hitlNode, "tool_whitelist", cfg.ToolWhitelist)
+	setStringInMap(hitlNode, "default_mode", cfg.EffectiveDefaultMode())
 	setStringInMap(hitlNode, "default_reviewer", cfg.EffectiveDefaultReviewer())
+	setIntInMap(hitlNode, "default_timeout_seconds", cfg.EffectiveDefaultTimeoutSeconds())
 	setIntInMap(hitlNode, "retention_days", cfg.RetentionDaysEffective())
 	setStringInMap(hitlNode, "audit_agent_prompt", cfg.AuditAgentPrompt)
 	setStringInMap(hitlNode, "audit_agent_prompt_review_edit", cfg.AuditAgentPromptReviewEdit)
 }
 
+// UpdateHitlDefaultConfig 更新全局默认人机协同配置并写入 config.yaml。
+func (h *ConfigHandler) UpdateHitlDefaultConfig(mode, reviewer string, timeoutSeconds int) error {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.config.Hitl.DefaultMode = config.HitlConfig{DefaultMode: mode}.EffectiveDefaultMode()
+	h.config.Hitl.DefaultReviewer = config.HitlConfig{DefaultReviewer: reviewer}.EffectiveDefaultReviewer()
+	if timeoutSeconds < 0 {
+		timeoutSeconds = 0
+	}
+	h.config.Hitl.DefaultTimeoutSeconds = &timeoutSeconds
+	if err := h.saveConfig(); err != nil {
+		return err
+	}
+	h.logger.Info("HITL 全局默认配置已写入配置文件",
+		zap.String("default_mode", h.config.Hitl.DefaultMode),
+		zap.String("default_reviewer", h.config.Hitl.DefaultReviewer),
+		zap.Int("default_timeout_seconds", timeoutSeconds),
+	)
+	return nil
+}
+
 // UpdateHitlDefaultReviewer 更新全局默认审批方并写入 config.yaml。
 func (h *ConfigHandler) UpdateHitlDefaultReviewer(reviewer string) error {
 	h.mu.Lock()

internal/handler/conversation.go

@@ -497,6 +497,7 @@ func summarizeProcessDetailData(eventType string, data interface{}) interface{}
 	}
 	allow := map[string]bool{
 		"toolName": true, "toolCallId": true, "index": true, "total": true,
+		"arguments": true, "argumentsObj": true,
 		"success": true, "isError": true, "executionId": true,
 		"einoAgent": true, "einoRole": true, "einoScope": true, "orchestration": true,
 		"agentFacing": true,

internal/handler/conversation_process_details_test.go

@@ -76,6 +76,65 @@ func TestProcessDetailsPageIncludesTerminalToolStatusAcrossPageBoundary(t *testi
 	}
 }
 
+func TestProcessDetailsPageUsesPersistedExecutionStatusAfterBackgroundCancel(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	db, err := database.NewDB(filepath.Join(t.TempDir(), "process-details-cancelled.db"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	conversation, err := db.CreateConversation("cancelled background", database.ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	message, err := db.AddMessage(conversation.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	execID := "exec-cancelled-after-background"
+	if err := db.AddProcessDetail(message.ID, conversation.ID, "tool_call", "call", map[string]interface{}{
+		"toolName": "exec", "toolCallId": "call-cancelled", "index": 1, "total": 1,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+	if err := db.AddProcessDetail(message.ID, conversation.ID, "tool_result", "background", map[string]interface{}{
+		"toolName": "exec", "toolCallId": "call-cancelled", "executionId": execID, "status": "background_running", "success": true,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_result): %v", err)
+	}
+	now := time.Now()
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID:        execID,
+		ToolName:  "exec",
+		Status:    mcp.ToolExecutionStatusCancelled,
+		StartTime: now,
+		EndTime:   &now,
+	}); err != nil {
+		t.Fatalf("SaveToolExecution: %v", err)
+	}
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/api/messages/"+message.ID+"/process-details?limit=10&offset=0", nil)
+	c.Params = gin.Params{{Key: "id", Value: message.ID}}
+	NewConversationHandler(db, zap.NewNop()).GetMessageProcessDetails(c)
+	if w.Code != 200 {
+		t.Fatalf("status = %d: %s", w.Code, w.Body.String())
+	}
+	var response struct {
+		ToolExecutions []database.ProcessDetailsToolExecution `json:"toolExecutions"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(response.ToolExecutions) != 1 {
+		t.Fatalf("tool executions = %d, want 1", len(response.ToolExecutions))
+	}
+	if got := response.ToolExecutions[0].Status; got != mcp.ToolExecutionStatusCancelled {
+		t.Fatalf("tool execution status = %q, want cancelled", got)
+	}
+}
+
 func TestProcessDetailsFullBackfillsEmptyToolCallArgumentsFromExecution(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	db, err := database.NewDB(filepath.Join(t.TempDir(), "process-details-args.db"), zap.NewNop())
@@ -139,3 +198,67 @@ func TestProcessDetailsFullBackfillsEmptyToolCallArgumentsFromExecution(t *testi
 		t.Fatalf("executionId = %#v, want exec-whoami", data["executionId"])
 	}
 }
+
+func TestProcessDetailsPageBackfillsEinoFilesystemArgumentsFromPrefixedExecution(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	db, err := database.NewDB(filepath.Join(t.TempDir(), "process-details-eino-fs-args.db"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	conversation, err := db.CreateConversation("eino fs args", database.ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	message, err := db.AddMessage(conversation.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	if err := db.AddProcessDetail(message.ID, conversation.ID, "tool_call", "calling read_file", map[string]interface{}{
+		"toolName": "read_file", "toolCallId": "call-read", "arguments": "", "argumentsObj": nil,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID:             "exec-read",
+		ToolName:       "eino_fs::read_file",
+		Arguments:      map[string]interface{}{"file_path": "/tmp/requirements.txt", "limit": float64(2000)},
+		Status:         "completed",
+		StartTime:      time.Now(),
+		ConversationID: conversation.ID,
+	}); err != nil {
+		t.Fatalf("SaveToolExecution: %v", err)
+	}
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/api/messages/"+message.ID+"/process-details?limit=50&offset=0", nil)
+	c.Params = gin.Params{{Key: "id", Value: message.ID}}
+	NewConversationHandler(db, zap.NewNop()).GetMessageProcessDetails(c)
+	if w.Code != 200 {
+		t.Fatalf("status = %d: %s", w.Code, w.Body.String())
+	}
+	var response struct {
+		ProcessDetails []map[string]interface{} `json:"processDetails"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(response.ProcessDetails) != 1 {
+		t.Fatalf("process details = %d, want 1", len(response.ProcessDetails))
+	}
+	data, ok := response.ProcessDetails[0]["data"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("data = %#v", response.ProcessDetails[0]["data"])
+	}
+	args, ok := data["argumentsObj"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("argumentsObj = %#v", data["argumentsObj"])
+	}
+	if args["file_path"] != "/tmp/requirements.txt" {
+		t.Fatalf("file_path = %#v, want /tmp/requirements.txt", args["file_path"])
+	}
+	if data["arguments"] == nil {
+		t.Fatalf("arguments should be preserved in summarized page data: %#v", data)
+	}
+}

internal/handler/eino_single_agent.go

@@ -192,6 +192,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	var emptyResponseContinueAttempt int
 	var finalizationAutoContinueAttempt int
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 
 	for {
 		segmentMainIterationMax := 0
@@ -268,6 +269,10 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 				continue
 			}
 			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback); len(cancelled) > 0 {
+				autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+				decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			}
 			if h.tryAutoContinueAfterFinalization(taskCtx, conversationID, result, decision, &finalizationAutoContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
 				mainIterationOffset += segmentMainIterationMax
 				timeoutCancel()
@@ -384,6 +389,10 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 
 	if decision.CompletionReason == "" {
 		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 	}
 	h.persistFinalizationDecision(conversationID, assistantMessageID, "eino_single", cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput), decision)
 
@@ -401,10 +410,11 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
 	}
 	sendEvent("response", responseText, finalizationResponsePayload(decision, map[string]interface{}{
-		"mcpExecutionIds": cumulativeMCPExecutionIDs,
-		"conversationId":  conversationID,
-		"messageId":       assistantMessageID,
-		"agentMode":       "eino_single",
+		"mcpExecutionIds":                  cumulativeMCPExecutionIDs,
+		"conversationId":                   conversationID,
+		"messageId":                        assistantMessageID,
+		"agentMode":                        "eino_single",
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 	sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 }
@@ -464,6 +474,7 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 	var emptyResponseContinueAttempt int
 	var finalizationAutoContinueAttempt int
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 	for {
 		result, runErr = multiagent.RunEinoSingleChatModelAgent(
 			taskCtx,
@@ -493,6 +504,10 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 			continue
 		}
 		decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, "eino_single", result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, prep.ConversationID, decision, progressCallback); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, "eino_single", result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 		if h.tryAutoContinueAfterFinalization(taskCtx, prep.ConversationID, result, decision, &finalizationAutoContinueAttempt, &curHist, &curMsg, progressCallback) {
 			continue
 		}
@@ -509,18 +524,19 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 		responseText = finalizationBlockedMessage(decision)
 	}
 	c.JSON(http.StatusOK, gin.H{
-		"response":            responseText,
-		"conversationId":      prep.ConversationID,
-		"mcpExecutionIds":     result.MCPExecutionIDs,
-		"assistantMessageId":  prep.AssistantMessageID,
-		"agentMode":           "eino_single",
-		"finalized":           decision.Finalized,
-		"finalizable":         decision.Finalizable,
-		"status":              decision.Status,
-		"completionReason":    decision.CompletionReason,
-		"evidenceVerified":    decision.EvidenceVerified,
-		"evidenceRefs":        decision.EvidenceRefs,
-		"pendingExecutionIds": decision.PendingExecutionIDs,
-		"missingChecks":       decision.MissingChecks,
+		"response":                         responseText,
+		"conversationId":                   prep.ConversationID,
+		"mcpExecutionIds":                  result.MCPExecutionIDs,
+		"assistantMessageId":               prep.AssistantMessageID,
+		"agentMode":                        "eino_single",
+		"finalized":                        decision.Finalized,
+		"finalizable":                      decision.Finalizable,
+		"status":                           decision.Status,
+		"completionReason":                 decision.CompletionReason,
+		"evidenceVerified":                 decision.EvidenceVerified,
+		"evidenceRefs":                     decision.EvidenceRefs,
+		"pendingExecutionIds":              decision.PendingExecutionIDs,
+		"missingChecks":                    decision.MissingChecks,
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	})
 }

internal/handler/finalization_auto_continue.go

@@ -2,16 +2,21 @@ package handler
 
 import (
 	"context"
+	"strings"
 	"time"
 
 	"cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/agentfinalizer"
+	"cyberstrike-ai/internal/mcp"
 	"cyberstrike-ai/internal/multiagent"
 
 	"go.uber.org/zap"
 )
 
 const finalizationAutoContinueMaxAttempts = 2
+const finalizationPendingToolCancelWait = 2 * time.Second
+const finalizationPendingToolCancelPoll = 50 * time.Millisecond
+const finalizationPendingToolCancelNote = "Agent 迭代已结束，最终回复前自动终止未完成的工具执行"
 
 func shouldAutoContinueAfterFinalization(d agentfinalizer.Decision, attempt int) bool {
 	if d.Finalizable || d.Finalized {
@@ -75,3 +80,105 @@ func finalizationAutoContinueBackoff(attempt int) time.Duration {
 	}
 	return time.Duration(attempt) * time.Second
 }
+
+func (h *AgentHandler) cleanupPendingToolExecutionsAfterIteration(
+	taskCtx context.Context,
+	conversationID string,
+	decision agentfinalizer.Decision,
+	progressCallback func(eventType, message string, data interface{}),
+) []string {
+	if h == nil || h.agent == nil || decision.CompletionReason != agentfinalizer.ReasonPendingTools {
+		return nil
+	}
+	pending := uniqueNonEmptyStrings(decision.PendingExecutionIDs)
+	if len(pending) == 0 {
+		return nil
+	}
+	cancelled := make([]string, 0, len(pending))
+	for _, executionID := range pending {
+		if h.agent.CancelMCPToolExecutionWithNote(executionID, finalizationPendingToolCancelNote) {
+			cancelled = append(cancelled, executionID)
+		} else if h.logger != nil {
+			h.logger.Warn("finalization pending tool cleanup could not cancel execution",
+				zap.String("conversationId", conversationID),
+				zap.String("executionId", executionID))
+		}
+	}
+	if len(cancelled) == 0 {
+		return nil
+	}
+	if progressCallback != nil {
+		progressCallback("finalization_pending_tools_cancelled", "迭代结束，已自动终止仍在运行的工具执行。", map[string]interface{}{
+			"conversationId":                   conversationID,
+			"source":                           "finalizer",
+			"autoCancelledPendingExecutionIds": cancelled,
+			"pendingExecutionIds":              pending,
+			"reason":                           agentfinalizer.ReasonPendingTools,
+		})
+	}
+	h.waitForToolExecutionsToLeavePending(taskCtx, cancelled, finalizationPendingToolCancelWait)
+	return cancelled
+}
+
+func (h *AgentHandler) waitForToolExecutionsToLeavePending(ctx context.Context, executionIDs []string, wait time.Duration) {
+	if h == nil || h.db == nil || len(executionIDs) == 0 || wait <= 0 {
+		return
+	}
+	timer := time.NewTimer(wait)
+	defer timer.Stop()
+	ticker := time.NewTicker(finalizationPendingToolCancelPoll)
+	defer ticker.Stop()
+	for {
+		if !h.hasPendingToolExecutions(executionIDs) {
+			return
+		}
+		select {
+		case <-contextDone(ctx):
+			return
+		case <-timer.C:
+			return
+		case <-ticker.C:
+		}
+	}
+}
+
+func (h *AgentHandler) hasPendingToolExecutions(executionIDs []string) bool {
+	if h == nil || h.db == nil {
+		return false
+	}
+	for _, executionID := range uniqueNonEmptyStrings(executionIDs) {
+		exec, err := h.db.GetToolExecution(executionID)
+		if err != nil || exec == nil {
+			continue
+		}
+		switch strings.TrimSpace(exec.Status) {
+		case mcp.ToolExecutionStatusQueued, mcp.ToolExecutionStatusRunning:
+			return true
+		}
+	}
+	return false
+}
+
+func uniqueNonEmptyStrings(values []string) []string {
+	seen := make(map[string]struct{}, len(values))
+	out := make([]string, 0, len(values))
+	for _, value := range values {
+		value = strings.TrimSpace(value)
+		if value == "" {
+			continue
+		}
+		if _, ok := seen[value]; ok {
+			continue
+		}
+		seen[value] = struct{}{}
+		out = append(out, value)
+	}
+	return out
+}
+
+func contextDone(ctx context.Context) <-chan struct{} {
+	if ctx == nil {
+		return nil
+	}
+	return ctx.Done()
+}

internal/handler/finalization_auto_continue_test.go

@@ -1,9 +1,18 @@
 package handler
 
 import (
+	"context"
+	"path/filepath"
 	"testing"
+	"time"
 
+	agentpkg "cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/agentfinalizer"
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
 )
 
 func TestShouldAutoContinueAfterFinalization(t *testing.T) {
@@ -57,3 +66,66 @@ func TestRequestRequiresExecutionEvidenceUsesExplicitPolicyOnly(t *testing.T) {
 		t.Fatal("explicit false policy should not require execution evidence")
 	}
 }
+
+func TestCleanupPendingToolExecutionsAfterIterationAllowsFinalization(t *testing.T) {
+	logger := zap.NewNop()
+	db, err := database.NewDB(filepath.Join(t.TempDir(), "cleanup-finalization.db"), logger)
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+
+	server := mcp.NewServerWithStorage(logger, db)
+	server.ConfigureToolWaitTimeoutSeconds(1)
+	server.RegisterTool(mcp.Tool{Name: "block", InputSchema: map[string]interface{}{"type": "object"}}, func(ctx context.Context, args map[string]interface{}) (*mcp.ToolResult, error) {
+		<-ctx.Done()
+		return nil, ctx.Err()
+	})
+	ag := agentpkg.NewAgent(&config.OpenAIConfig{}, &config.AgentConfig{}, server, nil, logger, 10)
+	h := &AgentHandler{agent: ag, db: db, logger: logger}
+
+	callCtx := mcp.WithMCPConversationID(context.Background(), "conv-cleanup")
+	result, execID, err := server.CallTool(callCtx, "block", nil)
+	if err != nil {
+		t.Fatalf("CallTool: %v", err)
+	}
+	if result == nil || !result.IsError || execID == "" {
+		t.Fatalf("expected background wait result, result=%#v execID=%q", result, execID)
+	}
+
+	decision := agentfinalizer.Decide(db, agentfinalizer.Input{
+		Response:        "基于已完成信息的阶段性总结。",
+		MCPExecutionIDs: []string{execID},
+	})
+	if decision.CompletionReason != agentfinalizer.ReasonPendingTools {
+		t.Fatalf("decision reason = %s, want pending tools: %+v", decision.CompletionReason, decision)
+	}
+
+	var eventType string
+	cancelled := h.cleanupPendingToolExecutionsAfterIteration(context.Background(), "conv-cleanup", decision, func(et, _ string, _ interface{}) {
+		eventType = et
+	})
+	if len(cancelled) != 1 || cancelled[0] != execID {
+		t.Fatalf("cancelled = %#v, want [%s]", cancelled, execID)
+	}
+	if eventType != "finalization_pending_tools_cancelled" {
+		t.Fatalf("event type = %q", eventType)
+	}
+
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		exec, err := db.GetToolExecution(execID)
+		if err == nil && exec != nil && exec.Status == mcp.ToolExecutionStatusCancelled {
+			after := agentfinalizer.Decide(db, agentfinalizer.Input{
+				Response:        "基于已完成信息的阶段性总结。",
+				MCPExecutionIDs: []string{execID},
+			})
+			if !after.Finalizable || !after.Finalized {
+				t.Fatalf("decision should finalize after cleanup: %+v", after)
+			}
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	t.Fatal("execution did not become cancelled")
+}

internal/handler/fofa.go

@@ -1,11 +1,13 @@
 package handler
 
 import (
+	"bytes"
 	"context"
 	"encoding/base64"
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"net"
 	"net/http"
 	"net/url"
@@ -87,6 +89,24 @@ type fofaSearchResponse struct {
 	Results       []map[string]interface{} `json:"results"`
 }
 
+type spaceSearchEnvelope struct {
+	Code       interface{}     `json:"code"`
+	Message    string          `json:"message"`
+	Error      string          `json:"error"`
+	Query      string          `json:"query"`
+	Total      int             `json:"total"`
+	TotalCount int             `json:"total_count"`
+	Page       int             `json:"page"`
+	PageSize   int             `json:"pagesize"`
+	Data       json.RawMessage `json:"data"`
+	Matches    json.RawMessage `json:"matches"`
+	Meta       struct {
+		Pagination struct {
+			Total int `json:"total"`
+		} `json:"pagination"`
+	} `json:"meta"`
+}
+
 func normalizeSpaceSearchProvider(provider string) string {
 	switch strings.ToLower(strings.TrimSpace(provider)) {
 	case "", "fofa":
@@ -156,19 +176,20 @@ func (h *FofaHandler) resolveAPIKey(provider string) string {
 }
 
 func (h *FofaHandler) resolveBaseURL(provider string) string {
+	provider = normalizeSpaceSearchProvider(provider)
 	if h.cfg != nil {
-		switch normalizeSpaceSearchProvider(provider) {
+		switch provider {
 		case "zoomeye":
 			if v := strings.TrimSpace(h.cfg.ZoomEye.BaseURL); v != "" {
-				return v
+				return canonicalizeSpaceSearchBaseURL(provider, v)
 			}
 		case "quake":
 			if v := strings.TrimSpace(h.cfg.Quake.BaseURL); v != "" {
-				return v
+				return canonicalizeSpaceSearchBaseURL(provider, v)
 			}
 		case "shodan":
 			if v := strings.TrimSpace(h.cfg.Shodan.BaseURL); v != "" {
-				return v
+				return canonicalizeSpaceSearchBaseURL(provider, v)
 			}
 		default:
 			if v := strings.TrimSpace(h.cfg.FOFA.BaseURL); v != "" {
@@ -176,11 +197,11 @@ func (h *FofaHandler) resolveBaseURL(provider string) string {
 			}
 		}
 	}
-	switch normalizeSpaceSearchProvider(provider) {
+	switch provider {
 	case "zoomeye":
-		return "https://api.zoomeye.org/v2/search"
+		return "https://api.zoomeye.ai/v2/search"
 	case "quake":
-		return "https://quake.360.cn/api/v3/search/quake_service"
+		return "https://quake.360.net/api/v3/search/quake_service"
 	case "shodan":
 		return "https://api.shodan.io"
 	default:
@@ -188,6 +209,20 @@ func (h *FofaHandler) resolveBaseURL(provider string) string {
 	}
 }
 
+func canonicalizeSpaceSearchBaseURL(provider, raw string) string {
+	v := strings.TrimSpace(raw)
+	if v == "" {
+		return v
+	}
+	switch normalizeSpaceSearchProvider(provider) {
+	case "zoomeye":
+		v = strings.Replace(v, "://api.zoomeye.org", "://api.zoomeye.ai", 1)
+	case "quake":
+		v = strings.Replace(v, "://quake.360.cn", "://quake.360.net", 1)
+	}
+	return v
+}
+
 // ParseNaturalLanguage 将自然语言解析为 FOFA 查询语法（仅生成，不执行查询）
 func (h *FofaHandler) ParseNaturalLanguage(c *gin.Context) {
 	var req fofaParseRequest
@@ -716,23 +751,17 @@ func (h *FofaHandler) searchZoomEye(c *gin.Context, req fofaSearchRequest, apiKe
 	if fields := strings.TrimSpace(req.Fields); fields != "" {
 		body["fields"] = fields
 	}
-	var apiResp struct {
-		Code     int                      `json:"code"`
-		Message  string                   `json:"message"`
-		Query    string                   `json:"query"`
-		Total    int                      `json:"total"`
-		Page     int                      `json:"page"`
-		PageSize int                      `json:"pagesize"`
-		Data     []map[string]interface{} `json:"data"`
-	}
+	var apiResp spaceSearchEnvelope
 	if !h.doJSONRequest(c, http.MethodPost, u.String(), apiKey, "API-KEY", body, &apiResp, "ZoomEye") {
 		return
 	}
-	if apiResp.Code != 60000 {
-		msg := strings.TrimSpace(apiResp.Message)
-		if msg == "" {
-			msg = "ZoomEye 返回错误"
-		}
+	rows, err := decodeSpaceSearchRows(apiResp.Data)
+	if err != nil {
+		c.JSON(http.StatusBadGateway, gin.H{"error": "解析 ZoomEye 响应失败: " + err.Error()})
+		return
+	}
+	if zoomEyeRequestFailed(apiResp.Code, apiResp.Message) {
+		msg := firstNonEmptySpaceSearchValue(apiResp.Message, messageFromRawObject(apiResp.Data), "ZoomEye 返回错误")
 		c.JSON(http.StatusBadGateway, gin.H{"error": msg})
 		return
 	}
@@ -744,8 +773,8 @@ func (h *FofaHandler) searchZoomEye(c *gin.Context, req fofaSearchRequest, apiKe
 		Page:         firstPositive(apiResp.Page, req.Page),
 		Total:        apiResp.Total,
 		Fields:       fields,
-		ResultsCount: len(apiResp.Data),
-		Results:      projectRows(apiResp.Data, fields),
+		ResultsCount: len(rows),
+		Results:      projectRows(rows, fields),
 	})
 }
 
@@ -766,25 +795,17 @@ func (h *FofaHandler) searchQuake(c *gin.Context, req fofaSearchRequest, apiKey
 	if len(fields) > 0 {
 		body["include"] = fields
 	}
-	var apiResp struct {
-		Code       interface{}              `json:"code"`
-		Message    string                   `json:"message"`
-		TotalCount int                      `json:"total_count"`
-		Data       []map[string]interface{} `json:"data"`
-		Meta       struct {
-			Pagination struct {
-				Total int `json:"total"`
-			} `json:"pagination"`
-		} `json:"meta"`
-	}
+	var apiResp spaceSearchEnvelope
 	if !h.doJSONRequest(c, http.MethodPost, u.String(), apiKey, "X-QuakeToken", body, &apiResp, "Quake") {
 		return
 	}
+	rows, err := decodeSpaceSearchRows(apiResp.Data)
+	if err != nil {
+		c.JSON(http.StatusBadGateway, gin.H{"error": "解析 Quake 响应失败: " + err.Error()})
+		return
+	}
 	if !isZeroSpaceSearchCode(apiResp.Code) {
-		msg := strings.TrimSpace(apiResp.Message)
-		if msg == "" {
-			msg = "Quake 返回错误"
-		}
+		msg := firstNonEmptySpaceSearchValue(apiResp.Message, messageFromRawObject(apiResp.Data), "Quake 返回错误")
 		c.JSON(http.StatusBadGateway, gin.H{"error": msg})
 		return
 	}
@@ -796,8 +817,8 @@ func (h *FofaHandler) searchQuake(c *gin.Context, req fofaSearchRequest, apiKey
 		Page:         req.Page,
 		Total:        total,
 		Fields:       fields,
-		ResultsCount: len(apiResp.Data),
-		Results:      projectRows(apiResp.Data, fields),
+		ResultsCount: len(rows),
+		Results:      projectRows(rows, fields),
 	})
 }
 
@@ -811,6 +832,9 @@ func isZeroSpaceSearchCode(code interface{}) bool {
 		return v == 0
 	case float64:
 		return v == 0
+	case json.Number:
+		n, err := v.Int64()
+		return err == nil && n == 0
 	case string:
 		return strings.TrimSpace(v) == "0"
 	default:
@@ -818,6 +842,128 @@ func isZeroSpaceSearchCode(code interface{}) bool {
 	}
 }
 
