The example executes a logging-only gesture planner on supplied numbers. Ratios 0.25,0.08,0.07,0.14,0.22,0.09 produce exactly two Click plans at observations 2 and 6. Pointer (0.25,0.75) in extents 101 by 81 maps to (25,60). The page calculates those plans in JavaScript; it neither runs Python nor recognizes a hand or moves the desktop cursor.
Define the coordinate contract
For a Python integration, keep normalized image points separate from world landmarks measured in meters. MediaPipe's documentation distinguishes those outputs. Define the original pinch ratio as contact-point distance divided by palm-reference distance using the same image-pixel units. Restore normalized x using image width and y using image height before measuring a nonsquare image; otherwise the distance ratio can be distorted. The thresholds here are authored exercise rules, not model defaults or a validated hand-recognition policy.
Validate finite pointer coordinates and positive integer screen extents before mapping. Use floor(clamp(x,0,1)*(W-1)) and the analogous expression for y. With W101/H81, maximum coordinates are 100/80 rather than 101/81. Values outside the normalized range clamp to the endpoint after validation; missing values are not yesterday's pointer. Confirm image orientation, one-hand scope, model artifact and landmark-role mapping before connecting actual model output to the planner. A synthetic named-point test does not establish those integration facts.
Arm, close once, and latch
Begin disarmed. A valid ratio above 0.2 arms the planner; a value below 0.1 clicks only if armed and then latches closed. The inclusive 0.1-to-0.2 band holds the current state with no click. In the six-observation sequence, the first open arms, the second closes and clicks, the third remains closed, and 0.14 does not count as release. The fifth open rearms and the sixth close clicks again. Record observation IDs alongside actions so a repeated-close bug is visible in the log.
An invalid or missing observation disarms and clears active coordinates. Under the single-hand contract, multiple hands and zero palm-reference distance also require that boundary in a native adapter. After interruption, a closed or hold-band observation cannot resume movement or clicks; a valid open must rearm first. Keep invalid separate from the valid hold band, which preserves the current latch. The example accepts a declared invalid token to make that interruption testable without a camera or fabricated landmark stream.
Keep the adapter outside the pure test
The default output is a plan log. Native Python tests should preserve ratio inputs, pointer mapping, latch transition, proposed move/click, source revision and interpreter. Before connecting a desktop adapter, check its permissions and dependencies, then test the actions it actually performs. Do not call a real click function just to demonstrate that the planner produced Click; that couples an algorithm check to an unrelated application. MediaPipe live-stream processing can drop submitted frames while busy, so an integration must track accepted observations instead of assuming every camera frame produced a result.
Change thresholds, orientation, screen bounds or landmark-role mapping only with a new session contract. Require fresh validation and open rearming after such a change. End by explaining why each click was planned and why intervening observations produced none. Keep camera recognition, operating-system behavior and accessibility claims outside this synthetic result. A correct two-click log is evidence about the authored planner, not about real-world gesture reliability or universal desktop safety.
Make your own working example
Create a logging-only virtual-mouse planner using editable comma-separated pinch ratios and normalized pointer x/y with integer extents W/H. Defaults ratios .25,.08,.07,.14,.22,.09; x.25 y.75 W101 H81. Map floor(clamp(x,0,1)*(W-1)), similarly y, giving 25,60.
Start disarmed. r>.2 arms. r<.1 while armed emits one Click plan then closes latch. Inclusive .1 through .2 holds without clicking. A new open rearms. Default Click observations are 2 and 6, total 2. Accept token invalid to represent an interruption; it disarms, clears current action and suppresses movement/clicks until a valid open. Closed or hold values alone cannot rearm.
Provide Run planner, Load interruption sequence, Reset controls. Interruption sequence .25,invalid,.08,.22,.09 yields one Click at 5. Reject empty fields, nonfinite/negative ratios and invalid screen extents. Edits clear previous log. Show pointer mapping and each observation/state/action, clearly synthetic JavaScript plans rather than Python/model/camera/OS execution. Include native integration reminders about one hand, unmirrored image, same pixel units for ratio and separately confirmed landmark roles.
Use one self-contained HTML file with no remote dependency. Keep the initial sample, output, and labeled controls near the top. Use at least 16px text, strong contrast, visible focus, and a stacked layout at 360px without horizontal overflow. Use explicit type=button click handlers and explicit DOM selectors; do not depend on native form submission. Clear stale feedback when task inputs change. Render entered text safely. Keep practice data local to the current page; do not imply saving, sending, or production integration. Return only the finished HTML.
Review the transition rule
Keep native Python and eventual adapter tests in their own workflow. For the HTML action explanation, NOS lets you revise eligible static text visually without asking AI to regenerate it. Clarify the release requirement next to the log before anyone connects a real desktop adapter.