feat(review): agent 侧虚拟工具 inspect_region —— 按可读分辨率放大指定区域
解决三个缺口：① 放大没复用标定/瓦片数学（LLM 一次 zoom 只出一张，宽区域仍读不清）；
②「看清没」由 VLM 自评（易过度自信/幻觉）；③ 只能按 handle / 手算 bbox 指认，
图上看到但没 handle 的东西指认不了。
核心思路：把「放大」升级成一个语义化工具——LLM 只指认「要看哪里」，由 agent 复用
tiling.py 的标定数学算出该切几张、每张覆盖多大，用构造保证可读。
新增 dwg_review/inspect.py（agent 侧，不编 DLL）：
- TOOL_SCHEMA：inspect_region(target, purpose)；target 三形态 {handle} / {window} /
{label,cx,cy,span}，purpose = text(8px) / line(6px)。
- InspectState：单图框内「截图标签 → 世界可见范围」登记表，支持逐级下钻
（返回的 inspect-N 可继续作为 label 指认）。
- resolve_target / _point_box：label+cx+cy → WCS。cy 是图像坐标（上→下），
世界 y 必须翻转（y = y2 - cy*H）——单测专门钉住。
- plan_tiles：复用 world_per_px + tiles_for_box + subdivide；超 INSPECT_MAX_TILES(4)
则硬失败并给出「缩小 target / 分多次请求」的指引（不静默降级）；拿不到比例/像素
则退回单张并如实告知。
关键几何：CAD 的 _.ZOOM _W 是等比适配（不拉伸），故「截图上的归一化坐标」必须按
max(W/ew, H/eh) 反算真实可见范围（新增 tiling.visible_box）——直接用送进去的窗口会偏。
zoom_to_window 返回的 view 是「加过 margin 的窗口」，不是可见范围，不能混用。
证据层配套：evidence_model.make_image(..., window=) 记录每张图的可见世界范围；
capture_images 用 view_px（overall 图原始像素）+ visible_box 逐张登记。
子循环接入：_run_readonly_tool 在只读白名单校验之前拦截 inspect_region（它不在
READ_ONLY_TOOLS 里，内部只调白名单内的只读 MCP 工具）；_review_tools 启用时用
inspect_region 取代三个裸视口工具；inspect_region 计入 _VIEW_TOOLS → 结束复位视口。
配置：INSPECT_ENABLED(True) / INSPECT_MAX_TILES(4)。
端到端实测（真实图纸 1 图框）：取证与审查两阶段均在最大化态、结束精确还原（4 项 PASS）；
LLM 的 11 次取证调用里 2 次是 inspect_region，且都显式给 window + purpose:"text"
（x[20000,115000] y[-6500,3000]，即 A0 图框底部横条＝标题栏所在）。
验收口径达成：报告报出标题栏乱码 ???????????????，而 get_text_content 对全图返回 0 条文字
（本图文字全在块/自定义实体内）→ 块内文字是被视觉放大读出来的。
pytest 343 passed。
This commit is contained in:
gjm committed
1 parent
3bc64b6c3e
commit
b757892c67
18 files changed
+1003
-45
No files matched your search
+1
-1
@@ -51,7 +51,7 @@ python verify_mcp_fault_injection.py # 工具异常保护（类型错误不再
|
||||
python verify_stdio_bridge.py # stdio ↔ TCP 桥（协议协商 / 转发 / 通知）
|
||||
python verify_annotate.py # 批注回写链路
|
||||
python verify_drawing_area.py # 绘图区最大化：基线 → max → restore 全回位（实测 3.68x）
|
||||
python verify_review_maximized.py # 审查全链路：取证+审查两阶段均在最大化态、结束精确还原
|
||||
python verify_review_maximized.py # 审查全链路：取证+审查两阶段均在最大化态、结束精确还原；记录 LLM 取证工具调用
|
||||
```
|
||||
|
||||
这些脚本通过 TCP 连接 `127.0.0.1:8080` 收发 JSON-RPC。
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
---
|
||||
name: Agent 路线与多源架构（2026-10-15）
|
||||
description: dwg_review agent 的总目标、多源图纸分层架构（视觉为最大公约数）、工具层与 agent 层的正交分工、多源证据层、截图去 UI、grounding 实测、热力图 + 自适应瓦片规划 + VLM 分辨率标定、绘图区最大化（set_drawing_area）与空瓦片过滤、审查阶段分辨率对齐 + 放大取证闭环、后续路线
|
||||
description: dwg_review agent 的总目标、多源图纸分层架构（视觉为最大公约数）、工具层与 agent 层的正交分工、多源证据层、截图去 UI、grounding 实测、热力图 + 自适应瓦片规划 + VLM 分辨率标定、绘图区最大化（set_drawing_area）与空瓦片过滤、审查阶段分辨率对齐、agent 侧虚拟工具 inspect_region（放大取证闭环）、后续路线
|
||||
type: project
|
||||
---
|
||||
|
||||
@@ -72,13 +72,9 @@ Agent 推理层（审查逻辑 / 清单 / 规则）
|
||||
|
||||
## 后续路线（未做）
|
||||
|
||||
1. **VLM 主动放大（自适应取证）**：复用受控只读取证子循环，让 VLM 请求放大看不清处。
|
||||
- **2026-10-16 部分落地**：审查阶段已对齐分辨率 + 提示词已明确放大协议 + 视口复位，
|
||||
实测 LLM 真的会连发 `zoom_to_window`+`get_viewport_screenshot`（见下节）。
|
||||
仍缺：① 按区域放大时的 **grounding→WCS 映射**（目前靠 LLM 自己用 bbox 算窗口，够用）；
|
||||
② 把「看清没」的判定**不交给 VLM**（防过度自信）——现靠回合上限 + 截图预算兜底。
|
||||
- 现有 `zoom_to_window`/`zoom_to_entity`/`get_viewport_screenshot` 已在只读白名单，
|
||||
**无需新增工具**。
|
||||
1. **VLM 主动放大（自适应取证）**：**2026-10-16 已落地**（见下节 `inspect_region`）。
|
||||
剩余可选打磨：把「看清没」的判定进一步形式化（目前靠 agent 用标定数学**构造保证**可读 +
|
||||
回合上限 + 截图预算兜底），以及 grounding→WCS 之外的其它指认方式。
|
||||
2. **PDF / 图片源**：在 `phases/evidence.collect()` 加分支，产出**同形** evidence
|
||||
（`window=None`、结构化空信封、`images` 带 label）；规划阶段抽象"审查单元"
|
||||
（CAD=图框 / PDF=页 / 图片=单）；批注对非 CAD 源降级（`annotator` 已天然跳过无 handle 的 issue）。
|
||||
@@ -202,3 +198,49 @@ LLM 在受控子循环里 zoom + 截图拿到的是**未最大化的 2110×462**
|
||||
→ **拍了 8 张只回灌 3 张**（其余被"已省略"占位）。改为配置项 `REVIEW_OBS_IMAGES`
|
||||
（默认 **6**，`config.py`），并把预算耗尽时的占位文案改成**明确告知模型"不要再请求截图，
|
||||
请给出结论"**（原先只说"已省略"）。实测模型截图次数不变（8），但**能看到 6 张而非 3 张**。
|
||||
|
||||
## A 收尾：agent 侧虚拟工具 `inspect_region`（2026-10-16）
|
||||
|
||||
**要解决的三个缺口**：① 放大**没复用标定/瓦片数学**（LLM 一次 zoom 只出一张，宽区域仍读不清）；
|
||||
② 「看清没」由 **VLM 自评**（易过度自信/幻觉）；③ 只能按 handle / 手算 bbox 指认，
|
||||
**图上看到但没 handle** 的东西指认不了。
|
||||
|
||||
**核心思路**：把「放大」升级成一个**语义化工具**——LLM 只指认「要看哪里」，
|
||||
由 agent 复用 `tiling.py` 的标定数学算出该切几张、每张覆盖多大，**用构造保证可读**。
|
||||
|
||||
**新增 `dwg_review/inspect.py`**（agent 侧，**不编 DLL**）：
|
||||
|
||||
- `TOOL_SCHEMA`：`inspect_region(target, purpose)`。`target` 三形态任选其一——
|
||||
`{handle}` / `{window}` / `{label, cx, cy, span}`；`purpose` = `text`(8px) / `line`(6px)。
|
||||
- `InspectState`：单图框内的「截图标签 → **世界可见范围**」登记表。初始登记取证阶段的
|
||||
`overall`/`tile-*`，每次放大再追加 `inspect-1`、`inspect-2`… → **LLM 可用上一轮返回的标签
|
||||
继续下钻**（逐级放大）。
|
||||
- `resolve_target` / `_point_box`：`label+cx+cy` → WCS。**cy 是图像坐标（上→下），
|
||||
世界 y 必须翻转**（`y = y2 - cy*H`）——最容易写错的一处，单测专门钉住。
|
||||
- `plan_tiles`：复用 `tiling.world_per_px` + `tiles_for_box` + `subdivide`。
|
||||
超出 `INSPECT_MAX_TILES`（默认 4）→ **硬失败**并给出「缩小 target / 分多次请求」的具体指引
|
||||
（不静默降级，否则等于给一堆读不清的图）；拿不到比例/像素 → 退回单张并如实告知。
|
||||
|
||||
**关键几何：`tiling.visible_box`**——CAD 的 `_.ZOOM _W` 是**等比适配（不拉伸）**，
|
||||
指定窗口按视口宽高比放大到刚好装下，多余一维露出更多内容。所以「截图上的归一化坐标」
|
||||
要准确映射回 WCS，必须按 `max(W/ew, H/eh)` 反算真实可见范围，**不能**直接用送进去的窗口。
|
||||
（`zoom_to_window` 返回的 `view` 是**加过 margin 的窗口**，不是可见范围，别混用。）
|
||||
|
||||
**证据层配套**：`evidence_model.make_image(..., window=)` 记录每张图的可见世界范围；
|
||||
`capture_images` 用 `view_px`（overall 图的原始像素）+ `visible_box` 逐张登记。
|
||||
|
||||
**子循环接入**（`phases/review.py`）：`_run_readonly_tool` 在**只读白名单校验之前**拦截
|
||||
`inspect_region`（它不在 `READ_ONLY_TOOLS` 里，内部只调白名单内的只读 MCP 工具）；
|
||||
`_review_tools` 启用时用 `inspect_region` **取代**三个裸视口工具（`zoom_to_window` /
|
||||
`zoom_to_entity` / `get_viewport_screenshot`）——一个语义化入口比三个裸工具更好用；
|
||||
`inspect_region` 计入 `_VIEW_TOOLS` → 结束时复位视口。
|
||||
配置：`INSPECT_ENABLED`(True) / `INSPECT_MAX_TILES`(4)。
|
||||
|
||||
**端到端实测（真实图纸 1 图框，`verify_review_maximized.py`）**：
|
||||
- 取证与审查两阶段均在最大化态、结束精确还原 → 4 项全 PASS。
|
||||
- LLM 的 11 次取证调用里 **2 次是 `inspect_region`**，且**都显式给了 `window` + `purpose:"text"`**：
|
||||
`x[20000,115000] y[-6500,3000]` 与 `x[-4000,30000] y[-6500,3000]`——A0 图框
|
||||
（`y[-6434,77666]`）的**底部横条**，即**标题栏**所在。模型自己把范围收窄到横条再放大。
|
||||
- **验收口径达成**：报告报出标题栏**乱码/占位字符 `???????????????`**，
|
||||
而 `get_text_content` 对全图返回 **0 条文字**（本图文字全在块/自定义实体内）——
|
||||
即**块内文字是被视觉放大读出来的**，不是结构化数据给的。
|
||||
@@ -47,7 +47,7 @@ type: reference
|
||||
| `verify_vlm_resolution_offline.py` | **离线**（纯 PIL 合成）标定 VLM 读图下限 `R_text`/`R_line`，`--dry-run` 只出图 |
|
||||