+func isZoomEyeSuccessCode(code interface{}) bool {
+	switch v := code.(type) {
+	case int:
+		return v == 60000
+	case int64:
+		return v == 60000
+	case float64:
+		return v == 60000
+	case json.Number:
+		n, err := v.Int64()
+		return err == nil && n == 60000
+	case string:
+		return strings.TrimSpace(v) == "60000"
+	default:
+		return false
+	}
+}
+
+func zoomEyeRequestFailed(code interface{}, message string) bool {
+	if isZoomEyeSuccessCode(code) {
+		return false
+	}
+	msg := strings.ToLower(strings.TrimSpace(message))
+	if code == nil || isZeroSpaceSearchCode(code) {
+		return msg != "" && msg != "success" && msg != "ok" && msg != "successful."
+	}
+	return true
+}
+
+func decodeSpaceSearchRows(raw json.RawMessage) ([]map[string]interface{}, error) {
+	raw = bytes.TrimSpace(raw)
+	if len(raw) == 0 || bytes.Equal(raw, []byte("null")) {
+		return nil, nil
+	}
+	switch raw[0] {
+	case '[':
+		var rows []map[string]interface{}
+		if err := json.Unmarshal(raw, &rows); err != nil {
+			return nil, err
+		}
+		if rows == nil {
+			return []map[string]interface{}{}, nil
+		}
+		return rows, nil
+	case '{':
+		var obj map[string]interface{}
+		if err := json.Unmarshal(raw, &obj); err != nil {
+			return nil, err
+		}
+		if len(obj) == 0 {
+			return []map[string]interface{}{}, nil
+		}
+		for _, key := range []string{"data", "items", "matches", "results", "list", "records"} {
+			nested, ok := obj[key]
+			if !ok {
+				continue
+			}
+			switch rows := nested.(type) {
+			case []map[string]interface{}:
+				return rows, nil
+			case []interface{}:
+				return interfaceSliceToRowMaps(rows), nil
+			}
+		}
+		return []map[string]interface{}{}, nil
+	default:
+		return nil, fmt.Errorf("unexpected JSON value")
+	}
+}
+
+func interfaceSliceToRowMaps(items []interface{}) []map[string]interface{} {
+	out := make([]map[string]interface{}, 0, len(items))
+	for _, item := range items {
+		if row, ok := item.(map[string]interface{}); ok {
+			out = append(out, row)
+		}
+	}
+	return out
+}
+
+func messageFromRawObject(raw json.RawMessage) string {
+	raw = bytes.TrimSpace(raw)
+	if len(raw) == 0 || raw[0] != '{' {
+		return ""
+	}
+	var obj map[string]interface{}
+	if err := json.Unmarshal(raw, &obj); err != nil {
+		return ""
+	}
+	for _, key := range []string{"message", "error", "errmsg", "msg"} {
+		if s, ok := obj[key].(string); ok {
+			if msg := strings.TrimSpace(s); msg != "" {
+				return msg
+			}
+		}
+	}
+	return ""
+}
+
+func extractRemoteAPIError(body []byte, statusCode int, label string) string {
+	trimmed := bytes.TrimSpace(body)
+	if len(trimmed) > 0 && trimmed[0] == '{' {
+		var obj map[string]interface{}
+		if err := json.Unmarshal(trimmed, &obj); err == nil {
+			for _, key := range []string{"error", "message", "errmsg", "msg"} {
+				if s, ok := obj[key].(string); ok {
+					if msg := strings.TrimSpace(s); msg != "" {
+						return msg
+					}
+				}
+			}
+		}
+	}
+	if len(trimmed) > 0 && trimmed[0] == '<' {
+		return fmt.Sprintf("%s 返回了网页而不是 JSON（HTTP %d），请检查 Base URL 或网络是否被拦截", label, statusCode)
+	}
+	if statusCode < 200 || statusCode >= 300 {
+		return fmt.Sprintf("%s 返回非 2xx: %d", label, statusCode)
+	}
+	return ""
+}
+
 func (h *FofaHandler) searchShodan(c *gin.Context, req fofaSearchRequest, apiKey string) {
 	baseURL := strings.TrimRight(h.resolveBaseURL("shodan"), "/") + "/shodan/host/search"
 	u, err := url.Parse(baseURL)
@@ -826,11 +972,7 @@ func (h *FofaHandler) searchShodan(c *gin.Context, req fofaSearchRequest, apiKey
 		return
 	}
 
-	var apiResp struct {
-		Total   int                      `json:"total"`
-		Matches []map[string]interface{} `json:"matches"`
-		Error   string                   `json:"error"`
-	}
+	var apiResp spaceSearchEnvelope
 	targetSize := req.Size
 	if targetSize <= 0 {
 		targetSize = 100
@@ -852,19 +994,23 @@ func (h *FofaHandler) searchShodan(c *gin.Context, req fofaSearchRequest, apiKey
 			params.Set("fields", fields)
 		}
 		pageURL.RawQuery = params.Encode()
-		apiResp.Matches = nil
-		apiResp.Error = ""
+		apiResp = spaceSearchEnvelope{}
 		if !h.doJSONRequest(c, http.MethodGet, pageURL.String(), "", "", nil, &apiResp, "Shodan") {
 			return
 		}
-		if strings.TrimSpace(apiResp.Error) != "" {
-			c.JSON(http.StatusBadGateway, gin.H{"error": apiResp.Error})
+		if errMsg := firstNonEmptySpaceSearchValue(apiResp.Error, apiResp.Message); errMsg != "" {
+			c.JSON(http.StatusBadGateway, gin.H{"error": errMsg})
 			return
 		}
-		if len(apiResp.Matches) == 0 {
+		pageMatches, err := decodeSpaceSearchRows(apiResp.Matches)
+		if err != nil {
+			c.JSON(http.StatusBadGateway, gin.H{"error": "解析 Shodan 响应失败: " + err.Error()})
+			return
+		}
+		if len(pageMatches) == 0 {
 			break
 		}
-		matches = append(matches, apiResp.Matches...)
+		matches = append(matches, pageMatches...)
 		if len(matches) >= targetSize {
 			matches = matches[:targetSize]
 			break
@@ -947,11 +1093,21 @@ func (h *FofaHandler) doJSONRequest(c *gin.Context, method, endpoint, apiKey, he
 		return false
 	}
 	defer resp.Body.Close()
-	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		c.JSON(http.StatusBadGateway, gin.H{"error": fmt.Sprintf("%s 返回非 2xx: %d", label, resp.StatusCode)})
+	respBody, err := io.ReadAll(io.LimitReader(resp.Body, 32<<20))
+	if err != nil {
+		c.JSON(http.StatusBadGateway, gin.H{"error": "读取 " + label + " 响应失败: " + err.Error()})
 		return false
 	}
-	if err := json.NewDecoder(resp.Body).Decode(out); err != nil {
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		msg := extractRemoteAPIError(respBody, resp.StatusCode, label)
+		c.JSON(http.StatusBadGateway, gin.H{"error": msg})
+		return false
+	}
+	if err := json.Unmarshal(respBody, out); err != nil {
+		if msg := extractRemoteAPIError(respBody, resp.StatusCode, label); strings.Contains(msg, "网页") {
+			c.JSON(http.StatusBadGateway, gin.H{"error": msg})
+			return false
+		}
 		c.JSON(http.StatusBadGateway, gin.H{"error": "解析 " + label + " 响应失败: " + err.Error()})
 		return false
 	}

internal/handler/fofa_test.go

@@ -164,7 +164,7 @@ func TestQuakeSearchHandlesStringErrorCode(t *testing.T) {
 			t.Fatalf("Quake token = %q, want test-quake-key", got)
 		}
 		w.Header().Set("Content-Type", "application/json")
-		_, _ = w.Write([]byte(`{"code":"q5000","message":"查询语法错误"}`))
+		_, _ = w.Write([]byte(`{"code":"q5000","message":"查询语法错误","data":{}}`))
 	}))
 	defer quakeServer.Close()
 
@@ -195,6 +195,144 @@ func TestQuakeSearchHandlesStringErrorCode(t *testing.T) {
 	}
 }
 
+func TestQuakeSearchAcceptsArrayData(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("QUAKE_API_KEY", "")
+
+	quakeServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"code":0,"message":"Successful.","data":[{"ip":"1.1.1.1","port":53}],"meta":{"pagination":{"total":1}}}`))
+	}))
+	defer quakeServer.Close()
+
+	h := NewFofaHandler(&config.Config{
+		Quake: config.SpaceSearchConfig{
+			BaseURL: quakeServer.URL,
+			APIKey:  "test-quake-key",
+		},
+	}, zap.NewNop())
+
+	recorder := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(recorder)
+	body := `{"provider":"quake","query":"ip:\"1.1.1.1\"","fields":"ip,port","size":10,"page":1}`
+	ctx.Request = httptest.NewRequest(http.MethodPost, "/api/fofa/search", strings.NewReader(body))
+	ctx.Request.Header.Set("Content-Type", "application/json")
+
+	h.Search(ctx)
+
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("Search() status = %d, body = %s", recorder.Code, recorder.Body.String())
+	}
+	var response fofaSearchResponse
+	if err := json.Unmarshal(recorder.Body.Bytes(), &response); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if response.ResultsCount != 1 || response.Total != 1 {
+		t.Fatalf("results_count=%d total=%d, want 1/1", response.ResultsCount, response.Total)
+	}
+}
+
+func TestZoomEyeSearchHandlesObjectDataError(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("ZOOMEYE_API_KEY", "")
+
+	zoomeyeServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if got := r.Header.Get("API-KEY"); got != "test-zoomeye-key" {
+			t.Fatalf("ZoomEye API-KEY = %q, want test-zoomeye-key", got)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"code":40001,"message":"invalid api key","data":{}}`))
+	}))
+	defer zoomeyeServer.Close()
+
+	h := NewFofaHandler(&config.Config{
+		ZoomEye: config.SpaceSearchConfig{
+			BaseURL: zoomeyeServer.URL,
+			APIKey:  "test-zoomeye-key",
+		},
+	}, zap.NewNop())
+
+	recorder := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(recorder)
+	body := `{"provider":"zoomeye","query":"ip=\"1.1.1.1\"","fields":"ip,port","size":10,"page":1}`
+	ctx.Request = httptest.NewRequest(http.MethodPost, "/api/fofa/search", strings.NewReader(body))
+	ctx.Request.Header.Set("Content-Type", "application/json")
+
+	h.Search(ctx)
+
+	if recorder.Code != http.StatusBadGateway {
+		t.Fatalf("Search() status = %d, want %d, body = %s", recorder.Code, http.StatusBadGateway, recorder.Body.String())
+	}
+	bodyText := recorder.Body.String()
+	if !strings.Contains(bodyText, "invalid api key") {
+		t.Fatalf("response should include ZoomEye error message, got %s", bodyText)
+	}
+	if strings.Contains(bodyText, "cannot unmarshal") {
+		t.Fatalf("response exposed JSON type decoding failure: %s", bodyText)
+	}
+}
+
+func TestShodanSearchSurfacesJSONErrorOnUnauthorized(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	t.Setenv("SHODAN_API_KEY", "")
+
+	shodanServer := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusUnauthorized)
+		_, _ = w.Write([]byte(`{"error":"Invalid API key"}`))
+	}))
+	defer shodanServer.Close()
+
+	h := NewFofaHandler(&config.Config{
+		Shodan: config.SpaceSearchConfig{
+			BaseURL: shodanServer.URL,
+			APIKey:  "test-shodan-key",
+		},
+	}, zap.NewNop())
+
+	recorder := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(recorder)
+	body := `{"provider":"shodan","query":"product:nginx","fields":"ip_str,port","size":10,"page":1}`
+	ctx.Request = httptest.NewRequest(http.MethodPost, "/api/fofa/search", strings.NewReader(body))
+	ctx.Request.Header.Set("Content-Type", "application/json")
+
+	h.Search(ctx)
+
+	if recorder.Code != http.StatusBadGateway {
+		t.Fatalf("Search() status = %d, want %d, body = %s", recorder.Code, http.StatusBadGateway, recorder.Body.String())
+	}
+	bodyText := recorder.Body.String()
+	if !strings.Contains(bodyText, "Invalid API key") {
+		t.Fatalf("response should include Shodan error message, got %s", bodyText)
+	}
+	if strings.Contains(bodyText, "非 2xx") {
+		t.Fatalf("response should not hide Shodan error behind generic status, got %s", bodyText)
+	}
+}
+
+func TestCanonicalizeSpaceSearchBaseURLMigratesLegacyHosts(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		provider string
+		in       string
+		want     string
+	}{
+		{provider: "zoomeye", in: "https://api.zoomeye.org/v2/search", want: "https://api.zoomeye.ai/v2/search"},
+		{provider: "quake", in: "https://quake.360.cn/api/v3/search/quake_service", want: "https://quake.360.net/api/v3/search/quake_service"},
+		{provider: "shodan", in: "https://api.shodan.io", want: "https://api.shodan.io"},
+	}
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.provider, func(t *testing.T) {
+			t.Parallel()
+			got := canonicalizeSpaceSearchBaseURL(tc.provider, tc.in)
+			if got != tc.want {
+				t.Fatalf("canonicalizeSpaceSearchBaseURL() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
+
 func TestExtractInfoCollectJSONObject(t *testing.T) {
 	t.Parallel()
 	cases := []struct {

internal/handler/hitl.go

@@ -289,6 +289,18 @@ func normalizeHitlMode(mode string) string {
 	}
 }
 
+func normalizeHitlDefaultMode(mode string) string {
+	v := strings.ToLower(strings.TrimSpace(mode))
+	switch v {
+	case "feedback", "followup":
+		return "approval"
+	case "approval", "review_edit":
+		return v
+	default:
+		return "off"
+	}
+}
+
 func (m *HITLManager) ActivateConversation(conversationID string, req *HITLRequest) {
 	if req == nil || !req.Enabled {
 		m.DeactivateConversation(conversationID)
@@ -629,7 +641,7 @@ func (h *AgentHandler) loadHITLConversationConfig(conversationID string) (*HITLR
 		return nil, err
 	}
 	if !has {
-		cfg.Reviewer = h.hitlEffectiveDefaultReviewer()
+		return h.hitlEffectiveDefaultRequest(), nil
 	}
 	return cfg, nil
 }
@@ -994,7 +1006,9 @@ func (h *AgentHandler) GetHITLConversationConfig(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{
 		"conversationId":          conversationID,
 		"hitl":                    cfg,
+		"defaultMode":             h.hitlEffectiveDefaultMode(),
 		"defaultReviewer":         h.hitlEffectiveDefaultReviewer(),
+		"defaultTimeoutSeconds":   h.hitlEffectiveDefaultTimeoutSeconds(),
 		"hitlGlobalToolWhitelist": h.hitlConfigGlobalToolWhitelist(),
 	})
 }
@@ -1051,11 +1065,64 @@ type setHitlDefaultReviewerReq struct {
 	Reviewer string `json:"reviewer"`
 }
 
+type setHitlDefaultConfigReq struct {
+	Mode           string `json:"mode"`
+	Reviewer       string `json:"reviewer"`
+	TimeoutSeconds int    `json:"timeoutSeconds"`
+}
+
+func (h *AgentHandler) hitlDefaultConfigResponse() gin.H {
+	return gin.H{
+		"defaultMode":             h.hitlEffectiveDefaultMode(),
+		"defaultReviewer":         h.hitlEffectiveDefaultReviewer(),
+		"defaultTimeoutSeconds":   h.hitlEffectiveDefaultTimeoutSeconds(),
+		"hitlGlobalToolWhitelist": h.hitlConfigGlobalToolWhitelist(),
+	}
+}
+
+// GetHITLDefaultConfig 返回 config.yaml 中的全局默认人机协同配置。
+func (h *AgentHandler) GetHITLDefaultConfig(c *gin.Context) {
+	c.JSON(http.StatusOK, h.hitlDefaultConfigResponse())
+}
+
+// UpdateHITLDefaultConfig 将全局默认人机协同配置写入 config.yaml。
+func (h *AgentHandler) UpdateHITLDefaultConfig(c *gin.Context) {
+	if h.hitlDefaultReviewerSaver == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "HITL 配置持久化不可用"})
+		return
+	}
+	var req setHitlDefaultConfigReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	mode := normalizeHitlDefaultMode(req.Mode)
+	reviewer := normalizeHitlReviewer(req.Reviewer)
+	timeoutSeconds := req.TimeoutSeconds
+	if timeoutSeconds < 0 {
+		timeoutSeconds = 0
+	}
+	if err := h.hitlDefaultReviewerSaver.UpdateHitlDefaultConfig(mode, reviewer, timeoutSeconds); err != nil {
+		h.logger.Warn("写入 HITL 默认配置到 config.yaml 失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if h.config != nil {
+		h.config.Hitl.DefaultMode = mode
+		h.config.Hitl.DefaultReviewer = reviewer
+		h.config.Hitl.DefaultTimeoutSeconds = &timeoutSeconds
+	}
+	if h.audit != nil {
+		h.audit.RecordOK(c, "hitl", "default_config_update", "HITL 全局默认配置更新", "hitl_config", "default", nil)
+	}
+	out := h.hitlDefaultConfigResponse()
+	out["ok"] = true
+	c.JSON(http.StatusOK, out)
+}
+
 // GetHITLDefaultReviewer 返回 config.yaml 中的全局默认审批方。
 func (h *AgentHandler) GetHITLDefaultReviewer(c *gin.Context) {
-	c.JSON(http.StatusOK, gin.H{
-		"defaultReviewer": h.hitlEffectiveDefaultReviewer(),
-	})
+	c.JSON(http.StatusOK, h.hitlDefaultConfigResponse())
 }
 
 // UpdateHITLDefaultReviewer 将全局默认审批方写入 config.yaml（未选会话时切换审批方）。
@@ -1081,10 +1148,9 @@ func (h *AgentHandler) UpdateHITLDefaultReviewer(c *gin.Context) {
 	if h.audit != nil {
 		h.audit.RecordOK(c, "hitl", "default_reviewer_update", "HITL 全局默认审批方更新", "hitl_config", "default_reviewer", nil)
 	}
-	c.JSON(http.StatusOK, gin.H{
-		"ok":              true,
-		"defaultReviewer": reviewer,
-	})
+	out := h.hitlDefaultConfigResponse()
+	out["ok"] = true
+	c.JSON(http.StatusOK, out)
 }
 
 // SetHITLGlobalToolWhitelist 整表替换 config.yaml 中的全局免审批工具白名单。

internal/handler/multi_agent.go

@@ -205,6 +205,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	}
 	agentMode := "eino_" + effectiveOrch
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 
 	for {
 		segmentMainIterationMax := 0
@@ -282,6 +283,10 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 				continue
 			}
 			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback); len(cancelled) > 0 {
+				autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+				decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			}
 			if h.tryAutoContinueAfterFinalization(taskCtx, conversationID, result, decision, &finalizationAutoContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
 				mainIterationOffset += segmentMainIterationMax
 				timeoutCancel()
@@ -398,6 +403,10 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 
 	if decision.CompletionReason == "" {
 		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 	}
 	h.persistFinalizationDecision(conversationID, assistantMessageID, agentMode, cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput), decision)
 
@@ -415,10 +424,11 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
 	}
 	sendEvent("response", responseText, finalizationResponsePayload(decision, map[string]interface{}{
-		"mcpExecutionIds": cumulativeMCPExecutionIDs,
-		"conversationId":  conversationID,
-		"messageId":       assistantMessageID,
-		"agentMode":       agentMode,
+		"mcpExecutionIds":                  cumulativeMCPExecutionIDs,
+		"conversationId":                   conversationID,
+		"messageId":                        assistantMessageID,
+		"agentMode":                        agentMode,
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 	sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 }
@@ -478,6 +488,7 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 	}
 	agentMode := "eino_" + effectiveOrch
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 	for {
 		result, runErr = multiagent.RunDeepAgent(
 			taskCtx,
@@ -514,6 +525,10 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 			continue
 		}
 		decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, agentMode, result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, prep.ConversationID, decision, progressCallback); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, agentMode, result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 		if h.tryAutoContinueAfterFinalization(taskCtx, prep.ConversationID, result, decision, &finalizationAutoContinueAttempt, &curHist, &curMsg, progressCallback) {
 			continue
 		}
@@ -533,18 +548,19 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 		responseText = finalizationBlockedMessage(decision)
 	}
 	c.JSON(http.StatusOK, ChatResponse{
-		Response:            responseText,
-		MCPExecutionIDs:     result.MCPExecutionIDs,
-		ConversationID:      prep.ConversationID,
-		Time:                time.Now(),
-		Finalizable:         decision.Finalizable,
-		Finalized:           decision.Finalized,
-		Status:              decision.Status,
-		CompletionReason:    decision.CompletionReason,
-		EvidenceVerified:    decision.EvidenceVerified,
-		EvidenceRefs:        decision.EvidenceRefs,
-		PendingExecutionIDs: decision.PendingExecutionIDs,
-		MissingChecks:       decision.MissingChecks,
+		Response:                         responseText,
+		MCPExecutionIDs:                  result.MCPExecutionIDs,
+		ConversationID:                   prep.ConversationID,
+		Time:                             time.Now(),
+		Finalizable:                      decision.Finalizable,
+		Finalized:                        decision.Finalized,
+		Status:                           decision.Status,
+		CompletionReason:                 decision.CompletionReason,
+		EvidenceVerified:                 decision.EvidenceVerified,
+		EvidenceRefs:                     decision.EvidenceRefs,
+		PendingExecutionIDs:              decision.PendingExecutionIDs,
+		MissingChecks:                    decision.MissingChecks,
+		AutoCancelledPendingExecutionIDs: autoCancelledPendingExecutionIDs,
 	})
 }
 

internal/handler/task_manager.go

@@ -3,6 +3,7 @@ package handler
 import (
 	"context"
 	"errors"
+	"sort"
 	"strings"
 	"sync"
 	"time"
@@ -484,7 +485,10 @@ func (m *AgentTaskManager) CancelTask(conversationID string, cause error) (bool,
 	if runtimeCancel != nil {
 		runtimeHandled = runtimeCancel(cause)
 	}
-	if cancel != nil && !runtimeHandled {
+	// 「彻底停止」必须同时取消宿主 context：原生 Agent Cancel 即使已受理，
+	// 也可能只在安全点返回或报告超时，不能据此让整条任务继续存活。
+	// 中断并继续仍保留原语义：原生取消已处理时由运行时负责恢复。
+	if cancel != nil && (!runtimeHandled || errors.Is(cause, ErrTaskCancelled)) {
 		cancel(cause)
 	}
 	if toolCanceler != nil {
@@ -591,6 +595,12 @@ func (m *AgentTaskManager) GetActiveTasks() []*AgentTask {
 			Status:         task.Status,
 		})
 	}
+	sort.Slice(result, func(i, j int) bool {
+		if result[i].StartedAt.Equal(result[j].StartedAt) {
+			return result[i].ConversationID < result[j].ConversationID
+		}
+		return result[i].StartedAt.Before(result[j].StartedAt)
+	})
 	return result
 }
 

internal/handler/task_manager_order_test.go

@@ -0,0 +1,31 @@
+package handler
+
+import (
+	"testing"
+	"time"
+)
+
+func TestGetActiveTasksUsesStableCreationOrder(t *testing.T) {
+	m := NewAgentTaskManager()
+	started := time.Date(2026, 8, 19, 10, 0, 0, 0, time.UTC)
+	m.mu.Lock()
+	m.tasks = map[string]*AgentTask{
+		"conversation-z":    {ConversationID: "conversation-z", StartedAt: started, Status: "running"},
+		"conversation-late": {ConversationID: "conversation-late", StartedAt: started.Add(time.Minute), Status: "running"},
+		"conversation-a":    {ConversationID: "conversation-a", StartedAt: started, Status: "running"},
+	}
+	m.mu.Unlock()
+
+	want := []string{"conversation-a", "conversation-z", "conversation-late"}
+	for attempt := 0; attempt < 20; attempt++ {
+		gotTasks := m.GetActiveTasks()
+		if len(gotTasks) != len(want) {
+			t.Fatalf("GetActiveTasks() length = %d, want %d", len(gotTasks), len(want))
+		}
+		for i, task := range gotTasks {
+			if task.ConversationID != want[i] {
+				t.Fatalf("attempt %d order[%d] = %q, want %q", attempt, i, task.ConversationID, want[i])
+			}
+		}
+	}
+}

internal/handler/task_manager_tool_cancel_test.go

@@ -32,7 +32,7 @@ func TestCancelTaskInvokesToolCancelerOnFullStop(t *testing.T) {
 	}
 }
 
-func TestCancelTaskUsesAgentRuntimeCancelAsPrimaryPath(t *testing.T) {
+func TestCancelTaskFullStopCancelsRuntimeAndParentContext(t *testing.T) {
 	tm := NewAgentTaskManager()
 	var order []string
 	tm.SetToolCanceler(func(conversationID string) {
@@ -61,7 +61,7 @@ func TestCancelTaskUsesAgentRuntimeCancelAsPrimaryPath(t *testing.T) {
 	if err != nil || !ok {
 		t.Fatalf("CancelTask: ok=%v err=%v", ok, err)
 	}
-	want := []string{"runtime", "tool"}
+	want := []string{"runtime", "context", "tool"}
 	if len(order) != len(want) {
 		t.Fatalf("order length got %d want %d: %#v", len(order), len(want), order)
 	}
@@ -72,6 +72,29 @@ func TestCancelTaskUsesAgentRuntimeCancelAsPrimaryPath(t *testing.T) {
 	}
 }
 
