# GeoAI Workbench Current Context

Last updated: 2026-08-19

This file is the current project snapshot for new Codex tasks. Keep it concise and replace stale facts instead of appending a conversation diary.

## Project Identity

- Repository: `E:\AllWorkProject\geoai-workbench`
- Technical foundation: [opengeos/geoai](https://github.com/opengeos/geoai), PyPI package `geoai-py`, import name `geoai`.
- Purpose: build independently runnable GeoAI capability Demos for drone and remote-sensing workflows, then evaluate integration with existing products.
- Product repositories in scope for later integration:
  - `E:\AllWorkProject\drone_project\ztzf-drone-web\applications\command-center-dashboard`
  - `E:\AllWorkProject\drone_project\ztzf-drone-web\applications\drone-web-manage`
- Capability classification: A = direct `geoai-py`; B = GeoAI plus ecosystem libraries; C = product/service capability consuming GeoAI results.

## Collaboration and Constraints

- The user is a beginner and can invest about 7-8 hours per day in running Demos and giving visual feedback.
- Codex is expected to handle environment setup, code, model selection, debugging, verification, and iteration.
- The machine runs Windows and PowerShell with 64 GB RAM and an AMD RX 590 GME 8 GB GPU. There is no NVIDIA CUDA, so current Demos use CPU inference.
- Use Python 3.12 for capability environments. The system also has Python 3.13, which must not replace or contaminate project environments.
- `geoai-py` is MIT-licensed, but every dependency, dataset, and model weight needs a separate commercial-license check.
- Current JPEG samples have no usable georeferencing. Their detections use pixel coordinates; accurate GeoJSON requires a georeferenced GeoTIFF.

## Environment Snapshot

Dedicated object-detection environment:

`E:\AllWorkProject\geoai-workbench\.venvs\01-object-detection`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| geoai-py | 0.42.0 |
| torch | 2.13.0+cpu |
| torchvision | 0.28.0+cpu |
| rasterio | 1.5.1 |
| geopandas | 1.1.4 |
| ultralytics | 8.4.118 |

`pip check` passes in this environment.

Dedicated anomaly-detection environment:

`D:\lgh\code\geoai-workbench\.venvs\09-anomaly-detection`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| geoai-py | 0.42.0 |
| scikit-learn | 1.9.0 |
| SciPy | 1.18.0 |
| rasterio | 1.5.1 |
| geopandas | 1.1.4 |
| OpenCV | 4.14.0 |

The environment is CPU-only, contains no model weights, and `pip check` passes.

Dedicated semantic-mapping environment:

`E:\AllWorkProject\geoai-workbench\.venvs\02-semantic-mapping`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| geoai-py | 0.42.0 |
| rasterio | 1.5.1 |
| geopandas | 1.1.4 |
| OpenCV | 5.0.0 |

The environment reuses the compatible package set from the verified object-detection environment after the isolated setup exceeded the dependency-download timeout. It has its own Python 3.12 interpreter and `pip check` passes.

Dedicated spatial-measurement environment:

`E:\AllWorkProject\geoai-workbench\.venvs\04-spatial-measurement`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| geoai-py | 0.42.0 |
| rasterio | 1.5.1 |
| geopandas | 1.1.4 |
| OpenCV | 4.14.0 |

The environment is CPU-only and `pip check` passes.

Dedicated trajectory-analysis environment:

`E:\AllWorkProject\geoai-workbench\.venvs\15-trajectory-analysis`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| pandas | 2.3.3 |
| geopandas | 1.1.4 |
| shapely | 2.1.2 |
| pyproj | 3.7.2 |
| scikit-learn | 1.9.0 |
| matplotlib | 3.11.1 |
| openpyxl | 3.1.5 |

`geoai-py` is intentionally not installed in this environment because version 0.42.0 has no trajectory, tracking, or behavior-recognition API. The capability consumes timestamped track results that may originate from an upstream GeoAI detection/export workflow. `pip check` passes.

