TangiEye | 实界之眼
Give your agent eyes on the physical world. 让 Agent 看见物理世界。
Requirements and routing
Use a host that can execute local commands and actually read local images.
Use the user's configured language. This Skill supplies images, not a vision model.
Windows is the initial target; read VALIDATION.md before claiming a tested host,
platform, camera, or protocol. Python 3.11+ and requirements.txt are required.
Resolve all Skill scripts relative to this SKILL.md directory. The examples below
assume that directory is the current working directory; elsewhere use absolute paths.
Use a project virtual environment, not a global package installation. Run:
python scripts/tangi_eye.py doctor
doctor does not open cameras unless given --source. A source probe captures and
discards a frame; do not describe it as passive device discovery. Exception: ADB
doctor --source adb checks authorization, lock state, camera package and folder
access without launching the camera or sending a shutter.
Use the image or camera source specified by the user or already established in
the session. If none is known, ask which camera/image to inspect. Do not enumerate
the network or silently switch to another camera. Keep an established source for
the task; do not repeatedly ask permission for already authorized captures.
Observe, then inspect
- Acquire one observation. Existing photos/screenshots use
ingest; cameras
usecapture. Specify a workspace output directory, normallywork/tangi-eye.
Camera output records the delivered dimensions, which may differ from a request.
powershell
python scripts/tangi_eye.py capture --source usb --device 0 --width 1920 --height 1080 --output work/tangi-eye
python scripts/tangi_eye.py ingest --input "board.jpg" --output work/tangi-eye
For HTTP/MJPEG/RTSP setup, credentials and USB backend options, read
references/cli.md. Do not put secret URLs in commands or chat;
use an existing environment variable with --url-env.
For an Android phone, read references/android-adb.md.
For a fresh observation, use capture --source adb --adb-serial SERIAL
--camera-package CAMERA_PACKAGE --timeout 30. The user unlocks and aims the
phone; the script opens the selected camera, sends one shutter, verifies one
newly saved photo and its transfer hash, then backgrounds the camera if still
in focus. Use the established device/package; only the documented Huawei setup
has real acceptance. Keep the camera app's current lens and settings.
No new photo or multiple new photos means failure, never an old-photo fallback.
Do not blindly retry after a timeout: a photo may already have been saved.
For an explicitly selected existing photo, use ingest --remote-path instead;
its capture time is unknown. screencap is a screenshot, not a camera photo.
-
Read the returned
overview_pathusing the host's image-reading tool.
File existence or JSON metadata is not visual evidence. Establish orientation,
the target component, and the context needed to interpret it. If the host cannot
read images, explain the limitation and request an image-capable host; do not guess. -
Crop a relevant region from that observation. Use normalized coordinates
[x0, y0, x1, y1]in the entire orientation-corrected image, not coordinates
in a letterboxed viewer, a previous crop, or the unrotated input. Divide target
positions by the displayed image's actual width and height, excluding UI margins.
Leave enough context to retain pin numbers, polarity, units, or component references.
powershell
python scripts/tangi_eye.py crop --manifest "ABSOLUTE_PATH_FROM_CAPTURE_OR_INGEST" --bbox 0.25 0.2 0.65 0.6
Read crop_path with the host image tool. Further regions must use the same
observation manifest and original coordinate space. Never crop the overview.
- Report only what the images support. Separate visible facts, possible
explanations, missing evidence, and the next useful action. Link to the overview
and relevant crop. For task-specific interpretation, read
references/inspection.md.
Evidence quality
- Original file bytes remain unchanged. Canonical PNG corrects EXIF orientation;
CMYK input is converted for PNG. Native crop pixels equal the canonical region.
JPEG/camera transport may already be lossy. “Lossless crop” does not mean optical
magnification or recovery of absent detail. - No generated repair, super-resolution, sharpening, or inferred character completion
may serve as diagnostic evidence. Do not infer missing digits from a likely part name. - When focus, glare, angle, or pixel count prevents a conclusion, request a specific
better view (move closer, refocus, light from the side, include the meter mode).
Repeating crops of the same insufficient pixels is not progress. - After rewiring or another state change, acquire a new observation. Import time is
not capture time. Acquisition timestamps indicate local receipt, not verified
sensor exposure time or server freshness. - A still LED image cannot establish a blink pattern. Visible discoloration cannot
establish an electrical failure, and a photo cannot prove continuity or voltage. - Source failure, timeout or an unreadable image must remain a failure. Do not
substitute an earlier observation without explicitly identifying it as old evidence.
Boundaries
The scripts perform local image processing and contact only the supplied source.
They do not upload images to a recognition API; the host agent's image processing
still follows that host's settings. No automatic publication or telemetry is included.
This Skill does not manipulate wiring, energize hardware, operate desktop mouse/keyboard,
record ongoing video, or recognize arbitrary proprietary camera APIs. Software UI
inspection accepts provided screenshots. Android control is limited to the selected
camera launch, one shutter, and conditional Back cleanup; no unlocking, permission
bypass, app installation or general UI automation. High-voltage work, changing connections,
and electrical measurement procedures require their own appropriate context.