+func TestCancelTaskInterruptContinueKeepsParentWhenRuntimeHandlesIt(t *testing.T) {
+	tm := NewAgentTaskManager()
+	ctx, cancel := context.WithCancelCause(context.Background())
+	if _, err := tm.StartTask("conv-interrupt-native", "hello", cancel); err != nil {
+		t.Fatalf("StartTask: %v", err)
+	}
+	unregister := tm.BindAgentRuntimeCancel("conv-interrupt-native", func(err error) bool {
+		if !errors.Is(err, multiagent.ErrInterruptContinue) {
+			t.Fatalf("runtime cancel got %v", err)
+		}
+		return true
+	})
+	defer unregister()
+
+	ok, err := tm.CancelTask("conv-interrupt-native", multiagent.ErrInterruptContinue)
+	if err != nil || !ok {
+		t.Fatalf("CancelTask: ok=%v err=%v", ok, err)
+	}
+	if cause := context.Cause(ctx); cause != nil {
+		t.Fatalf("interrupt-continue parent context cause = %v, want nil", cause)
+	}
+}
+
 func TestCancelTaskFallsBackToContextWhenAgentRuntimeCancelMisses(t *testing.T) {
 	tm := NewAgentTaskManager()
 	var order []string

internal/handler/token_usage.go

@@ -0,0 +1,92 @@
+package handler
+
+import (
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/security"
+
+	"github.com/gin-gonic/gin"
+	"go.uber.org/zap"
+)
+
+// GetTokenUsageStats returns model token usage aggregates for dashboard views.
+func (h *ConversationHandler) GetTokenUsageStats(c *gin.Context) {
+	filter := tokenUsageFilterFromQuery(c)
+	if session, ok := security.CurrentSession(c); ok {
+		filter.Access = database.RBACListAccess{UserID: session.UserID, Scope: session.Scope}
+	}
+	stats, err := h.db.GetModelTokenUsageStats(filter)
+	if err != nil {
+		h.logger.Error("获取Token用量统计失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, stats)
+}
+
+// GetConversationTokenUsageStats returns token usage scoped to one conversation.
+func (h *ConversationHandler) GetConversationTokenUsageStats(c *gin.Context) {
+	filter := tokenUsageFilterFromQuery(c)
+	filter.ConversationID = strings.TrimSpace(c.Param("id"))
+	if session, ok := security.CurrentSession(c); ok {
+		filter.Access = database.RBACListAccess{UserID: session.UserID, Scope: session.Scope}
+	}
+	stats, err := h.db.GetModelTokenUsageStats(filter)
+	if err != nil {
+		h.logger.Error("获取对话Token用量统计失败", zap.Error(err), zap.String("conversationId", filter.ConversationID))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, stats)
+}
+
+func tokenUsageFilterFromQuery(c *gin.Context) database.ModelTokenUsageFilter {
+	days, _ := strconv.Atoi(strings.TrimSpace(c.DefaultQuery("days", "7")))
+	if days <= 0 {
+		days = 7
+	}
+	if days > 365 {
+		days = 365
+	}
+	limit, _ := strconv.Atoi(strings.TrimSpace(c.DefaultQuery("limit", "10")))
+	if limit <= 0 {
+		limit = 10
+	}
+	if limit > 500 {
+		limit = 500
+	}
+	filter := database.ModelTokenUsageFilter{
+		ConversationID: strings.TrimSpace(c.Query("conversation_id")),
+		ProjectID:      strings.TrimSpace(c.Query("project_id")),
+		Days:           days,
+		Limit:          limit,
+	}
+	if since := parseTokenUsageQueryTime(c.Query("since")); !since.IsZero() {
+		filter.Since = since
+	} else if days > 0 {
+		now := time.Now()
+		start := time.Date(now.Year(), now.Month(), now.Day(), 0, 0, 0, 0, now.Location()).AddDate(0, 0, -(days - 1))
+		filter.Since = start
+	}
+	if until := parseTokenUsageQueryTime(c.Query("until")); !until.IsZero() {
+		filter.Until = until
+	}
+	return filter
+}
+
+func parseTokenUsageQueryTime(raw string) time.Time {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return time.Time{}
+	}
+	for _, layout := range []string{time.RFC3339Nano, time.RFC3339, "2006-01-02"} {
+		if t, err := time.Parse(layout, raw); err == nil {
+			return t
+		}
+	}
+	return time.Time{}
+}

internal/knowledge/wire_retriever.go

@@ -8,9 +8,11 @@ import (
 	"time"
 
 	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/llm"
 	"cyberstrike-ai/internal/openai"
 
 	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
+	"github.com/cloudwego/eino/components/model"
 	"github.com/cloudwego/eino/flow/retriever/multiquery"
 	"go.uber.org/zap"
 )
@@ -29,21 +31,38 @@ func WireRetrieverPipeline(ctx context.Context, r *Retriever, openAI *config.Ope
 	}
 	r.wireOpenAI = openAI
 
-	httpClient := openai.NewEinoHTTPClient(openAI, &http.Client{Timeout: 120 * time.Second})
-	maxCompletionTokens := openAI.MaxCompletionTokensEffective()
-	chatCfg := &einoopenai.ChatModelConfig{
-		APIKey:              strings.TrimSpace(openAI.APIKey),
-		BaseURL:             strings.TrimSuffix(strings.TrimSpace(openAI.BaseURL), "/"),
-		Model:               strings.TrimSpace(openAI.Model),
-		HTTPClient:          httpClient,
-		MaxCompletionTokens: &maxCompletionTokens,
-	}
-	if chatCfg.Model == "" {
-		chatCfg.Model = "gpt-4o"
-	}
-	rewriteLLM, err := einoopenai.NewChatModel(ctx, chatCfg)
-	if err != nil {
-		return fmt.Errorf("multi_query rewrite model: %w", err)
+	baseHTTPClient := &http.Client{Timeout: 120 * time.Second}
+	var rewriteLLM model.ChatModel
+	if llm.IsClaudeProvider(openAI.Provider) {
+		nativeModel, err := llm.NewClaudeAgenticModel(
+			ctx,
+			*openAI,
+			baseHTTPClient,
+			openAI.MaxCompletionTokensEffective(),
+			nil,
+		)
+		if err != nil {
+			return fmt.Errorf("multi_query native Claude rewrite model: %w", err)
+		}
+		rewriteLLM = llm.NewAgenticChatModelAdapter(nativeModel)
+	} else {
+		httpClient := openai.NewEinoHTTPClient(openAI, baseHTTPClient)
+		maxCompletionTokens := openAI.MaxCompletionTokensEffective()
+		chatCfg := &einoopenai.ChatModelConfig{
+			APIKey:              strings.TrimSpace(openAI.APIKey),
+			BaseURL:             strings.TrimSuffix(strings.TrimSpace(openAI.BaseURL), "/"),
+			Model:               strings.TrimSpace(openAI.Model),
+			HTTPClient:          httpClient,
+			MaxCompletionTokens: &maxCompletionTokens,
+		}
+		if chatCfg.Model == "" {
+			chatCfg.Model = "gpt-4o"
+		}
+		var err error
+		rewriteLLM, err = einoopenai.NewChatModel(ctx, chatCfg)
+		if err != nil {
+			return fmt.Errorf("multi_query rewrite model: %w", err)
+		}
 	}
 
 	reranker, err := NewHTTPReranker(&r.config.Rerank, openAI, r.logger)

internal/llm/agentic_chat_adapter.go

@@ -0,0 +1,125 @@
+package llm
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+)
+
+// AgenticChatModelAdapter exposes a text-oriented AgenticModel as a classic
+// BaseChatModel for Eino components that have not adopted AgenticMessage yet.
+// It adapts only Eino's in-memory message shape; no HTTP protocol is translated.
+type AgenticChatModelAdapter struct {
+	model model.AgenticModel
+	tools []*schema.ToolInfo
+}
+
+func NewAgenticChatModelAdapter(agenticModel model.AgenticModel) model.ChatModel {
+	return &AgenticChatModelAdapter{model: agenticModel}
+}
+
+func (a *AgenticChatModelAdapter) BindTools(tools []*schema.ToolInfo) error {
+	if a == nil || a.model == nil {
+		return fmt.Errorf("agentic chat adapter: model is nil")
+	}
+	a.tools = append([]*schema.ToolInfo(nil), tools...)
+	return nil
+}
+
+func (a *AgenticChatModelAdapter) Generate(
+	ctx context.Context,
+	input []*schema.Message,
+	opts ...model.Option,
+) (*schema.Message, error) {
+	if a == nil || a.model == nil {
+		return nil, fmt.Errorf("agentic chat adapter: model is nil")
+	}
+	out, err := a.model.Generate(ctx, classicMessagesToAgentic(input), commonAgenticOptions(a.tools, opts...)...)
+	if err != nil {
+		return nil, err
+	}
+	return agenticMessageToClassic(out), nil
+}
+
+func (a *AgenticChatModelAdapter) Stream(
+	ctx context.Context,
+	input []*schema.Message,
+	opts ...model.Option,
+) (*schema.StreamReader[*schema.Message], error) {
+	if a == nil || a.model == nil {
+		return nil, fmt.Errorf("agentic chat adapter: model is nil")
+	}
+	stream, err := a.model.Stream(ctx, classicMessagesToAgentic(input), commonAgenticOptions(a.tools, opts...)...)
+	if err != nil {
+		return nil, err
+	}
+	return schema.StreamReaderWithConvert(stream, func(msg *schema.AgenticMessage) (*schema.Message, error) {
+		return agenticMessageToClassic(msg), nil
+	}), nil
+}
+
+func classicMessagesToAgentic(input []*schema.Message) []*schema.AgenticMessage {
+	out := make([]*schema.AgenticMessage, 0, len(input))
+	for _, msg := range input {
+		if msg == nil {
+			continue
+		}
+		role := schema.AgenticRoleTypeUser
+		switch msg.Role {
+		case schema.System:
+			role = schema.AgenticRoleTypeSystem
+		case schema.Assistant:
+			role = schema.AgenticRoleTypeAssistant
+		}
+		agentic := &schema.AgenticMessage{Role: role}
+		if msg.Role == schema.Assistant {
+			if msg.Content != "" {
+				agentic.ContentBlocks = append(agentic.ContentBlocks, schema.NewContentBlock(&schema.AssistantGenText{Text: msg.Content}))
+			}
+		} else if msg.Content != "" {
+			agentic.ContentBlocks = append(agentic.ContentBlocks, schema.NewContentBlock(&schema.UserInputText{Text: msg.Content}))
+		}
+		out = append(out, agentic)
+	}
+	return out
+}
+
+func agenticMessageToClassic(msg *schema.AgenticMessage) *schema.Message {
+	if msg == nil {
+		return nil
+	}
+	content, reasoning := AgenticText(msg)
+	return &schema.Message{
+		Role:             schema.Assistant,
+		Content:          content,
+		ReasoningContent: reasoning,
+	}
+}
+
+func commonAgenticOptions(boundTools []*schema.ToolInfo, opts ...model.Option) []model.Option {
+	common := model.GetCommonOptions(&model.Options{
+		Tools: append([]*schema.ToolInfo(nil), boundTools...),
+	}, opts...)
+	out := make([]model.Option, 0, 6)
+	if common.Temperature != nil {
+		out = append(out, model.WithTemperature(*common.Temperature))
+	}
+	if common.Model != nil {
+		out = append(out, model.WithModel(*common.Model))
+	}
+	if common.TopP != nil {
+		out = append(out, model.WithTopP(*common.TopP))
+	}
+	if common.MaxTokens != nil {
+		out = append(out, model.WithMaxTokens(*common.MaxTokens))
+	}
+	if len(common.Stop) > 0 {
+		out = append(out, model.WithStop(common.Stop))
+	}
+	if common.Tools != nil {
+		out = append(out, model.WithTools(common.Tools))
+	}
+	return out
+}

internal/llm/claude.go

@@ -0,0 +1,60 @@
+package llm
+
+import (
+	"context"
+	"net/http"
+	"strings"
+
+	"cyberstrike-ai/internal/config"
+
+	agenticclaude "github.com/cloudwego/eino-ext/components/model/agenticclaude"
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+)
+
+func IsClaudeProvider(provider string) bool {
+	provider = strings.ToLower(strings.TrimSpace(provider))
+	return provider == "claude" || provider == "anthropic"
+}
+
+func NewClaudeAgenticModel(
+	ctx context.Context,
+	cfg config.OpenAIConfig,
+	httpClient *http.Client,
+	maxTokens int,
+	extraFields map[string]any,
+) (model.AgenticModel, error) {
+	if maxTokens <= 0 {
+		maxTokens = cfg.MaxCompletionTokensEffective()
+	}
+	if cfg.IsDeepSeekEndpointOrModel() {
+		httpClient = newDeepSeekAnthropicCompatibleClient(httpClient)
+	}
+	return agenticclaude.New(ctx, &agenticclaude.Config{
+		APIKey:      strings.TrimSpace(cfg.APIKey),
+		BaseURL:     strings.TrimSuffix(strings.TrimSpace(cfg.BaseURL), "/"),
+		Model:       strings.TrimSpace(cfg.Model),
+		MaxTokens:   maxTokens,
+		HTTPClient:  httpClient,
+		ExtraFields: extraFields,
+	})
+}
+
+func AgenticText(msg *schema.AgenticMessage) (content, reasoning string) {
+	if msg == nil {
+		return "", ""
+	}
+	var contentParts, reasoningParts []string
+	for _, block := range msg.ContentBlocks {
+		if block == nil {
+			continue
+		}
+		switch {
+		case block.AssistantGenText != nil:
+			contentParts = append(contentParts, block.AssistantGenText.Text)
+		case block.Reasoning != nil:
+			reasoningParts = append(reasoningParts, block.Reasoning.Text)
+		}
+	}
+	return strings.Join(contentParts, ""), strings.Join(reasoningParts, "")
+}

internal/llm/deepseek_anthropic_compat.go

@@ -0,0 +1,76 @@
+package llm
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+)
+
+// newDeepSeekAnthropicCompatibleClient compensates for DeepSeek's Anthropic
+// endpoint lagging behind the current Anthropic SDK. The SDK emits
+// {"type":"custom"} for function tools, while DeepSeek expects the older
+// name/input_schema/description shape without that discriminator.
+//
+// This is a field-level compatibility fix; requests still originate from
+// Eino's native agenticclaude model and remain Anthropic Messages API requests.
+func newDeepSeekAnthropicCompatibleClient(base *http.Client) *http.Client {
+	if base == nil {
+		base = http.DefaultClient
+	}
+	cloned := *base
+	transport := base.Transport
+	if transport == nil {
+		transport = http.DefaultTransport
+	}
+	cloned.Transport = &deepSeekAnthropicCompatRoundTripper{base: transport}
+	return &cloned
+}
+
+type deepSeekAnthropicCompatRoundTripper struct {
+	base http.RoundTripper
+}
+
+func (rt *deepSeekAnthropicCompatRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
+	if req == nil || req.Body == nil || req.Method != http.MethodPost {
+		return rt.base.RoundTrip(req)
+	}
+	body, err := io.ReadAll(req.Body)
+	if err != nil {
+		return nil, fmt.Errorf("read DeepSeek Anthropic request: %w", err)
+	}
+	_ = req.Body.Close()
+
+	var payload map[string]any
+	if err := json.Unmarshal(body, &payload); err != nil {
+		req.Body = io.NopCloser(bytes.NewReader(body))
+		return rt.base.RoundTrip(req)
+	}
+	tools, ok := payload["tools"].([]any)
+	if !ok {
+		req.Body = io.NopCloser(bytes.NewReader(body))
+		return rt.base.RoundTrip(req)
+	}
+	changed := false
+	for _, rawTool := range tools {
+		tool, ok := rawTool.(map[string]any)
+		if !ok || tool["type"] != "custom" {
+			continue
+		}
+		delete(tool, "type")
+		changed = true
+	}
+	if changed {
+		body, err = json.Marshal(payload)
+		if err != nil {
+			return nil, fmt.Errorf("marshal DeepSeek Anthropic request: %w", err)
+		}
+	}
+	req.Body = io.NopCloser(bytes.NewReader(body))
+	req.ContentLength = int64(len(body))
+	req.GetBody = func() (io.ReadCloser, error) {
+		return io.NopCloser(bytes.NewReader(body)), nil
+	}
+	return rt.base.RoundTrip(req)
+}

internal/llm/deepseek_anthropic_compat_test.go

@@ -0,0 +1,55 @@
+package llm
+
+import (
+	"bytes"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+type captureRoundTripper struct {
+	body string
+}
+
+func (rt *captureRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
+	body, err := io.ReadAll(req.Body)
+	if err != nil {
+		return nil, err
+	}
+	rt.body = string(body)
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     make(http.Header),
+		Body:       io.NopCloser(bytes.NewReader(nil)),
+		Request:    req,
+	}, nil
+}
+
+func TestDeepSeekAnthropicCompatStripsOnlyCustomToolType(t *testing.T) {
+	t.Parallel()
+	capture := &captureRoundTripper{}
+	client := newDeepSeekAnthropicCompatibleClient(&http.Client{Transport: capture})
+	req, err := http.NewRequest(
+		http.MethodPost,
+		"https://api.deepseek.com/anthropic/v1/messages",
+		strings.NewReader(`{"tools":[{"type":"custom","name":"mcp_tool","input_schema":{"type":"object"}},{"type":"web_search_20260209","name":"web_search"}]}`),
+	)
+	if err != nil {
+		t.Fatalf("NewRequest: %v", err)
+	}
+	resp, err := client.Do(req)
+	if err != nil {
+		t.Fatalf("Do: %v", err)
+	}
+	_ = resp.Body.Close()
+	if strings.Contains(capture.body, `"type":"custom"`) {
+		t.Fatalf("custom discriminator was not removed: %s", capture.body)
+	}
+	if !strings.Contains(capture.body, `"type":"web_search_20260209"`) {
+		t.Fatalf("server tool discriminator was removed: %s", capture.body)
+	}
+	if !strings.Contains(capture.body, `"name":"mcp_tool"`) {
+		t.Fatalf("custom tool definition was removed: %s", capture.body)
+	}
+}

internal/multiagent/agentic_orphan_tool_pruner_middleware.go

@@ -0,0 +1,126 @@
+package multiagent
+
+import (
+	"context"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// agenticOrphanToolPrunerMiddleware is the AgenticMessage equivalent of
+// orphanToolPrunerMiddleware. It removes user-role messages whose content
+// blocks are exclusively FunctionToolResult entries with CallIDs that do not
+// match any FunctionToolCall in the history.
+//
+// This is a defense-in-depth layer after agenticToolPairReconcilerMiddleware;
+// the reconciler handles the common case (assistant followed by its results)
+// while this pruner catches stray results that appear before their assistant
+// or in non-adjacent positions (e.g. after summarization rewriting).
+type agenticOrphanToolPrunerMiddleware struct {
+	*adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]
+	logger *zap.Logger
+	phase  string
+}
+
+func newAgenticOrphanToolPrunerMiddleware(logger *zap.Logger, phase string) adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage] {
+	return &agenticOrphanToolPrunerMiddleware{
+		TypedBaseChatModelAgentMiddleware: &adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]{},
+		logger:                            logger,
+		phase:                             phase,
+	}
+}
+
+func (m *agenticOrphanToolPrunerMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.TypedChatModelAgentState[*schema.AgenticMessage],
+	mc *adk.TypedModelContext[*schema.AgenticMessage],
+) (context.Context, *adk.TypedChatModelAgentState[*schema.AgenticMessage], error) {
+	_ = mc
+	if m == nil || state == nil || len(state.Messages) == 0 {
+		return ctx, state, nil
+	}
+
+	// Pass 1: collect all provided CallIDs from assistant FunctionToolCall blocks.
+	provided := make(map[string]struct{}, 8)
+	for _, msg := range state.Messages {
+		if msg == nil || msg.Role != schema.AgenticRoleTypeAssistant {
+			continue
+		}
+		for _, block := range msg.ContentBlocks {
+			if block != nil && block.FunctionToolCall != nil && block.FunctionToolCall.CallID != "" {
+				provided[block.FunctionToolCall.CallID] = struct{}{}
+			}
+		}
+	}
+
+	// Fast path: check if any orphan exists.
+	hasOrphan := false
+	for _, msg := range state.Messages {
+		if msg == nil || !isPureAgenticToolResult(msg) {
+			continue
+		}
+		for _, id := range agenticToolResultCallIDs(msg) {
+			if _, ok := provided[id]; !ok {
+				hasOrphan = true
+				break
+			}
+		}
+		if hasOrphan {
+			break
+		}
+	}
+	if !hasOrphan {
+		return ctx, state, nil
+	}
+
+	// Pass 2: build pruned list.
+	pruned := make([]*schema.AgenticMessage, 0, len(state.Messages))
+	var droppedIDs []string
+	var droppedNames []string
+	for _, msg := range state.Messages {
+		if msg == nil {
+			continue
+		}
+		if !isPureAgenticToolResult(msg) {
+			pruned = append(pruned, msg)
+			continue
+		}
+		// Check if ALL result call IDs are orphans. If any is matched, keep the
+		// message (the reconciler already handled partial mismatches).
+		allOrphan := true
+		for _, id := range agenticToolResultCallIDs(msg) {
+			if _, ok := provided[id]; ok {
+				allOrphan = false
+				break
+			}
+		}
+		if allOrphan {
+			for _, block := range msg.ContentBlocks {
+				if block != nil && block.FunctionToolResult != nil {
+					droppedIDs = append(droppedIDs, block.FunctionToolResult.CallID)
+					droppedNames = append(droppedNames, block.FunctionToolResult.Name)
+				}
+			}
+			continue
+		}
+		pruned = append(pruned, msg)
+	}
+
+	if len(droppedIDs) == 0 {
+		return ctx, state, nil
+	}
+	if m.logger != nil {
+		m.logger.Warn("agentic orphan tool messages pruned before model call",
+			zap.String("phase", m.phase),
+			zap.Int("dropped_count", len(droppedIDs)),
+			zap.Strings("dropped_tool_call_ids", droppedIDs),
+			zap.Strings("dropped_tool_names", droppedNames),
+			zap.Int("messages_before", len(state.Messages)),
+			zap.Int("messages_after", len(pruned)),
+		)
+	}
+	ns := *state
+	ns.Messages = pruned
+	return ctx, &ns, nil
+}

internal/multiagent/agentic_tool_pair_reconciler_middleware.go

@@ -0,0 +1,247 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// agenticToolPairReconcilerMiddleware is the AgenticMessage equivalent of
+// toolPairReconcilerMiddleware. It ensures every assistant FunctionToolCall
+// block is followed by a matching FunctionToolResult message, patching or
+// dropping as needed so the downstream model never receives an unpaired
+// tool-call history.
+//
+// In the AgenticMessage protocol:
+//   - Assistant tool calls: Role=AgenticRoleTypeAssistant with FunctionToolCall content blocks.
+//   - Tool results: Role=AgenticRoleTypeUser with FunctionToolResult content blocks.
+//
+// This middleware runs after summarization which may truncate history and
+// break pairings.
+type agenticToolPairReconcilerMiddleware struct {
+	*adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]
+	logger *zap.Logger
+	phase  string
+}
+
+func newAgenticToolPairReconcilerMiddleware(logger *zap.Logger, phase string) adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage] {
+	return &agenticToolPairReconcilerMiddleware{
+		TypedBaseChatModelAgentMiddleware: &adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]{},
+		logger:                            logger,
+		phase:                             phase,
+	}
+}
+
+func (m *agenticToolPairReconcilerMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.TypedChatModelAgentState[*schema.AgenticMessage],
+	mc *adk.TypedModelContext[*schema.AgenticMessage],
+) (context.Context, *adk.TypedChatModelAgentState[*schema.AgenticMessage], error) {
+	_ = mc
+	if m == nil || state == nil || len(state.Messages) == 0 {
+		return ctx, state, nil
+	}
+
+	usedIDs := make(map[string]struct{}, 16)
+	changed := false
+	patched := 0
+	dropped := 0
+	out := make([]*schema.AgenticMessage, 0, len(state.Messages))
+
+	for i := 0; i < len(state.Messages); {
+		msg := state.Messages[i]
+		if msg == nil {
+			changed = true
+			i++
+			continue
+		}
+
+		calls := agenticFunctionToolCalls(msg)
+
+		// Non-assistant or assistant without tool calls — but check for orphan
+		// tool-result messages (user role with only FunctionToolResult blocks).
+		if len(calls) == 0 {
+			if isPureAgenticToolResult(msg) {
+				// Orphan tool result not preceded by its assistant; drop it.
+				changed = true
+				dropped++
+				i++
+				continue
+			}
+			out = append(out, msg)
+			i++
+			continue
+		}
+
+		// Deduplicate / fix empty call IDs.
+		idsChanged := false
+		for ci := range calls {
+			id := calls[ci].CallID
+			_, duplicate := usedIDs[id]
+			if id == "" || duplicate {
+				base := fmt.Sprintf("patched_agentic_call_%d_%d", i, ci)
+				id = base
+				for suffix := 1; ; suffix++ {
+					if _, exists := usedIDs[id]; !exists {
+						break
+					}
+					id = fmt.Sprintf("%s_%d", base, suffix)
+				}
+				calls[ci].CallID = id
+				idsChanged = true
+				changed = true
+			}
+			usedIDs[id] = struct{}{}
+		}
+
+		assistant := msg
+		if idsChanged {
+			assistant = cloneAgenticMessageWithCalls(msg, calls)
+		}
+		out = append(out, assistant)
+
+		// Build expected set.
+		expected := make(map[string]*schema.FunctionToolCall, len(calls))
+		for ci := range calls {
+			expected[calls[ci].CallID] = calls[ci]
+		}
+
+		// Consume following tool-result messages.
+		results := make(map[string]*schema.AgenticMessage, len(calls))
+		j := i + 1
+		for j < len(state.Messages) {
+			next := state.Messages[j]
+			if next == nil {
+				changed = true
+				j++
+				continue
+			}
+			if !isPureAgenticToolResult(next) {
+				break
+			}
+			resultCallIDs := agenticToolResultCallIDs(next)
+			consumed := false
+			for _, rid := range resultCallIDs {
+				if _, wanted := expected[rid]; !wanted {
+					continue
+				}
+				if _, dup := results[rid]; dup {
+					continue
+				}
+				results[rid] = next
+				consumed = true
+			}
+			if !consumed {
+				changed = true
+				dropped++
+			}
+			j++
+		}
+
+		// Emit results in call order, patching missing ones.
+		for _, tc := range calls {
+			if result, ok := results[tc.CallID]; ok {
+				out = append(out, result)
+				continue
+			}
+			out = append(out, makeAgenticPatchedToolResult(tc.CallID, tc.Name))
+			changed = true
+			patched++
+		}
+		i = j
+	}
+
+	if !changed {
+		return ctx, state, nil
+	}
+	if m.logger != nil {
+		m.logger.Warn("agentic tool-call/result pairs reconciled before model call",
+			zap.String("phase", m.phase),
+			zap.Int("patched_results", patched),
+			zap.Int("dropped_results", dropped),
+			zap.Int("messages_before", len(state.Messages)),
+			zap.Int("messages_after", len(out)),
+		)
+	}
+	ns := *state
+	ns.Messages = out
+	return ctx, &ns, nil
+}
+
+// agenticFunctionToolCalls extracts FunctionToolCall pointers from an
+// assistant message's content blocks. Returns nil for non-assistant messages.
+func agenticFunctionToolCalls(msg *schema.AgenticMessage) []*schema.FunctionToolCall {
+	if msg == nil || msg.Role != schema.AgenticRoleTypeAssistant {
+		return nil
+	}
+	var out []*schema.FunctionToolCall
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolCall != nil {
+			out = append(out, block.FunctionToolCall)
+		}
+	}
+	return out
+}
+
+// isPureAgenticToolResult returns true when the message is a user-role
+// message whose content blocks are exclusively FunctionToolResult entries.
+func isPureAgenticToolResult(msg *schema.AgenticMessage) bool {
+	if msg == nil || msg.Role != schema.AgenticRoleTypeUser || len(msg.ContentBlocks) == 0 {
+		return false
+	}
+	for _, block := range msg.ContentBlocks {
+		if block == nil {
+			continue
+		}
+		if block.FunctionToolResult == nil {
+			return false
+		}
+	}
+	return true
+}
+
+// agenticToolResultCallIDs extracts all CallIDs from FunctionToolResult blocks.
+func agenticToolResultCallIDs(msg *schema.AgenticMessage) []string {
+	if msg == nil {
+		return nil
+	}
+	var ids []string
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolResult != nil && block.FunctionToolResult.CallID != "" {
+			ids = append(ids, block.FunctionToolResult.CallID)
+		}
+	}
+	return ids
+}
+
+func cloneAgenticMessageWithCalls(msg *schema.AgenticMessage, calls []*schema.FunctionToolCall) *schema.AgenticMessage {
+	cloned := *msg
+	cloned.ContentBlocks = make([]*schema.ContentBlock, 0, len(msg.ContentBlocks))
+	callIdx := 0
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolCall != nil && callIdx < len(calls) {
+			cloned.ContentBlocks = append(cloned.ContentBlocks, schema.NewContentBlock(calls[callIdx]))
+			callIdx++
+		} else {
+			cloned.ContentBlocks = append(cloned.ContentBlocks, block)
+		}
+	}
+	return &cloned
+}
+
+func makeAgenticPatchedToolResult(callID, name string) *schema.AgenticMessage {
+	return &schema.AgenticMessage{
+		Role: schema.AgenticRoleTypeUser,
+		ContentBlocks: []*schema.ContentBlock{schema.NewContentBlock(&schema.FunctionToolResult{
+			CallID: callID,
+			Name:   name,
+			Content: []*schema.FunctionToolResultContentBlock{{
+				Type: schema.FunctionToolResultContentBlockTypeText,
+				Text: &schema.UserInputText{Text: patchedMissingToolResult},
+			}},
+		})},
+	}
+}