Dedicated change-detection environment:

`E:\AllWorkProject\geoai-workbench\.venvs\00-change-detection`

| Component | Version |
| --- | --- |
| Python | 3.12.10 |
| geoai-py | 0.42.0 |
| torchange | 0.0.4 |
| ever-beta | 0.6.1 |
| torch | 2.13.0+cpu |
| rasterio | 1.5.1 |
| geopandas | 1.1.4 |
| OpenCV | 4.14.0 |

The environment reuses the compatible 3.12 package set from the verified semantic-mapping
environment through a local `.pth` file and adds ChangeStar dependencies. `pip check` passes.

## Local Experiment Console

- Location: `apps/workbench-console/`
- Frontend: Vue 3 + Vite + Ant Design Vue + Pinia. Trajectory maps use Cesium `1.126.0`, public ArcGIS World Imagery, and optional TianDiTu image/label layers configured by `apps/workbench-console/.env.local` `VITE_TIANDITU_TOKEN`.
- Build: `Set-Location .\apps\workbench-console; npm install; npm run build`
- Start command: `py -3.12 .\scripts\serve_workbench_console.py`
- URL: `http://127.0.0.1:6173` (built console); Vite development is `http://127.0.0.1:6174/apps/workbench-console/`. Only `6xxx` ports are accepted.
- Scope: independent local experiment workbench for this repository. It has no code, account, or product API link to the two drone-product repositories. The map client directly requests public ArcGIS tiles and, only when configured, TianDiTu tiles.
- Current workflow: `00-change-detection`, `01-object-detection`, `02-semantic-mapping`, `04-spatial-measurement`, `09-anomaly-detection`, and `15-trajectory-analysis` support new local runs, searchable case libraries, and visual result workspaces. Anomaly detection accepts separate normal-reference and inspection groups through bounded binary uploads, runs asynchronously, and shows references, input/output comparison, heatmaps, candidates and downloads. Change detection shows two-date imagery, a ChangeStar raster overlay, pixel-coordinate GeoJSON, per-feature probability stats and downloads. Spatial measurement shows raster/vector results, per-object metrics and downloads. The server creates a new run ID for every submission, only accepts allowlisted input types and enabled tasks, and calls fixed capability scripts in fixed virtual environments.
- File exposure: the static handler permits only console assets, `shared/outputs`, and source images required for result comparison. Upload APIs write new raw and processed run directories but do not expose the rest of the repository.
- Styling: `apps/workbench-console/src/styles.css` is organized into design variables, application shell, shared workbench components, capability workspaces, maps, and responsive rules. New capability pages must use a scoped workspace class and shared spacing variables instead of global override patches.

## Capability Status

| Capability | Status | Current conclusion |
| --- | --- | --- |
| `00-change-detection` | Runnable CPU Demo verified | A capability: `geoai.ChangeStarDetection` plus `geoai.masks_to_vector`; ORB registration and JPG-to-GeoTIFF conversion are input preparation. Auto/image mode handles ordinary imagery in pixel coordinates; GeoTIFF mode preserves spatial reference for inputs with matching CRS. When two GeoTIFF grids differ, the second image is reprojected to the first grid only in the processing copy; original files are unchanged. Real pair `07-16.jpg` -> `07-19.jpg` completed in about 41.6 seconds at 1024 px, with 755/769 registration inliers, 98.52% valid area, 674 changed pixels and one GeoJSON polygon. CLI/API/console runs accept per-run thresholds and processing modes/resolution options, all recorded in `run_metadata.json`. |
| `01-object-detection` | In progress; runnable people and vehicle experiments | Tiled YOLO helps people; GeoAI NWPU substantially improves top-down vehicles; tree detection is not implemented. |
| `02-semantic-mapping` | Runnable CPU baseline verified | B capability: deterministic RGB/HSV segmentation plus `geoai.masks_to_vector`; raster and vector outputs are inspectable, but this is not a trained GeoAI model. |
| `04-spatial-measurement` | Runnable CPU Demo verified | B capability: `geoai.masks_to_vector` plus ecosystem measurement; counts labelled raster objects and reports area/perimeter in projected or explicitly non-metric pixel/coordinate units. |
| `09-anomaly-detection` | Runnable CPU CLI Demo verified | B capability: interpretable image features plus robust rules and Isolation Forest score visual outliers; `geoai-py` supplies `masks_to_vector`. Results are review candidates, not business-event conclusions. |
| `15-trajectory-analysis` | Runnable CPU Demo verified | C capability consuming timestamped tracks: spatial metrics, DBSCAN and explicit behavior rules; not a built-in `geoai-py` function. |
| `03`, `05` through `08`, `10` through `14`, `16` through `18` | Directory and initial README only | No verified local Demo yet. Start each one through `$geoai-capability-builder`. |

