Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

Declare a finger gone when the CST9217 stops acknowledging it; drop the goal hold

0xdiego2536d0f073001f
diff --git a/device/esp32-circle/main/ui/display.c b/device/esp32-circle/main/ui/display.c
index 0fa91c43d..94ed3421c 100644
--- a/device/esp32-circle/main/ui/display.c
+++ b/device/esp32-circle/main/ui/display.c
@@ -96,7 +96,7 @@ static void heartbeat(void)
touch_stats(&t);
lv_mem_monitor_t m;
lv_mem_monitor(&m);
- ESP_LOGI(TAG, "alive up=%lus heap=%u/%u psram=%u lv=%u%%used frag=%u%% touches=%lu last_press=%lus%s%s%s",
+ ESP_LOGI(TAG, "alive up=%lus heap=%u/%u psram=%u lv=%u%%used frag=%u%% touches=%lu last_press=%lus%s%s%s%s",
(unsigned long)(lv_tick_get() / 1000),
(unsigned)heap_caps_get_free_size(MALLOC_CAP_INTERNAL),
(unsigned)heap_caps_get_largest_free_block(MALLOC_CAP_INTERNAL),
@@ -105,6 +105,7 @@ static void heartbeat(void)
(unsigned long)t.presses, (unsigned long)(t.last_press_ms_ago / 1000),
t.controller_ok ? "" : " TOUCH-DEAD",
t.read_failures ? " read-fails" : "",
+ t.inferred_releases ? " inferred-releases" : "",
t.held_now ? " PRESS-HELD" : "");
}
diff --git a/device/esp32-circle/main/ui/touch.c b/device/esp32-circle/main/ui/touch.c
index 4bba478d7..4dd091e89 100644
--- a/device/esp32-circle/main/ui/touch.c
+++ b/device/esp32-circle/main/ui/touch.c
@@ -29,6 +29,10 @@ static esp_lcd_panel_io_handle_t s_tp_io; // kept so a dead controller can be
// every recognizer in this file — holding a finger that is no longer there. Which is the reported shape.
static uint32_t s_read_fail_total;
static uint32_t s_read_fail_run; // consecutive; reset by any good read
+// No-ACK reads in a row while the driver still reports a finger. See the INVALID_RESPONSE branch.
+static uint32_t s_noack_run;
+static uint32_t s_inferred_releases; // lifetime, for the heartbeat: how often the fix above fired
+#define NOACK_RELEASE_READS 3
#define READ_FAIL_REINIT 50 // ~1s of nothing but errors → tear the controller down and bring it back
#define READ_FAIL_LOG_EVERY 100
#define REINIT_RETRY_MS 5000 // a reinit that failed is tried again this often
@@ -41,8 +45,8 @@ static bool s_stuck_warned; // the ">held 5s" line for this press alre
#define STUCK_PRESS_MS 5000
#define STUCK_PRESS_EVERY_MS 30000
-// How far a finger may drift and still count as "held still". Shared by the near-still TAP test and the
-// GOAL hold — it outlived the create-project long-press that introduced it.
+// How far a finger may drift and still count as "held still" — the near-still TAP test. It outlived the
+// create-project long-press and the goal hold that used it.
#define LONG_MOVE_PX 30
#define DOUBLE_TAP_MS 500 // two taps within this window (and close in position) = a double-tap. Also the
@@ -111,12 +115,8 @@ static uint32_t s_tap_pending_ms;
// Where the press landed. The tap is deferred past the double-tap window, by which time the finger has
// gone, so the point has to be carried with it — ui_tap uses it to decide what was actually touched.
static int16_t s_tap_x, s_tap_y;
-// GOAL gesture: a still hold ≥2s (awake, not already recording) starts a GOAL voice command.
-// It "consumes" the press: swallows the rest so it lands as no UI tap.
-#define GOAL_HOLD_MS 2000
-static bool s_goal_swallow;
-// Swallow all input until the finger lifts — set after a wake tap / double-tap-voice / goal-hold so the
-// consumed gesture doesn't also drive the UI. File-scope so swipe_track can be gated off while it's set
+// Swallow all input until the finger lifts — set after a wake tap so the consumed gesture doesn't also
+// drive the UI. File-scope so swipe_track can be gated off while it's set
// (otherwise swipe_track starts mid-gesture on the read after wake and classifies the release as a tap →
// stray "open detail" on the very tap that woke the screen).
static bool s_swallow_until_release;
@@ -246,7 +246,6 @@ static void swipe_track(bool pressed, uint16_t x, uint16_t y)
// The DETAIL READER has its own tight gesture set (the rest is native vertical scroll):
// • up-swipe from the bottom ~20px → Overview • horizontal swipe → back to the agent screen
// Everything else on the reader (mid-screen vertical drag, taps) is left to LVGL scroll / does nothing.