internal/multiagent/agentic_tool_pair_reconciler_middleware_test.go

@@ -0,0 +1,157 @@
+package multiagent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestAgenticToolPairReconcilerPatchesMissing(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{"q":"x"}`),
+			agenticAssistantToolCall("c2", "execute", `{"cmd":"ls"}`),
+			// c1 result present, c2 missing
+			agenticToolResult("c1", "search", "found it"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Expected: assistant(c1) -> result(c1) -> assistant(c2) -> patched_result(c2)
+	if len(out.Messages) != 4 {
+		t.Fatalf("messages = %d, want 4", len(out.Messages))
+	}
+	// c1 assistant
+	if calls := agenticFunctionToolCalls(out.Messages[0]); len(calls) != 1 || calls[0].CallID != "c1" {
+		t.Fatal("msg[0] should be assistant(c1)")
+	}
+	// c1 result
+	if ids := agenticToolResultCallIDs(out.Messages[1]); len(ids) != 1 || ids[0] != "c1" {
+		t.Fatal("msg[1] should be result(c1)")
+	}
+	// c2 assistant
+	if calls := agenticFunctionToolCalls(out.Messages[2]); len(calls) != 1 || calls[0].CallID != "c2" {
+		t.Fatal("msg[2] should be assistant(c2)")
+	}
+	// c2 patched result
+	if ids := agenticToolResultCallIDs(out.Messages[3]); len(ids) != 1 || ids[0] != "c2" {
+		t.Fatal("msg[3] should be patched result(c2)")
+	}
+	resultText := out.Messages[3].ContentBlocks[0].FunctionToolResult.Content[0].Text.Text
+	if resultText != patchedMissingToolResult {
+		t.Fatalf("patched text = %q", resultText)
+	}
+}
+
+func TestAgenticToolPairReconcilerDropsOrphan(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			// Orphan tool result with no preceding assistant
+			agenticToolResult("orphan", "deleted_tool", "stale data"),
+			{Role: schema.AgenticRoleTypeUser, ContentBlocks: []*schema.ContentBlock{
+				schema.NewContentBlock(&schema.UserInputText{Text: "hello"}),
+			}},
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 1 {
+		t.Fatalf("messages = %d, want 1 (orphan dropped)", len(out.Messages))
+	}
+	if out.Messages[0].ContentBlocks[0].UserInputText == nil {
+		t.Fatal("remaining message should be the user text")
+	}
+}
+
+func TestAgenticToolPairReconcilerNoopWhenPaired(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Should return original state unchanged
+	if &out.Messages[0] == &state.Messages[0] {
+		// pointer equality on slice — state not cloned
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2", len(out.Messages))
+	}
+}
+
+func TestAgenticToolPairReconcilerFixesEmptyCallID(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			{
+				Role: schema.AgenticRoleTypeAssistant,
+				ContentBlocks: []*schema.ContentBlock{schema.NewContentBlock(&schema.FunctionToolCall{
+					CallID: "", Name: "search", Arguments: `{}`,
+				})},
+			},
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	calls := agenticFunctionToolCalls(out.Messages[0])
+	if len(calls) != 1 || calls[0].CallID == "" {
+		t.Fatalf("empty call ID should be patched, got %q", calls[0].CallID)
+	}
+}
+
+func TestAgenticOrphanToolPrunerRemovesOrphan(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticOrphanToolPrunerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+			// Orphan: no assistant has call_id "c_orphan"
+			agenticToolResult("c_orphan", "deleted", "stale"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2 (orphan pruned)", len(out.Messages))
+	}
+}
+
+func TestAgenticOrphanToolPrunerNoopWhenClean(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticOrphanToolPrunerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2", len(out.Messages))
+	}
+}

internal/multiagent/eino_adk_run_loop.go

@@ -48,6 +48,48 @@ func isInterruptContinue(ctx context.Context) bool {
 	return errors.Is(context.Cause(ctx), ErrInterruptContinue)
 }
 
+func isEinoStreamCanceled(err error) bool {
+	if err == nil {
+		return false
+	}
+	if errors.Is(err, adk.ErrStreamCanceled) {
+		return true
+	}
+	var streamCanceled *adk.StreamCanceledError
+	return errors.As(err, &streamCanceled)
+}
+
+func isEinoCancelError(err error) bool {
+	if err == nil {
+		return false
+	}
+	var cancelErr *adk.CancelError
+	return errors.As(err, &cancelErr)
+}
+
+// isEinoVoluntaryCancelErr reports cancel signals produced by Agent Cancel /
+// TurnLoop preempt (CancelError, ErrStreamCanceled, context.Canceled).
+func isEinoVoluntaryCancelErr(err error) bool {
+	if err == nil {
+		return false
+	}
+	return isEinoCancelError(err) || isEinoStreamCanceled(err) || errors.Is(err, context.Canceled)
+}
+
+// isEinoTurnLoopPreemptErr is true when a cancel/stream-cancel leaked from the
+// current agent turn while the host task context is still alive. TurnLoop
+// interrupt-continue does not cancel the parent context; treating that leak as
+// fatal would abort the whole run instead of starting the queued next turn.
+func isEinoTurnLoopPreemptErr(ctx context.Context, err error) bool {
+	if err == nil || !isEinoVoluntaryCancelErr(err) {
+		return false
+	}
+	if ctx != nil && ctx.Err() != nil {
+		return false
+	}
+	return true
+}
+
 func isEinoIterationLimitError(err error) bool {
 	if err == nil {
 		return false
@@ -401,22 +443,41 @@ func nextAgentEventWithContext(ctx context.Context, iter *adk.AsyncIterator[*adk
 
 // recvSchemaMessageStream 消费 ADK Tool 流式结果；ctx 取消时立即返回，避免 amass 等无输出时永久阻塞。
 func recvSchemaMessageStream(ctx context.Context, stream *schema.StreamReader[*schema.Message]) (content, toolCallID, toolName string, recvErr error) {
-	if stream == nil {
-		return "", "", "", nil
+	msgs, recvErr := recvSchemaToolResultMessages(ctx, stream)
+	if len(msgs) == 0 {
+		return "", "", "", recvErr
 	}
-	var buf strings.Builder
-	recvErr = recvEinoSchemaMessageStreamWithContext(ctx, stream, 8, func(chunk *schema.Message) {
-		if chunk.Content != "" {
-			buf.WriteString(chunk.Content)
+	parts := make([]string, 0, len(msgs))
+	for _, msg := range msgs {
+		if msg == nil {
+			continue
 		}
-		if tid := strings.TrimSpace(chunk.ToolCallID); tid != "" {
-			toolCallID = tid
+		parts = append(parts, msg.Content)
+		if id := strings.TrimSpace(msg.ToolCallID); id != "" {
+			toolCallID = id
 		}
-		if name := strings.TrimSpace(chunk.ToolName); name != "" {
+		if name := strings.TrimSpace(msg.ToolName); name != "" {
 			toolName = name
 		}
+	}
+	return strings.Join(parts, ""), toolCallID, toolName, recvErr
+}
+
+// recvSchemaToolResultMessages 先收齐 Tool 流，再用 Eino ConcatMessages 合并。
+// EventSender 一 call 一条流时走 ConcatMessages；并行结果被摊平进同一条流时按 CallID 分列再合并。
+func recvSchemaToolResultMessages(ctx context.Context, stream *schema.StreamReader[*schema.Message]) (msgs []*schema.Message, recvErr error) {
+	if stream == nil {
+		return nil, nil
+	}
+	var chunks []*schema.Message
+	recvErr = recvEinoSchemaMessageStreamWithContext(ctx, stream, 8, func(chunk *schema.Message) {
+		chunks = append(chunks, chunk)
 	})
-	return buf.String(), toolCallID, toolName, recvErr
+	msgs, concatErr := concatToolResultChunks(chunks)
+	if concatErr != nil && recvErr == nil {
+		return nil, concatErr
+	}
+	return msgs, recvErr
 }
 
 func buildEinoCheckpointID(orchMode string) string {

internal/multiagent/eino_adk_run_loop_stream_test.go

@@ -30,6 +30,29 @@ func TestRecvSchemaMessageStream_EOF(t *testing.T) {
 	}
 }
 
+func TestRecvSchemaToolResultMessages_SplitsParallelIDs(t *testing.T) {
+	sr, sw := schema.Pipe[*schema.Message](8)
+	_ = sw.Send(schema.ToolMessage("one-", "tc-1", schema.WithToolName("nmap")), nil)
+	_ = sw.Send(schema.ToolMessage("two-", "tc-2", schema.WithToolName("nmap")), nil)
+	_ = sw.Send(schema.ToolMessage("a", "tc-1", schema.WithToolName("nmap")), nil)
+	_ = sw.Send(schema.ToolMessage("b", "tc-2", schema.WithToolName("nmap")), nil)
+	sw.Close()
+
+	msgs, err := recvSchemaToolResultMessages(context.Background(), sr)
+	if err != nil {
+		t.Fatalf("unexpected err: %v", err)
+	}
+	if len(msgs) != 2 {
+		t.Fatalf("msgs = %#v, want 2", msgs)
+	}
+	if msgs[0].ToolCallID != "tc-1" || msgs[0].Content != "one-a" {
+		t.Fatalf("msg 0 = %#v", msgs[0])
+	}
+	if msgs[1].ToolCallID != "tc-2" || msgs[1].Content != "two-b" {
+		t.Fatalf("msg 1 = %#v", msgs[1])
+	}
+}
+
 func TestRecvSchemaMessageStream_CapturesToolName(t *testing.T) {
 	sr, sw := schema.Pipe[*schema.Message](4)
 	_ = sw.Send(schema.ToolMessage("hello", "tc-1", schema.WithToolName("execute")), nil)

internal/multiagent/eino_agentic_chat_model_tail_middleware.go

@@ -20,8 +20,13 @@ func appendEinoAgenticChatModelTailMiddlewares(
 	handlers = append(handlers, newAgenticSystemMessageNormalizerMiddleware(cfg.logger, cfg.phase))
 	handlers = append(handlers, newAgenticContinuationUserDedupMiddleware(cfg.logger, cfg.phase))
 	if cfg.agenticSummarization != nil {
+		handlers = append(handlers, newAgenticToolPairReconcilerMiddleware(cfg.logger, cfg.phase+"_pre_summarization"))
 		handlers = append(handlers, cfg.agenticSummarization)
 	}
+	handlers = append(handlers, newAgenticToolPairReconcilerMiddleware(cfg.logger, cfg.phase))
+	if !cfg.skipOrphanPruner {
+		handlers = append(handlers, newAgenticOrphanToolPrunerMiddleware(cfg.logger, cfg.phase))
+	}
 	if !cfg.skipTrace && cfg.trace != nil {
 		if capMw := newAgenticModelFacingTraceMiddleware(cfg.trace); capMw != nil {
 			handlers = append(handlers, capMw)

internal/multiagent/eino_agentic_chat_model_tail_middleware_test.go

@@ -106,7 +106,8 @@ func TestAppendEinoAgenticChatModelTailMiddlewares(t *testing.T) {
 		phase: "agentic",
 		trace: holder,
 	})
-	if len(handlers) != 3 {
-		t.Fatalf("handlers = %d, want system + continuation + trace", len(handlers))
+	// system + continuation + reconciler + orphan_pruner + trace
+	if len(handlers) != 5 {
+		t.Fatalf("handlers = %d, want system + continuation + reconciler + orphan_pruner + trace", len(handlers))
 	}
 }

internal/multiagent/eino_agentic_event_adapter.go

@@ -31,6 +31,11 @@ func adaptAgenticEventToEinoEvents(ev *adk.TypedAgentEvent[*schema.AgenticMessag
 		return []*adk.AgentEvent{base(&adk.AgentOutput{CustomizedOutput: customized})}
 	}
 	if mv.IsStreaming {
+		// Tool 流保持 1 event ↔ 1 MessageStream，对齐 ADK EventSenderToolWrapper：
+		// 每个 CallID 在工具包装层就已经是独立事件。这里不能再按 CallID 现场拆成
+		// 多条 live pipe——drain 会阻塞读完当前流，交错的并行 chunk 会把另一列写满后死锁。
+		// 若上游仍把 ToolsNode 的 MergeStreamReaders 摊成一条流，由
+		// concatToolResultChunks 按列 ConcatMessages 恢复。
 		return []*adk.AgentEvent{base(&adk.AgentOutput{
 			MessageOutput: &adk.MessageVariant{
 				IsStreaming:   true,

internal/multiagent/eino_agentic_message.go

@@ -110,10 +110,16 @@ func AgenticMessageToEino(msg *schema.AgenticMessage) []*schema.Message {
 			base.Role = schema.Assistant
 			base.Content += block.AssistantGenText.Text
 		case block.FunctionToolCall != nil:
+			var index *int
+			if block.StreamingMeta != nil {
+				i := block.StreamingMeta.Index
+				index = &i
+			}
 			base.Role = schema.Assistant
 			base.ToolCalls = append(base.ToolCalls, schema.ToolCall{
-				ID:   block.FunctionToolCall.CallID,
-				Type: "function",
+				Index: index,
+				ID:    block.FunctionToolCall.CallID,
+				Type:  "function",
 				Function: schema.FunctionCall{
 					Name:      block.FunctionToolCall.Name,
 					Arguments: block.FunctionToolCall.Arguments,

internal/multiagent/eino_agentic_summarize.go

@@ -10,7 +10,6 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/database"
 
-	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/adk/middlewares/summarization"
 	"github.com/cloudwego/eino/components/model"
@@ -109,9 +108,7 @@ func newEinoAgenticSummarizationMiddleware(
 	retryPolicy := einoTransientRunRetryPolicyFromMW(mwCfg)
 	retryMax := retryPolicy.maxAttempts
 	var summaryOverflowRetries int
-	summaryModelOpts := []model.Option{
-		einoopenai.WithMaxCompletionTokens(outputReserve),
-	}
+	summaryModelOpts := newEinoSummarizationModelOptions(outputReserve, modelName, "agentic", &appCfg.OpenAI, logger)
 
 	mw, err := summarization.NewTyped[*schema.AgenticMessage](ctx, &summarization.TypedConfig[*schema.AgenticMessage]{
 		Model:        summaryModel,

internal/multiagent/eino_agentic_tool_calling_adapter.go

@@ -0,0 +1,118 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+)
+
+// agenticToolCallingChatModelAdapter lets Eino's classic plan_execute
+// Planner/Replanner consume a native AgenticModel without translating the HTTP
+// protocol. Only the in-memory Eino message and option shapes are adapted.
+type agenticToolCallingChatModelAdapter struct {
+	model model.AgenticModel
+	tools []*schema.ToolInfo
+}
+
+func newAgenticToolCallingChatModelAdapter(agenticModel model.AgenticModel) model.ToolCallingChatModel {
+	return &agenticToolCallingChatModelAdapter{model: agenticModel}
+}
+
+func (m *agenticToolCallingChatModelAdapter) WithTools(tools []*schema.ToolInfo) (model.ToolCallingChatModel, error) {
+	if m == nil || m.model == nil {
+		return nil, fmt.Errorf("agentic tool-calling adapter: model is nil")
+	}
+	clonedTools := append([]*schema.ToolInfo(nil), tools...)
+	return &agenticToolCallingChatModelAdapter{model: m.model, tools: clonedTools}, nil
+}
+
+func (m *agenticToolCallingChatModelAdapter) Generate(
+	ctx context.Context,
+	input []*schema.Message,
+	opts ...model.Option,
+) (*schema.Message, error) {
+	if m == nil || m.model == nil {
+		return nil, fmt.Errorf("agentic tool-calling adapter: model is nil")
+	}
+	out, err := m.model.Generate(ctx, EinoMessagesToAgentic(input), m.agenticOptions(opts...)...)
+	if err != nil {
+		return nil, err
+	}
+	converted := AgenticMessageToEino(out)
+	if len(converted) == 0 {
+		return nil, fmt.Errorf("agentic tool-calling adapter: model returned no message")
+	}
+	return converted[0], nil
+}
+
+func (m *agenticToolCallingChatModelAdapter) Stream(
+	ctx context.Context,
+	input []*schema.Message,
+	opts ...model.Option,
+) (*schema.StreamReader[*schema.Message], error) {
+	if m == nil || m.model == nil {
+		return nil, fmt.Errorf("agentic tool-calling adapter: model is nil")
+	}
+	stream, err := m.model.Stream(ctx, EinoMessagesToAgentic(input), m.agenticOptions(opts...)...)
+	if err != nil {
+		return nil, err
+	}
+	return agenticStreamToEinoStream(stream), nil
+}
+
+func (m *agenticToolCallingChatModelAdapter) agenticOptions(opts ...model.Option) []model.Option {
+	common := model.GetCommonOptions(&model.Options{
+		Tools: append([]*schema.ToolInfo(nil), m.tools...),
+	}, opts...)
+	out := make([]model.Option, 0, 8)
+	if common.Temperature != nil {
+		out = append(out, model.WithTemperature(*common.Temperature))
+	}
+	if common.Model != nil {
+		out = append(out, model.WithModel(*common.Model))
+	}
+	if common.TopP != nil {
+		out = append(out, model.WithTopP(*common.TopP))
+	}
+	if common.MaxTokens != nil {
+		out = append(out, model.WithMaxTokens(*common.MaxTokens))
+	}
+	if len(common.Stop) > 0 {
+		out = append(out, model.WithStop(common.Stop))
+	}
+	if common.Tools != nil {
+		out = append(out, model.WithTools(common.Tools))
+	}
+	if common.AgenticToolChoice != nil {
+		out = append(out, model.WithAgenticToolChoice(common.AgenticToolChoice))
+	} else if common.ToolChoice != nil {
+		out = append(out, model.WithAgenticToolChoice(classicToolChoiceToAgentic(
+			*common.ToolChoice,
+			common.AllowedToolNames,
+		)))
+	}
+	return out
+}
+
+func classicToolChoiceToAgentic(choice schema.ToolChoice, allowedNames []string) *schema.AgenticToolChoice {
+	allowed := make([]*schema.AllowedTool, 0, len(allowedNames))
+	for _, name := range allowedNames {
+		if name != "" {
+			allowed = append(allowed, &schema.AllowedTool{FunctionName: name})
+		}
+	}
+	out := &schema.AgenticToolChoice{Type: choice}
+	switch choice {
+	case schema.ToolChoiceAllowed:
+		if len(allowed) > 0 {
+			out.Allowed = &schema.AgenticAllowedToolChoice{Tools: allowed}
+		}
+	case schema.ToolChoiceForced:
+		if len(allowed) > 0 {
+			out.Forced = &schema.AgenticForcedToolChoice{Tools: allowed}
+		}
+	}
+	return out
+}

internal/multiagent/eino_agentic_tool_calling_adapter_test.go

@@ -0,0 +1,118 @@
+package multiagent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+)
+
+type capturingAgenticToolCallingModel struct {
+	input   []*schema.AgenticMessage
+	options *model.Options
+}
+
+func (m *capturingAgenticToolCallingModel) Generate(
+	_ context.Context,
+	input []*schema.AgenticMessage,
+	opts ...model.Option,
+) (*schema.AgenticMessage, error) {
+	m.input = input
+	m.options = model.GetCommonOptions(nil, opts...)
+	return &schema.AgenticMessage{
+		Role: schema.AgenticRoleTypeAssistant,
+		ContentBlocks: []*schema.ContentBlock{
+			schema.NewContentBlock(&schema.FunctionToolCall{
+				CallID:    "call-1",
+				Name:      "emit_plan",
+				Arguments: `{"steps":["inspect"]}`,
+			}),
+		},
+	}, nil
+}
+
+func (m *capturingAgenticToolCallingModel) Stream(
+	context.Context,
+	[]*schema.AgenticMessage,
+	...model.Option,
+) (*schema.StreamReader[*schema.AgenticMessage], error) {
+	return schema.StreamReaderFromArray([]*schema.AgenticMessage{
+		{
+			Role: schema.AgenticRoleTypeAssistant,
+			ContentBlocks: []*schema.ContentBlock{
+				schema.NewContentBlockChunk(&schema.FunctionToolCall{
+					CallID:    "call-1",
+					Name:      "emit_plan",
+					Arguments: `{"steps":[`,
+				}, &schema.StreamingMeta{Index: 0}),
+			},
+		},
+		{
+			Role: schema.AgenticRoleTypeAssistant,
+			ContentBlocks: []*schema.ContentBlock{
+				schema.NewContentBlockChunk(&schema.FunctionToolCall{
+					Arguments: `"inspect"]}`,
+				}, &schema.StreamingMeta{Index: 0}),
+			},
+		},
+	}), nil
+}
+
+func TestAgenticToolCallingAdapterConvertsForcedToolChoice(t *testing.T) {
+	t.Parallel()
+	native := &capturingAgenticToolCallingModel{}
+	adapter, err := newAgenticToolCallingChatModelAdapter(native).WithTools([]*schema.ToolInfo{{
+		Name: "emit_plan",
+		Desc: "emit a structured plan",
+	}})
+	if err != nil {
+		t.Fatalf("WithTools: %v", err)
+	}
+
+	out, err := adapter.Generate(
+		context.Background(),
+		[]*schema.Message{schema.UserMessage("plan this task")},
+		model.WithToolChoice(schema.ToolChoiceForced),
+	)
+	if err != nil {
+		t.Fatalf("Generate: %v", err)
+	}
+	if len(native.input) != 1 || native.input[0].Role != schema.AgenticRoleTypeUser {
+		t.Fatalf("native input = %#v", native.input)
+	}
+	if native.options == nil || len(native.options.Tools) != 1 || native.options.Tools[0].Name != "emit_plan" {
+		t.Fatalf("native tools = %#v", native.options)
+	}
+	if native.options.AgenticToolChoice == nil || native.options.AgenticToolChoice.Type != schema.ToolChoiceForced {
+		t.Fatalf("agentic tool choice = %#v", native.options.AgenticToolChoice)
+	}
+	if len(out.ToolCalls) != 1 || out.ToolCalls[0].Function.Name != "emit_plan" {
+		t.Fatalf("classic output = %#v", out)
+	}
+}
+
+func TestAgenticToolCallingAdapterPreservesStreamingToolCallIndex(t *testing.T) {
+	t.Parallel()
+	adapter := newAgenticToolCallingChatModelAdapter(&capturingAgenticToolCallingModel{})
+	stream, err := adapter.Stream(context.Background(), []*schema.Message{
+		schema.UserMessage("plan this task"),
+	})
+	if err != nil {
+		t.Fatalf("Stream: %v", err)
+	}
+	out, err := schema.ConcatMessageStream(stream)
+	if err != nil {
+		t.Fatalf("ConcatMessageStream: %v", err)
+	}
+	if len(out.ToolCalls) != 1 {
+		t.Fatalf("tool calls = %#v, want one merged call", out.ToolCalls)
+	}
+	call := out.ToolCalls[0]
+	if call.Index == nil || *call.Index != 0 {
+		t.Fatalf("tool call index = %#v", call.Index)
+	}
+	if call.Function.Name != "emit_plan" || call.Function.Arguments != `{"steps":["inspect"]}` {
+		t.Fatalf("merged tool call = %#v", call)
+	}
+}

internal/multiagent/eino_assistant_stream_event_handler.go

@@ -2,7 +2,6 @@ package multiagent
 
 import (
 	"context"
-	"errors"
 