## Semantic Mapping Snapshot

## Anomaly Detection Snapshot

- Script: `capabilities/09-anomaly-detection/run_anomaly_detection.py`; reproducible input generator: `generate_validation_inputs.py` in the same directory.
- Boundary: B. Seventeen interpretable RGB/HSV/texture features feed a robust MAD rule baseline and a 200-tree Isolation Forest; fixed-view inputs additionally use class-agnostic corresponding-location Lab appearance and grayscale structure differences after image-wide normalization. `geoai-py 0.42.0` only supplies `masks_to_vector`. There are no model weights or colour/object-specific rules.
- Contract: image or directory input; minimum 50 reference windows; outputs are two heatmaps, comparison overlay, encoded mask GeoTIFF, candidate GeoJSON, tile CSV and run metadata. Ordinary images use pixel coordinates; valid GeoTIFF CRS/transform is preserved.
- Validation: controlled color, blur and dark anomalies were injected into a copy of `07-16.jpg`; `07-19.jpg` was used unchanged as a difficult real sample. Source-copy SHA256 values match the originals.
- Measured result: 588 reference windows and three inputs completed in 41.051 seconds CPU. The rule method hit 3/3 controlled regions with precision 0.583, recall 0.672 and IoU 0.454; Isolation Forest hit 2/3 with precision 0.831, recall 0.500 and IoU 0.454. Normal-image coverage was 0.33% and 0.00%, respectively. Every image completed within 15 seconds.
- Visual limitation: the normal image has a small rule false positive at a vegetation/rock transition. Difficult-image candidates occur around concrete structure, a dark opening and a top dark region, consistent with viewpoint/material/lighting shift; no real anomaly or business accuracy is claimed.
- Console: `GET/POST /api/anomaly-detection/runs`, binary `PUT /api/anomaly-detection/uploads/<upload-id>` and asynchronous `GET /api/anomaly-detection/jobs/<job-id>`. A real one-reference/one-input upload produced run `anomaly-20260818-102845-61f55b` with six review candidates. Desktop and 390 px layouts were visually checked; tables use internal horizontal scrolling and the collapsed menu trigger remains inside the viewport.
- Synthetic fixed-view validation: five ImageGen normal references cover overcast, sunny, after-rain, morning-haze and late-afternoon conditions; one test image adds a localized branch/leaf/mud/rock blockage. The original pooled-feature run returned zero candidates and is retained as `ImageGen 五参考图堵塞复核（漏检案例）`.
- Fixed-view optimization: `--spatial-mode auto` now selects same-position robust rules when 3+ references and all target images share dimensions, after normalizing each image's global feature profile; otherwise it falls back to the original global rule. At the unchanged 0.995 quantile, the synthetic blockage produced one correctly located rule candidate with 9.375% coverage. Five leave-one-normal-out checks produced 7.292%, 0%, 0%, 0%, and 0% coverage. The successful output is retained as `ImageGen 五参考图堵塞复核（同位置优化）`; Isolation Forest still missed this sample.
- Generic local-change enhancement: same-position mode now also calibrates nearest-normal-reference Lab appearance and grayscale intensity/edge/contrast distances. A previously missed small added pattern became one 9,467-pixel rule candidate with 100% coverage of its 6,710 validation pixels and 0.602% full-image coverage; it is retained as `通用局部异常增强（小目标漏检修复）`. The blockage regression remains one candidate at 9.594% rule coverage. Five leave-one-normal-out combined coverages are 7.292%, 0%, 0%, 0.460%, and 1.209%. No sample colour, object class, or fixed coordinate is encoded.
- Next decision: add image registration and/or a channel ROI or specialist blockage detector before accepting camera motion or claiming blockage semantics. Obtain real normal/labelled anomaly pairs; the synthetic result is only a pipeline check.