| `verify_text_resolution.py` | **CAD 实测**标定（临时加字 + 平行线，测完 `delete_entities` 清理）；`--keep` 保留 |
|
||||
| `verify_drawing_area.py` | 绘图区最大化验收：基线 → `set_drawing_area(max)` → `restore`，断言窗口矩形/最大化态/绘图区尺寸全回位（实测 3.68×） |
|
||||
| `verify_review_maximized.py` | 审查全链路验收：在 `collect`/`review_frame` 入口各采一次窗口矩形，断言**取证与审查两阶段都在最大化态**、结束精确还原；并统计子循环里的 zoom/截图调用次数 |
|
||||
| `verify_review_maximized.py` | 审查全链路验收：在 `collect`/`review_frame` 入口各采一次窗口矩形，断言**取证与审查两阶段都在最大化态**、结束精确还原；并统计子循环 zoom/截图次数 + 记录 LLM 实际调用的取证工具（含虚拟工具 `inspect_region`） |
|
||||
|
||||
> 历史：根目录 `test.py` / `test_autoagent.py` / `countOfCircle.py` / `agent_loop.py` /
|
||||
> `dwg_review_agent.py` 五个遗留调试脚本已于 2026-10-14 删除（前三个含硬编码 API Key，
|
||||
|
||||
+15
-7
@@ -53,13 +53,21 @@ def build_system_prompt(checklist, *, allow_tools=True):
|
||||
|
||||
if allow_tools:
|
||||
lines += [
|
||||
"你可以调用**只读**取证工具进一步确认细节；**不要凭猜测下结论**：",
|
||||
" - **截图里的文字/标注/细节看不清时**：用 `zoom_to_window` 放大该区域"
|
||||
"（窗口坐标用图框 bbox 或实体 bbox 计算，可适当留边距），"
|
||||
"再 `get_viewport_screenshot` 复核；已知 handle 时也可直接 `zoom_to_entity` 放大该实体。",
|
||||
" - 结构化数据不足以判断时：用 `get_entity_detail` / `find_duplicate_entities` /"
|
||||
" `find_zero_length` / `get_layers` 等查证后再下结论。",
|
||||
" - 放大截图是**复核手段**：确认清楚后即写入结论，不要为看图而反复放大。",
|
||||
"你可以调用**只读**取证工具进一步确认细节；**不要凭猜测下结论**。",
|
||||
"",
|
||||
"**看不清某处时，调用 `inspect_region` 指认位置**，它会自动按可读分辨率放大并返回清晰截图。",
|
||||
"`target` 三种指认方式任选其一：",
|
||||
' - `{"handle": "3A4"}` —— 已知实体句柄；',
|
||||
' - `{"label": "tile-3", "cx": 0.4, "cy": 0.6, "span": 0.2}` —— 某张已给截图上的一个位置',
|
||||
" （`cx` 从左到右、`cy` **从上到下**，均为 0~1 归一化；`span` 为要看的范围占该截图的比例，默认 0.25）；",
|
||||
' - `{"window": {"min_x": …, "min_y": …, "max_x": …, "max_y": …}}` —— 直接给世界坐标框。',
|
||||
'`purpose` 选 `"text"`（要读文字/标注，默认）或 `"line"`（只看走向/连接，覆盖更广）。',
|
||||
"返回的截图会带 `inspect-N` 标签，可继续作为 `label` 指认更细的位置（逐级下钻）。",
|
||||
"**你不需要判断截图够不够清楚**——`inspect_region` 会保证返回的图达到可读分辨率；",
|
||||
"若它提示区域过大，请缩小 `target`（或用 `span` 缩小范围）后重试。",
|
||||
"",
|
||||
"结构化数据不足以判断时：用 `get_entity_detail` / `find_duplicate_entities` /",
|
||||
"`find_zero_length` / `get_layers` 等查证后再下结论。",
|
||||
"",
|
||||
]
|
||||
|
||||
|
||||
@@ -94,6 +94,11 @@ REVIEW_MAX_ITERS = int(os.environ.get("REVIEW_MAX_ITERS", "4"))
|
||||
# 实测：提示词鼓励「看不清→放大重看」后，模型一轮内会发多次 zoom+截图，
|
||||
# 上限过小会让它白拍（拍了却看不到）。6 可覆盖多数复核需求。
|
||||
REVIEW_OBS_IMAGES = int(os.environ.get("REVIEW_OBS_IMAGES", "6"))
|
||||
# 是否启用 agent 侧虚拟工具 inspect_region（按可读分辨率放大指定区域）。
|
||||
# 关闭则退回「LLM 自己组合 zoom_to_window + get_viewport_screenshot」的旧行为。
|
||||
INSPECT_ENABLED = _read_bool("INSPECT_ENABLED", True)
|
||||
# inspect_region 单次请求最多返回几张瓦片（超出则中止并要求缩小 target）
|
||||
INSPECT_MAX_TILES = int(os.environ.get("INSPECT_MAX_TILES", "4"))
|
||||
# 报告输出目录
|
||||
REPORT_DIR = _resolve_path(os.environ.get("REPORT_DIR", "reports"))
|
||||
# 批注回写使用的图层
|
||||
|
||||
@@ -16,9 +16,16 @@ SOURCE_IMAGE = "image"
|
||||
STRUCTURED_SOURCES = ("entities", "texts", "dimensions", "blocks", "polylines")
|
||||
|
||||
|
||||
def make_image(b64, mime="image/png", label=None):
|
||||
"""构造 images 列表元素。"""
|
||||
return {"b64": b64, "mime": mime or "image/png", "label": label}
|
||||
def make_image(b64, mime="image/png", label=None, window=None):
|
||||
"""构造 images 列表元素。
|
||||
|
||||
`window` 为该截图**实际可见的世界范围** `(min_x,min_y,max_x,max_y)`（可选）。
|
||||
审查阶段的 `inspect_region` 靠它把「截图上的归一化坐标」映射回 WCS（逐级下钻）。
|
||||
"""
|
||||
img = {"b64": b64, "mime": mime or "image/png", "label": label}
|
||||
if window is not None:
|
||||
img["window"] = window
|
||||
return img
|
||||
|
||||
|
||||
def source_type_of(container):
|
||||
@@ -37,7 +44,7 @@ def evidence_images(evidence):
|
||||
"""规范化证据里的图片列表（单点归一）。
|
||||
|
||||
优先读 `images`；否则回退旧的 `image_b64` / `image_mime` 单图字段。
|
||||
统一返回 `[{"b64","mime","label"}]`，过滤掉空 b64。
|
||||
统一返回 `[{"b64","mime","label","window"}]`，过滤掉空 b64。
|
||||
"""
|
||||
if not isinstance(evidence, dict):
|
||||
return []
|
||||
@@ -55,6 +62,7 @@ def evidence_images(evidence):
|
||||
"b64": b64,
|
||||
"mime": item.get("mime") or "image/png",
|
||||
"label": item.get("label"),
|
||||
"window": item.get("window"),
|
||||
}
|
||||
)
|
||||
if images:
|
||||
@@ -67,6 +75,7 @@ def evidence_images(evidence):
|
||||
"b64": b64,
|
||||
"mime": evidence.get("image_mime") or "image/png",
|
||||
"label": None,
|
||||
"window": None,
|
||||
}
|
||||
]
|
||||
return []
|
||||
|
||||
@@ -0,0 +1,288 @@
|
||||
"""agent 侧虚拟工具 `inspect_region`：按**可读分辨率**放大指定区域。
|
||||
|
||||
动机（A 阶段收尾）：
|
||||
让 LLM 自己组合 `zoom_to_window` + `get_viewport_screenshot` 时——
|
||||
① 单张截图可能仍读不清（宽区域下采样后字太小）；
|
||||
② 「看清没」只能由 VLM 自评（易过度自信 / 幻觉）。
|
||||
本模块把「放大」升级为**一个语义化工具**：LLM 只指认「要看哪里」，由 agent 复用
|
||||
`tiling.py` 的标定数学算出该切几张、每张覆盖多大，**用构造保证可读**。
|
||||
|
||||
不变量：
|
||||
- 只调用**白名单内**的只读 MCP 工具（`zoom_to_window` / `get_viewport_screenshot`），
|
||||
不写库、不改图元、不动图层；
|
||||
- 会改动视口 → 调用方（`phases/review.py`）负责在该图框审查结束后复位。
|
||||
|
||||
坐标约定：
|
||||
- 世界坐标：x 向右、y **向上**；
|
||||
- 截图内归一化坐标：cx 左→右、cy **上→下**（图像坐标，左上原点）——
|
||||
`_point_box` 会翻转 y，这是最容易写错的一处。
|
||||
"""
|
||||
|
||||
from . import config, evidence_model, tiling
|
||||
from .llm_client import LlmClient, downsample_image_b64, image_size_b64
|
||||
from .phases.evidence import zoom_and_shot
|
||||
|
||||
TOOL_NAME = "inspect_region"
|
||||
|
||||
# zoom 时四周留边比例（与 `evidence.zoom_and_shot` 的默认 margin 一致）
|
||||
_ZOOM_MARGIN = 0.02
|
||||
|
||||
# 送进瓦片规划前先把目标框外扩一点，避免把边界上的字切掉
|
||||
_EDGE_PAD = 0.05
|
||||
|
||||
TOOL_SCHEMA = {
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": TOOL_NAME,
|
||||
"description": (
|
||||
"放大并返回指定区域的**可读分辨率**截图。当你看不清已给截图里的文字/标注/细节时调用；"
|
||||
"会自动计算需要几张、每张覆盖多大，**无需你判断清晰度**。"
|
||||
"target 三选一：handle（实体句柄）/ window（世界坐标框）/ "
|
||||
"label+cx+cy（某张已给截图上的归一化位置）。"
|
||||
),
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"target": {
|
||||
"type": "object",
|
||||
"description": "要放大的位置（三种形态任选其一）",
|
||||
"properties": {
|
||||
"handle": {
|
||||
"type": "string",
|
||||
"description": '实体句柄，如 "3A4"',
|
||||
},
|
||||
"window": {
|
||||
"type": "object",
|
||||
"description": "世界坐标框",
|
||||
"properties": {
|
||||
"min_x": {"type": "number"},
|
||||
"min_y": {"type": "number"},
|
||||
"max_x": {"type": "number"},
|
||||
"max_y": {"type": "number"},
|
||||
},
|
||||
"required": ["min_x", "min_y", "max_x", "max_y"],
|
||||
},
|
||||
"label": {
|
||||
"type": "string",
|
||||
"description": '已给截图的标签，如 "tile-3" 或之前返回的 "inspect-1"',
|
||||
},
|
||||
"cx": {
|
||||
"type": "number",
|
||||
"description": "该截图内的归一化横坐标 0~1（左→右）",