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/schema"
@@ -139,7 +138,7 @@ func (h *einoAssistantStreamEventHandler) Handle(mv *adk.MessageVariant, agentNa
 			streamUsage = maxEinoTokenUsage(streamUsage, chunk.ResponseMeta.Usage)
 		}
 	})
-	if recvErr != nil && !errors.Is(recvErr, context.Canceled) && h.logger != nil {
+	if recvErr != nil && !isEinoVoluntaryCancelErr(recvErr) && h.logger != nil {
 		h.logger.Warn("eino stream recv error, flushing incomplete stream",
 			zap.Error(recvErr),
 			zap.String("agent", agentName),

internal/multiagent/eino_exit_fallback_test.go

@@ -55,6 +55,70 @@ func TestEinoExtractFallbackAssistantFromMsgs_prefersToolOverEarlierAssistant(t
 	}
 }
 
+func TestEinoExtractFallbackAssistantFromMsgs_plainAssistant(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("plain answer", nil),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "plain answer" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestEinoExtractFallbackAssistantFromMsgs_finalAssistantAfterToolResult(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("", []schema.ToolCall{{
+			ID:   "call-1",
+			Type: "function",
+			Function: schema.FunctionCall{
+				Name:      "execute",
+				Arguments: `{"command":"pwd"}`,
+			},
+		}}),
+		schema.ToolMessage("/tmp", "call-1", schema.WithToolName("execute")),
+		schema.AssistantMessage("final after tool", nil),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "final after tool" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestEinoExtractFallbackAssistantFromMsgs_doesNotUseAssistantBeforeUnfinishedToolResult(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("I will inspect that.", nil),
+		schema.AssistantMessage("", []schema.ToolCall{{
+			ID:   "call-1",
+			Type: "function",
+			Function: schema.FunctionCall{
+				Name:      "execute",
+				Arguments: `{"command":"pwd"}`,
+			},
+		}}),
+		schema.ToolMessage("/tmp", "call-1", schema.WithToolName("execute")),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "" {
+		t.Fatalf("got %q, want empty", got)
+	}
+}
+
+func TestEinoRunResultBuilderFinalFallsBackToPlainAssistantTrace(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator(nil)
+	runMessages.Append(schema.UserMessage("hi"))
+	runMessages.Append(schema.AssistantMessage("plain answer", nil))
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "deep",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "plain answer" {
+		t.Fatalf("response = %q, want plain answer", got.Response)
+	}
+}
+
 func toolExitMsg(content, callID string) *schema.Message {
 	m := schema.ToolMessage(content, callID)
 	m.ToolName = "exit"

internal/multiagent/eino_filesystem_tool_monitor.go

@@ -64,6 +64,17 @@ func toolCallArgsFromAccumulated(msgs []adk.Message, toolCallID, expectToolName
 	return map[string]interface{}{}
 }
 
+func mustMarshalToolArguments(args map[string]interface{}) string {
+	if len(args) == 0 {
+		return "{}"
+	}
+	raw, err := json.Marshal(args)
+	if err != nil {
+		return "{}"
+	}
+	return string(raw)
+}
+
 // beginEinoADKFilesystemToolMonitor 在 Eino ADK filesystem 工具开始调用时写入 running 状态。
 func beginEinoADKFilesystemToolMonitor(
 	ctx context.Context,
@@ -71,6 +82,7 @@ func beginEinoADKFilesystemToolMonitor(
 	rec einomcp.ExecutionRecorder,
 	binder *MCPExecutionBinder,
 	toolCallID, toolName string,
+	args map[string]interface{},
 ) {
 	if ag == nil || rec == nil {
 		return
@@ -87,7 +99,7 @@ func beginEinoADKFilesystemToolMonitor(
 		return
 	}
 	storedName := "eino_fs::" + strings.ToLower(name)
-	id := ag.BeginLocalToolExecution(ctx, storedName, map[string]interface{}{})
+	id := ag.BeginLocalToolExecution(ctx, storedName, args)
 	if id == "" {
 		return
 	}
@@ -108,18 +120,21 @@ func recordEinoADKFilesystemToolMonitor(
 	msgs []adk.Message,
 	resultText string,
 	isErr bool,
-) {
+) string {
 	if ag == nil || rec == nil {
-		return
+		return ""
 	}
 	name := strings.TrimSpace(toolName)
 	if name == "" || strings.EqualFold(name, "execute") {
-		return
+		return ""
 	}
 	if !isBuiltinEinoADKFilesystemToolName(name) {
-		return
+		return ""
 	}
 	args := toolCallArgsFromAccumulated(msgs, toolCallID, name)
+	if len(args) == 0 && binder != nil {
+		args = binder.Arguments(toolCallID)
+	}
 	storedName := "eino_fs::" + strings.ToLower(name)
 	var invErr error
 	if isErr {
@@ -138,4 +153,5 @@ func recordEinoADKFilesystemToolMonitor(
 	if id != "" && execID == "" {
 		rec(id, toolCallID)
 	}
+	return id
 }

internal/multiagent/eino_filesystem_tool_monitor_test.go

@@ -2,6 +2,7 @@ package multiagent
 
 import (
 	"context"
+	"strings"
 	"testing"
 
 	"cyberstrike-ai/internal/agent"
@@ -10,6 +11,7 @@ import (
 	"cyberstrike-ai/internal/mcp"
 
 	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/components/tool"
 	"github.com/cloudwego/eino/schema"
 	"go.uber.org/zap"
 )
@@ -26,7 +28,7 @@ func TestEinoADKFilesystemToolMonitorBindsFinishesAndUpdatesDisplayResult(t *tes
 		recorded = append(recorded, executionID+"|"+toolCallID)
 	})
 
-	beginEinoADKFilesystemToolMonitor(ctx, ag, rec, binder, "call-read", "read_file")
+	beginEinoADKFilesystemToolMonitor(ctx, ag, rec, binder, "call-read", "read_file", map[string]interface{}{"path": "/tmp/secret.txt"})
 	execID := binder.ExecutionID("call-read")
 	if execID == "" {
 		t.Fatal("expected begin to bind execution id")
@@ -38,6 +40,9 @@ func TestEinoADKFilesystemToolMonitorBindsFinishesAndUpdatesDisplayResult(t *tes
 	if len(recorded) != 1 || recorded[0] != execID+"|call-read" {
 		t.Fatalf("recorded begin ids = %#v", recorded)
 	}
+	if got, _ := exec.Arguments["path"].(string); got != "/tmp/secret.txt" {
+		t.Fatalf("begin execution args = %#v", exec.Arguments)
+	}
 
 	runMessages := newEinoRunMessageAccumulator([]adk.Message{
 		&schema.Message{
@@ -80,3 +85,86 @@ func TestEinoADKFilesystemToolMonitorBindsFinishesAndUpdatesDisplayResult(t *tes
 		t.Fatalf("finish should reuse existing execution without recording a second id, got %#v", recorded)
 	}
 }
+
+func TestEinoADKFilesystemToolMonitorSpillsLargeReadFileResultForProgress(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	logger := zap.NewNop()
+	server := mcp.NewServer(logger)
+	server.ConfigureToolResultMaxBytes(400)
+	server.ConfigureToolResultSpillRoot(t.TempDir())
+	ag := agent.NewAgent(&config.OpenAIConfig{}, &config.AgentConfig{}, server, nil, logger, 1)
+	binder := NewMCPExecutionBinder()
+	rec := einomcp.ExecutionRecorder(func(executionID, toolCallID string) {})
+	var event map[string]interface{}
+
+	runMessages := newEinoRunMessageAccumulator([]adk.Message{
+		&schema.Message{
+			Role: schema.Assistant,
+			ToolCalls: []schema.ToolCall{{
+				ID:   "call-read",
+				Type: "function",
+				Function: schema.FunctionCall{
+					Name:      "read_file",
+					Arguments: `{"path":"/tmp/large.txt"}`,
+				},
+			}},
+		},
+	})
+	beginEinoADKFilesystemToolMonitor(ctx, ag, rec, binder, "call-read", "read_file", map[string]interface{}{"path": "/tmp/large.txt"})
+	emitter := newEinoToolResultProgressEmitter(einoToolResultProgressEmitterConfig{
+		ConversationID:          "conv-1",
+		RunMessages:             runMessages,
+		FilesystemMonitorAgent:  ag,
+		FilesystemMonitorRecord: rec,
+		MCPExecutionBinder:      binder,
+		Progress: func(eventType, _ string, data interface{}) {
+			if eventType == "tool_result" {
+				event, _ = data.(map[string]interface{})
+			}
+		},
+	})
+
+	if !emitter.Emit(ctx, "read_file", strings.Repeat("0123456789", 100), "call-read", false, "lead") {
+		t.Fatal("expected tool result emit")
+	}
+	result, _ := event["result"].(string)
+	if !strings.Contains(result, "<persisted-output>") || !strings.Contains(result, "Full output saved to:") {
+		t.Fatalf("large read_file result was not spilled in progress event: %q", result)
+	}
+	if len(result) > 400 {
+		t.Fatalf("progress result exceeded configured max: len=%d text=%q", len(result), result)
+	}
+}
+
+func TestEinoAgenticFilesystemWrapperCapturesArgumentsAndSpillsResult(t *testing.T) {
+	t.Parallel()
+	binder := NewMCPExecutionBinder()
+	mw := &einoAgenticFilesystemToolMiddleware{
+		TypedChatModelAgentMiddleware: &adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]{},
+		conversationID:                "conv-1",
+		toolMaxBytes:                  400,
+		reductionRootDir:              t.TempDir(),
+		binder:                        binder,
+	}
+	endpoint, err := mw.WrapInvokableToolCall(context.Background(), func(ctx context.Context, argumentsInJSON string, opts ...tool.Option) (string, error) {
+		return strings.Repeat("0123456789", 100), nil
+	}, &adk.ToolContext{Name: "read_file", CallID: "call-read"})
+	if err != nil {
+		t.Fatalf("WrapInvokableToolCall: %v", err)
+	}
+	result, err := endpoint(context.Background(), `{"file_path":"/tmp/requirements.txt","limit":2000}`)
+	if err != nil {
+		t.Fatalf("endpoint: %v", err)
+	}
+	args := binder.Arguments("call-read")
+	if args["file_path"] != "/tmp/requirements.txt" {
+		t.Fatalf("captured args = %#v", args)
+	}
+	if !strings.Contains(result, "<persisted-output>") || !strings.Contains(result, "Full output saved to:") {
+		t.Fatalf("expected persisted-output summary, got %q", result)
+	}
+	if len(result) > 400 {
+		t.Fatalf("summary exceeded max bytes: len=%d", len(result))
+	}
+}

internal/multiagent/eino_model_resilience.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/llm"
 	"cyberstrike-ai/internal/openai"
 	"cyberstrike-ai/internal/reasoning"
 
@@ -49,7 +50,7 @@ func newEinoBaseHTTPClient() *http.Client {
 	}
 }
 
-func newEinoOpenAIChatModelFactory(
+func newEinoToolCallingChatModelFactory(
 	baseHTTPClient *http.Client,
 	reasoningClient *reasoning.ClientIntent,
 	logger *zap.Logger,
@@ -58,6 +59,13 @@ func newEinoOpenAIChatModelFactory(
 		baseHTTPClient = newEinoBaseHTTPClient()
 	}
 	return func(ctx context.Context, oa config.OpenAIConfig, mode einoModelMode) (model.ToolCallingChatModel, error) {
+		if isEinoAgenticClaudeProvider(oa.Provider) {
+			nativeModel, err := newEinoClaudeAgenticChatModel(ctx, oa, mode, baseHTTPClient, reasoningClient)
+			if err != nil {
+				return nil, err
+			}
+			return newAgenticToolCallingChatModelAdapter(nativeModel), nil
+		}
 		httpClient := openai.NewEinoHTTPClient(&oa, baseHTTPClient)
 		openai.AttachSummarizationDiagTransport(httpClient, logger)
 		maxCompletionTokens := oa.MaxCompletionTokensEffective()
@@ -81,7 +89,7 @@ func newEinoOpenAIChatModelFactory(
 	}
 }
 
-func newEinoOpenAIAgenticChatModelFactory(
+func newEinoAgenticChatModelFactory(
 	baseHTTPClient *http.Client,
 	reasoningClient *reasoning.ClientIntent,
 	logger *zap.Logger,
@@ -90,8 +98,11 @@ func newEinoOpenAIAgenticChatModelFactory(
 		baseHTTPClient = newEinoBaseHTTPClient()
 	}
 	return func(ctx context.Context, oa config.OpenAIConfig, mode einoModelMode) (model.AgenticModel, error) {
-		if !supportsEinoAgenticOpenAIBackend(oa) {
-			return nil, fmt.Errorf("eino agentic model: provider %q is not enabled for agenticopenai backend", strings.TrimSpace(oa.Provider))
+		if !supportsEinoAgenticBackend(oa) {
+			return nil, fmt.Errorf("eino agentic model: provider %q is not supported", strings.TrimSpace(oa.Provider))
+		}
+		if isEinoAgenticClaudeProvider(oa.Provider) {
+			return newEinoClaudeAgenticChatModel(ctx, oa, mode, baseHTTPClient, reasoningClient)
 		}
 		httpClient := openai.NewEinoHTTPClient(&oa, baseHTTPClient)
 		openai.AttachSummarizationDiagTransport(httpClient, logger)
@@ -111,9 +122,36 @@ func newEinoOpenAIAgenticChatModelFactory(
 	}
 }
 
-func supportsEinoAgenticOpenAIBackend(oa config.OpenAIConfig) bool {
+func newEinoClaudeAgenticChatModel(
+	ctx context.Context,
+	oa config.OpenAIConfig,
+	mode einoModelMode,
+	httpClient *http.Client,
+	reasoningClient *reasoning.ClientIntent,
+) (model.AgenticModel, error) {
+	extraFields := reasoning.AgenticOpenAIExtraFields(&oa, reasoningClient)
+	if mode == einoModelModePlanner {
+		extraFields = reasoning.AgenticOpenAIPlannerExtraFields(&oa)
+	}
+	return llm.NewClaudeAgenticModel(
+		ctx,
+		oa,
+		httpClient,
+		oa.MaxCompletionTokensEffective(),
+		extraFields,
+	)
+}
+
+func supportsEinoAgenticBackend(oa config.OpenAIConfig) bool {
 	provider := strings.ToLower(strings.TrimSpace(oa.Provider))
-	return provider == "" || provider == "openai" || provider == "openai_compatible"
+	return provider == "" ||
+		provider == "openai" ||
+		provider == "openai_compatible" ||
+		isEinoAgenticClaudeProvider(provider)
+}
+
+func isEinoAgenticClaudeProvider(provider string) bool {
+	return llm.IsClaudeProvider(provider)
 }
 
 func agenticModelGateFactory(factory einoAgenticModelConfigFactory, oa config.OpenAIConfig, mode einoModelMode) einoAgenticModelFactory {

internal/multiagent/eino_model_resilience_test.go

@@ -8,6 +8,7 @@ import (
 
 	"cyberstrike-ai/internal/config"
 
+	agenticclaude "github.com/cloudwego/eino-ext/components/model/agenticclaude"
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/components/model"
 	"github.com/cloudwego/eino/schema"
@@ -307,9 +308,9 @@ func TestNewEinoAgenticModelFailoverConfigEmitsProgressEvent(t *testing.T) {
 	}
 }
 
-func TestNewEinoOpenAIAgenticChatModelFactoryBuildsBackend(t *testing.T) {
+func TestNewEinoAgenticChatModelFactoryBuildsOpenAIBackend(t *testing.T) {
 	t.Parallel()
-	factory := newEinoOpenAIAgenticChatModelFactory(newEinoBaseHTTPClient(), nil, nil)
+	factory := newEinoAgenticChatModelFactory(newEinoBaseHTTPClient(), nil, nil)
 	m, err := factory(context.Background(), config.OpenAIConfig{
 		Provider: "openai",
 		APIKey:   "test-key",
@@ -338,16 +339,23 @@ func TestNewEinoOpenAIAgenticChatModelFactoryBuildsBackend(t *testing.T) {
 	}
 }
 
-func TestNewEinoOpenAIAgenticChatModelFactoryRejectsUnsupportedProvider(t *testing.T) {
+func TestNewEinoAgenticChatModelFactoryBuildsNativeClaudeBackend(t *testing.T) {
 	t.Parallel()
-	factory := newEinoOpenAIAgenticChatModelFactory(newEinoBaseHTTPClient(), nil, nil)
-	if _, err := factory(context.Background(), config.OpenAIConfig{
+	factory := newEinoAgenticChatModelFactory(newEinoBaseHTTPClient(), nil, nil)
+	m, err := factory(context.Background(), config.OpenAIConfig{
 		Provider: "claude",
 		APIKey:   "test-key",
 		BaseURL:  "https://api.anthropic.com/v1",
 		Model:    "claude-sonnet-4",
-	}, einoModelModeNormal); err == nil {
-		t.Fatal("expected unsupported provider error")
+	}, einoModelModeNormal)
+	if err != nil {
+		t.Fatalf("claude agentic factory: %v", err)
+	}
+	if m == nil {
+		t.Fatal("claude agentic factory returned nil model")
+	}
+	if _, ok := m.(*agenticclaude.Model); !ok {
+		t.Fatalf("claude agentic factory returned %T, want native agenticclaude.Model", m)
 	}
 	gate := evaluateEinoAgenticModelGate(agenticModelGateFactory(factory, config.OpenAIConfig{
 		Provider: "claude",
@@ -355,8 +363,25 @@ func TestNewEinoOpenAIAgenticChatModelFactoryRejectsUnsupportedProvider(t *testi
 		BaseURL:  "https://api.anthropic.com/v1",
 		Model:    "claude-sonnet-4",
 	}, einoModelModeNormal), einoAgenticRuntimeSupportV0914())
-	if gate.Ready || !containsString(gate.Missing, "model.AgenticModel backend") {
-		t.Fatalf("gate = %#v, want backend missing for unsupported provider", gate)
+	if !gate.Ready {
+		t.Fatalf("gate = %#v, want ready with native Claude backend", gate)
+	}
+}
+
+func TestNewEinoToolCallingChatModelFactoryUsesNativeClaudeAdapter(t *testing.T) {
+	t.Parallel()
+	factory := newEinoToolCallingChatModelFactory(newEinoBaseHTTPClient(), nil, nil)
+	m, err := factory(context.Background(), config.OpenAIConfig{
+		Provider: "claude",
+		APIKey:   "test-key",
+		BaseURL:  "https://api.anthropic.com",
+		Model:    "claude-sonnet-4",
+	}, einoModelModePlanner)
+	if err != nil {
+		t.Fatalf("claude planner factory: %v", err)
+	}
+	if _, ok := m.(*agenticToolCallingChatModelAdapter); !ok {
+		t.Fatalf("claude planner factory returned %T, want native agentic adapter", m)
 	}
 }
 

internal/multiagent/eino_run_error_handler.go

@@ -3,6 +3,8 @@ package multiagent
 import (
 	"context"
 	"errors"
+	"fmt"
+	"strings"
 
 	"github.com/cloudwego/eino/adk"
 )
@@ -82,12 +84,109 @@ func (h *einoRunErrorHandler) emitError(err error, kind string) {
 	if h == nil || h.progress == nil || err == nil {
 		return
 	}
+	userErr := einoUserFacingRunError(err)
 	data := map[string]interface{}{
 		"conversationId": h.conversationID,
 		"source":         "eino",
+		"error":          err.Error(),
 	}
 	if kind != "" {
 		data["errorKind"] = kind
+	} else if userErr.kind != "" {
+		data["errorKind"] = userErr.kind
 	}
-	h.progress("error", err.Error(), data)
+	if userErr.summary != "" {
+		data["errorSummary"] = userErr.summary
+	}
+	if userErr.retryExhausted {
+		data["retryExhausted"] = true
+		if userErr.totalRetries > 0 {
+			data["totalRetries"] = userErr.totalRetries
+		}
+	}
+	if userErr.rawLastError != "" {
+		data["lastError"] = userErr.rawLastError
+	}
+	if userErr.technicalError != "" {
+		data["technicalError"] = userErr.technicalError
+	}
+	if userErr.hasModelOriginalError {
+		data["modelOriginalError"] = userErr.rawLastError
+	} else if userErr.retryExhausted {
+		data["hasModelOriginalError"] = false
+	}
+	message := err.Error()
+	if userErr.message != "" {
+		message = userErr.message
+	}
+	h.progress("error", message, data)
+}
+
+type einoRunUserError struct {
+	message               string
+	kind                  string
+	summary               string
+	rawLastError          string
+	technicalError        string
+	retryExhausted        bool
+	totalRetries          int
+	hasModelOriginalError bool
+}
+
+func einoUserFacingRunError(err error) einoRunUserError {
+	var out einoRunUserError
+	if err == nil {
+		return out
+	}
+	var retryErr *adk.RetryExhaustedError
+	if !errors.As(err, &retryErr) {
+		return out
+	}
+	out.retryExhausted = true
+	out.totalRetries = retryErr.TotalRetries
+	lastErr := retryErr.LastErr
+	if lastErr == nil {
+		out.kind = "model_retry_exhausted"
+		out.summary = "模型调用多次重试后仍未成功。"
+		out.message = out.summary
+		return out
+	}
+	out.rawLastError = strings.TrimSpace(lastErr.Error())
+	if isEinoShouldRetryOutputRejected(lastErr) {
+		out.kind = "model_output_rejected"
+		out.summary = "模型未返回原始错误；输出被重试策略拒绝。"
+		out.technicalError = out.rawLastError
+		out.message = formatEinoRetryExhaustedMessage(out.summary, retryErr.TotalRetries)
+		return out
+	}
+	kind, summary := einoTransientRunErrorUserDetail(lastErr)
+	if strings.TrimSpace(summary) == "" {
+		summary = einoTrimRetryErrorSummary(lastErr.Error())
+	}
+	if kind == "" {
+		kind = "model_retry_exhausted"
+	}
+	out.kind = kind
+	out.summary = summary
+	out.hasModelOriginalError = out.rawLastError != ""
+	out.message = formatEinoRetryExhaustedMessage(summary, retryErr.TotalRetries)
+	return out
+}
+
+func isEinoShouldRetryOutputRejected(err error) bool {
+	if err == nil {
+		return false
+	}
+	return strings.Contains(strings.ToLower(err.Error()), "model output rejected by shouldretry")
+}
+
+func formatEinoRetryExhaustedMessage(summary string, totalRetries int) string {
+	summary = strings.TrimSpace(summary)
+	if summary == "" {
+		summary = "模型调用多次重试后仍未成功。"
+	}
+	if totalRetries > 0 {
+		return fmt.Sprintf("模型调用重试已耗尽（已重试 %d 次）：%s", totalRetries, summary)
+	}
+	return "模型调用重试已耗尽：" + summary
 }

internal/multiagent/eino_run_error_handler_test.go

@@ -3,6 +3,7 @@ package multiagent
 import (
 	"context"
 	"errors"
+	"strings"
 	"testing"
 
 	"github.com/cloudwego/eino/adk"
@@ -61,6 +62,99 @@ func TestEinoRunErrorHandlerTimeoutAndGeneralErrorProgress(t *testing.T) {
 	}
 }
 
+func TestEinoRunErrorHandlerRetryExhaustedEmptyOutputProgress(t *testing.T) {
+	err := &adk.RetryExhaustedError{
+		LastErr:      errors.New("model output rejected by ShouldRetry at attempt 5"),
+		TotalRetries: 4,
+	}
+	var message string
+	var data map[string]interface{}
+
+	got := newEinoRunErrorHandler(einoRunErrorHandlerConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, msg string, raw interface{}) {
+			if eventType == "error" {
+				message = msg
+				data, _ = raw.(map[string]interface{})
+			}
+		},
+	}).Handle(err)
+
+	if !errors.Is(got, err) {
+		t.Fatalf("err = %v", got)
+	}
+	if !strings.Contains(message, "模型调用重试已耗尽") ||
+		!strings.Contains(message, "模型未返回原始错误；输出被重试策略拒绝。") ||
+		strings.Contains(message, "model output rejected by ShouldRetry at attempt 5") {
+		t.Fatalf("message = %q", message)
+	}
+	if data["errorKind"] != "model_output_rejected" {
+		t.Fatalf("errorKind = %#v", data["errorKind"])
+	}
+	if data["errorSummary"] != "模型未返回原始错误；输出被重试策略拒绝。" {
+		t.Fatalf("errorSummary = %#v", data["errorSummary"])
+	}
+	if data["hasModelOriginalError"] != false {
+		t.Fatalf("hasModelOriginalError = %#v", data["hasModelOriginalError"])
+	}
+	if data["retryExhausted"] != true || data["totalRetries"] != 4 {
+		t.Fatalf("retry metadata = %#v", data)
+	}
+	if data["lastError"] != "model output rejected by ShouldRetry at attempt 5" {
+		t.Fatalf("lastError = %#v", data["lastError"])
+	}
+	if data["technicalError"] != "model output rejected by ShouldRetry at attempt 5" {
+		t.Fatalf("technicalError = %#v", data["technicalError"])
+	}
+	if _, ok := data["modelOriginalError"]; ok {
+		t.Fatalf("modelOriginalError should be absent for ShouldRetry rejection, got %#v", data["modelOriginalError"])
+	}
+	if data["error"] != err.Error() {
+		t.Fatalf("raw error = %#v, want %#v", data["error"], err.Error())
+	}
+}
+
+func TestEinoRunErrorHandlerRetryExhaustedOriginalErrorProgress(t *testing.T) {
+	err := &adk.RetryExhaustedError{
+		LastErr:      errors.New("HTTP 429 Too Many Requests"),
+		TotalRetries: 3,
+	}
+	var message string
+	var data map[string]interface{}
+
+	got := newEinoRunErrorHandler(einoRunErrorHandlerConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, msg string, raw interface{}) {
+			if eventType == "error" {
+				message = msg
+				data, _ = raw.(map[string]interface{})
+			}
+		},
+	}).Handle(err)
+
+	if !errors.Is(got, err) {
+		t.Fatalf("err = %v", got)
+	}
+	if !strings.Contains(message, "HTTP 429 Too Many Requests") {
+		t.Fatalf("message = %q", message)
+	}
+	if data["errorKind"] != "rate_limit" {
+		t.Fatalf("errorKind = %#v", data["errorKind"])
+	}
+	if data["errorSummary"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("errorSummary = %#v", data["errorSummary"])
+	}
+	if data["lastError"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("lastError = %#v", data["lastError"])
+	}
+	if data["modelOriginalError"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("modelOriginalError = %#v", data["modelOriginalError"])
+	}
+	if _, ok := data["hasModelOriginalError"]; ok {
+		t.Fatalf("hasModelOriginalError should be absent when original error is present, got %#v", data["hasModelOriginalError"])
+	}
+}
+
 func TestEinoRunErrorHandlerIterationLimitProgress(t *testing.T) {
 	var events []string
 	var errorKind interface{}

internal/multiagent/eino_run_event_drain.go

@@ -228,6 +228,7 @@ func (d *einoRunEventDrain) markPendingWithMonitor(tc toolCallPendingInfo) {
 		d.cfg.MCPExecutionBinder,
 		tc.ToolCallID,
 		tc.ToolName,
+		tc.Arguments,
 	)
 }
 

internal/multiagent/eino_run_message_accumulator.go

@@ -57,6 +57,16 @@ func (a *einoRunMessageAccumulator) Messages() []adk.Message {
 	return a.msgs
 }
 