## Change Detection Snapshot

- Script: `capabilities/00-change-detection/run_change_detection.py`.
- Boundary: A. The core inference and vector conversion call `geoai.ChangeStarDetection` and `geoai.masks_to_vector`; OpenCV ORB homography is only input registration for ordinary JPGs.
- Inputs: user-provided `baseData/07-16.jpg` and `baseData/07-19.jpg`, both 3664 x 2748 RGB with no usable CRS. Staged copies are under `shared/data/raw/00-change-detection/validation-20260817/`.
- Outputs: `shared/outputs/00-change-detection/validation-real-20260817-v4/`, `validation-self-20260817-v3/` and `resolution-2048-urban-validation/` contain registered GeoTIFFs, probability/mask rasters, overlays, GeoJSON, feature stats and `run_metadata.json`.
- Measured result: real pair at 1024 px / tile 512 / overlap 64 took 41.556 seconds on CPU, found 755 inliers from 769 ORB matches, 98.52% valid area, 674 changed pixels (0.0857%) and one polygon. Self-comparison at 512 px took 25.902 seconds and found zero changed pixels and zero polygons. On the urban parking sample, increasing from 1024 to 2048 produced 16 polygons instead of one and covered more vehicle/parking changes, but took about 84 seconds CPU and still had misses/false positives.
- GeoTIFF grid-alignment validation: uploaded `1.tif` (6217 x 5525) and `2.tif` (6218 x 5497) share `EPSG:32650` but have slightly different origins and pixel sizes. The second image was reprojected bilinearly to the first grid in the processing copy, then completed at 1024 px in 80.882 seconds with georeferenced outputs. SHA256 checks confirmed the source staging files and their raw run copies are byte-identical.
- Honest limitation: ChangeStar weights are trained for Changen2/S1 building change and the close-up rock/concrete scene is out of distribution. No manual truth exists, so no precision, recall, IoU, change type, or engineering alert is claimed. Pixel GeoJSON has no CRS; the downloaded model is CC BY-NC-SA 4.0 and not cleared for commercial use.
- Console: `GET/POST /api/change-detection/runs` on local `6xxx` ports. Real uploads at `6176` verified thresholds and selectable processing resolution; auto/image and GeoTIFF mode handling is covered by unit tests. Separate raw before/after directories and a separate processed directory are used; unsafe basename sanitization plus invalid/oversized request rejection were verified.
- Next decision: obtain georeferenced same-GSD orthophotos and manually labelled change masks for the intended business class, then compare registration error, false positives, misses and IoU before any batch or product use.