|
||||
},
|
||||
"cy": {
|
||||
"type": "number",
|
||||
"description": "该截图内的归一化纵坐标 0~1（上→下）",
|
||||
},
|
||||
"span": {
|
||||
"type": "number",
|
||||
"description": "以该点为中心的视野跨度（占该截图宽/高的比例），默认 0.25",
|
||||
},
|
||||
},
|
||||
},
|
||||
"purpose": {
|
||||
"type": "string",
|
||||
"enum": ["text", "line"],
|
||||
"description": "放大目的：text=读文字/标注（默认，更细）；line=看走向/连接（更粗）",
|
||||
},
|
||||
},
|
||||
"required": ["target"],
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
class InspectState:
|
||||
"""单图框内的「截图标签 → 世界可见范围」登记表。
|
||||
|
||||
初始登记取证阶段给的 `overall` / `tile-*`（它们的 `window` 字段），
|
||||
之后每次 `inspect_region` 的新截图再追加 `inspect-1`、`inspect-2`…，
|
||||
于是 LLM 可以用**上一轮返回的标签**继续下钻。
|
||||
"""
|
||||
|
||||
def __init__(self, evidence):
|
||||
self.labels = {}
|
||||
for img in evidence_model.evidence_images(evidence):
|
||||
label = img.get("label")
|
||||
box = tiling.norm_box(img.get("window"))
|
||||
if label and box is not None:
|
||||
self.labels[label] = box
|
||||
self.seq = 0
|
||||
|
||||
def register(self, box):
|
||||
"""登记一张新截图，返回它的标签。"""
|
||||
self.seq += 1
|
||||
label = f"inspect-{self.seq}"
|
||||
self.labels[label] = box
|
||||
return label
|
||||
|
||||
|
||||
def _text(s):
|
||||
return {"type": "text", "text": s}
|
||||
|
||||
|
||||
def readable_px(purpose):
|
||||
"""按用途取 VLM 可读下限（像素）。"""
|
||||
return config.TEXT_READABLE_PX if purpose == "text" else config.LINE_READABLE_PX
|
||||
|
||||
|
||||
def _view_px(evidence):
|
||||
"""绘图区**原始像素**（取 overall 图；无图/无 Pillow → None）。"""
|
||||
img = evidence_model.primary_image(evidence)
|
||||
if not img:
|
||||
return None
|
||||
return image_size_b64(img.get("b64"))
|
||||
|
||||
|
||||
def _bbox_for_handle(evidence, handle):
|
||||
"""在已采集的结构化证据里按 handle 找 bbox；找不到返回 None。"""
|
||||
want = handle.strip().upper()
|
||||
for key in evidence_model.STRUCTURED_SOURCES:
|
||||
env = evidence.get(key)
|
||||
items = env.get("items") if isinstance(env, dict) else None
|
||||
for item in items or []:
|
||||
if not isinstance(item, dict):
|
||||
continue
|
||||
got = item.get("handle")
|
||||
if isinstance(got, str) and got.strip().upper() == want:
|
||||
box = tiling.norm_box(item.get("bbox"))
|
||||
if box is not None:
|
||||
return box
|
||||
return None
|
||||
|
||||
|
||||
def _point_box(base, cx, cy, span):
|
||||
"""截图内归一化点 → 世界框。
|
||||
|
||||
`cx` 左→右；`cy` **上→下**（图像坐标），故世界 y 需翻转：`y = y2 - cy * H`。
|
||||
"""
|
||||
x1, y1, x2, y2 = base
|
||||
w, h = x2 - x1, y2 - y1
|
||||
cx = min(max(float(cx), 0.0), 1.0)
|
||||
cy = min(max(float(cy), 0.0), 1.0)
|
||||
span = min(max(float(span), 0.02), 1.0)
|
||||
px = x1 + cx * w
|
||||
py = y2 - cy * h
|
||||
hw, hh = w * span / 2.0, h * span / 2.0
|
||||
return (px - hw, py - hh, px + hw, py + hh)
|
||||
|
||||
|
||||
def resolve_target(evidence, target, state):
|
||||
"""解析 `target` → `(世界框, 人类可读说明)`；失败返回 `(None, 错误说明)`。"""
|
||||
if not isinstance(target, dict):
|
||||
return None, "target 必须是对象"
|
||||
|
||||
handle = target.get("handle")
|
||||
if isinstance(handle, str) and handle.strip():
|
||||
box = _bbox_for_handle(evidence, handle)
|
||||
if box is None:
|
||||
return None, (
|
||||
f"未在已采集证据里找到 handle {handle} 的包围盒；"
|
||||
f"请改用 window 或 label 指认位置"
|
||||
)
|
||||
return box, f"实体 {handle} 的包围盒"
|
||||
|
||||
if target.get("window") is not None:
|
||||
box = tiling.window_to_box(target.get("window"))
|
||||
if box is None:
|
||||
return None, "window 非法：需要 min_x/min_y/max_x/max_y 四个数值"
|
||||
return box, "给定的世界坐标框"
|
||||
|
||||
label = target.get("label")
|
||||
if isinstance(label, str) and label.strip():
|
||||
label = label.strip()
|
||||
base = state.labels.get(label)
|
||||
if base is None:
|
||||
known = "、".join(sorted(state.labels)) or "（无）"
|
||||
return None, f"未知截图标签 {label}；当前可用：{known}"
|
||||
try:
|
||||
cx = float(target.get("cx", 0.5))
|
||||
cy = float(target.get("cy", 0.5))
|
||||
span = float(target.get("span", 0.25))
|
||||
except (TypeError, ValueError):
|
||||
return None, "cx / cy / span 必须是数值"
|
||||
return (
|
||||
_point_box(base, cx, cy, span),
|
||||
f"截图 {label} 上 ({cx:.2f}, {cy:.2f}) 附近",
|
||||
)
|
||||
|
||||
return None, "target 需给出 handle / window / label 之一"
|
||||
|
||||
|
||||
def plan_tiles(box, scale, purpose, view_px, *, max_tiles, overlap, margin):
|
||||
"""把目标框规划成若干 zoom 窗口。
|
||||
|
||||
返回 `(windows, note)`；`windows is None` 表示**硬失败**（区域过大，已中止）。
|
||||
拿不到比例 / 绘图区像素时**退回单张**并如实告知（无法保证可读）。
|
||||
"""
|
||||
s = tiling.world_per_px(scale, config.TEXT_HEIGHT_MM, readable_px(purpose))
|
||||
eff = tiling.effective_px(view_px, config.IMAGE_MAX_EDGE)
|
||||
padded = tiling.expand_box(box, _EDGE_PAD)
|
||||
if padded is None or padded[2] <= padded[0] or padded[3] <= padded[1]:
|
||||
return None, "目标区域退化（宽或高为 0），无法放大"
|
||||
|
||||
if s <= 0 or eff[0] <= 0 or eff[1] <= 0:
|
||||
win = tiling.box_to_window(padded)
|
||||
return [win], "（拿不到出图比例或绘图区像素，已退回单张；文字可能仍读不清）"
|
||||
|
||||
cols, rows = tiling.tiles_for_box(padded, s * eff[0], s * eff[1], overlap)
|
||||
need = cols * rows
|
||||
if need > max_tiles:
|
||||
return None, (
|
||||
f"该区域按可读分辨率需 {need} 张截图，超过单次上限 {max_tiles}；"
|
||||
f"请缩小 target（例如用 label + cx + cy + span 指定更小范围）或分多次请求"
|
||||
)
|
||||
wins = tiling.subdivide(padded, cols, rows, overlap, frame_win=None)
|
||||
if not wins:
|
||||
return None, "目标区域退化，无法放大"
|
||||
return wins, f"按 {purpose} 判据切为 {cols}×{rows} = {need} 张"
|
||||
|
||||
|
||||
def inspect_region(ctx, evidence, args, image_budget, state):
|
||||
"""执行一次放大取证，返回 `(content, image_budget)`。"""
|
||||
args = args if isinstance(args, dict) else {}
|
||||
purpose = str(args.get("purpose") or "text").strip().lower()
|
||||
if purpose not in ("text", "line"):
|
||||
purpose = "text"
|
||||
|
||||
box, note = resolve_target(evidence, args.get("target"), state)
|
||||
if box is None:
|
||||
return [_text(f"inspect_region 未执行：{note}")], image_budget
|
||||
|
||||
view_px = _view_px(evidence)
|
||||
scale = tiling.frame_scale(evidence.get("frame"))
|
||||
wins, plan_note = plan_tiles(
|
||||
box, scale, purpose, view_px,
|
||||
max_tiles=config.INSPECT_MAX_TILES,
|
||||
overlap=config.TILE_OVERLAP,
|
||||
margin=_ZOOM_MARGIN,
|
||||
)
|
||||
if wins is None:
|
||||
return [_text(f"inspect_region 未执行：{plan_note}")], image_budget
|
||||
|
||||
content = [_text(f"inspect_region：{note}；{plan_note}。")]
|
||||
taken = 0
|
||||
for i, win in enumerate(wins, 1):
|
||||
if image_budget[0] <= 0:
|
||||