+func (a *einoRunMessageAccumulator) NewMessages() []adk.Message {
+	if a == nil {
+		return nil
+	}
+	if a.baseCount < 0 || a.baseCount >= len(a.msgs) {
+		return nil
+	}
+	return a.msgs[a.baseCount:]
+}
+
 func (a *einoRunMessageAccumulator) BaseCount() int {
 	if a == nil {
 		return 0

internal/multiagent/eino_run_message_accumulator_test.go

@@ -27,6 +27,10 @@ func TestEinoRunMessageAccumulatorTracksBaseAndAppends(t *testing.T) {
 	if len(msgs) != 2 || msgs[1].Role != schema.Assistant || msgs[1].Content != "hello" {
 		t.Fatalf("messages = %#v", msgs)
 	}
+	newMsgs := acc.NewMessages()
+	if len(newMsgs) != 1 || newMsgs[0].Content != "hello" {
+		t.Fatalf("new messages = %#v", newMsgs)
+	}
 }
 
 func TestEinoRunMessageAccumulatorToolMessage(t *testing.T) {

internal/multiagent/eino_run_progress_tracker_test.go

@@ -98,6 +98,39 @@ func TestEinoRunProgressTrackerDedupesToolCalls(t *testing.T) {
 	}
 }
 
+func TestEinoRunProgressTrackerDedupesSameToolCallIDsWithDifferentArgs(t *testing.T) {
+	var toolCalls int
+	progress := func(eventType, _ string, _ interface{}) {
+		if eventType == "tool_call" {
+			toolCalls++
+		}
+	}
+	tracker := newEinoRunProgressTracker("deep", "lead", "conv-1", progress, nil, nil)
+	first := &schema.Message{ToolCalls: []schema.ToolCall{{
+		ID:   "call-1",
+		Type: "function",
+		Function: schema.FunctionCall{
+			Name:      "nmap",
+			Arguments: `{"host":"10.0.0.1"}`,
+		},
+	}}}
+	second := &schema.Message{ToolCalls: []schema.ToolCall{{
+		ID:   "call-1",
+		Type: "function",
+		Function: schema.FunctionCall{
+			Name:      "nmap",
+			Arguments: `{"host":"10.0.0.1","ports":"1-1024"}`,
+		},
+	}}}
+
+	tracker.EmitToolCalls(first, "lead", nil)
+	tracker.EmitToolCalls(second, "lead", nil)
+
+	if toolCalls != 1 {
+		t.Fatalf("tool call events = %d, want 1", toolCalls)
+	}
+}
+
 func TestEinoRunProgressTrackerHidesModelOutputRecoveryToolCalls(t *testing.T) {
 	var eventTypes []string
 	var marked []toolCallPendingInfo

internal/multiagent/eino_run_result_builder.go

@@ -45,7 +45,7 @@ func (b *einoRunResultBuilder) BuildFinal() *RunResult {
 func (b *einoRunResultBuilder) build(partial bool) *RunResult {
 	var runMsgs []adk.Message
 	if b.cfg.RunMessages != nil {
-		runMsgs = b.cfg.RunMessages.Messages()
+		runMsgs = b.cfg.RunMessages.NewMessages()
 	}
 	var lastAssistant string
 	var lastPlanExecuteExecutor string
@@ -107,6 +107,9 @@ func buildEinoRunResultFromAccumulated(
 	if cleaned == "" {
 		if fb := strings.TrimSpace(einoExtractFallbackAssistantFromMsgs(runAccumulatedMsgs)); fb != "" {
 			cleaned = fb
+			if orchMode == "plan_execute" {
+				cleaned = UnwrapPlanExecuteUserText(cleaned)
+			}
 		}
 	}
 	cleaned = dedupeRepeatedParagraphs(cleaned, 80)
@@ -146,32 +149,38 @@ func markModelFacingTraceForPersistence(msgs []adk.Message) []adk.Message {
 	return out
 }
 
-// einoExtractFallbackAssistantFromMsgs 在「主通道未产出助手正文」时，从 Eino ADK 轨迹中回填用户可见回复。
-// 典型场景：监督者仅调用 exit（final_result 落在 Tool 消息中），或工具结果已写入历史但 lastAssistant 未更新。
+// einoExtractFallbackAssistantFromMsgs 在「主通道未产出助手正文」时，从 Eino ADK
+// 原生消息轨迹中回填用户可见回复。这里保持克制：只采纳倒序最近的可交付终态，
+// 避免把工具调用前的过渡语或子任务过程误升为最终回复。
 //
-// 优先级：最后一次 exit 工具输出 → 最后一条含 exit 的助手 tool_calls 参数中的 final_result。
+// 可交付终态：
+// - exit 工具输出；
+// - assistant 调用 exit 时 arguments.final_result；
+// - 没有后续普通工具结果截断的纯 assistant 正文。
 func einoExtractFallbackAssistantFromMsgs(msgs []adk.Message) string {
 	for i := len(msgs) - 1; i >= 0; i-- {
 		m := msgs[i]
-		if m == nil || m.Role != schema.Tool {
+		if m == nil {
 			continue
 		}
-		if !strings.EqualFold(strings.TrimSpace(m.ToolName), adk.ToolInfoExit.Name) {
-			continue
-		}
-		content := strings.TrimSpace(m.Content)
-		if content == "" || strings.HasPrefix(content, einomcp.ToolErrorPrefix) {
-			continue
-		}
-		return content
-	}
-	for i := len(msgs) - 1; i >= 0; i-- {
-		m := msgs[i]
-		if m == nil || m.Role != schema.Assistant {
-			continue
-		}
-		if s := einoExtractExitFinalFromAssistantToolCalls(m); s != "" {
-			return s
+		switch m.Role {
+		case schema.Tool:
+			if strings.EqualFold(strings.TrimSpace(m.ToolName), adk.ToolInfoExit.Name) {
+				content := strings.TrimSpace(m.Content)
+				if content != "" && !strings.HasPrefix(content, einomcp.ToolErrorPrefix) {
+					return content
+				}
+			}
+			return ""
+		case schema.Assistant:
+			if s := einoExtractExitFinalFromAssistantToolCalls(m); s != "" {
+				return s
+			}
+			if len(m.ToolCalls) == 0 {
+				if content := strings.TrimSpace(m.Content); content != "" {
+					return content
+				}
+			}
 		}
 	}
 	return ""

internal/multiagent/eino_run_result_builder_test.go

@@ -55,6 +55,24 @@ func TestEinoRunResultBuilderFinalUsesSnapshots(t *testing.T) {
 	}
 }
 
+func TestEinoRunResultBuilderFallbackIgnoresBaseHistory(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator([]adk.Message{
+		schema.UserMessage("previous request"),
+		schema.AssistantMessage("previous answer", nil),
+		schema.UserMessage("new request"),
+	})
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "deep",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "empty" {
+		t.Fatalf("response = %q, want empty hint", got.Response)
+	}
+}
+
 func TestEinoRunResultBuilderPlanExecutePrefersExecutorOutput(t *testing.T) {
 	runMessages := newEinoRunMessageAccumulator(nil)
 	runMessages.Append(schema.AssistantMessage(`{"response":"planner text"}`, nil))
@@ -73,3 +91,18 @@ func TestEinoRunResultBuilderPlanExecutePrefersExecutorOutput(t *testing.T) {
 		t.Fatalf("response = %q, want executor text", got.Response)
 	}
 }
+
+func TestEinoRunResultBuilderPlanExecuteUnwrapsFallbackAssistant(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator(nil)
+	runMessages.Append(schema.AssistantMessage(`{"response":"fallback executor text"}`, nil))
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "plan_execute",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "fallback executor text" {
+		t.Fatalf("response = %q, want fallback executor text", got.Response)
+	}
+}

internal/multiagent/eino_run_runtime_session.go

@@ -126,6 +126,9 @@ func (s *einoRunRuntimeSession) HandleRunError(runErr error) einoRunRuntimeError
 	if s == nil || runErr == nil {
 		return einoRunRuntimeErrorResult{}
 	}
+	if isEinoTurnLoopPreemptErr(s.ctx, runErr) {
+		return einoRunRuntimeErrorResult{}
+	}
 	restarted, fatal := s.maybeRestart(runErr)
 	if fatal != nil {
 		result, err := s.takePartial(fatal)
@@ -368,5 +371,9 @@ func (s *einoRunRuntimeSession) emitUsageSummary(reason string) bool {
 	if s == nil || s.usage == nil {
 		return false
 	}
-	return s.usage.EmitOnce(s.conversationID, s.orchMode, reason, s.progress, s.logger)
+	modelName := ""
+	if s.args != nil {
+		modelName = s.args.ModelName
+	}
+	return s.usage.EmitOnce(s.conversationID, s.orchMode, reason, modelName, s.progress, s.logger)
 }

internal/multiagent/eino_run_runtime_session_test.go

@@ -153,6 +153,44 @@ func TestEinoRunRuntimeSessionCancellationReturnsPartialError(t *testing.T) {
 	}
 }
 
+func TestEinoRunRuntimeSessionHandleRunErrorSwallowsTurnLoopPreempt(t *testing.T) {
+	agent := &fakeRuntimeSessionAgent{}
+	var events []string
+	drain := newEinoRunEventDrain(einoRunEventDrainConfig{
+		ConversationID:   "conv-1",
+		OrchMode:         "deep",
+		OrchestratorName: "lead",
+		Progress: func(eventType, _ string, _ interface{}) {
+			events = append(events, eventType)
+		},
+		BaseMessages: []adk.Message{schema.UserMessage("base")},
+	})
+	session := newEinoRunRuntimeSession(einoRunRuntimeSessionConfig{
+		Context: context.Background(),
+		Args: &einoADKRunLoopArgs{
+			ConversationID:   "conv-1",
+			OrchMode:         "deep",
+			OrchestratorName: "lead",
+			Progress: func(eventType, _ string, _ interface{}) {
+				events = append(events, eventType)
+			},
+			DA: agent,
+		},
+		Drain:        drain,
+		BaseMessages: []adk.Message{schema.UserMessage("base")},
+		EmptyHint:    "empty",
+	})
+	defer session.Close()
+
+	got := session.HandleRunError(adk.ErrStreamCanceled)
+	if got.Restarted || got.Result != nil || got.Err != nil {
+		t.Fatalf("result = %+v, want swallowed preempt", got)
+	}
+	if containsString(events, "error") || containsString(events, "eino_usage_summary") {
+		t.Fatalf("events = %#v, want no fatal/partial events", events)
+	}
+}
+
 func TestEinoRunRuntimeSessionBuildFinalEmitsUsageSummary(t *testing.T) {
 	agent := &fakeRuntimeSessionAgent{}
 	var usageEvent map[string]interface{}

internal/multiagent/eino_run_usage_accumulator.go

@@ -61,6 +61,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 	conversationID string,
 	orchestration string,
 	reason string,
+	modelName string,
 	progress func(eventType, message string, data interface{}),
 	logger *zap.Logger,
 ) bool {
@@ -81,6 +82,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 		"source":           "eino",
 		"orchestration":    orchestration,
 		"reason":           reason,
+		"model":            modelName,
 		"modelCalls":       s.ModelCalls,
 		"promptTokens":     s.PromptTokens,
 		"completionTokens": s.CompletionTokens,
@@ -96,6 +98,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 			zap.String("conversationId", conversationID),
 			zap.String("orchestration", orchestration),
 			zap.String("reason", reason),
+			zap.String("model", modelName),
 			zap.Int("modelCalls", s.ModelCalls),
 			zap.Int("promptTokens", s.PromptTokens),
 			zap.Int("completionTokens", s.CompletionTokens),

internal/multiagent/eino_run_usage_accumulator_test.go

@@ -49,16 +49,16 @@ func TestEinoRunUsageAccumulatorEmitOnce(t *testing.T) {
 		}
 	}
 
-	if !acc.EmitOnce("conv-1", "deep", "final", progress, nil) {
+	if !acc.EmitOnce("conv-1", "deep", "final", "gpt-test", progress, nil) {
 		t.Fatal("first emit should return true")
 	}
-	if acc.EmitOnce("conv-1", "deep", "partial", progress, nil) {
+	if acc.EmitOnce("conv-1", "deep", "partial", "gpt-test", progress, nil) {
 		t.Fatal("second emit should return false")
 	}
 	if len(events) != 1 {
 		t.Fatalf("events = %#v, want one usage summary", events)
 	}
-	if events[0]["conversationId"] != "conv-1" || events[0]["orchestration"] != "deep" || events[0]["reason"] != "final" || events[0]["totalTokens"] != 3 {
+	if events[0]["conversationId"] != "conv-1" || events[0]["orchestration"] != "deep" || events[0]["reason"] != "final" || events[0]["model"] != "gpt-test" || events[0]["totalTokens"] != 3 {
 		t.Fatalf("event = %#v", events[0])
 	}
 }

internal/multiagent/eino_single_runner.go

@@ -90,7 +90,7 @@ func RunEinoSingleChatModelAgent(
 	}
 
 	baseHTTPClient := newEinoBaseHTTPClient()
-	agenticModelFactory := newEinoOpenAIAgenticChatModelFactory(baseHTTPClient, reasoningClient, logger)
+	agenticModelFactory := newEinoAgenticChatModelFactory(baseHTTPClient, reasoningClient, logger)
 	mainModel, err := agenticModelFactory(ctx, appCfg.OpenAI, einoModelModeNormal)
 	if err != nil {
 		return nil, fmt.Errorf("eino single agentic 模型: %w", err)
@@ -120,7 +120,7 @@ func RunEinoSingleChatModelAgent(
 	}
 	if einoSkillMW != nil {
 		if einoFSTools && einoLoc != nil {
-			fsMw, fsErr := subAgentAgenticFilesystemMiddleware(ctx, einoLoc, toolInvokeNotify, einoSingleAgentName, einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
+			fsMw, fsErr := subAgentAgenticFilesystemMiddleware(ctx, einoLoc, toolInvokeNotify, einoSingleAgentName, conversationID, projectID, ma.EinoMiddleware.ReductionRootDir, toolMaxBytesFromMW(&ma.EinoMiddleware), mcpExecBinder, einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
 			if fsErr != nil {
 				return nil, fmt.Errorf("eino single filesystem 中间件: %w", fsErr)
 			}
@@ -203,15 +203,18 @@ func RunEinoSingleChatModelAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                "eino_single",
-		OrchestratorName:        einoSingleAgentName,
-		ConversationID:          conversationID,
-		Progress:                progress,
-		Logger:                  logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		OrchMode:             "eino_single",
+		OrchestratorName:     einoSingleAgentName,
+		ConversationID:       conversationID,
+		Progress:             progress,
+		Logger:               logger,
+		SnapshotMCPIDs:       snapshotMCPIDs,
+		StreamsMainAssistant: streamsMainAssistant,
+		EinoRoleTag:          einoRoleTag,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,

internal/multiagent/eino_skills.go

@@ -2,6 +2,7 @@ package multiagent
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -10,11 +11,13 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/einomcp"
 	"cyberstrike-ai/internal/security"
+	"cyberstrike-ai/internal/tooloutput"
 
 	localbk "github.com/cloudwego/eino-ext/adk/backend/local"
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/adk/middlewares/filesystem"
 	"github.com/cloudwego/eino/adk/middlewares/skill"
+	"github.com/cloudwego/eino/components/tool"
 	"github.com/cloudwego/eino/schema"
 	"go.uber.org/zap"
 )
@@ -100,6 +103,11 @@ func subAgentAgenticFilesystemMiddleware(
 	loc *localbk.Local,
 	invokeNotify *einomcp.ToolInvokeNotifyHolder,
 	einoAgentName string,
+	conversationID string,
+	projectID string,
+	reductionRootDir string,
+	toolMaxBytes int,
+	binder *MCPExecutionBinder,
 	beginMonitor func(toolCallID, command string) string,
 	appendPartialMonitor func(executionID, toolCallID, chunk string),
 	registerCancelMonitor func(executionID string, cancel context.CancelFunc),
@@ -113,7 +121,7 @@ func subAgentAgenticFilesystemMiddleware(
 	if loc == nil {
 		return nil, nil
 	}
-	return filesystem.NewTyped[*schema.AgenticMessage](ctx, &filesystem.MiddlewareConfig{
+	mw, err := filesystem.NewTyped[*schema.AgenticMessage](ctx, &filesystem.MiddlewareConfig{
 		Backend: loc,
 		StreamingShell: &einoStreamingShellWrap{
 			inner:                   security.NewEinoStreamingShell(),
@@ -130,6 +138,71 @@ func subAgentAgenticFilesystemMiddleware(
 			shellNoOutputTimeoutSec: shellNoOutputTimeoutSec,
 		},
 	})
+	if err != nil {
+		return nil, err
+	}
+	return &einoAgenticFilesystemToolMiddleware{
+		TypedChatModelAgentMiddleware: mw,
+		conversationID:                conversationID,
+		projectID:                     projectID,
+		reductionRootDir:              reductionRootDir,
+		toolMaxBytes:                  toolMaxBytes,
+		binder:                        binder,
+	}, nil
+}
+
+type einoAgenticFilesystemToolMiddleware struct {
+	adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage]
+	conversationID   string
+	projectID        string
+	reductionRootDir string
+	toolMaxBytes     int
+	binder           *MCPExecutionBinder
+}
+
+func (m *einoAgenticFilesystemToolMiddleware) WrapInvokableToolCall(ctx context.Context, endpoint adk.InvokableToolCallEndpoint, tCtx *adk.ToolContext) (adk.InvokableToolCallEndpoint, error) {
+	wrapped, err := m.TypedChatModelAgentMiddleware.WrapInvokableToolCall(ctx, endpoint, tCtx)
+	if err != nil {
+		return nil, err
+	}
+	if tCtx == nil || !isBuiltinEinoADKFilesystemToolName(tCtx.Name) {
+		return wrapped, nil
+	}
+	return func(ctx context.Context, argumentsInJSON string, opts ...tool.Option) (string, error) {
+		args := parseToolArgumentsObject(argumentsInJSON)
+		if len(args) > 0 && m.binder != nil {
+			m.binder.BindArguments(tCtx.CallID, args)
+		}
+		result, runErr := wrapped(ctx, argumentsInJSON, opts...)
+		if runErr != nil {
+			return result, runErr
+		}
+		return m.boundToolResult(tCtx.CallID, result), nil
+	}, nil
+}
+
+func (m *einoAgenticFilesystemToolMiddleware) boundToolResult(toolCallID, result string) string {
+	if m == nil || m.toolMaxBytes <= 0 || len(result) <= m.toolMaxBytes {
+		return result
+	}
+	return tooloutput.BoundWithSpill(result, m.toolMaxBytes, tooloutput.SpillOpts{
+		RootDir:        m.reductionRootDir,
+		ProjectID:      m.projectID,
+		ConversationID: m.conversationID,
+		ExecutionID:    toolCallID,
+	})
+}
+
+func parseToolArgumentsObject(raw string) map[string]interface{} {
+	raw = strings.TrimSpace(raw)
+	if raw == "" || raw == "{}" || raw == "null" {
+		return nil
+	}
+	var args map[string]interface{}
+	if err := json.Unmarshal([]byte(raw), &args); err != nil || len(args) == 0 {
+		return nil
+	}
+	return args
 }
 
 // agentToolTimeoutMinutes 返回 agent.tool_timeout_minutes（与 executeToolViaMCP 一致）；cfg 为 nil 时 0。

internal/multiagent/eino_stream_error_handler.go

@@ -46,6 +46,12 @@ func (h *einoStreamErrorHandler) Handle(streamErr error, agentName string) einoS
 	if h == nil || streamErr == nil {
 		return einoStreamErrorHandleResult{}
 	}
+	if isEinoTurnLoopPreemptErr(h.ctx, streamErr) {
+		// Host context is still alive: TurnLoop preempt canceled the in-flight
+		// model/tool stream. Keep the outer iterator open so the queued
+		// interrupt_continue turn can start.
+		return einoStreamErrorHandleResult{Handled: true}
+	}
 	if isInterruptContinue(h.ctx) {
 		result, err := h.partial(streamErr)
 		return einoStreamErrorHandleResult{Handled: true, Result: result, Err: err}

internal/multiagent/eino_stream_error_handler_test.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"errors"
 	"testing"
+
+	"github.com/cloudwego/eino/adk"
 )
 
 func TestEinoStreamErrorHandlerEmitsProgressAndRestarts(t *testing.T) {
@@ -70,6 +72,34 @@ func TestEinoStreamErrorHandlerRetryFatalUsesPartial(t *testing.T) {
 	}
 }
 
+func TestEinoStreamErrorHandlerTurnLoopPreemptSwallowsStreamCanceled(t *testing.T) {
+	var progressCalled bool
+	var retryCalled bool
+	var partialCalled bool
+	handler := newEinoStreamErrorHandler(
+		context.Background(),
+		"conv-1",
+		func(string, string, interface{}) { progressCalled = true },
+		nil,
+		func(error) (bool, error) {
+			retryCalled = true
+			return false, errors.New("should not retry")
+		},
+		func(error) (*RunResult, error) {
+			partialCalled = true
+			return nil, errors.New("should not take partial")
+		},
+	)
+
+	got := handler.Handle(adk.ErrStreamCanceled, "lead")
+	if !got.Handled || got.Restarted || got.Result != nil || got.Err != nil {
+		t.Fatalf("result = %+v, want swallowed preempt", got)
+	}
+	if progressCalled || retryCalled || partialCalled {
+		t.Fatalf("progressCalled=%v retryCalled=%v partialCalled=%v, want all false", progressCalled, retryCalled, partialCalled)
+	}
+}
+
 func TestEinoStreamErrorHandlerInterruptContinueUsesPartialWithoutProgress(t *testing.T) {
 	base := context.Background()
 	ctx, cancel := context.WithCancelCause(base)
@@ -103,6 +133,23 @@ func TestEinoStreamErrorHandlerInterruptContinueUsesPartialWithoutProgress(t *te
 	}
 }
 
+func TestIsEinoTurnLoopPreemptErr(t *testing.T) {
+	if !isEinoTurnLoopPreemptErr(context.Background(), adk.ErrStreamCanceled) {
+		t.Fatal("alive host + stream canceled should be treated as TurnLoop preempt")
+	}
+	if !isEinoTurnLoopPreemptErr(context.Background(), context.Canceled) {
+		t.Fatal("alive host + context.Canceled should be treated as TurnLoop preempt")
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if isEinoTurnLoopPreemptErr(ctx, adk.ErrStreamCanceled) {
+		t.Fatal("canceled host should not be treated as TurnLoop preempt")
+	}
+	if isEinoTurnLoopPreemptErr(context.Background(), errors.New("boom")) {
+		t.Fatal("regular errors must stay fatal")
+	}
+}
+
 func TestEinoStreamErrorHandlerNilError(t *testing.T) {
 	got := newEinoStreamErrorHandler(context.Background(), "conv", nil, nil, nil, nil).Handle(nil, "lead")
 	if got.Handled || got.Restarted || got.Result != nil || got.Err != nil {

internal/multiagent/eino_summarize.go

@@ -164,24 +164,7 @@ func newEinoSummarizationMiddleware(
 	retryMax := retryPolicy.maxAttempts
 	var summaryOverflowRetries int
 
-	// ModelOptions apply only to summarization Generate (same ChatModel instance as the agent).
-	// Strip thinking/reasoning on this call path; mark requests for empty-choices diagnostics.
-	summaryModelOpts := []model.Option{
-		einoopenai.WithMaxCompletionTokens(outputReserve),
-		einoopenai.WithExtraHeader(map[string]string{
-			copenai.SummarizationRequestHeader: "1",
-		}),
-		einoopenai.WithRequestPayloadModifier(func(_ context.Context, in []*schema.Message, rawBody []byte) ([]byte, error) {
-			if logger != nil {
-				logger.Info("eino summarization generate request",
-					zap.Int("input_messages", len(in)),
-					zap.Int("payload_bytes", len(rawBody)),
-					zap.String("model", modelName),
-				)
-			}
-			return stripReasoningFromSummarizationPayload(rawBody)
-		}),
-	}
+	summaryModelOpts := newEinoSummarizationModelOptions(outputReserve, modelName, "classic", &appCfg.OpenAI, logger)
 
 	mw, err := summarization.New(ctx, &summarization.Config{
 		Model:        summaryModel,
@@ -308,6 +291,34 @@ func newEinoSummarizationMiddleware(
 	return mw, nil
 }
 
