aiko/denpa

Fork 0

Desktop transparent pal — no window chrome, corner-scalable #18

New issue

Closed

opened 2026-07-29 17:32:08 +00:00 by aiko · 3 comments

aiko commented

2026-07-29 17:32:08 +00:00

Owner

What to build

Card 1b, plus 7e. The window has no background and no titlebar: Tauri transparent: true, decorations: false, alwaysOnTop: true. Set a click-through region (ignoreCursorEvents outside hit areas) so only the model, bubbles and bar are hit-testable — empty space must not eat desktop clicks.

She is most of the window; the bar is 40px and only as wide as she is.

Transmission bubble tucked at her shoulder, with a tail pointing at her; a cyan user counter-bubble below it. Keep only the last two turns — older ones fade to the log.
The bar is the whole chrome, and doubles as the drag handle (data-tauri-drag-region): 最前面 grip, carrier meter, an invocation line that always states what caused the current turn (自発 · trigger / 送信 / 待機), then 文 / 受信 / 割り込み keys.
No text field in the bar. 文 reveals a composer panel above it: ⏎ sends, ⇧⏎ newlines.
Corner grip (拡縮) scales the whole pal as one group from transform-origin: bottom left, so she grows up and out and never drifts off-screen. Clamp 0.68×–1.32×, show a live readout and dashed bounds while dragging, and persist scale per monitor. Same stored value as the 相手 scale slider.
Global hotkey ⌥ held = push to talk, so the window never needs focus.

7e — faults at pal scale. At 0.7× a 25px headline and a two-hop chain will not fit, so a fault becomes one line plus a way in: the meter dies, the bar border turns acid, 不通 replaces the station name, she says one short line in her own voice, and a small なぜ button opens the full failure screen in her window. Per fault: 401 → 鍵が通らない; no model → 姿がない and she keeps talking; mic lost → the 受信 key turns acid and nothing else changes.

Design reference: docs/design/denpa-receiver/README.md (card ids 1a, 4a… index into Denpa Receiver.dc.html, openable in a browser).

Acceptance criteria

Window is fully transparent and undecorated; desktop clicks pass through empty space
Bar drags the window; corner grip scales the whole pal within the clamp
Scale persists per monitor and matches the settings slider
Only the last two turns show; older ones are in the log
Composer sends on ⏎, newlines on ⇧⏎
Invocation line always reflects the current turn's cause
Global PTT hotkey works without window focus
Each fault renders its one-line pal treatment, with なぜ opening the full screen
Tests cover the click-through region and the scale clamp

Blocked by

## What to build Card 1b, plus 7e. **The window has no background and no titlebar**: Tauri `transparent: true`, `decorations: false`, `alwaysOnTop: true`. Set a **click-through region** (`ignoreCursorEvents` outside hit areas) so only the model, bubbles and bar are hit-testable — empty space must not eat desktop clicks. She is most of the window; the bar is 40px and only as wide as she is. - **Transmission bubble** tucked at her shoulder, with a tail pointing at her; a cyan user counter-bubble below it. **Keep only the last two turns** — older ones fade to the log. - **The bar is the whole chrome**, and doubles as the drag handle (`data-tauri-drag-region`): 最前面 grip, carrier meter, an **invocation line that always states what caused the current turn** (自発 · trigger / 送信 / 待機), then 文 / 受信 / 割り込み keys. - **No text field in the bar.** 文 reveals a composer panel above it: ⏎ sends, ⇧⏎ newlines. - **Corner grip** (拡縮) scales **the whole pal as one group** from `transform-origin: bottom left`, so she grows up and out and never drifts off-screen. Clamp 0.68×–1.32×, show a live readout and dashed bounds while dragging, and **persist scale per monitor**. Same stored value as the 相手 scale slider. - Global hotkey `⌥` held = push to talk, so the window never needs focus. **7e — faults at pal scale.** At 0.7× a 25px headline and a two-hop chain will not fit, so a fault becomes **one line plus a way in**: the meter dies, the bar border turns acid, 不通 replaces the station name, she says one short line in her own voice, and a small なぜ button opens the full failure screen in her window. Per fault: 401 → 鍵が通らない; no model → 姿がない and **she keeps talking**; mic lost → the 受信 key turns acid and nothing else changes. Design reference: `docs/design/denpa-receiver/README.md` (card ids `1a`, `4a`… index into `Denpa Receiver.dc.html`, openable in a browser). ## Acceptance criteria - [ ] Window is fully transparent and undecorated; desktop clicks pass through empty space - [ ] Bar drags the window; corner grip scales the whole pal within the clamp - [ ] Scale persists per monitor and matches the settings slider - [ ] Only the last two turns show; older ones are in the log - [ ] Composer sends on ⏎, newlines on ⇧⏎ - [ ] Invocation line always reflects the current turn's cause - [ ] Global PTT hotkey works without window focus - [ ] Each fault renders its one-line pal treatment, with なぜ opening the full screen - [ ] Tests cover the click-through region and the scale clamp ## Blocked by - #10 - #15