content.append(_text(
|
||||
f"（截图预算已用完，剩余 {len(wins) - i + 1} 张未拍；"
|
||||
f"请基于已有信息给出结论，不要再请求截图。）"
|
||||
))
|
||||
break
|
||||
shot = zoom_and_shot(ctx, win, margin=_ZOOM_MARGIN)
|
||||
if not shot:
|
||||
continue
|
||||
image_budget[0] -= 1
|
||||
visible = tiling.visible_box(tiling.window_to_box(win), view_px, _ZOOM_MARGIN)
|
||||
label = state.register(visible)
|
||||
content.append(_text(
|
||||
f"（放大截图 {label}：可见世界范围 "
|
||||
f"x[{visible[0]:.1f}, {visible[2]:.1f}] y[{visible[1]:.1f}, {visible[3]:.1f}]）"
|
||||
))
|
||||
b64, mime = downsample_image_b64(shot[0], shot[1])
|
||||
content.append(LlmClient.image_content(b64, mime))
|
||||
taken += 1
|
||||
|
||||
if taken == 0:
|
||||
content.append(_text("（未能取得截图。）"))
|
||||
return content, image_budget
|
||||
@@ -20,6 +20,9 @@ _LABEL_OVERALL = "overall"
|
||||
# 绘图区最大化用的工具名（旧版插件没有此工具 → 静默降级）
|
||||
_TOOL_DRAWING_AREA = "set_drawing_area"
|
||||
|
||||
# zoom 时四周留边比例（与 `zoom_to_window` 的默认 margin 一致）
|
||||
_ZOOM_MARGIN = 0.02
|
||||
|
||||
|
||||
@contextlib.contextmanager
|
||||
def maximized_drawing_area(ctx):
|
||||
@@ -95,7 +98,7 @@ def sub_windows(window, tiles):
|
||||
return out
|
||||
|
||||
|
||||
def _zoom_and_shot(ctx, window, margin=0.02):
|
||||
def zoom_and_shot(ctx, window, margin=_ZOOM_MARGIN):
|
||||
"""缩放到窗口、等待渲染、截一张图，返回 (b64, mime) 或 None。"""
|
||||
ctx.tools.call("zoom_to_window", dict(window, margin=margin))
|
||||
time.sleep(config.RENDER_WAIT)
|
||||
@@ -209,6 +212,8 @@ def capture_images(ctx, window, *, tiles=0, max_images=None, frame=None,
|
||||
- 局部窗口优先由 `_plan_tiles` 按**实体分布**给出（空白区不出图）；
|
||||
不可用时退回 `sub_windows(window, tiles)`，**同样按内容过滤**空白象限
|
||||
（过滤后为空则退回未过滤集合，保证不漏拍）。
|
||||
- 每张图登记其**实际可见世界范围**（`window` 字段）——审查阶段的 `inspect_region`
|
||||
靠它把「截图上的归一化坐标」映射回 WCS（逐级下钻）。
|
||||
- 结束时**复位到整体视口**，保证后续流程从整幅视图起手。
|
||||
- 受 max_images（默认 config.EVIDENCE_MAX_IMAGES）截断。
|
||||
"""
|
||||
@@ -216,14 +221,23 @@ def capture_images(ctx, window, *, tiles=0, max_images=None, frame=None,
|
||||
max_images = config.EVIDENCE_MAX_IMAGES
|
||||
images = []
|
||||
|
||||
def _add(shot, label):
|
||||
if shot and len(images) < max_images:
|
||||
images.append(evidence_model.make_image(shot[0], shot[1], label))
|
||||
|
||||
# 整体图：既是证据，也用于测量绘图区**实际像素**（决定单张截图的世界覆盖范围）
|
||||
overall = _zoom_and_shot(ctx, window)
|
||||
overall = zoom_and_shot(ctx, window)
|
||||
view_px = image_size_b64(overall[0]) if overall else None
|
||||
_add(overall, _LABEL_OVERALL)
|
||||
|
||||
def _add(shot, label, win):
|
||||
"""登记截图；`win` 为送给 CAD 的窗口，用于反算实际可见世界范围。"""
|
||||
if shot and len(images) < max_images:
|
||||
images.append(
|
||||
evidence_model.make_image(
|
||||
shot[0], shot[1], label,
|
||||
window=tiling.visible_box(
|
||||
tiling.window_to_box(win), view_px, _ZOOM_MARGIN
|
||||
),
|
||||
)
|
||||
)
|
||||
|
||||
_add(overall, _LABEL_OVERALL, window)
|
||||
|
||||
# 内容过滤：空白区不出图。按**实体 bbox** 判定，规划与回退两条路径共用同一谓词；
|
||||
# 实体数据被截断时不可信，放弃过滤（宁可多拍也不漏）。
|
||||
@@ -237,10 +251,10 @@ def capture_images(ctx, window, *, tiles=0, max_images=None, frame=None,
|
||||
for i, sub in enumerate(regions, 1):
|
||||
if len(images) >= max_images:
|
||||
break
|
||||
_add(_zoom_and_shot(ctx, sub), f"tile-{i}")
|
||||
_add(zoom_and_shot(ctx, sub), f"tile-{i}", sub)
|
||||
|
||||
# 复位整体视口（不截图）
|
||||
ctx.tools.call("zoom_to_window", dict(window, margin=0.02))
|
||||
ctx.tools.call("zoom_to_window", dict(window, margin=_ZOOM_MARGIN))
|
||||
return images
|
||||
|
||||
|
||||
|
||||
@@ -14,7 +14,7 @@
|
||||
|
||||
import json
|
||||
|
||||
from .. import config, evidence_model
|
||||
from .. import config, evidence_model, inspect
|
||||
from ..cad_tools import READ_ONLY_TOOLS, is_write_tool
|
||||
from ..findings import parse_conclusion
|
||||
from ..llm_client import LlmClient, downsample_image_b64
|
||||
@@ -24,7 +24,13 @@ from ..mcp_client import McpError, McpToolError, result_images, result_texts
|
||||
_MAX_OBS_IMAGES = config.REVIEW_OBS_IMAGES
|
||||
|
||||
# 会改动视口的只读取证工具：用过之后须把视口复位，别把用户的 CAD 留在放大处
|
||||
_VIEW_TOOLS = frozenset({"zoom_to_window", "zoom_to_entity"})
|
||||
_VIEW_TOOLS = frozenset({"zoom_to_window", "zoom_to_entity", inspect.TOOL_NAME})
|
||||
|
||||
# 启用 inspect_region 时，这三个裸视口工具不再直接提供给 LLM（由 inspect_region 统一代理），
|
||||
# 但仍留在只读白名单里，作为兜底仍然可被调用。
|
||||
_RAW_VIEW_TOOLS = frozenset(
|
||||
{"zoom_to_window", "zoom_to_entity", "get_viewport_screenshot"}
|
||||
)
|
||||
|
||||
|
||||
def evidence_payload(evidence):
|
||||
@@ -45,8 +51,17 @@ def evidence_payload(evidence):
|
||||
}
|
||||
|
||||
|
||||
def _run_readonly_tool(ctx, name, args, image_budget):
|
||||
"""执行一个只读取证工具，返回回灌给 LLM 的 content 列表。"""
|
||||
def _run_readonly_tool(ctx, name, args, image_budget, *, evidence=None, state=None):
|
||||
"""执行一个只读取证工具，返回回灌给 LLM 的 content 列表。
|
||||
|
||||
`inspect_region` 是 **agent 侧虚拟工具**（不走 MCP），在只读白名单校验**之前**拦截：
|
||||
它内部只调白名单内的只读 MCP 工具。
|
||||
"""
|
||||
if name == inspect.TOOL_NAME:
|
||||
if not config.INSPECT_ENABLED or evidence is None or state is None:
|
||||
return [{"type": "text", "text": "inspect_region 当前不可用。"}], image_budget
|
||||
return inspect.inspect_region(ctx, evidence, args, image_budget, state)
|
||||
|
||||
if name not in READ_ONLY_TOOLS or is_write_tool(name):
|
||||
return [{"type": "text", "text": f"工具 {name} 不在只读取证白名单内，已拒绝。"}], image_budget
|
||||
|
||||
@@ -75,6 +90,19 @@ def _run_readonly_tool(ctx, name, args, image_budget):
|
||||
return content, image_budget
|
||||
|
||||
|
||||
def _review_tools(ctx):
|
||||
"""审查子循环提供给 LLM 的工具 schema。
|
||||
|
||||
启用 `inspect_region` 时，用它**取代**三个裸视口工具（`zoom_to_window` /
|
||||
`zoom_to_entity` / `get_viewport_screenshot`）——一个语义化入口比三个裸工具更好用，
|
||||
且可读性由 agent 保证；关闭时维持原样（退回旧行为）。
|
||||
"""
|
||||
if config.INSPECT_ENABLED:
|
||||
names = READ_ONLY_TOOLS - _RAW_VIEW_TOOLS
|
||||
return ctx.tools.openai_tools(names=names) + [inspect.TOOL_SCHEMA]
|
||||
return ctx.tools.openai_tools(names=READ_ONLY_TOOLS)
|
||||
|
||||
|
||||
def _reset_viewport(ctx, evidence):
|
||||
"""把视口复位到该图框的整幅视图（审查中的放大不应残留给用户）。"""
|
||||
win = evidence.get("window")
|
||||
@@ -121,8 +149,9 @@ def _review_frame(ctx, evidence, system_prompt, log, view_changed):
|
||||
{"role": "user", "content": user_content},
|
||||
]
|
||||
|
||||
tools_schema = ctx.tools.openai_tools(names=READ_ONLY_TOOLS)
|
||||
tools_schema = _review_tools(ctx)
|
||||
image_budget = [_MAX_OBS_IMAGES]
|
||||
state = inspect.InspectState(evidence)
|
||||
|
||||
for _ in range(max(0, ctx.max_iters)):