+// newEinoSummarizationModelOptions applies only to summarization Generate calls
+// on the shared main model. Summary generation should be plain-text and cheap:
+// strip provider reasoning/thinking controls so DeepSeek/OpenAI-compatible
+// endpoints do not spend the reserved output budget on invisible reasoning.
+func newEinoSummarizationModelOptions(outputReserve int, modelName, kind string, oa *config.OpenAIConfig, logger *zap.Logger) []model.Option {
+	label := "eino summarization generate request"
+	if strings.TrimSpace(kind) != "" && kind != "classic" {
+		label = "eino " + kind + " summarization generate request"
+	}
+	return []model.Option{
+		model.WithMaxTokens(outputReserve),
+		einoopenai.WithMaxCompletionTokens(outputReserve),
+		einoopenai.WithExtraHeader(map[string]string{
+			copenai.SummarizationRequestHeader: "1",
+		}),
+		einoopenai.WithRequestPayloadModifier(func(_ context.Context, in []*schema.Message, rawBody []byte) ([]byte, error) {
+			if logger != nil {
+				logger.Info(label,
+					zap.Int("input_messages", len(in)),
+					zap.Int("payload_bytes", len(rawBody)),
+					zap.String("model", modelName),
+				)
+			}
+			return stripReasoningFromSummarizationPayload(rawBody, oa)
+		}),
+	}
+}
+
 // summarizationInputBudgetOpts controls spill/truncation behavior when a round alone exceeds budget.
 type summarizationInputBudgetOpts struct {
 	toolMaxBytes int

internal/multiagent/eino_summarize_payload.go

@@ -1,12 +1,33 @@
 package multiagent
 
 import (
+	"strings"
+
+	"cyberstrike-ai/internal/config"
 	copenai "cyberstrike-ai/internal/openai"
 )
 
 // stripReasoningFromSummarizationPayload removes thinking / reasoning fields from a
 // chat-completions JSON body. Applied only to summarization Generate calls via
 // model.ModelOptions on the shared ChatModel — main-agent requests are unchanged.
-func stripReasoningFromSummarizationPayload(rawBody []byte) ([]byte, error) {
+func stripReasoningFromSummarizationPayload(rawBody []byte, oa *config.OpenAIConfig) ([]byte, error) {
+	if shouldDisableDeepSeekThinkingForSummarization(oa) {
+		return copenai.DisableThinkingForChatCompletionBody(rawBody)
+	}
 	return copenai.StripReasoningFromChatCompletionBody(rawBody)
 }
+
+func shouldDisableDeepSeekThinkingForSummarization(oa *config.OpenAIConfig) bool {
+	if oa == nil {
+		return false
+	}
+	profile := strings.ToLower(strings.TrimSpace(oa.Reasoning.ProfileEffective()))
+	switch profile {
+	case "deepseek", "deepseek_compat":
+		return true
+	case "", "auto":
+		return oa.IsDeepSeekEndpointOrModel()
+	default:
+		return false
+	}
+}

internal/multiagent/eino_summarize_payload_test.go

@@ -3,11 +3,15 @@ package multiagent
 import (
 	"strings"
 	"testing"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/cloudwego/eino/components/model"
 )
 
 func TestStripReasoningFromSummarizationPayload(t *testing.T) {
 	in := []byte(`{"model":"deepseek-chat","messages":[],"thinking":{"type":"enabled"},"reasoning_effort":"high"}`)
-	out, err := stripReasoningFromSummarizationPayload(in)
+	out, err := stripReasoningFromSummarizationPayload(in, nil)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -20,7 +24,7 @@ func TestStripReasoningFromSummarizationPayload(t *testing.T) {
 	}
 
 	plain := []byte(`{"model":"gpt-4o","messages":[]}`)
-	out2, err := stripReasoningFromSummarizationPayload(plain)
+	out2, err := stripReasoningFromSummarizationPayload(plain, nil)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -28,3 +32,53 @@ func TestStripReasoningFromSummarizationPayload(t *testing.T) {
 		t.Fatalf("expected unchanged payload, got %s", out2)
 	}
 }
+
+func TestStripReasoningFromSummarizationPayloadDisablesDeepSeekThinking(t *testing.T) {
+	in := []byte(`{"model":"deepseek-v4-flash","messages":[],"thinking":{"type":"enabled"},"reasoning_effort":"high"}`)
+	oa := &config.OpenAIConfig{
+		BaseURL: "https://api.deepseek.com/v1",
+		Model:   "deepseek-v4-flash",
+	}
+	out, err := stripReasoningFromSummarizationPayload(in, oa)
+	if err != nil {
+		t.Fatal(err)
+	}
+	s := string(out)
+	if strings.Contains(s, "reasoning_effort") {
+		t.Fatalf("expected reasoning_effort stripped, got %s", s)
+	}
+	if !strings.Contains(s, `"thinking":{"type":"disabled"}`) {
+		t.Fatalf("expected DeepSeek thinking disabled, got %s", s)
+	}
+}
+
+func TestStripReasoningFromSummarizationPayloadHonorsOpenAICompatProfile(t *testing.T) {
+	in := []byte(`{"model":"deepseek-v4-flash","messages":[],"thinking":{"type":"enabled"},"reasoning_effort":"high"}`)
+	oa := &config.OpenAIConfig{
+		BaseURL: "https://api.deepseek.com/v1",
+		Model:   "deepseek-v4-flash",
+		Reasoning: config.OpenAIReasoningConfig{
+			Profile: "openai_compat",
+		},
+	}
+	out, err := stripReasoningFromSummarizationPayload(in, oa)
+	if err != nil {
+		t.Fatal(err)
+	}
+	s := string(out)
+	if strings.Contains(s, "thinking") || strings.Contains(s, "reasoning_effort") {
+		t.Fatalf("expected OpenAI-compatible profile to strip reasoning fields, got %s", s)
+	}
+}
+
+func TestEinoSummarizationModelOptionsSetCommonMaxTokens(t *testing.T) {
+	const outputReserve = 4096
+	opts := newEinoSummarizationModelOptions(outputReserve, "minimax-m3", "agentic", nil, nil)
+	common := model.GetCommonOptions(nil, opts...)
+	if common == nil || common.MaxTokens == nil {
+		t.Fatal("expected summarization options to set common max_tokens")
+	}
+	if *common.MaxTokens != outputReserve {
+		t.Fatalf("max_tokens = %d, want %d", *common.MaxTokens, outputReserve)
+	}
+}

internal/multiagent/eino_tool_result_concat.go

@@ -0,0 +1,91 @@
+package multiagent
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/cloudwego/eino/schema"
+)
+
+// concatToolResultChunks 按 Eino 原生语义合并工具结果流：
+//   - 同一 CallID（EventSender 一 call 一 event）：schema.ConcatMessages
+//   - 并行工具被摊进同一条流（ToolsNode MergeStreamReaders 扁平化后）：
+//     按 CallID 分列后再 ConcatMessages，等价于 schema.ConcatMessageArray
+func concatToolResultChunks(chunks []*schema.Message) ([]*schema.Message, error) {
+	if len(chunks) == 0 {
+		return nil, nil
+	}
+	if toolResultChunksShareCallID(chunks) {
+		merged, err := schema.ConcatMessages(chunks)
+		if err != nil {
+			return nil, err
+		}
+		return []*schema.Message{merged}, nil
+	}
+	return concatToolResultChunksByCallID(chunks)
+}
+
+func toolResultChunksShareCallID(chunks []*schema.Message) bool {
+	id := ""
+	for _, chunk := range chunks {
+		if chunk == nil {
+			continue
+		}
+		got := strings.TrimSpace(chunk.ToolCallID)
+		if got == "" {
+			continue
+		}
+		if id == "" {
+			id = got
+			continue
+		}
+		if got != id {
+			return false
+		}
+	}
+	return true
+}
+
+func concatToolResultChunksByCallID(chunks []*schema.Message) ([]*schema.Message, error) {
+	type column struct {
+		key    string
+		chunks []*schema.Message
+	}
+	var ordered []column
+	index := make(map[string]int)
+	lastKey := ""
+	anon := 0
+	for _, chunk := range chunks {
+		if chunk == nil {
+			continue
+		}
+		key := strings.TrimSpace(chunk.ToolCallID)
+		if key == "" {
+			if lastKey != "" {
+				key = lastKey
+			} else {
+				key = fmt.Sprintf("\x00anon-%d", anon)
+				anon++
+			}
+		}
+		if idx, ok := index[key]; ok {
+			ordered[idx].chunks = append(ordered[idx].chunks, chunk)
+		} else {
+			index[key] = len(ordered)
+			ordered = append(ordered, column{key: key, chunks: []*schema.Message{chunk}})
+		}
+		lastKey = key
+	}
+	out := make([]*schema.Message, 0, len(ordered))
+	for _, col := range ordered {
+		merged, err := schema.ConcatMessages(col.chunks)
+		if err != nil {
+			return nil, err
+		}
+		if strings.HasPrefix(col.key, "\x00anon-") {
+			merged.ToolCallID = ""
+		}
+		out = append(out, merged)
+	}
+	return out, nil
+}

internal/multiagent/eino_tool_result_concat_test.go

@@ -0,0 +1,41 @@
+package multiagent
+
+import (
+	"testing"
+
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestConcatToolResultChunksUsesEinoConcatForSingleCall(t *testing.T) {
+	got, err := concatToolResultChunks([]*schema.Message{
+		schema.ToolMessage("hel", "call-1", schema.WithToolName("execute")),
+		schema.ToolMessage("lo", "call-1", schema.WithToolName("execute")),
+	})
+	if err != nil {
+		t.Fatalf("concat: %v", err)
+	}
+	if len(got) != 1 || got[0].ToolCallID != "call-1" || got[0].Content != "hello" || got[0].ToolName != "execute" {
+		t.Fatalf("got = %#v, want one ConcatMessages result", got)
+	}
+}
+
+func TestConcatToolResultChunksSplitsParallelCalls(t *testing.T) {
+	got, err := concatToolResultChunks([]*schema.Message{
+		schema.ToolMessage("nmap 1/2 ", "call-1", schema.WithToolName("nmap")),
+		schema.ToolMessage("nmap 2/2 ", "call-2", schema.WithToolName("nmap")),
+		schema.ToolMessage("22/tcp", "call-1", schema.WithToolName("nmap")),
+		schema.ToolMessage("80/tcp", "call-2", schema.WithToolName("nmap")),
+	})
+	if err != nil {
+		t.Fatalf("concat: %v", err)
+	}
+	if len(got) != 2 {
+		t.Fatalf("got = %#v, want two calls", got)
+	}
+	if got[0].ToolCallID != "call-1" || got[0].Content != "nmap 1/2 22/tcp" {
+		t.Fatalf("call-1 = %#v", got[0])
+	}
+	if got[1].ToolCallID != "call-2" || got[1].Content != "nmap 2/2 80/tcp" {
+		t.Fatalf("call-2 = %#v", got[1])
+	}
+}

internal/multiagent/eino_tool_result_event_handler.go

@@ -42,24 +42,42 @@ func (h *einoToolResultEventHandler) HandleStreaming(mv *adk.MessageVariant, age
 	if h == nil || mv == nil || !mv.IsStreaming || mv.MessageStream == nil || mv.Role != schema.Tool {
 		return false
 	}
-	toolName := strings.TrimSpace(mv.ToolName)
-	content, streamToolCallID, streamToolName, recvErr := recvSchemaMessageStream(h.ctx, mv.MessageStream)
-	if toolName == "" {
-		toolName = streamToolName
+	defaultName := strings.TrimSpace(mv.ToolName)
+	msgs, recvErr := recvSchemaToolResultMessages(h.ctx, mv.MessageStream)
+	if isEinoVoluntaryCancelErr(recvErr) && len(msgs) == 0 {
+		msgs = []*schema.Message{schema.ToolMessage("已中断并继续，当前工具调用已停止。", "", schema.WithToolName(defaultName))}
 	}
-	isErr := einoToolResultIsError(toolName, content)
-	content = einoToolResultBody(content)
-	if streamToolCallID != "" && h.runMessages != nil {
-		h.runMessages.AppendToolMessage(content, streamToolCallID, schema.WithToolName(toolName))
+	if len(msgs) == 0 {
+		msgs = []*schema.Message{schema.ToolMessage("", "", schema.WithToolName(defaultName))}
 	}
-	if h.emitter != nil {
-		h.emitter.Emit(h.ctx, toolName, content, streamToolCallID, isErr, agentName)
-	}
-	if recvErr != nil && h.logger != nil {
-		h.logger.Warn("eino tool result stream recv error",
-			zap.Error(recvErr),
-			zap.String("agent", agentName),
-			zap.String("tool", toolName))
+	for _, msg := range msgs {
+		if msg == nil {
+			continue
+		}
+		toolName := strings.TrimSpace(msg.ToolName)
+		if toolName == "" {
+			toolName = defaultName
+		}
+		content := msg.Content
+		if isEinoVoluntaryCancelErr(recvErr) && strings.TrimSpace(content) == "" {
+			content = "已中断并继续，当前工具调用已停止。"
+		}
+		isErr := einoToolResultIsError(toolName, content) || isEinoVoluntaryCancelErr(recvErr)
+		content = einoToolResultBody(content)
+		toolCallID := strings.TrimSpace(msg.ToolCallID)
+		if toolCallID != "" && h.runMessages != nil {
+			h.runMessages.AppendToolMessage(content, toolCallID, schema.WithToolName(toolName))
+		}
+		if h.emitter != nil {
+			h.emitter.Emit(h.ctx, toolName, content, toolCallID, isErr, agentName)
+		}
+		if recvErr != nil && !isEinoVoluntaryCancelErr(recvErr) && h.logger != nil {
+			h.logger.Warn("eino tool result stream recv error",
+				zap.Error(recvErr),
+				zap.String("agent", agentName),
+				zap.String("tool", toolName),
+				zap.String("toolCallId", toolCallID))
+		}
 	}
 	if recvErr == nil && h.confirmRecovery != nil {
 		h.confirmRecovery()

internal/multiagent/eino_tool_result_event_handler_test.go

@@ -59,6 +59,54 @@ func TestEinoToolResultEventHandlerHandlesStreamingToolResult(t *testing.T) {
 	}
 }
 
+func TestEinoToolResultEventHandlerSplitsParallelStreamingResults(t *testing.T) {
+	var events []map[string]interface{}
+	runMessages := newEinoRunMessageAccumulator(nil)
+	emitter := newEinoToolResultProgressEmitter(einoToolResultProgressEmitterConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, _ string, data interface{}) {
+			if eventType != "tool_result" {
+				return
+			}
+			m, _ := data.(map[string]interface{})
+			events = append(events, m)
+		},
+	})
+	handler := newEinoToolResultEventHandler(einoToolResultEventHandlerConfig{
+		RunMessages: runMessages,
+		Emitter:     emitter,
+	})
+	stream := schema.StreamReaderFromArray([]*schema.Message{
+		{Role: schema.Tool, Content: "nmap 1/2 start ", ToolCallID: "call-1", ToolName: "nmap"},
+		{Role: schema.Tool, Content: "nmap 2/2 start ", ToolCallID: "call-2", ToolName: "nmap"},
+		{Role: schema.Tool, Content: "22/tcp open", ToolCallID: "call-1", ToolName: "nmap"},
+		{Role: schema.Tool, Content: "80/tcp open", ToolCallID: "call-2", ToolName: "nmap"},
+	})
+	mv := &adk.MessageVariant{
+		IsStreaming:   true,
+		Role:          schema.Tool,
+		ToolName:      "nmap",
+		MessageStream: stream,
+	}
+
+	if !handler.HandleStreaming(mv, "worker") {
+		t.Fatal("streaming tool result was not handled")
+	}
+	if len(events) != 2 {
+		t.Fatalf("events = %#v, want two tool_result", events)
+	}
+	if events[0]["toolCallId"] != "call-1" || events[0]["result"] != "nmap 1/2 start 22/tcp open" {
+		t.Fatalf("first event = %#v", events[0])
+	}
+	if events[1]["toolCallId"] != "call-2" || events[1]["result"] != "nmap 2/2 start 80/tcp open" {
+		t.Fatalf("second event = %#v", events[1])
+	}
+	msgs := runMessages.Messages()
+	if len(msgs) != 2 || msgs[0].ToolCallID != "call-1" || msgs[1].ToolCallID != "call-2" {
+		t.Fatalf("run messages = %#v", msgs)
+	}
+}
+
 func TestEinoToolResultEventHandlerHandlesMaterializedToolResult(t *testing.T) {
 	var event map[string]interface{}
 	emitter := newEinoToolResultProgressEmitter(einoToolResultProgressEmitterConfig{
@@ -93,3 +141,39 @@ func TestEinoToolResultEventHandlerIgnoresNonToolOutput(t *testing.T) {
 		t.Fatal("assistant message should not be handled as tool result")
 	}
 }
+
+func TestEinoToolResultEventHandlerMarksCanceledStreamAsInterrupted(t *testing.T) {
+	var event map[string]interface{}
+	emitter := newEinoToolResultProgressEmitter(einoToolResultProgressEmitterConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, _ string, data interface{}) {
+			if eventType == "tool_result" {
+				event, _ = data.(map[string]interface{})
+			}
+		},
+	})
+	handler := newEinoToolResultEventHandler(einoToolResultEventHandlerConfig{Emitter: emitter})
+	reader, writer := schema.Pipe[*schema.Message](1)
+	writer.Send(nil, adk.ErrStreamCanceled)
+	writer.Close()
+	mv := &adk.MessageVariant{
+		IsStreaming:   true,
+		Role:          schema.Tool,
+		ToolName:      "http-framework-test",
+		MessageStream: reader,
+	}
+
+	if !handler.HandleStreaming(mv, "penetration") {
+		t.Fatal("canceled tool stream was not handled")
+	}
+	if event["toolName"] != "http-framework-test" {
+		t.Fatalf("event = %#v", event)
+	}
+	if event["isError"] != true || event["success"] != false {
+		t.Fatalf("event flags = %#v, want interrupted tool result", event)
+	}
+	result, _ := event["result"].(string)
+	if result == "stream canceled" || result == "" {
+		t.Fatalf("result = %q, want interrupt-continue notice", result)
+	}
+}

internal/multiagent/eino_tool_result_progress_emitter.go

@@ -115,7 +115,22 @@ func (e *einoToolResultProgressEmitter) Emit(ctx context.Context, toolName, cont
 	if e.executeStdoutDup != nil {
 		e.executeStdoutDup.Record(toolName, content, displayIsErr)
 	}
-	recordEinoADKFilesystemToolMonitor(ctx, e.filesystemMonitorAgent, e.filesystemMonitorRecord, e.mcpExecutionBinder, toolName, toolCallID, e.messages(), content, displayIsErr)
+	if args := e.toolCallArguments(toolCallID, toolName); len(args) > 0 {
+		data["argumentsObj"] = args
+		data["arguments"] = mustMarshalToolArguments(args)
+	}
+	if execID := recordEinoADKFilesystemToolMonitor(ctx, e.filesystemMonitorAgent, e.filesystemMonitorRecord, e.mcpExecutionBinder, toolName, toolCallID, e.messages(), content, displayIsErr); execID != "" {
+		if stored := e.filesystemMonitorAgent.MCPExecutionResultText(execID); strings.TrimSpace(stored) != "" {
+			content = stored
+			if len(content) > 200 {
+				preview = content[:200] + "..."
+			} else {
+				preview = content
+			}
+			data["result"] = content
+			data["resultPreview"] = preview
+		}
+	}
 	if e.filesystemMonitorAgent != nil && e.mcpExecutionBinder != nil {
 		if execID := e.mcpExecutionBinder.ExecutionID(toolCallID); execID != "" {
 			e.filesystemMonitorAgent.UpdateMCPExecutionDisplayResult(execID, content)
@@ -152,3 +167,15 @@ func (e *einoToolResultProgressEmitter) messages() []adk.Message {
 	}
 	return e.runMessages.Messages()
 }
+
+func (e *einoToolResultProgressEmitter) toolCallArguments(toolCallID, toolName string) map[string]interface{} {
+	if e == nil {
+		return nil
+	}
+	if e.mcpExecutionBinder != nil {
+		if args := e.mcpExecutionBinder.Arguments(toolCallID); len(args) > 0 {
+			return args
+		}
+	}
+	return toolCallArgsFromAccumulated(e.messages(), toolCallID, toolName)
+}

internal/multiagent/eino_tool_result_progress_emitter_test.go

@@ -1,6 +1,11 @@
 package multiagent
 
-import "testing"
+import (
+	"testing"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
 
 func TestEinoToolResultProgressEmitterInfersPendingAndDedupes(t *testing.T) {
 	var events []map[string]interface{}
@@ -135,3 +140,44 @@ func TestEinoToolResultProgressEmitterTruncatesPreview(t *testing.T) {
 		t.Fatalf("preview = %q len=%d", got, len(got))
 	}
 }
+
+func TestEinoToolResultProgressEmitterBackfillsArgumentsFromRunMessages(t *testing.T) {
+	var data map[string]interface{}
+	progress := func(eventType, _ string, raw interface{}) {
+		if eventType == "tool_result" {
+			data, _ = raw.(map[string]interface{})
+		}
+	}
+	runMessages := newEinoRunMessageAccumulator([]adk.Message{
+		&schema.Message{
+			Role: schema.Assistant,
+			ToolCalls: []schema.ToolCall{{
+				ID:   "call-read",
+				Type: "function",
+				Function: schema.FunctionCall{
+					Name:      "read_file",
+					Arguments: `{"path":"/tmp/requirements.txt"}`,
+				},
+			}},
+		},
+	})
+	emitter := newEinoToolResultProgressEmitter(einoToolResultProgressEmitterConfig{
+		ConversationID: "conv-1",
+		Progress:       progress,
+		RunMessages:    runMessages,
+	})
+
+	if !emitter.Emit(nil, "read_file", "ok", "call-read", false, "lead") {
+		t.Fatal("expected tool result emit")
+	}
+	args, ok := data["argumentsObj"].(map[string]interface{})
+	if !ok {
+		t.Fatalf("argumentsObj = %#v", data["argumentsObj"])
+	}
+	if args["path"] != "/tmp/requirements.txt" {
+		t.Fatalf("path = %#v, want /tmp/requirements.txt", args["path"])
+	}
+	if data["arguments"] != `{"path":"/tmp/requirements.txt"}` {
+		t.Fatalf("arguments = %#v", data["arguments"])
+	}
+}

internal/multiagent/eino_transient_retry.go

@@ -52,6 +52,9 @@ func isEinoTransientRunError(err error) bool {
 	if msg == "" {
 		return false
 	}
+	if isEinoEmptySummaryContentErrorText(msg) {
+		return true
+	}
 	if status := httpStatusFromErrorText(msg); status > 0 {
 		return isRetryableHTTPStatus(status)
 	}
@@ -94,6 +97,11 @@ func isEinoTransientRunError(err error) bool {
 	return false
 }
 
+func isEinoEmptySummaryContentErrorText(msg string) bool {
+	return strings.Contains(msg, "summary content is empty") ||
+		strings.Contains(msg, "agentic summarization returned empty summary")
+}
+
 func isRetryableHTTPStatus(status int) bool {
 	switch status {
 	case 408, 409, 425, 429:

internal/multiagent/eino_transient_retry_test.go

@@ -36,6 +36,7 @@ func TestIsEinoTransientRunError(t *testing.T) {
 		{"http2 goaway", errors.New("failed to receive stream chunk: error, http2: server sent GOAWAY and closed the connection; LastStreamID=791, ErrCode=NO_ERROR"), true},
 		{"unexpected internal stream chunk", errors.New("failed to receive stream chunk: error, The service encountered an unexpected internal error. Request id: 0217851391106464f01ec66621d0980a42fd45436ed75957a6a0a"), true},
 		{"unexpected eof", errors.New("unexpected EOF"), true},
+		{"empty summarization output", errors.New("[NodeRunError] summary content is empty\nnode path: [node_1, ChatModel]"), true},
 		{"503", errors.New("upstream returned 503"), true},
 		{"iteration limit", errors.New("max iteration reached"), false},
 		{"canceled", context.Canceled, false},

internal/multiagent/eino_turn_loop_bridge_test.go

@@ -107,6 +107,76 @@ func TestRunEinoADKAgentLoopUsesTurnLoopInterruptPush(t *testing.T) {
 	}
 }
 
+func TestRunEinoADKAgentLoopInterruptContinueSurvivesStreamCanceled(t *testing.T) {
+	baseCtx, cancel := context.WithTimeout(context.Background(), 3*time.Second)
+	defer cancel()
+
+	pushCh := make(chan func(string) bool, 1)
+	ctx := WithAgentTurnLoopInterruptRegistrar(baseCtx, func(push func(string) bool) func() {
+		pushCh <- push
+		return func() {}
+	})
+
+	mockModel := newTurnLoopHangingStreamModel()
+	agent, err := adk.NewChatModelAgent(ctx, &adk.ChatModelAgentConfig{
+		Name:  "turn-loop-agent",
+		Model: mockModel,
+	})
+	if err != nil {
+		t.Fatalf("NewChatModelAgent: %v", err)
+	}
+
+	done := make(chan struct{})
+	var result *RunResult
+	var runErr error
+	go func() {
+		defer close(done)
+		result, runErr = runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
+			OrchMode:                 "deep",
+			OrchestratorName:         "turn-loop-agent",
+			ConversationID:           "conv-stream-cancel",
+			DA:                       agent,
+			EmptyResponseMessage:     "empty",
+			TurnLoopInterruptTimeout: 20 * time.Millisecond,
+		}, []*schema.Message{schema.UserMessage("initial task")})
+	}()
+
+	select {
+	case <-mockModel.started:
+	case <-ctx.Done():
+		t.Fatal("first stream did not start")
+	}
+	var push func(string) bool
+	select {
+	case push = <-pushCh:
+	case <-ctx.Done():
+		t.Fatal("turn loop interrupt hook was not registered")
+	}
+	if !push("focus emobile") {
+		t.Fatal("turn loop interrupt push was rejected")
+	}
+
+	select {
+	case <-mockModel.started:
+	case <-ctx.Done():
+		t.Fatal("second stream did not start after interrupt")
+	}
+	select {
+	case <-done:
+	case <-ctx.Done():
+		t.Fatal("run loop did not finish")
+	}
+	if runErr != nil {
+		t.Fatalf("runErr = %v, want interrupt-continue to survive stream canceled", runErr)
+	}
+	if result == nil || result.Response != "done" {
+		t.Fatalf("result = %#v, want continued turn output", result)
+	}
+	if isEinoStreamCanceled(runErr) {
+		t.Fatal("stream canceled leaked as the run error")
+	}
+}
+
 func containsString(items []string, target string) bool {
 	for _, item := range items {
 		if item == target {

internal/multiagent/eino_turn_loop_event_bridge.go

@@ -2,7 +2,6 @@ package multiagent
 
 import (
 	"context"
-	"errors"
 	"strings"
 	"sync/atomic"
 