aiko added the

ready-for-agent

label

2026-07-29 17:32:08 +00:00

aiko commented

2026-07-30 03:24:15 +00:00

Author

Owner

Branch: issue/denpa-18-desktop-transparent-pal (off main at 875a261). npm test — 17 files, 138 passed (baseline was 60 in 8; nothing deleted or skipped). npm run typecheck still reports exactly 585 errors and none of them name a path I touched.

What landed

components/pal/hit-regions.ts — the click-through region. Tauri's set_ignore_cursor_events has no equivalent of Electron's { forward: true }, so a window that is ignoring the cursor never receives the mouseover that would hand control back. The pal therefore polls cursorPosition() and hit-tests it against the rects of the elements it marked data-denpa-hit; the root also sets pointer-events: none so the poll lag cannot let empty space take a desktop click. It applies the flag through the existing update_component_hover command rather than set_ignore_mouse_events, because that one also makes the window focusable again.
pal-scale.ts — the 0.68×–1.32× clamp, the bottom-left grip arithmetic, the 1.14× readout and per-monitor persistence. This is the only store for the scale, so 5b's 大きさ slider reads and writes the same value.
pal-turns.ts (last two turns only, relay rows excluded), invocation-line.ts (自発 · trigger / 送信 / 待機, never blank), pal-composer.ts (⏎ sends, ⇧⏎ newlines, IME commit does not send), pal-fault.ts (7e, one line plus a way in, per fault).
desktop-pal.tsx + bar / bubbles / composer / corner grip, consuming #10's tokens, CarrierMeter and useReducedMotion. Three tokens added to denpa-tokens.ts in the same style: hairline.acid, text.onSignal, and a wash.* group for the translucent panels a window with no background needs.
PalContainer replaces InputSubtitle as pet mode's UI, wired to the existing contexts and send path. src-tauri: a bare-⌥ global hotkey emitting ptt-press/ptt-release, and four core:window permissions for the cursor/monitor reads.

Deliberately left out

The fault treatments are not driven by anything. They render from the fault prop and are covered by tests, but PalContainer passes none — 401 / no-model / upstream-down are not reported to the client until the adapter route lands (#20/#21), and なぜ has no failure screen to open yet. A fault that cannot happen would be a disabled affordance.
自発 never appears on the invocation line. ProactiveSpeakContext does not carry the trigger yet, per the design's own deferral.

Assumptions (all five are in .denpa-work/QUESTIONS.md)

A bare AltLeft is registrable as a global shortcut; failure is logged and PTT is simply absent, as with the other two hotkeys.
The Live2D canvas stays where App.tsx puts it and takes the group's scale from a --denpa-pal-scale CSS variable, rather than being remounted inside the pal group.
不通 replaces the station name only for faults that actually stop the transmission; no-model and mic-lost leave it alone, since she keeps working in both.