|
||||
resp = ctx.llm.chat(messages, tools=tools_schema)
|
||||
@@ -142,7 +171,9 @@ def _review_frame(ctx, evidence, system_prompt, log, view_changed):
|
||||
log(f" [取证] {name} {args if args else ''}")
|
||||
if name in _VIEW_TOOLS:
|
||||
view_changed[0] = True
|
||||
content, image_budget = _run_readonly_tool(ctx, name, args, image_budget)
|
||||
content, image_budget = _run_readonly_tool(
|
||||
ctx, name, args, image_budget, evidence=evidence, state=state
|
||||
)
|
||||
messages.append(
|
||||
{"role": "tool", "tool_call_id": tool_call.id, "content": content}
|
||||
)
|
||||
|
||||
@@ -163,6 +163,66 @@ def boxes_intersect(a, b, eps=0.0):
|
||||
)
|
||||
|
||||
|
||||
def box_to_window(box):
|
||||
"""bbox 元组 → window 对象（`zoom_to_window` 的入参）；非法返回 None。"""
|
||||
b = norm_box(box)
|
||||
if b is None:
|
||||
return None
|
||||
return {"min_x": b[0], "min_y": b[1], "max_x": b[2], "max_y": b[3]}
|
||||
|
||||
|
||||
def window_to_box(win):
|
||||
"""window 对象 → bbox 元组；非法返回 None。"""
|
||||
if not isinstance(win, dict):
|
||||
return None
|
||||
return norm_box(
|
||||
[win.get("min_x"), win.get("min_y"), win.get("max_x"), win.get("max_y")]
|
||||
)
|
||||
|
||||
|
||||
def expand_box(box, frac):
|
||||
"""把 bbox 四周按 `frac` 比例外扩；非法输入返回 None。"""
|
||||
b = norm_box(box)
|
||||
if b is None:
|
||||
return None
|
||||
try:
|
||||
f = float(frac or 0.0)
|
||||
except (TypeError, ValueError):
|
||||
f = 0.0
|
||||
mx, my = (b[2] - b[0]) * f, (b[3] - b[1]) * f
|
||||
return (b[0] - mx, b[1] - my, b[2] + mx, b[3] + my)
|
||||
|
||||
|
||||
def visible_box(box, view_px, margin=0.0):
|
||||
"""`zoom_to_window(box, margin)` 之后**实际可见**的世界范围。
|
||||
|
||||
CAD 的 `_.ZOOM _W` 是**等比适配**（不拉伸）：指定窗口按视口宽高比放大到刚好装下，
|
||||
多余的一维露出更多内容。故可见范围 = 以窗口中心为中心、按 `max(W/ew, H/eh)`
|
||||
放大后的矩形。`view_px` 为截图的**原始像素**（未下采样）；拿不到时退回
|
||||
「窗口外扩 margin」的近似值。
|
||||
|
||||
用途：审查阶段把「截图上的归一化坐标」准确映射回 WCS（逐级下钻）。
|
||||
"""
|
||||
b = expand_box(box, margin)
|
||||
if b is None:
|
||||
return None
|
||||
if not view_px or len(view_px) != 2:
|
||||
return b
|
||||
try:
|
||||
ew, eh = int(view_px[0]), int(view_px[1])
|
||||
except (TypeError, ValueError):
|
||||
return b
|
||||
if ew <= 0 or eh <= 0:
|
||||
return b
|
||||
w, h = b[2] - b[0], b[3] - b[1]
|
||||
if w <= 0 or h <= 0:
|
||||
return b
|
||||
k = max(w / ew, h / eh)
|
||||
vw, vh = ew * k, eh * k
|
||||
cx, cy = (b[0] + b[2]) / 2.0, (b[1] + b[3]) / 2.0
|
||||
return (cx - vw / 2.0, cy - vh / 2.0, cx + vw / 2.0, cy + vh / 2.0)
|
||||
|
||||
|
||||
def cluster_text_flags(clusters, text_boxes):
|
||||
"""标记每个簇是否含文字类实体（bbox 相交即算，保守）。
|
||||
|
||||
|
||||
@@ -65,12 +65,15 @@ def test_build_system_prompt_tool_hint_toggle():
|
||||
assert "只读" not in without
|
||||
|
||||
|
||||
def test_build_system_prompt_instructs_zoom_when_unreadable():
|
||||
def test_build_system_prompt_instructs_inspect_region():
|
||||
prompt = build_system_prompt(DEFAULT_CHECKLIST, allow_tools=True)
|
||||
assert "看不清" in prompt
|
||||
assert "zoom_to_window" in prompt
|
||||
assert "zoom_to_entity" in prompt
|
||||
assert "get_viewport_screenshot" in prompt
|
||||
# 放大取证走 inspect_region 这一个语义化入口（不再让 LLM 自己组合裸视口工具）
|
||||
assert "inspect_region" in prompt
|
||||
assert "handle" in prompt and "label" in prompt and "window" in prompt
|
||||
assert "cx" in prompt and "cy" in prompt
|
||||
assert "从上到下" in prompt # cy 方向必须写清（图像坐标）
|
||||
assert "inspect-N" in prompt # 返回标签可继续下钻
|
||||
assert "不需要判断截图够不够清楚" in prompt
|
||||
|
||||
|
||||
def test_build_system_prompt_requires_structured_json_output():
|
||||
|
||||
@@ -106,6 +106,12 @@ def test_review_obs_images_default_is_sane():
|
||||
assert config.REVIEW_OBS_IMAGES >= 3
|
||||
|
||||
|
||||
def test_inspect_defaults_are_sane():
|
||||
# agent 侧虚拟工具默认启用；单次瓦片上限为正
|
||||
assert config.INSPECT_ENABLED is True
|
||||
assert config.INSPECT_MAX_TILES >= 1
|
||||
|
||||
|
||||
def test_drawing_area_constants_have_expected_types():
|
||||
assert isinstance(config.DRAWING_AREA_MAXIMIZE, bool)
|
||||
assert isinstance(config.DRAWING_AREA_ENTER_WAIT, float)
|
||||
|
||||
@@ -77,7 +77,15 @@ def test_collect_frame_assembles_evidence():
|
||||
assert ev["blocks"]["total"] == 0
|
||||
assert ev["polylines"]["total"] == 3
|
||||
assert ev["source_type"] == "cad"
|
||||
assert ev["images"] == [{"b64": "QUJD", "mime": "image/png", "label": "overall"}]
|
||||
# 截图登记「实际可见世界范围」：拿不到绘图区像素时退回「窗口外扩 margin」的近似值
|
||||
assert ev["images"] == [
|
||||
{
|
||||
"b64": "QUJD",
|
||||
"mime": "image/png",
|
||||
"label": "overall",
|
||||
"window": (-2.0, -1.0, 102.0, 51.0),
|
||||
}
|
||||
]
|
||||
|
||||
|
||||
def test_collect_frame_calls_expected_tools():
|
||||
|
||||
@@ -52,7 +52,7 @@ def test_evidence_images_filters_empty_b64():
|
||||
def test_evidence_images_legacy_fallback():
|
||||
ev = {"image_b64": "LEG", "image_mime": "image/jpeg"}
|
||||
assert em.evidence_images(ev) == [
|
||||
{"b64": "LEG", "mime": "image/jpeg", "label": None}
|
||||
{"b64": "LEG", "mime": "image/jpeg", "label": None, "window": None}
|
||||
]
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,348 @@
|
||||
"""`dwg_review.inspect` 单元测试（agent 侧虚拟工具 `inspect_region`）。"""
|
||||
|
||||
import base64
|
||||
import io
|
||||
|
||||
import pytest
|
||||
|
||||
from dwg_review import config, inspect
|
||||
from dwg_review.inspect import (
|
||||
InspectState,
|
||||
_point_box,
|
||||
inspect_region,
|
||||
plan_tiles,
|
||||
readable_px,
|
||||
resolve_target,
|
||||
)
|
||||
from dwg_review.phases.review import review_frame
|
||||
from tests.fakes import (
|
||||
FakeLlmClient,
|
||||
FakeMcpClient,
|
||||
frame,
|
||||
llm_calls,
|
||||
llm_text,
|
||||
make_ctx,
|
||||
make_evidence,
|
||||
ok_image,
|
||||
)
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def inspect_env(monkeypatch):
|
||||
"""钉死 inspect 相关配置，使用例不受外部环境变量影响。"""
|
||||
for name, value in {
|
||||
"TEXT_HEIGHT_MM": 2.0,
|
||||
"TEXT_READABLE_PX": 8,
|
||||
"LINE_READABLE_PX": 6,
|
||||
"INSPECT_MAX_TILES": 4,
|
||||
"INSPECT_ENABLED": True,
|
||||
"TILE_OVERLAP": 0.0,
|
||||
"IMAGE_MAX_EDGE": 1280,
|
||||
"RENDER_WAIT": 0.0,
|
||||
}.items():
|
||||
monkeypatch.setattr(config, name, value)
|
||||
return config
|
||||
|
||||
|
||||
def _png_b64(w=2000, h=1414):
|
||||
from PIL import Image
|
||||
|
||||
buf = io.BytesIO()
|
||||
Image.new("RGB", (w, h), "black").save(buf, format="PNG")
|
||||
return base64.b64encode(buf.getvalue()).decode("ascii")
|
||||
|
||||
|
||||