- Script: `capabilities/02-semantic-mapping/run_semantic_segmentation.py`.
- Boundary: B. `geoai-py 0.42.0` supplies class-wise `masks_to_vector`; deterministic RGB/HSV rules supply the first offline segmentation mask because no general pretrained weight is bundled.
- Classes: `other`, `vegetation`, `water`, and `impervious`. Outputs are mask PNG, overlay PNG, mask GeoTIFF, class GeoJSON, and `run_metadata.json`.
- Validation input: two existing real 3840 x 2160 drone JPEGs copied to `shared/data/raw/02-semantic-mapping/validation-20260817/`. Both are ordinary non-georeferenced images, so their vector output is in pixel coordinates.
- Measured result: two images completed on CPU in 15.738 seconds. The vegetation-heavy image produced 1,707,028 vegetation pixels; the difficult road image produced 8,167,327 impervious pixels. Both used `geoai.masks_to_vector`, and raster/vector artifacts plus both overlays were inspected.
- Honest limitation: blue roofs and other blue objects can be false water, and lighting/material similarity can change all rule classes. There is no manual ground truth, so no accuracy, IoU, precision, or recall is claimed.
- Task catalog: `capabilities/02-semantic-mapping/configs/task-catalog.json`. Only `color_baseline` is selectable; project classes such as drainage blockage are recorded but disabled until data, labels, weights, licenses, and acceptance thresholds are available.
- Console: `GET /api/semantic-mapping/tasks` and `GET/POST /api/semantic-mapping/runs`, at most six allowlisted images per run. The server validates `taskId` before creating a run directory. Raw, processed, and output paths are separate. Run `semantic-20260817-063122-5f20ef` verified a real upload and persisted `color_baseline` / `通用颜色规则基线` in structured metadata.
- Next decision: choose one business class, obtain normal and difficult georeferenced GeoTIFFs plus manual masks, then license and evaluate a lightweight pretrained model through `geoai.semantic_segmentation` before any batch or product use.

## 项目算法方案层

- 新增 `solutions/`：它是 C 类项目算法编排层，引用 `capabilities/` 中已验证的原子能力，不复制模型实现。
- `solutions/mining-inspection/requirements-map.json` 将需求文档中的 8 条记录归一化为 7 个方案域；“排洪沟巡查”和“排洪沟堵塞识别”合并，避免重复模型、标注和规则。
- 七个方案域为：人员与个体防护、爆破警戒热成像、排洪沟与箱涵堵塞、边坡变化、竖井巡查、索道巡查、尾矿库异常筛查。优先级、就绪度和 A/B/C 实现边界分别记录，不能相互替代。
- 当前两个专项目录 `01-person-safety` 和 `02-drainage-blockage` 仍为 `scaffold`，没有运行脚本、模型下载或产品告警接入；其余方案域保留在需求映射和方案索引中。
- 项目数据流必须区分模型观察、空间/测量分析、事件候选和平台动作。安全帽以人员与安全帽检测及实例关联为主，属性分类只作为可选辅助。
- 爆破区热成像人员识别是受数据条件限制的 P1，不等同于 RGB 人员检测的成熟度；当前颜色规则基线不是排洪沟模型，RGB 浑浊或颜色异常也不能证明矿浆渗漏。
- `shared/configs/solution-result.schema.json` 规定项目算法输出必须区分模型观察结果与供人工复核的事件候选。

## geoai-py 能力目录

- `capabilities/GEOAI_PY_DIRECT_CAPABILITIES.md` 已按 `geoai-py 0.42.0` 实际导出接口整理能力边界。
- 算法人员交付模型时，必须同时提供权重/许可证、输入波段与 CRS、类别或掩膜规范、推理参数、输出样例和评估结果。
- 当前对象检测环境已具备 Transformers、timm、Lightning 和 OpenCV；`torchange`、`omniwatermask`、`rfdetr`、`terratorch`、`segment-geospatial`、`detectron2` 和 `onnxruntime` 尚未安装。
- 目标检测、分割、分类、变化检测、水体/树冠/深度/嵌入等是算法产出可被 GeoAI 工作流承接的方向；电子围栏、告警、无人机控制、工单和调度仍属于项目产品能力。