Not verified: I could not run the app, so every layout and visual claim here rests on reading 1b/7e, not on seeing it rendered. Specifically unverified: whether the bubble actually meets her shoulder, whether the declared model hit rect coincides with the drawn canvas, and whether the click-through poll feels immediate in practice.

Branch: `issue/denpa-18-desktop-transparent-pal` (off `main` at 875a261). `npm test` — **17 files, 138 passed** (baseline was 60 in 8; nothing deleted or skipped). `npm run typecheck` still reports exactly 585 errors and none of them name a path I touched. **What landed** - `components/pal/hit-regions.ts` — the click-through region. Tauri's `set_ignore_cursor_events` has no equivalent of Electron's `{ forward: true }`, so a window that is ignoring the cursor never receives the mouseover that would hand control back. The pal therefore polls `cursorPosition()` and hit-tests it against the rects of the elements it marked `data-denpa-hit`; the root also sets `pointer-events: none` so the poll lag cannot let empty space take a desktop click. It applies the flag through the existing `update_component_hover` command rather than `set_ignore_mouse_events`, because that one also makes the window focusable again. - `pal-scale.ts` — the 0.68×–1.32× clamp, the bottom-left grip arithmetic, the `1.14×` readout and per-monitor persistence. This is the only store for the scale, so 5b's 大きさ slider reads and writes the same value. - `pal-turns.ts` (last two turns only, relay rows excluded), `invocation-line.ts` (自発 · trigger / 送信 / 待機, never blank), `pal-composer.ts` (⏎ sends, ⇧⏎ newlines, IME commit does not send), `pal-fault.ts` (7e, one line plus a way in, per fault). - `desktop-pal.tsx` + bar / bubbles / composer / corner grip, consuming #10's tokens, `CarrierMeter` and `useReducedMotion`. Three tokens added to `denpa-tokens.ts` in the same style: `hairline.acid`, `text.onSignal`, and a `wash.*` group for the translucent panels a window with no background needs. - `PalContainer` replaces `InputSubtitle` as pet mode's UI, wired to the existing contexts and send path. `src-tauri`: a bare-⌥ global hotkey emitting `ptt-press`/`ptt-release`, and four `core:window` permissions for the cursor/monitor reads. **Deliberately left out** - **The fault treatments are not driven by anything.** They render from the `fault` prop and are covered by tests, but `PalContainer` passes none — 401 / no-model / upstream-down are not reported to the client until the adapter route lands (#20/#21), and なぜ has no failure screen to open yet. A fault that cannot happen would be a disabled affordance. - **自発 never appears on the invocation line.** `ProactiveSpeakContext` does not carry the trigger yet, per the design's own deferral. **Assumptions** (all five are in `.denpa-work/QUESTIONS.md`) - A bare `AltLeft` is registrable as a global shortcut; failure is logged and PTT is simply absent, as with the other two hotkeys. - The Live2D canvas stays where App.tsx puts it and takes the group's scale from a `--denpa-pal-scale` CSS variable, rather than being remounted inside the pal group. - 不通 replaces the station name only for faults that actually stop the transmission; no-model and mic-lost leave it alone, since she keeps working in both. **Not verified:** I could not run the app, so every layout and visual claim here rests on reading 1b/7e, not on seeing it rendered. Specifically unverified: whether the bubble actually meets her shoulder, whether the declared model hit rect coincides with the drawn canvas, and whether the click-through poll feels immediate in practice.

aiko commented

2026-07-30 03:30:36 +00:00

Author

Owner

Branch: issue/denpa-18-desktop-transparent-pal (pushed, no PR, issue left open).

What landed

The desktop pal as a background-less window. App.tsx now renders PalContainer where it rendered InputSubtitle, composed of DesktopPal, PalBar, PalBubbles, PalComposerPanel and CornerGrip. Logic was kept out of the components and put in testable modules: hit-regions.ts (which rects are opaque to the cursor), pal-scale.ts (the shared scale store), pal-turns.ts (grouping messages into turns), invocation-line.ts, pal-composer.ts, and pal-fault.ts (the 7e fault treatment table).