def _evidence(images=None, entities=None, sheet_scale=100.0, sheet_size="A0"):
|
||||
"""图框 (0,0)-(120000,40000)、1:100、A0 的 evidence（含 overall 截图）。"""
|
||||
ev = make_evidence()
|
||||
ev["frame"] = frame([0, 0, 120000, 40000], sheet_size=sheet_size, sheet_scale=sheet_scale)
|
||||
ev["window"] = {"min_x": 0.0, "min_y": 0.0, "max_x": 120000.0, "max_y": 40000.0}
|
||||
ev["entities"] = entities or {
|
||||
"total": 1,
|
||||
"items": [{"handle": "A1", "type": "LINE", "bbox": [0, 0, 1000, 500]}],
|
||||
}
|
||||
if images is None:
|
||||
images = [
|
||||
{
|
||||
"b64": _png_b64(),
|
||||
"mime": "image/png",
|
||||
"label": "overall",
|
||||
"window": (0.0, 0.0, 120000.0, 40000.0),
|
||||
}
|
||||
]
|
||||
ev["images"] = images
|
||||
return ev
|
||||
|
||||
|
||||
def _client(b64=None):
|
||||
return FakeMcpClient(
|
||||
responses={"get_viewport_screenshot": ok_image(b64 or _png_b64())}
|
||||
)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# InspectState
|
||||
# ---------------------------------------------------------------------------
|
||||
def test_inspect_state_registers_evidence_labels_with_window():
|
||||
ev = _evidence(images=[
|
||||
{"b64": "A", "mime": "image/png", "label": "overall", "window": (0, 0, 10, 10)},
|
||||
{"b64": "B", "mime": "image/png", "label": "tile-1", "window": [0, 0, 5, 5]},
|
||||
{"b64": "C", "mime": "image/png", "label": "tile-2"}, # 无 window → 不登记
|
||||
])
|
||||
state = InspectState(ev)
|
||||
assert set(state.labels) == {"overall", "tile-1"}
|
||||
assert state.labels["tile-1"] == (0.0, 0.0, 5.0, 5.0)
|
||||
|
||||
|
||||
def test_inspect_state_register_assigns_incrementing_labels():
|
||||
state = InspectState(_evidence())
|
||||
assert state.register((1, 2, 3, 4)) == "inspect-1"
|
||||
assert state.register((0, 0, 1, 1)) == "inspect-2"
|
||||
assert state.labels["inspect-1"] == (1, 2, 3, 4)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# _point_box / resolve_target
|
||||
# ---------------------------------------------------------------------------
|
||||
def test_point_box_flips_y_because_cy_is_image_down():
|
||||
base = (0.0, 0.0, 100.0, 50.0)
|
||||
top = _point_box(base, 0.5, 0.0, 0.5) # 图像顶部
|
||||
bottom = _point_box(base, 0.5, 1.0, 0.5) # 图像底部
|
||||
assert (top[1] + top[3]) / 2 == pytest.approx(50.0) # → 世界 y 大的一侧
|
||||
assert (bottom[1] + bottom[3]) / 2 == pytest.approx(0.0)
|
||||
|
||||
|
||||
def test_point_box_centers_and_clamps_span():
|
||||
base = (0.0, 0.0, 100.0, 50.0)
|
||||
box = _point_box(base, 0.5, 0.5, 0.5)
|
||||
assert box == pytest.approx((25.0, 12.5, 75.0, 37.5))
|
||||
assert _point_box(base, 0.5, 0.5, 5.0) == pytest.approx(base) # span 上限 1
|
||||
tiny = _point_box(base, 0.5, 0.5, 0.0)
|
||||
assert (tiny[2] - tiny[0]) == pytest.approx(100.0 * 0.02) # span 下限 0.02
|
||||
|
||||
|
||||
def test_resolve_target_by_handle_is_case_insensitive():
|
||||
ev = _evidence()
|
||||
box, note = resolve_target(ev, {"handle": "a1"}, InspectState(ev))
|
||||
assert box == (0.0, 0.0, 1000.0, 500.0)
|
||||
assert "a1" in note
|
||||
|
||||
|
||||
def test_resolve_target_handle_not_found():
|
||||
ev = _evidence()
|
||||
box, note = resolve_target(ev, {"handle": "ZZ"}, InspectState(ev))
|
||||
assert box is None
|
||||
assert "ZZ" in note and "window" in note
|
||||
|
||||
|
||||
def test_resolve_target_by_window():
|
||||
ev = _evidence()
|
||||
box, note = resolve_target(
|
||||
ev, {"window": {"min_x": 1, "min_y": 2, "max_x": 3, "max_y": 4}}, InspectState(ev)
|
||||
)
|
||||
assert box == (1.0, 2.0, 3.0, 4.0)
|
||||
assert note
|
||||
|
||||
|
||||
def test_resolve_target_bad_window():
|
||||
ev = _evidence()
|
||||
box, note = resolve_target(ev, {"window": {"min_x": 1}}, InspectState(ev))
|
||||
assert box is None
|
||||
assert "window 非法" in note
|
||||
|
||||
|
||||
def test_resolve_target_by_label_point():
|
||||
ev = _evidence()
|
||||
state = InspectState(ev)
|
||||
box, note = resolve_target(
|
||||
ev, {"label": "overall", "cx": 0.5, "cy": 0.5, "span": 0.5}, state
|
||||
)
|
||||
# overall 覆盖 (0,0)-(120000,40000) → 中心 (60000, 20000)，跨度为半幅
|
||||
assert box == pytest.approx((30000.0, 10000.0, 90000.0, 30000.0))
|
||||
assert "overall" in note
|
||||
|
||||
|
||||
def test_resolve_target_unknown_label_lists_known():
|
||||
ev = _evidence()
|
||||
box, note = resolve_target(ev, {"label": "tile-9"}, InspectState(ev))
|
||||
assert box is None
|
||||
assert "tile-9" in note and "overall" in note
|
||||
|
||||
|
||||
def test_resolve_target_rejects_bad_inputs():
|
||||
ev = _evidence()
|
||||
state = InspectState(ev)
|
||||
assert resolve_target(ev, "nope", state)[0] is None
|
||||
assert resolve_target(ev, {}, state)[0] is None
|
||||
assert resolve_target(ev, {"label": "overall", "cx": "x"}, state)[0] is None
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# plan_tiles
|
||||
# ---------------------------------------------------------------------------
|
||||
def test_readable_px_by_purpose(inspect_env):
|
||||
assert readable_px("text") == inspect_env.TEXT_READABLE_PX
|
||||
assert readable_px("line") == inspect_env.LINE_READABLE_PX
|
||||
|
||||
|
||||
def test_plan_tiles_single_tile_for_small_region(inspect_env):
|
||||
wins, note = plan_tiles(
|
||||
(0, 0, 1000, 500), 100.0, "text", (2000, 1414),
|
||||
max_tiles=4, overlap=0.0, margin=0.02,
|
||||
)
|
||||
assert len(wins) == 1
|
||||
assert "1×1" in note
|
||||
# 目标框被外扩 5% 留边
|
||||
assert wins[0]["min_x"] < 0 and wins[0]["max_x"] > 1000
|
||||
|
||||
|
||||
def test_plan_tiles_splits_wide_region(inspect_env):
|
||||
# 66000 世界单位宽（含 5% 留边）/ 单张 32000 → 3 张
|
||||
wins, note = plan_tiles(
|
||||
(0, 0, 60000, 20000), 100.0, "text", (2000, 1414),
|
||||
max_tiles=9, overlap=0.0, margin=0.02,
|
||||
)
|
||||
assert len(wins) == 3
|
||||
assert "3 张" in note
|
||||
|
||||
|
||||
def test_plan_tiles_fatal_when_over_cap(inspect_env):
|
||||
wins, note = plan_tiles(
|
||||
(0, 0, 100000, 100000), 100.0, "text", (2000, 1414),
|
||||
max_tiles=4, overlap=0.0, margin=0.02,
|
||||
)
|
||||
assert wins is None
|
||||
assert "超过单次上限 4" in note
|
||||
assert "span" in note # 给出缩小 target 的具体指引
|
||||
|
||||
|
||||
def test_plan_tiles_falls_back_to_single_without_scale(inspect_env):
|
||||
wins, note = plan_tiles(
|
||||
(0, 0, 1000, 500), 0.0, "text", (2000, 1414),
|
||||
max_tiles=4, overlap=0.0, margin=0.02,
|
||||
)
|
||||
assert len(wins) == 1
|
||||
assert "退回单张" in note
|
||||
|
||||
|
||||