@@ -45,9 +44,16 @@ func (b *einoTurnLoopEventBridge) OnAgentEvents(
 		if ev == nil {
 			continue
 		}
-		if ev.Err != nil && isEinoTurnLoopPreemptCancel(tc, ev.Err) {
-			b.emitPreempted()
-			continue
+		if ev.Err != nil && isEinoVoluntaryCancelErr(ev.Err) {
+			// TurnLoop owns cancel routing. Returning CancelError /
+			// ErrStreamCanceled from OnAgentEvents aborts the whole loop and
+			// races with the Preempted signal becoming observable. A preempt
+			// must return nil so the queued user supplement can start the next
+			// turn; a terminal stop is surfaced through TurnLoopExitState.
+			if isEinoTurnLoopPreemptCancel(tc, ev.Err) {
+				b.emitPreempted()
+			}
+			return nil
 		}
 		if b.gen != nil {
 			b.gen.Send(ev)
@@ -79,11 +85,7 @@ func (b *einoTurnLoopEventBridge) emitPreempted() {
 }
 
 func isEinoTurnLoopPreemptCancel(tc *adk.TurnContext[EinoTurnLoopItem, *schema.Message], err error) bool {
-	if tc == nil || err == nil {
-		return false
-	}
-	var cancelErr *adk.CancelError
-	if !errors.As(err, &cancelErr) {
+	if tc == nil || !isEinoVoluntaryCancelErr(err) {
 		return false
 	}
 	select {

internal/multiagent/eino_turn_loop_event_bridge_test.go

@@ -41,6 +41,52 @@ func TestEinoTurnLoopEventBridgeSwallowsPreemptCancel(t *testing.T) {
 	}
 }
 
+func TestEinoTurnLoopEventBridgeNeverForwardsStreamCanceled(t *testing.T) {
+	iter, gen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
+	outIter, outGen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
+	bridge := newEinoTurnLoopEventBridge("conv", "eino_single", nil, outGen)
+
+	gen.Send(&adk.AgentEvent{Err: adk.ErrStreamCanceled})
+	gen.Close()
+
+	err := bridge.OnAgentEvents(context.Background(), &adk.TurnContext[EinoTurnLoopItem, *schema.Message]{
+		Preempted: make(chan struct{}),
+	}, iter)
+	if err != nil {
+		t.Fatalf("ErrStreamCanceled must be owned by TurnLoop, got %v", err)
+	}
+	if bridge.ForwardedError() {
+		t.Fatal("ErrStreamCanceled must not be marked as forwarded")
+	}
+	outGen.Close()
+	if ev, ok := outIter.Next(); ok || ev != nil {
+		t.Fatalf("ErrStreamCanceled must not be forwarded, got ok=%v ev=%#v", ev, ok)
+	}
+}
+
+func TestEinoTurnLoopEventBridgeNeverForwardsCancelError(t *testing.T) {
+	iter, gen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
+	outIter, outGen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
+	bridge := newEinoTurnLoopEventBridge("conv", "eino_single", nil, outGen)
+
+	gen.Send(&adk.AgentEvent{Err: &adk.CancelError{Info: &adk.AgentCancelInfo{}}})
+	gen.Close()
+
+	err := bridge.OnAgentEvents(context.Background(), &adk.TurnContext[EinoTurnLoopItem, *schema.Message]{
+		Preempted: make(chan struct{}),
+	}, iter)
+	if err != nil {
+		t.Fatalf("CancelError must be owned by TurnLoop, got %v", err)
+	}
+	if bridge.ForwardedError() {
+		t.Fatal("CancelError must not be marked as forwarded")
+	}
+	outGen.Close()
+	if ev, ok := outIter.Next(); ok || ev != nil {
+		t.Fatalf("CancelError must not be forwarded, got ok=%v ev=%#v", ok, ev)
+	}
+}
+
 func TestEinoTurnLoopEventBridgeForwardsRegularError(t *testing.T) {
 	iter, gen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()
 	outIter, outGen := adk.NewAsyncIteratorPair[*adk.AgentEvent]()

internal/multiagent/eino_turn_loop_runtime_test.go

@@ -65,6 +65,73 @@ func (m *turnLoopBlockingModel) snapshotInputs() [][]*schema.Message {
 	return out
 }
 
+// turnLoopHangingStreamModel emits one stream chunk then blocks until the
+// agent cancel scope ends. That matches production ChatModel streams that
+// receive ErrStreamCanceled when TurnLoop preempt escalates to CancelImmediate.
+type turnLoopHangingStreamModel struct {
+	started chan struct{}
+	mu      sync.Mutex
+	inputs  [][]*schema.Message
+}
+
+func newTurnLoopHangingStreamModel() *turnLoopHangingStreamModel {
+	return &turnLoopHangingStreamModel{started: make(chan struct{}, 8)}
+}
+
+func (m *turnLoopHangingStreamModel) Generate(ctx context.Context, input []*schema.Message, opts ...model.Option) (*schema.Message, error) {
+	sr, err := m.Stream(ctx, input, opts...)
+	if err != nil {
+		return nil, err
+	}
+	defer sr.Close()
+	var last *schema.Message
+	for {
+		msg, rerr := sr.Recv()
+		if rerr != nil {
+			if last != nil {
+				return last, nil
+			}
+			return nil, rerr
+		}
+		if msg != nil {
+			last = msg
+		}
+	}
+}
+
+func (m *turnLoopHangingStreamModel) Stream(ctx context.Context, input []*schema.Message, _ ...model.Option) (*schema.StreamReader[*schema.Message], error) {
+	m.mu.Lock()
+	m.inputs = append(m.inputs, cloneSchemaMessages(input))
+	callNo := len(m.inputs)
+	m.mu.Unlock()
+
+	select {
+	case m.started <- struct{}{}:
+	default:
+	}
+
+	if callNo == 1 {
+		reader, writer := schema.Pipe[*schema.Message](1)
+		go func() {
+			writer.Send(schema.AssistantMessage("partial", nil), nil)
+			<-ctx.Done()
+			writer.Close()
+		}()
+		return reader, nil
+	}
+	return schema.StreamReaderFromArray([]*schema.Message{schema.AssistantMessage("done", nil)}), nil
+}
+
+func (m *turnLoopHangingStreamModel) snapshotInputs() [][]*schema.Message {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	out := make([][]*schema.Message, len(m.inputs))
+	for i := range m.inputs {
+		out[i] = cloneSchemaMessages(m.inputs[i])
+	}
+	return out
+}
+
 func TestEinoTurnLoopRuntimePushInterruptStartsNextTurn(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
 	defer cancel()

internal/multiagent/mcp_execution_binder.go

@@ -7,12 +7,16 @@ import (
 
 // MCPExecutionBinder maps ADK toolCallID → MCP monitor execution ID for a single agent run.
 type MCPExecutionBinder struct {
-	mu         sync.RWMutex
-	byToolCall map[string]string
+	mu             sync.RWMutex
+	byToolCall     map[string]string
+	argsByToolCall map[string]map[string]interface{}
 }
 
 func NewMCPExecutionBinder() *MCPExecutionBinder {
-	return &MCPExecutionBinder{byToolCall: make(map[string]string)}
+	return &MCPExecutionBinder{
+		byToolCall:     make(map[string]string),
+		argsByToolCall: make(map[string]map[string]interface{}),
+	}
 }
 
 func (b *MCPExecutionBinder) Bind(toolCallID, executionID string) {
@@ -29,6 +33,19 @@ func (b *MCPExecutionBinder) Bind(toolCallID, executionID string) {
 	b.mu.Unlock()
 }
 
+func (b *MCPExecutionBinder) BindArguments(toolCallID string, args map[string]interface{}) {
+	if b == nil || len(args) == 0 {
+		return
+	}
+	tid := strings.TrimSpace(toolCallID)
+	if tid == "" {
+		return
+	}
+	b.mu.Lock()
+	b.argsByToolCall[tid] = cloneToolArgs(args)
+	b.mu.Unlock()
+}
+
 func (b *MCPExecutionBinder) ExecutionID(toolCallID string) string {
 	if b == nil {
 		return ""
@@ -38,3 +55,24 @@ func (b *MCPExecutionBinder) ExecutionID(toolCallID string) string {
 	defer b.mu.RUnlock()
 	return b.byToolCall[tid]
 }
+
+func (b *MCPExecutionBinder) Arguments(toolCallID string) map[string]interface{} {
+	if b == nil {
+		return nil
+	}
+	tid := strings.TrimSpace(toolCallID)
+	b.mu.RLock()
+	defer b.mu.RUnlock()
+	return cloneToolArgs(b.argsByToolCall[tid])
+}
+
+func cloneToolArgs(args map[string]interface{}) map[string]interface{} {
+	if len(args) == 0 {
+		return nil
+	}
+	out := make(map[string]interface{}, len(args))
+	for k, v := range args {
+		out[k] = v
+	}
+	return out
+}

internal/multiagent/mcp_execution_binder_test.go

@@ -9,9 +9,13 @@ import (
 func TestMCPExecutionBinder(t *testing.T) {
 	b := NewMCPExecutionBinder()
 	b.Bind("call-1", "exec-1")
+	b.BindArguments("call-1", map[string]interface{}{"file_path": "/tmp/a.txt"})
 	if got := b.ExecutionID("call-1"); got != "exec-1" {
 		t.Fatalf("expected exec-1, got %q", got)
 	}
+	if got := b.Arguments("call-1"); got["file_path"] != "/tmp/a.txt" {
+		t.Fatalf("arguments = %#v", got)
+	}
 	if got := b.ExecutionID("missing"); got != "" {
 		t.Fatalf("expected empty, got %q", got)
 	}

internal/multiagent/runner.go

@@ -52,6 +52,7 @@ type RunResult struct {
 type toolCallPendingInfo struct {
 	ToolCallID string
 	ToolName   string
+	Arguments  map[string]interface{}
 	EinoAgent  string
 	EinoRole   string
 }
@@ -154,8 +155,8 @@ func RunDeepAgent(
 	mainDefs := ag.ToolsForRole(roleTools)
 
 	baseHTTPClient := newEinoBaseHTTPClient()
-	modelFactory := newEinoOpenAIChatModelFactory(baseHTTPClient, reasoningClient, logger)
-	agenticModelFactory := newEinoOpenAIAgenticChatModelFactory(baseHTTPClient, reasoningClient, logger)
+	modelFactory := newEinoToolCallingChatModelFactory(baseHTTPClient, reasoningClient, logger)
+	agenticModelFactory := newEinoAgenticChatModelFactory(baseHTTPClient, reasoningClient, logger)
 	agenticModelRetryCfg := newEinoAgenticModelRetryConfig(&ma.EinoMiddleware, logger, "multiagent")
 	agenticModelFailoverCfg, err := newEinoAgenticModelFailoverConfig(ctx, appCfg, &ma.EinoMiddleware, einoModelModeNormal, agenticModelFactory, logger, "multiagent", progress, orchMode, conversationID)
 	if err != nil {
@@ -232,7 +233,7 @@ func RunDeepAgent(
 			}
 			if agenticSkillMW != nil {
 				if agenticFSTools && agenticLoc != nil {
-					subFs, fsErr := subAgentAgenticFilesystemMiddleware(ctx, agenticLoc, toolInvokeNotify, id, einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
+					subFs, fsErr := subAgentAgenticFilesystemMiddleware(ctx, agenticLoc, toolInvokeNotify, id, conversationID, projectID, ma.EinoMiddleware.ReductionRootDir, toolMaxBytesFromMW(&ma.EinoMiddleware), mcpExecBinder, einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
 					if fsErr != nil {
 						return nil, fmt.Errorf("子代理 %q filesystem 中间件: %w", id, fsErr)
 					}
@@ -492,7 +493,7 @@ func RunDeepAgent(
 		}
 		var peFsMw adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage]
 		if agenticSkillMW != nil && agenticFSTools && agenticLoc != nil {
-			peFsMw, err = subAgentAgenticFilesystemMiddleware(ctx, agenticLoc, toolInvokeNotify, "executor", einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
+			peFsMw, err = subAgentAgenticFilesystemMiddleware(ctx, agenticLoc, toolInvokeNotify, "executor", conversationID, projectID, ma.EinoMiddleware.ReductionRootDir, toolMaxBytesFromMW(&ma.EinoMiddleware), mcpExecBinder, einoExecBegin, einoExecAppendPartial, einoExecRegisterCancel, einoExecUnregisterCancel, einoExecFinish, agentToolTimeoutMinutes(appCfg), agentToolWaitTimeoutSeconds(appCfg), agentShellNoOutputTimeoutSeconds(appCfg), nil)
 			if err != nil {
 				return nil, fmt.Errorf("plan_execute agentic filesystem 中间件: %w", err)
 			}
@@ -613,15 +614,18 @@ func RunDeepAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                orchMode,
-		OrchestratorName:        orchestratorName,
-		ConversationID:          conversationID,
-		Progress:                progress,
-		Logger:                  logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		OrchMode:             orchMode,
+		OrchestratorName:     orchestratorName,
+		ConversationID:       conversationID,
+		Progress:             progress,
+		Logger:               logger,
+		SnapshotMCPIDs:       snapshotMCPIDs,
+		StreamsMainAssistant: streamsMainAssistant,
+		EinoRoleTag:          einoRoleTag,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,
@@ -907,10 +911,37 @@ func tryEmitToolCallsOnce(
 	if _, ok := seen[sig]; ok {
 		return
 	}
+	if idSig := toolCallsStableIDSignature(msg); idSig != "" {
+		idKey := agentName + "\x1eids\x1e" + idSig
+		if _, ok := seen[idKey]; ok {
+			return
+		}
+		seen[idKey] = struct{}{}
+	}
 	seen[sig] = struct{}{}
 	emitToolCallsFromMessage(msg, agentName, orchestratorName, conversationID, orchMode, progress, subAgentToolStep, mainAgentToolStep, markPending)
 }
 
+func toolCallsStableIDSignature(msg *schema.Message) string {
+	if msg == nil || len(msg.ToolCalls) == 0 {
+		return ""
+	}
+	visible := filterVisibleToolCallsForProgress(msg.ToolCalls)
+	ids := make([]string, 0, len(visible))
+	for _, tc := range visible {
+		id := strings.TrimSpace(tc.ID)
+		if id == "" {
+			continue
+		}
+		ids = append(ids, id)
+	}
+	if len(ids) == 0 {
+		return ""
+	}
+	sort.Strings(ids)
+	return strings.Join(ids, ";")
+}
+
 func emitToolCallsFromMessage(
 	msg *schema.Message,
 	agentName, orchestratorName, conversationID, orchMode string,
@@ -994,6 +1025,7 @@ func emitToolCallsFromMessage(
 			markPending(toolCallPendingInfo{
 				ToolCallID: toolCallID,
 				ToolName:   display,
+				Arguments:  argsObj,
 				EinoAgent:  agentName,
 				EinoRole:   role,
 			})

internal/openai/claude_native.go

@@ -0,0 +1,245 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"strings"
+
+	"cyberstrike-ai/internal/llm"
+
+	"github.com/cloudwego/eino/components/model"
+	"github.com/cloudwego/eino/schema"
+)
+
+type claudeNativePayload struct {
+	Model               string                `json:"model"`
+	Messages            []claudeNativeMessage `json:"messages"`
+	Temperature         *float32              `json:"temperature,omitempty"`
+	TopP                *float32              `json:"top_p,omitempty"`
+	MaxCompletionTokens int                   `json:"max_completion_tokens,omitempty"`
+	MaxTokens           int                   `json:"max_tokens,omitempty"`
+	Thinking            any                   `json:"thinking,omitempty"`
+	OutputConfig        any                   `json:"output_config,omitempty"`
+}
+
+type claudeNativeMessage struct {
+	Role    string          `json:"role"`
+	Content json.RawMessage `json:"content"`
+}
+
+func (c *Client) isClaude() bool {
+	return c != nil && c.config != nil && llm.IsClaudeProvider(c.config.Provider)
+}
+
+func (c *Client) claudeNativeChatCompletion(ctx context.Context, payload, out any) error {
+	req, err := c.parseClaudeNativePayload(payload)
+	if err != nil {
+		return err
+	}
+	nativeModel, err := llm.NewClaudeAgenticModel(
+		ctx,
+		*c.config,
+		c.httpClient,
+		req.maxTokens(c.config.MaxCompletionTokensEffective()),
+		req.extraFields(),
+	)
+	if err != nil {
+		return fmt.Errorf("create native Claude model: %w", err)
+	}
+	resp, err := nativeModel.Generate(ctx, req.agenticMessages(), req.options()...)
+	if err != nil {
+		return fmt.Errorf("native Claude generate: %w", err)
+	}
+	return marshalClaudeNativeResponse(resp, req.model(c.config.Model), out)
+}
+
+func (c *Client) claudeNativeChatCompletionStream(
+	ctx context.Context,
+	payload any,
+	onDelta func(delta string) error,
+) (string, error) {
+	req, err := c.parseClaudeNativePayload(payload)
+	if err != nil {
+		return "", err
+	}
+	nativeModel, err := llm.NewClaudeAgenticModel(
+		ctx,
+		*c.config,
+		c.httpClient,
+		req.maxTokens(c.config.MaxCompletionTokensEffective()),
+		req.extraFields(),
+	)
+	if err != nil {
+		return "", fmt.Errorf("create native Claude model: %w", err)
+	}
+	stream, err := nativeModel.Stream(ctx, req.agenticMessages(), req.options()...)
+	if err != nil {
+		return "", fmt.Errorf("native Claude stream: %w", err)
+	}
+	defer stream.Close()
+
+	var full strings.Builder
+	for {
+		chunk, recvErr := stream.Recv()
+		if errors.Is(recvErr, io.EOF) {
+			return full.String(), nil
+		}
+		if recvErr != nil {
+			return full.String(), fmt.Errorf("native Claude stream receive: %w", recvErr)
+		}
+		content, _ := llm.AgenticText(chunk)
+		if content == "" {
+			continue
+		}
+		full.WriteString(content)
+		if onDelta != nil {
+			if err := onDelta(content); err != nil {
+				return full.String(), err
+			}
+		}
+	}
+}
+
+func (c *Client) parseClaudeNativePayload(payload any) (*claudeNativePayload, error) {
+	raw, err := json.Marshal(payload)
+	if err != nil {
+		return nil, fmt.Errorf("marshal Claude payload: %w", err)
+	}
+	var req claudeNativePayload
+	if err := json.Unmarshal(raw, &req); err != nil {
+		return nil, fmt.Errorf("unmarshal Claude payload: %w", err)
+	}
+	if strings.TrimSpace(req.model(c.config.Model)) == "" {
+		return nil, fmt.Errorf("native Claude model is empty")
+	}
+	if len(req.Messages) == 0 {
+		return nil, fmt.Errorf("native Claude messages are empty")
+	}
+	return &req, nil
+}
+
+func (p *claudeNativePayload) model(fallback string) string {
+	if modelName := strings.TrimSpace(p.Model); modelName != "" {
+		return modelName
+	}
+	return strings.TrimSpace(fallback)
+}
+
+func (p *claudeNativePayload) maxTokens(fallback int) int {
+	if p.MaxCompletionTokens > 0 {
+		return p.MaxCompletionTokens
+	}
+	if p.MaxTokens > 0 {
+		return p.MaxTokens
+	}
+	return fallback
+}
+
+func (p *claudeNativePayload) extraFields() map[string]any {
+	fields := make(map[string]any, 2)
+	if p.Thinking != nil {
+		fields["thinking"] = p.Thinking
+	}
+	if p.OutputConfig != nil {
+		fields["output_config"] = p.OutputConfig
+	}
+	if len(fields) == 0 {
+		return nil
+	}
+	return fields
+}
+
+func (p *claudeNativePayload) options() []model.Option {
+	opts := make([]model.Option, 0, 3)
+	if p.Temperature != nil {
+		opts = append(opts, model.WithTemperature(*p.Temperature))
+	}
+	if p.TopP != nil {
+		opts = append(opts, model.WithTopP(*p.TopP))
+	}
+	if p.MaxCompletionTokens > 0 || p.MaxTokens > 0 {
+		opts = append(opts, model.WithMaxTokens(p.maxTokens(0)))
+	}
+	return opts
+}
+
+func (p *claudeNativePayload) agenticMessages() []*schema.AgenticMessage {
+	out := make([]*schema.AgenticMessage, 0, len(p.Messages))
+	for _, msg := range p.Messages {
+		text := claudeNativeTextContent(msg.Content)
+		role := schema.AgenticRoleTypeUser
+		var block *schema.ContentBlock
+		switch strings.ToLower(strings.TrimSpace(msg.Role)) {
+		case "system":
+			role = schema.AgenticRoleTypeSystem
+			block = schema.NewContentBlock(&schema.UserInputText{Text: text})
+		case "assistant":
+			role = schema.AgenticRoleTypeAssistant
+			block = schema.NewContentBlock(&schema.AssistantGenText{Text: text})
+		default:
+			block = schema.NewContentBlock(&schema.UserInputText{Text: text})
+		}
+		out = append(out, &schema.AgenticMessage{
+			Role:          role,
+			ContentBlocks: []*schema.ContentBlock{block},
+		})
+	}
+	return out
+}
+
+func claudeNativeTextContent(raw json.RawMessage) string {
+	var text string
+	if err := json.Unmarshal(raw, &text); err == nil {
+		return text
+	}
+	var parts []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	}
+	if err := json.Unmarshal(raw, &parts); err == nil {
+		var out strings.Builder
+		for _, part := range parts {
+			if part.Type == "" || part.Type == "text" {
+				out.WriteString(part.Text)
+			}
+		}
+		return out.String()
+	}
+	return strings.TrimSpace(string(raw))
+}
+
+func marshalClaudeNativeResponse(resp *schema.AgenticMessage, modelName string, out any) error {
+	if out == nil {
+		return nil
+	}
+	content, reasoning := llm.AgenticText(resp)
+	id := ""
+	if resp != nil && resp.ResponseMeta != nil && resp.ResponseMeta.ClaudeExtension != nil {
+		id = resp.ResponseMeta.ClaudeExtension.ID
+	}
+	wire := map[string]any{
+		"id":     id,
+		"object": "chat.completion",
+		"model":  modelName,
+		"choices": []any{map[string]any{
+			"index": 0,
+			"message": map[string]any{
+				"role":              "assistant",
+				"content":           content,
+				"reasoning_content": reasoning,
+			},
+			"finish_reason": "stop",
+		}},
+	}
+	raw, err := json.Marshal(wire)
+	if err != nil {
+		return fmt.Errorf("marshal native Claude response: %w", err)
+	}
+	if err := json.Unmarshal(raw, out); err != nil {
+		return fmt.Errorf("unmarshal native Claude response: %w", err)
+	}
+	return nil
+}

internal/openai/claude_native_test.go

@@ -0,0 +1,70 @@
+package openai
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"cyberstrike-ai/internal/config"
+)
+
+type roundTripFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}
+
+func TestChatCompletionUsesNativeClaudeMessagesAPI(t *testing.T) {
+	t.Parallel()
+	httpClient := &http.Client{Transport: roundTripFunc(func(req *http.Request) (*http.Response, error) {
+		if req.URL.Path != "/v1/messages" {
+			t.Fatalf("request path = %q, want /v1/messages", req.URL.Path)
+		}
+		if req.Header.Get("x-api-key") != "test-key" {
+			t.Fatalf("x-api-key header = %q", req.Header.Get("x-api-key"))
+		}
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     http.Header{"Content-Type": []string{"application/json"}},
+			Body: io.NopCloser(strings.NewReader(`{
+				"id":"msg_1",
+				"type":"message",
+				"role":"assistant",
+				"model":"claude-test",
+				"content":[{"type":"text","text":"native ok"}],
+				"stop_reason":"end_turn",
+				"usage":{"input_tokens":1,"output_tokens":2}
+			}`)),
+			Request: req,
+		}, nil
+	})}
+	client := NewClient(&config.OpenAIConfig{
+		Provider: "claude",
+		BaseURL:  "https://example.test",
+		APIKey:   "test-key",
+		Model:    "claude-test",
+	}, httpClient, nil)
+
+	var out struct {
+		Choices []struct {
+			Message struct {
+				Content string `json:"content"`
+			} `json:"message"`
+		} `json:"choices"`
+	}
+	err := client.ChatCompletion(context.Background(), map[string]any{
+		"model": "claude-test",
+		"messages": []map[string]string{
+			{"role": "user", "content": "hello"},
+		},
+		"max_completion_tokens": 16,
+	}, &out)
+	if err != nil {
+		t.Fatalf("ChatCompletion: %v", err)
+	}
+	if len(out.Choices) != 1 || out.Choices[0].Message.Content != "native ok" {
+		t.Fatalf("response = %#v", out)
+	}
+}

internal/openai/claude_reasoning_legacy.go

@@ -0,0 +1,21 @@
+package openai
+
+import "strings"
+
+// claudeReasoningRoundTripSep is retained only to render historical traces
+// written by the removed OpenAI-to-Claude HTTP bridge.
+const claudeReasoningRoundTripSep = "\n---CSAI_CLAUDE_THINKING_BLOCKS---\n"
+
+// DisplayReasoningContent strips the obsolete bridge metadata suffix from
+// historical records. Native AgenticMessage reasoning does not add this suffix.
+func DisplayReasoningContent(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return ""
+	}
+	i := strings.LastIndex(s, claudeReasoningRoundTripSep)
+	if i < 0 {
+		return s
+	}
+	return strings.TrimSpace(s[:i])
+}

internal/openai/claude_reasoning_legacy_test.go

@@ -0,0 +1,10 @@
+package openai
+
+import "testing"
+
+func TestDisplayReasoningContentStripsLegacyClaudeSuffix(t *testing.T) {
+	raw := "hello" + claudeReasoningRoundTripSep + `[{"type":"thinking"}]`
+	if got := DisplayReasoningContent(raw); got != "hello" {
+		t.Fatalf("DisplayReasoningContent() = %q, want hello", got)
+	}
+}