## Spatial Measurement Snapshot

- Script: `capabilities/04-spatial-measurement/run_spatial_measurement.py`.
- Boundary: B. `geoai-py 0.42.0` supplies `masks_to_vector`; Rasterio, GeoPandas and Shapely supply CRS handling and count/area/perimeter/centroid/bounds measurements. It consumes labelled rasters and does not infer classes from RGB imagery.
- Inputs: single-band PNG/GeoTIFF labels (`0` background, positive integers as classes). Outputs are object GeoTIFF, PNG preview, measured GeoJSON, CSV and `run_metadata.json`.
- Validation inputs: normal and difficult 3840 x 2160 label GeoTIFFs derived from the existing real drone-image semantic-mapping validation set, stored under `shared/data/raw/04-spatial-measurement/validation-20260817/`.
- Measured result: the normal raster produced 1,210 objects in 25.190 seconds; the difficult road-dominant raster produced 89 objects in 9.474 seconds. Both call `geoai.masks_to_vector`, detect incomplete complex contours against a Rasterio reference, and apply a recorded completeness repair; raster/vector/table artifacts and both previews were inspected.
- Unit boundary: both real validation rasters have no CRS, so results are explicitly `pixel^2` and `pixel`. A projected-CRS unit test confirms the map-unit branch. No metre or accuracy claim is made for the real samples.
- Console: `GET/POST /api/spatial-measurement/runs`, at most four allowlisted PNG/TIF/TIFF label rasters. A valid upload, unsafe-name sanitization, invalid extension, oversized request, and script failure were verified at local port 6175; the default remains 6173. Browser screenshot verification was unavailable because no browser backend was exposed.
- Next decision: obtain a georeferenced projected label GeoTIFF plus surveyed or manually measured truth, then evaluate count false positives/misses and area/perimeter error before product or batch use.

## Object Detection Snapshot

Target classes: people, vehicles, and trees.

Inputs:

- Directory: `shared/data/raw/01-object-detection`
- 20 files total: 19 ordinary JPEG images and one internal DJI MPO with a `.jpeg` extension.
- Most images are 3840 x 2160. The MPO is skipped safely.

Environment and scripts:

- General tiled detector: `capabilities/01-object-detection/run_detection.py`
- GeoAI aerial vehicle detector: `capabilities/01-object-detection/run_geoai_vehicle_detection.py`
- General model: Ultralytics `yolo11n.pt`, CPU, 1024-pixel tiles, 20% overlap, confidence 0.20.
- Aerial vehicle model: `giswqs/nwpu-vhr10-maskrcnn:best_model.pth`, called through `geoai-py`, 512-pixel windows, 128-pixel overlap, confidence 0.30.

Measured results:

- Whole-image YOLO baseline across 19 JPEGs: 13 detections, including 3 people and 10 cars; this misses many small targets.
- `DJI_20260713102047_0001_V_19.jpeg`: tiled YOLO found five person candidates, including duplicates and a false positive; the visible red vehicle was still missed by both YOLO and NWPU.
- `DJI_20260810092727_0001_V_10.jpeg`: general YOLO found only 2-3 cars. GeoAI NWPU produced 33 raw vehicle detections and 32 after containment deduplication, with about 90 seconds CPU inference.
- NWPU is much better for top-down aerial vehicles but still misses vehicles and can produce partial-box duplicates or other aerial-class false positives.

Current technical decisions:

- Do not batch-run the final pipeline across all images yet.
- Use separate model branches: tiled YOLO for people, aerial-specific detection for vehicles, and a future tree-crown detection or segmentation model for trees.
- Do not claim that the current COCO or NWPU model detects trees.
- Establish a fixed manually annotated validation set before tuning or training.
- Product use of Ultralytics and NWPU weights remains blocked on license review and accuracy evaluation.

## Trajectory Analysis Snapshot

Inputs and artifacts:

- Generator: `capabilities/15-trajectory-analysis/generate_demo_inputs.py`
- Analyzer: `capabilities/15-trajectory-analysis/run_trajectory_analysis.py`
- Input contract: WGS84 timestamped CSV observations plus GeoJSON reference routes and zones, connected by a `*.case.json` manifest.
- Generated representative inputs: `shared/data/raw/15-trajectory-analysis`
- Inspectable outputs: `shared/outputs/15-trajectory-analysis`
- User-provided real source data is staged, unchanged, in `shared/data/raw/15-trajectory-analysis/tian-dun-demo-20260814/`: no-fly GeoJSON, flyable-area Gzip binary source, DJI WPMZ KMZ, and an actual-flight XLSX log. `prepare_real_flight.py` produces a single WGS84 case from all four sources at `shared/data/processed/15-trajectory-analysis/tian-dun-flight-19578/`. The Gzip was verified as 46,518 closed `int32 / 1e7` coordinate rings and yields two local flyable polygons for display only; the zone rule still recognizes only `zone_type=restricted`.