def test_plan_tiles_degenerate_box_is_fatal(inspect_env):
|
||||
wins, note = plan_tiles(
|
||||
(0, 0, 0, 0), 100.0, "text", (2000, 1414),
|
||||
max_tiles=4, overlap=0.0, margin=0.02,
|
||||
)
|
||||
assert wins is None
|
||||
assert "退化" in note
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# inspect_region（执行体）
|
||||
# ---------------------------------------------------------------------------
|
||||
def test_inspect_region_handle_success(inspect_env):
|
||||
pytest.importorskip("PIL")
|
||||
client = _client()
|
||||
ctx = make_ctx(client=client)
|
||||
ev = _evidence()
|
||||
state = InspectState(ev)
|
||||
|
||||
content, budget = inspect_region(ctx, ev, {"target": {"handle": "A1"}}, [6], state)
|
||||
|
||||
assert budget[0] == 5 # 只用掉 1 张
|
||||
assert "实体 A1" in content[0]["text"]
|
||||
assert any(c.get("type") == "image_url" for c in content)
|
||||
assert "zoom_to_window" in client.called_names()
|
||||
assert "inspect-1" in state.labels
|
||||
# 登记的可见范围应**包含**目标框（ZOOM W 等比适配会露出更多）
|
||||
vis = state.labels["inspect-1"]
|
||||
assert vis[0] <= 0.0 and vis[2] >= 1000.0
|
||||
|
||||
|
||||
def test_inspect_region_label_target_reports_label_and_range(inspect_env):
|
||||
pytest.importorskip("PIL")
|
||||
ctx = make_ctx(client=_client())
|
||||
ev = _evidence()
|
||||
state = InspectState(ev)
|
||||
|
||||
content, budget = inspect_region(
|
||||
ctx, ev, {"target": {"label": "overall", "cx": 0.5, "cy": 0.5, "span": 0.5}}, [6], state
|
||||
)
|
||||
assert budget[0] == 3 # 中心半幅需 3 张瓦片
|
||||
assert "overall" in content[0]["text"]
|
||||
assert any("可见世界范围" in c.get("text", "") for c in content)
|
||||
|
||||
|
||||
def test_inspect_region_purpose_line_is_offered(inspect_env):
|
||||
pytest.importorskip("PIL")
|
||||
ctx = make_ctx(client=_client())
|
||||
ev = _evidence()
|
||||
content, _ = inspect_region(
|
||||
ctx, ev, {"target": {"handle": "A1"}, "purpose": "line"}, [6], InspectState(ev)
|
||||
)
|
||||
assert "line" in content[0]["text"]
|
||||
|
||||
|
||||
def test_inspect_region_bad_target_makes_no_mcp_calls(inspect_env):
|
||||
client = _client()
|
||||
ctx = make_ctx(client=client)
|
||||
ev = _evidence()
|
||||
content, budget = inspect_region(
|
||||
ctx, ev, {"target": {"label": "nope"}}, [6], InspectState(ev)
|
||||
)
|
||||
assert budget[0] == 6
|
||||
assert "未执行" in content[0]["text"]
|
||||
assert client.called_names() == []
|
||||
|
||||
|
||||
def test_inspect_region_oversized_target_makes_no_mcp_calls(inspect_env):
|
||||
client = _client()
|
||||
ctx = make_ctx(client=client)
|
||||
ev = _evidence()
|
||||
content, budget = inspect_region(
|
||||
ctx, ev,
|
||||
{"target": {"window": {"min_x": 0, "min_y": 0, "max_x": 200000, "max_y": 200000}}},
|
||||
[6], InspectState(ev),
|
||||
)
|
||||
assert budget[0] == 6
|
||||
assert "超过单次上限" in content[0]["text"]
|
||||
assert client.called_names() == []
|
||||
|
||||
|
||||
def test_inspect_region_budget_exhausted_returns_no_images(inspect_env):
|
||||
pytest.importorskip("PIL")
|
||||
ctx = make_ctx(client=_client())
|
||||
ev = _evidence()
|
||||
content, budget = inspect_region(
|
||||
ctx, ev, {"target": {"handle": "A1"}}, [0], InspectState(ev)
|
||||
)
|
||||
assert budget[0] == 0
|
||||
assert all(c.get("type") == "text" for c in content)
|
||||
assert any("预算已用完" in c.get("text", "") for c in content)
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 与审查子循环的集成
|
||||
# ---------------------------------------------------------------------------
|
||||
def test_review_frame_can_call_inspect_region(inspect_env):
|
||||
pytest.importorskip("PIL")
|
||||
llm = FakeLlmClient(script=[
|
||||
llm_calls([("inspect_region", {"target": {"handle": "A1"}})]),
|
||||
llm_text('{"summary": "结论", "issues": []}'),
|
||||
])
|
||||
client = _client()
|
||||
ctx = make_ctx(client=client, llm=llm)
|
||||
ev = _evidence()
|
||||
|
||||
out = review_frame(ctx, ev, "sys", log=lambda *a: None)
|
||||
|
||||
assert out["summary"] == "结论"
|
||||
names = client.called_names()
|
||||
assert "zoom_to_window" in names # inspect_region 内部真的缩放了
|
||||
assert names.count("zoom_to_window") == 2 # 放大 1 次 + 结束复位 1 次
|
||||
|
||||
|
||||
def test_review_frame_offers_inspect_region_to_llm(inspect_env):
|
||||
llm = FakeLlmClient(script=[llm_text("x")])
|
||||
ctx = make_ctx(client=FakeMcpClient(), llm=llm)
|
||||
review_frame(ctx, make_evidence(), "sys", log=lambda *a: None)
|
||||
names = {t["function"]["name"] for t in llm.calls[0]["tools"]}
|
||||
assert "inspect_region" in names
|
||||
assert "zoom_to_window" not in names
|
||||
@@ -2,9 +2,13 @@
|
||||
|
||||
import json
|
||||
|
||||
import pytest
|
||||
|
||||
from dwg_review import config
|
||||
from dwg_review.cad_tools import READ_ONLY_TOOLS
|
||||
from dwg_review.mcp_client import McpToolError
|
||||
from dwg_review.phases.review import (
|
||||
_review_tools,
|
||||
_run_readonly_tool,
|
||||
evidence_payload,
|
||||
review_frame,
|
||||
@@ -339,3 +343,60 @@ def test_review_frame_no_viewport_reset_without_window():
|
||||
ev["window"] = None # 非 CAD 源（无窗口）→ 无从复位
|
||||
review_frame(ctx, ev, "sys", log=lambda *a: None)
|
||||
assert client.called_names().count("zoom_to_window") == 1
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 虚拟工具 inspect_region 的接入
|
||||
# ---------------------------------------------------------------------------
|
||||
def _tool_names(ctx):
|
||||
return {t["function"]["name"] for t in _review_tools(ctx)}
|
||||
|
||||
|
||||
def test_review_tools_swap_raw_view_tools_for_inspect_region(monkeypatch):
|
||||
monkeypatch.setattr(config, "INSPECT_ENABLED", True)
|
||||
names = _tool_names(make_ctx())
|
||||
assert "inspect_region" in names
|
||||
assert "query_entities" in names # 其余只读工具照旧
|
||||
assert "zoom_to_window" not in names
|
||||
assert "zoom_to_entity" not in names
|
||||
assert "get_viewport_screenshot" not in names
|
||||
|
||||
|
||||
def test_review_tools_keep_raw_view_tools_when_disabled(monkeypatch):
|
||||
monkeypatch.setattr(config, "INSPECT_ENABLED", False)
|
||||
names = _tool_names(make_ctx())
|
||||
assert "inspect_region" not in names
|
||||
assert {"zoom_to_window", "zoom_to_entity", "get_viewport_screenshot"} <= names
|
||||
|
||||
|
||||
def test_readonly_intercepts_inspect_region_before_whitelist(monkeypatch):
|
||||
# inspect_region 不在 READ_ONLY_TOOLS 里，但必须被**先**拦截（不能判成越权）
|
||||
monkeypatch.setattr(config, "INSPECT_ENABLED", True)
|
||||
ctx = make_ctx(client=FakeMcpClient())
|
||||