- // (double-tap = voice and the ≥2s hold = goal are handled in touch_read, independent of this.)
bool reader = ui_reader_is_open();
int home_edge = reader ? READER_HOME_EDGE_PX : BOTTOM_EDGE_PX;
// HOME gesture: an upward swipe that STARTED at the bottom edge → jump to Overview. (y grows downward;
@@ -314,18 +313,31 @@ static void touch_read(lv_indev_t *indev, lv_indev_data_t *data)
static uint32_t ok_reads, stale_reads; // per press, for the release line — see below
if (rc == ESP_OK) {
s_read_fail_run = 0;
+ s_noack_run = 0;
pressed = esp_lcd_touch_get_coordinates(s_tp, &x, &y, &strength, &cnt, 1) && cnt > 0;
if (pressed) ok_reads++;
} else if (rc == ESP_ERR_INVALID_RESPONSE) {
- // The chip's report has no ACK byte — measured on the desk, this is what EVERY idle read returns
- // (~30/s with no finger near the glass), so it is not an error and never a reason to reinit. The
- // driver leaves its point table untouched on this path, which is the part that matters: what
- // LVGL sees is the LAST ACKED frame. Kept as such here, for now, and counted per press so the
- // release line says how many stale reads a held finger produces — that number decides whether
- // "no ACK" can be read as "no finger" (the fix for a release frame the chip showed once and this
- // task missed, which leaves the glass held forever).
+ // The chip's report has no ACK byte. MEASURED, not assumed (2026-09-15, this log): every idle
+ // read answers this way (~30/s with no finger near the glass), and NOT ONE read does while a
+ // finger is down — a still 2.9s hold read `acked=84 stale=0`. So no ACK is no finger.
+ //
+ // The driver leaves its point table untouched on this path, and until now so did this file:
+ // what LVGL saw was the LAST ACKED frame. A release the chip showed once and this task read a
+ // beat late — a busy frame, an I2C reset — therefore left the glass held forever: no swipe, no
+ // Voice, the tile still following the app. That is the stuck-dial report, and this is its fix.
+ // Three misses in a row (~100ms) before the finger is declared gone, so one dropped sample
+ // mid-drag cannot split a stroke in two.
+ s_read_fail_run = 0;
pressed = esp_lcd_touch_get_coordinates(s_tp, &x, &y, &strength, &cnt, 1) && cnt > 0;
- if (pressed) stale_reads++;
+ if (pressed && ++s_noack_run >= NOACK_RELEASE_READS) {
+ ESP_LOGW(TAG, "release inferred: no ACK for %u reads while held at (%u,%u) — the chip's release frame was missed",
+ (unsigned)s_noack_run, x, y);
+ pressed = false;
+ s_noack_run = 0;
+ s_inferred_releases++;
+ } else if (pressed) {
+ stale_reads++;
+ }
} else {
// A real I2C failure. The driver has already reset the chip (with 60ms of delays, on this task).
// No answer is no finger: the point table is whatever it was, and trusting it is how a finger
@@ -364,9 +376,9 @@ static void touch_read(lv_indev_t *indev, lv_indev_data_t *data)
if (!s_stuck_warned || lv_tick_elaps(warned_at) >= STUCK_PRESS_EVERY_MS) {
s_stuck_warned = true;
warned_at = lv_tick_get();
- ESP_LOGW(TAG, "press held %us at (%u,%u): swallow=%d goal=%d action=%d band=%d voice=%d asleep=%d switch=%d notif=%d",
+ ESP_LOGW(TAG, "press held %us at (%u,%u): swallow=%d action=%d band=%d voice=%d asleep=%d switch=%d notif=%d",
(unsigned)(lv_tick_elaps(s_press_t0) / 1000), x, y,
- s_swallow_until_release, s_goal_swallow, s_action_capture, s_band_drag,
+ s_swallow_until_release, s_action_capture, s_band_drag,
ui_voice_is_active(), display_is_asleep(), ui_switch_is_open(), ui_notif_is_open());
}
}
@@ -437,31 +449,17 @@ static void touch_read(lv_indev_t *indev, lv_indev_data_t *data)
}
}
- // Swipes + tap-to-stop-voice. Skip while swallowing a consumed gesture (wake / voice / goal) so the
+ // Swipes + tap-to-stop-voice. Skip while swallowing a consumed gesture (the wake tap) so the
// release of that gesture isn't misread as a fresh tap.