Click-through is a 60ms renderer-side poll that hit-tests data-denpa-hit rects and applies the result through the existing update_component_hover Tauri command; use-click-through.ts and lib/tauri-pal.ts are the seams. Rust side: a held-AltLeft global shortcut for push-to-talk in src-tauri/src/lib.rs plus the matching src-tauri/capabilities/default.json entries. Three tokens were added to theme/denpa-tokens.ts in the existing style rather than invented locally: hairline.acid (7e's acid bar border), text.onSignal (1b's #c8f7f4 counter-bubble text), and a wash.* group for the translucent panel/edge values a background-less window needs.

Tests

npm test — 17 files, 138 tests, all passing:

 RUN  v4.1.10 F:/Projects/denpa-repos/denpa-client/.claude/worktrees/wf_9f79f977-7a2-3
 Test Files  17 passed (17)
      Tests  138 passed (138)
   Duration  3.07s

They cover hit-region geometry and the rects the pal declares; the scale store; turn grouping, including that a message with no question in front of it becomes a turn of its own and that relay rows never open or extend a turn; invocation-line and composer state; and the whole fault treatment table (station name, meter, keeps-talking, line text, bar border) per fault. The component tests assert structure and attributes — which element carries the drag region, which carry data-denpa-hit, the transform and its origin, which strings appear per fault. They assert nothing about appearance.

npm run typecheck still reports exactly 585 errors, the pre-existing count, and grepping its output for pal, tauri-pal, App.tsx and theme returns nothing. Two new errors did appear mid-work in pal-bar.tsx and pal-composer-panel.tsx (Chakra Box as="button" rejects type/disabled); they were fixed by switching to chakra('button') before committing. npm run lint is broken on main and was not run.

No Rust tests were added or run. cargo is not part of the gate and building the Tauri tree in a worktree with no target dir of its own was not worth the turn, so the ⌥ hotkey and the new capability entries are unexecuted code.

Visual claims never observed rendering

I never rendered the pal. Every layout claim below rests on reading design 1b/7e, not on seeing it:

the bubble meets her shoulder at bottom: 296px
the bar is only as wide as she is
the model's declared hit rect coincides with the Live2D canvas App.tsx draws
the composer sits above the bar
the corner grip lands at the top-right of the pal bounds
scaling the pal group and the out-of-subtree canvas through one shared CSS variable keeps them aligned

Also unconfirmed at runtime: that the click-through poll feels immediate, and that a bare AltLeft registers as a global hotkey at all.

Deliberately left out

The 7e fault treatments render from a table and are tested, but nothing drives them — PalContainer passes no fault. The two-hop status that names 401 / no-model / upstream-down is not reported to the client until the adapter route lands (#20/#21), and なぜ has no failure screen (5a/6a–6d) to open. Rendering a fault that cannot occur would be a disabled affordance, which the design forbids. 自発 does not reach the invocation line either: ProactiveSpeakContext does not carry the trigger yet, per the design's own deferral.

The 5b 大きさ slider is not built — different card, different issue. pal-scale.ts is the single store both surfaces are meant to share, so the slider is a read/write against it.

InputSubtitle is no longer rendered, but the file and its hooks were left untouched rather than deleted, since a sibling branch may reference them.

Assumptions

A bare Code::AltLeft is registrable as a global shortcut for held-modifier PTT. Registration failure is logged and PTT is simply absent, matching how the two existing hotkeys behave. Unverified on a running app, and a global Alt hold may interfere with other apps' menu activation on Windows.
The corner grip must scale the canvas too, but App.tsx owns the Live2D container outside PalContainer's subtree — so the pal publishes --denpa-pal-scale on the document element and pet mode's canvas style reads it with the same transform-origin: left bottom, rather than remounting the Cubism/WebGL canvas inside the pal group.
不通 replaces the station name only for faults that stop the transmission (upstream-down, 401). no-model and mic-lost leave the name and the live meter alone, because 6c says both hops are up and 6d says nothing but the 受信 key changes.
mic-lost has no line and no なぜ button — "nothing else changes" taken literally, so the acid 受信 key is itself the way into the 6d screen.
Click-through is polled at 60ms and hit-tested in the renderer, applied via the existing update_component_hover command (which also restores focusability), rather than a new Rust cursor watcher.
A turn is one transmission plus the message that caused it; an agent message with no question in front of it becomes its own turn, which is how an unprompted transmission will appear once it exists. Relay rows never open or extend a turn.
Pet mode is the pal, so PalContainer replaces InputSubtitle in App.tsx.
The three missing tokens belong in denpa-tokens.ts, not local to the pal.