ev = make_evidence()
|
||||
from dwg_review.inspect import InspectState
|
||||
|
||||
content, budget = _run_readonly_tool(
|
||||
ctx, "inspect_region", {"target": {}}, [6], evidence=ev, state=InspectState(ev)
|
||||
)
|
||||
assert "不在只读取证白名单内" not in content[0]["text"]
|
||||
assert "未执行" in content[0]["text"] # 空 target → 明确的用法错误
|
||||
assert budget == [6]
|
||||
|
||||
|
||||
def test_readonly_inspect_region_unavailable_when_disabled(monkeypatch):
|
||||
monkeypatch.setattr(config, "INSPECT_ENABLED", False)
|
||||
ctx = make_ctx(client=FakeMcpClient())
|
||||
ev = make_evidence()
|
||||
from dwg_review.inspect import InspectState
|
||||
|
||||
content, _ = _run_readonly_tool(
|
||||
ctx, "inspect_region", {}, [6], evidence=ev, state=InspectState(ev)
|
||||
)
|
||||
assert "不可用" in content[0]["text"]
|
||||
|
||||
|
||||
def test_readonly_inspect_region_unavailable_without_evidence():
|
||||
ctx = make_ctx(client=FakeMcpClient())
|
||||
content, _ = _run_readonly_tool(ctx, "inspect_region", {}, [6])
|
||||
assert "不可用" in content[0]["text"]
|
||||
@@ -1,5 +1,7 @@
|
||||
"""`dwg_review.tiling` 单元测试（纯函数，不依赖 CAD / LLM / Pillow）。"""
|
||||
|
||||
import pytest
|
||||
|
||||
from dwg_review import tiling as tl
|
||||
|
||||
|
||||
@@ -221,3 +223,60 @@ def test_plan_regions_clamps_to_frame():
|
||||
for w in wins:
|
||||
assert w["min_x"] >= 0 and w["max_x"] <= 1000
|
||||
assert w["min_y"] >= 0 and w["max_y"] <= 500
|
||||
|
||||
|
||||
# --- box / window 互转与可见范围 -------------------------------------------
|
||||
|
||||
def test_box_window_roundtrip():
|
||||
box = (1.0, 2.0, 3.0, 4.0)
|
||||
win = tl.box_to_window(box)
|
||||
assert win == {"min_x": 1.0, "min_y": 2.0, "max_x": 3.0, "max_y": 4.0}
|
||||
assert tl.window_to_box(win) == box
|
||||
|
||||
|
||||
def test_box_to_window_invalid():
|
||||
assert tl.box_to_window(None) is None
|
||||
assert tl.box_to_window([1, 2, 3]) is None
|
||||
|
||||
|
||||
def test_window_to_box_invalid():
|
||||
assert tl.window_to_box(None) is None
|
||||
assert tl.window_to_box({}) is None
|
||||
assert tl.window_to_box({"min_x": 1, "min_y": 2}) is None
|
||||
|
||||
|
||||
def test_expand_box_by_fraction():
|
||||
assert tl.expand_box((0, 0, 100, 50), 0.05) == pytest.approx((-5.0, -2.5, 105.0, 52.5))
|
||||
assert tl.expand_box((0, 0, 100, 50), 0.0) == (0.0, 0.0, 100.0, 50.0)
|
||||
assert tl.expand_box(None, 0.1) is None
|
||||
# 非法 frac 视为 0（不抛异常）
|
||||
assert tl.expand_box((0, 0, 10, 10), "x") == (0.0, 0.0, 10.0, 10.0)
|
||||
|
||||
|
||||
def test_visible_box_widens_to_viewport_aspect():
|
||||
# 1000x500 的窗口在 1280x280（约 4.57:1）视口下等比适配 → 横向露出更多，纵向刚好贴合
|
||||
vis = tl.visible_box((0, 0, 1000, 500), (1280, 280), 0.0)
|
||||
assert vis[0] < 0 and vis[2] > 1000
|
||||
assert vis[1] == pytest.approx(0.0)
|
||||
assert vis[3] == pytest.approx(500.0)
|
||||
assert (vis[0] + vis[2]) / 2 == pytest.approx(500.0) # 中心不变
|
||||
assert (vis[1] + vis[3]) / 2 == pytest.approx(250.0)
|
||||
|
||||
|
||||
def test_visible_box_tall_window_is_fit_by_height():
|
||||
# 100x1000（很窄很高）在 1280x280 视口下按**高**适配 → 纵向刚好贴合，横向大幅露出
|
||||
vis = tl.visible_box((0, 0, 100, 1000), (1280, 280), 0.0)
|
||||
assert vis[1] == pytest.approx(0.0) and vis[3] == pytest.approx(1000.0)
|
||||
assert vis[0] < 0 and vis[2] > 100
|
||||
assert (vis[0] + vis[2]) / 2 == pytest.approx(50.0) # 中心不变
|
||||
|
||||
|
||||
def test_visible_box_without_view_px_falls_back_to_margin():
|
||||
assert tl.visible_box((0, 0, 100, 50), None, 0.1) == pytest.approx(
|
||||
(-10.0, -5.0, 110.0, 55.0)
|
||||
)
|
||||
assert tl.visible_box((0, 0, 100, 50), (0, 0), 0.0) == (0.0, 0.0, 100.0, 50.0)
|
||||
|
||||
|
||||
def test_visible_box_invalid_box():
|
||||
assert tl.visible_box(None, (100, 100), 0.1) is None
|
||||
@@ -1,11 +1,13 @@
|
||||
"""验收：审查阶段是否也在「绘图区最大化」状态下进行，且结束后精确还原。
|
||||
|
||||
做法：在取证 `collect` 与审查 `review_frame` 入口各采一次窗口矩形，
|
||||
并在 `McpClient.call_tool` 上挂计数器，统计 LLM 子循环是否真的用了放大工具。
|
||||
并在 `McpClient.call_tool` 上挂计数器统计子循环的 zoom/截图次数，
|
||||
同时记录 LLM 实际调用的**取证工具**（含 agent 侧虚拟工具 `inspect_region`）。
|
||||
用法：先启动已加载插件的 CAD 并打开图纸，然后 `python verify_review_maximized.py`。
|
||||
"""
|
||||
|
||||
import ctypes
|
||||
import json
|
||||
import sys
|
||||
from ctypes import wintypes
|
||||
|
||||
@@ -60,10 +62,12 @@ def area(r):
|
||||
# --- 采样点 ---------------------------------------------------------------
|
||||
phase_rects = {"evidence": [], "review": []}
|
||||
zoom_calls = []
|
||||
llm_tools = [] # LLM 在审查子循环里实际调用的取证工具（名称 + 参数）
|
||||
|
||||
_orig_collect = ev.collect
|
||||
_orig_review = rv.review_frame
|
||||
_orig_call = McpClient.call_tool
|
||||
_orig_run_tool = rv._run_readonly_tool
|
||||
|
||||
|
||||
def collect(ctx, index, frame):
|
||||
@@ -82,9 +86,15 @@ def call_tool(self, name, arguments=None, **kw):
|
||||
return _orig_call(self, name, arguments, **kw)
|
||||
|
||||
|
||||
def run_tool(ctx, name, args, budget, **kw):
|
||||
llm_tools.append((name, args))
|
||||
return _orig_run_tool(ctx, name, args, budget, **kw)
|
||||
|
||||
|
||||
ev.collect = collect
|
||||
rv.review_frame = review
|
||||
McpClient.call_tool = call_tool
|
||||
rv._run_readonly_tool = run_tool
|
||||
|
||||
before = rect()
|
||||
print(f"起始窗口矩形 = {before} 面积={area(before)}")
|
||||
@@ -97,12 +107,18 @@ print("=== 结果 ===")
|
||||
print(f"取证阶段采样 = {phase_rects['evidence']}")
|
||||
print(f"审查阶段采样 = {phase_rects['review']}")
|
||||
print(f"结束窗口矩形 = {after} 面积={area(after)}")
|
||||
print(f"子循环取证调用 = {zoom_calls}")
|
||||
print(f"子循环 zoom/截图调用 = {len(zoom_calls)} 次")
|
||||
print(f"LLM 取证工具调用 = {len(llm_tools)} 次")
|
||||
for name, args in llm_tools:
|
||||
brief = json.dumps(args, ensure_ascii=False)
|
||||
print(f" - {name} {brief[:160]}")
|
||||
|
||||
inspect_n = sum(1 for name, _ in llm_tools if name == "inspect_region")
|
||||
ok_ev = all(area(r) > area(before) for r in phase_rects["evidence"]) if phase_rects["evidence"] else False
|
||||
ok_rv = all(area(r) > area(before) for r in phase_rects["review"]) if phase_rects["review"] else False
|
||||
print()
|
||||
print(f"[{'PASS' if ok_ev else 'FAIL'}] 取证阶段处于最大化（面积大于基线）")
|
||||
print(f"[{'PASS' if ok_rv else 'FAIL'}] 审查阶段处于最大化（面积大于基线）")
|
||||
print(f"[{'PASS' if after == before else 'FAIL'}] 结束精确还原窗口矩形")
|
||||
print(f"[{'PASS' if inspect_n else 'INFO'}] LLM 使用了 inspect_region（{inspect_n} 次）")
|
||||
sys.exit(0 if (ok_ev and ok_rv and after == before and rc == 0) else 1)
|
||||
Reference in new issue
Block a user