# Development history The first two phases were not explicitly numbered on the display; v1/v2 below are retrospective labels. Versions include experiments and failed diagnostics, not 24 successful product releases. This public repository preserves available source snapshots and does not fabricate historical Git commits. | Version | Problem and change | Evidence / remaining limitation | |---|---|---| | v1 phase | Established Arduino/M5Stack environment, backed up UIFlow locally, read native gestures and displayed them | Hardware/compile/upload path established; no media control | | v2 phase | Corrected the physical mounting's 90-degree direction mapping | Physical directions remapped; installation-dependent | | v3 | Removed UP/DOWN, enabled Gaming mode, local display updates and an area-growth poke fallback | Natural wrist-turn/two-finger movements still missed | | v4 | Added timestamped center/area/presence/flag traces and nonblocking serial diagnostics | Could distinguish missing input from rejected recognition | | v5 | Added trajectory-based lateral tracker, one event per sweep and ACTION counter | Recorded left/right tests; four actual right motions yielded four RIGHT events | | v6 | Replaced sustained area growth with short area-pulse/wrist-motion approximation | Replay improved; actual turns exceeded displacement bounds | | v7 | Relaxed wrist displacement and retained the action latch across short visibility gaps | Complete motion detected once, but fist-only control falsely triggered | | v8 | Required area pulse plus wrist trajectory, removed area-only branch | Small controlled set passed; later natural motions still missed | | v9 | Extended 180 ms window to 400 ms and improved outlier handling | One three-poke set detected all three; not universal reliability | | v10 | Added steady-hand-then-turn branch | Some held-start movements recovered, partial practical success | | v11 | Added separate larger-turn branch | Later three-motion test detected only one | | v12 | Split reverse preparation and forward extension into phases | Three motions detected; old steady branch misfired on withdrawal | | v13 | Removed STEADY_POINT, retained short/large/phase rules | Twelve replay checks; fist-only withdrawal and separate circle hardware checks passed | | v14 | Added BLE media keys, automatic release, connection guards and button controls | Windows media operation confirmed; continuous circles latched as one action | | v15 | Added continuous heading-based circle counting | Three-turn replay and noncircle controls; iPad connected but did not act | | v16 | Tried combined keyboard/media HID, corrected PnP byte order, added subscription diagnostics | WAIT REPORT; compatibility not solved | | v17 | Changed fixed identity to exclude cached layout and logged map/subscription | Found later same-UUID report objects were absent from GATT | | v18 | Local duplicate-UUID registration fix plus report-handle guard and reversible build patch | Xiaomi volume worked; iPad subscribed but still did not act | | v19 | Explicit 16-bit consumer bitmap; direct volume/held-key/letter diagnostics | Removing Smart Keyboard did not fix iPad behavior | | v20 | Encrypted report and Report Reference access, security diagnostics | User confirmed iPad media operation; text-keyboard identity affected typing UX | | v21 | Removed keyboard/LED reports; consumer-only layout, refreshed fixed identity | Descriptor checks passed; screen-keyboard recovery not separately confirmed at that stage | | v22 | Replaced two-finger motion with GripTracker for grasp | User reported missed grasps and withdrawal-triggered track changes | | v23 | Fixed moving grasp baseline; recent area peak, curl fallback, endpoint settling and withdrawal suppression | User confirmed grasp, clean withdrawal and left/right track changes | | v24 | Double-click host switching with old-peer rejection and preserved bonds | Compiled/flashed; user confirmed another host READY and actual media control | | v25 draft | Added area-pulse and asymmetric late-turn filtering to reduce fist sweeps | **Uncompiled, unflashed and untested**; patch only, not default | ## Open issue after v24 Continuous clenched-fist sweeps and intended slaps can produce overlapping centroid/area/time signals. The user selected uninterrupted slap interaction and accepted approximate filtering instead of a new dwell prerequisite. Reliable hand-shape discrimination is not established. Initial Xiaomi Home / Home Assistant integration remains unimplemented.