Measured validation:

- The normal synthetic case has 2 tracks and 0 events.
- The difficult synthetic case has 3 tracks, drops 1 duplicate timestamp, and finds exactly one stop, one route deviation, one restricted-zone event, and one gathering event.
- Directory processing handled 5 tracks and 78 cleaned observations in about 2.8 seconds on CPU.
- Missing required CSV columns fail cleanly with exit code 2. Both 1400 x 980 PNG results were visually inspected.
- Real Tian Dun flight 19578: 233 observations became one WGS84 drone track; the 10-waypoint WPMZ route, one locally selected no-fly source area and two local flyable polygons were prepared and analyzed in 0.362 seconds. It produced two route-deviation runs (77 and 175 seconds) and one 464-second source-zone intersection, with no stop or gathering event. The static result map was visually inspected.

Current decisions and limitations:

- Keep the boundary explicit: the current Demo is C because it does not call `geoai-py`; using geospatial ecosystem libraries alone does not make it B. `geoai-py` is only a possible upstream source.
- This is rule-based behavior detection, not learned video action recognition, and it assumes track IDs already exist.
- Synthetic data verifies logic only. Do not claim real-world accuracy or batch-run operational data yet.
- The Tian Dun zone intersection is a spatial result against an unverified third-party source, not a violation conclusion. The reference route does not encode takeoff, return-to-home or landing, so its two deviation runs require flight-phase-aware validation.
- The flyable-area binary structure has been decoded, but its authority, publication time and operational semantics remain unverified; it is visual-only until that validation is complete.
- Next decision: obtain manual truth labels for this real flight and another difficult real flight, then measure false positives, misses, GPS sensitivity and phase-aware thresholds.

## Starting a New Capability

Open a new Codex task with the repository as the working directory and use:

```text
使用 $geoai-capability-builder，创建下一个能力：<能力名称>。
先读取 AGENTS.md 和 PROJECT_CONTEXT.md，再准备环境、Demo 输入输出和验收方式。
```

When the capability needs direct user uploads, interactive execution, maps, image comparison, or an inspectable result workspace, add:

```text
同时使用 $geoai-workbench-console 接入独立本地实验控制台。
```

`$geoai-capability-builder` owns the capability workflow. `$geoai-workbench-console` owns the independent local console workflow, including run discovery, safe uploads, output presentation, UI layout, and visual verification.

After completing verified work, update this snapshot and the capability README. Change `AGENTS.md` only when a long-lived rule changes.
## 2026-08-14 工作台更新

- 本地控制台已从只读结果页升级为独立实验工作台：能力页包含新建运行入口、服务端自动发现的案例库和结果工作区，仍不连接任何无人机产品。
- `scripts/serve_workbench_console.py` 仅监听 `127.0.0.1` 等回环地址，支持轨迹与目标检测的受限 JSON/Base64 上传 API；每次生成唯一运行目录，不覆盖原始数据或既有输出。
- 轨迹上传入口接受 XLSX、KMZ、禁飞区 GeoJSON 和可选适飞区 Gzip，预处理器已支持缺少适飞区时继续生成可运行 case。
- 目标检测页面默认并列展示原图与标注图；案例数据从输出目录扫描，历史多模型检测结果可以在同一案例库中选择。
- 异常检测案例库支持把历史案例的原始参考图、待检测图和运行参数回填到新运行表单，供用户调整后创建新的检测运行；回填操作不自动运行，也不覆盖历史案例。
- 轨迹汇总的 `normal`、`stop`、`route_deviation`、`restricted_zone`、`gathering` 标签在 UI 显示中文，原始英文值保留在 CSV/GeoJSON。