// ...and not under the agent picker, whose rows are LVGL's to dispatch: swipe_track would read a row
// tap as the tap that opens the detail reader, and a scroll of the list as a carousel swipe, both
// happening to the tile UNDERNEATH the overlay.
- if (!display_is_asleep() && !s_swallow_until_release && !s_goal_swallow && !ui_switch_is_open())
+ if (!display_is_asleep() && !s_swallow_until_release && !ui_switch_is_open())
swipe_track(gesture_pressed, x, y);
- // GOAL: a still hold ≥3s (awake, not already recording) → start a GOAL voice command. Consumes the press
- // (swallow the rest so it doesn't land as a UI tap on release).
- {
- static bool gprev; static uint32_t gt0; static int16_t gsx, gsy; static bool gmoved, gfired;
- if (gesture_pressed && !gprev) { gt0 = lv_tick_get(); gsx = x; gsy = y; gmoved = false; gfired = false; }
- else if (gesture_pressed && !gfired) {
- int dx = (int)x - gsx, dy = (int)y - gsy;
- if (dx * dx + dy * dy > LONG_MOVE_PX * LONG_MOVE_PX) gmoved = true;
- if (!gmoved && !display_is_asleep() && !ui_voice_is_active() && lv_tick_elaps(gt0) >= GOAL_HOLD_MS) {
- gfired = true;
- s_goal_swallow = true; // swallow until the finger lifts (no stray UI tap)
- ESP_LOGI(TAG, "goal hold at (%u,%u) → voice goal, swallowing until release", x, y);
- ui_voice_start_goal();
- }
- }
- gprev = gesture_pressed;
- }
+ // A still hold used to start a GOAL voice command here (owner, 2026-09-15: "remove luôn cái long
+ // hold ra goal voice"). Holding the glass now means nothing beyond what LVGL makes of it; Voice is
+ // the button, and only the button.
// Resolve a deferred single tap. Anchored to the PRESS and only fires once the double-tap window has
// passed — so a double-tap always cancels it first. Extra guard: skip if a turn just started (voice).
@@ -476,11 +474,10 @@ static void touch_read(lv_indev_t *indev, lv_indev_data_t *data)
// After a double-tap starts voice, the finger is usually STILL down. Keep swallowing until it lifts,
// so the two taps don't also land as UI presses (a stray tile/settings click).
- if (s_swallow_until_release || s_goal_swallow) {
+ if (s_swallow_until_release) {
if (pressed) { data->state = LV_INDEV_STATE_RELEASED; return; }
- ESP_LOGI(TAG, "swallow cleared (wake=%d goal=%d)", s_swallow_until_release, s_goal_swallow);
+ ESP_LOGI(TAG, "swallow cleared (wake tap consumed)");
s_swallow_until_release = false;
- s_goal_swallow = false; // finger lifted → goal-hold press fully consumed
}
if (display_is_asleep()) {
@@ -567,6 +564,7 @@ void touch_stats(touch_stats_t *out)
out->presses = s_presses;
out->last_press_ms_ago = s_presses ? lv_tick_elaps(s_last_press_tick) : 0;
out->read_failures = s_read_fail_total;
+ out->inferred_releases = s_inferred_releases;
out->controller_ok = s_tp != NULL;
out->held_now = s_stuck_warned;
}
diff --git a/device/esp32-circle/main/ui/touch.h b/device/esp32-circle/main/ui/touch.h
index 04cecfead..0a76ba862 100644
--- a/device/esp32-circle/main/ui/touch.h
+++ b/device/esp32-circle/main/ui/touch.h
@@ -19,6 +19,7 @@ typedef struct {
uint32_t presses;
uint32_t last_press_ms_ago;
uint32_t read_failures;
+ uint32_t inferred_releases; // releases the chip never reported, declared from its silence
bool controller_ok;
bool held_now; // a press has been down past the stuck threshold
} touch_stats_t;