Review findings

Two findings, both addressed, none skipped.

Blocker — pal-fault.ts: no-model (6c) set barBorderAcid: true. The code contradicted its own adjacent comment ("both hops are up … the carrier is alive"), and design row 6c states "Both hops ✓" with 声だけで続ける as "a real path, not a dead end". The acid bar edge is the signal that the transmission stopped (README:497, under the base fault treatment), so firing it on a live carrier mis-draws the fault. Flipped to false. Verified the acid-bordered bubble that carries the 姿がない line is computed independently in pal-bubbles.tsx and is unaffected — barBorderAcid is consumed only at pal-bar.tsx:122, for the bar edge colour. The adjacent comment was tightened to say why the bar edge stays cyan.

Major — pal-fault.test.ts: the no-model case asserted station name, meter, keeps-talking and line but not barBorderAcid, which is how the above passed. Added expect(treatment.barBorderAcid).toBe(false);. Ran it against the old behaviour first and watched it fail for the right reason:

 FAIL  pal-fault.test.ts > palFaultTreatment > keeps her talking with no model, and leaves the meter alive
 AssertionError: expected true to be false // Object.is equality
 ❯ src/renderer/src/components/pal/pal-fault.test.ts:53:37

Then it passes with the fix. The 138/138 tail above is post-fix.

One process gap to be honest about: the caveman:cavecrew-reviewer pass the global instructions require was not run — no agent-spawning tool is available to me as a subagent. Treat this branch as having had one pair of eyes, not two.

Branch: `issue/denpa-18-desktop-transparent-pal` (pushed, no PR, issue left open). ## What landed The desktop pal as a background-less window. `App.tsx` now renders `PalContainer` where it rendered `InputSubtitle`, composed of `DesktopPal`, `PalBar`, `PalBubbles`, `PalComposerPanel` and `CornerGrip`. Logic was kept out of the components and put in testable modules: `hit-regions.ts` (which rects are opaque to the cursor), `pal-scale.ts` (the shared scale store), `pal-turns.ts` (grouping messages into turns), `invocation-line.ts`, `pal-composer.ts`, and `pal-fault.ts` (the 7e fault treatment table). Click-through is a 60ms renderer-side poll that hit-tests `data-denpa-hit` rects and applies the result through the existing `update_component_hover` Tauri command; `use-click-through.ts` and `lib/tauri-pal.ts` are the seams. Rust side: a held-`AltLeft` global shortcut for push-to-talk in `src-tauri/src/lib.rs` plus the matching `src-tauri/capabilities/default.json` entries. Three tokens were added to `theme/denpa-tokens.ts` in the existing style rather than invented locally: `hairline.acid` (7e's acid bar border), `text.onSignal` (1b's `#c8f7f4` counter-bubble text), and a `wash.*` group for the translucent panel/edge values a background-less window needs. ## Tests `npm test` — 17 files, 138 tests, all passing: ``` RUN v4.1.10 F:/Projects/denpa-repos/denpa-client/.claude/worktrees/wf_9f79f977-7a2-3 Test Files 17 passed (17) Tests 138 passed (138) Duration 3.07s ``` They cover hit-region geometry and the rects the pal declares; the scale store; turn grouping, including that a message with no question in front of it becomes a turn of its own and that relay rows never open or extend a turn; invocation-line and composer state; and the whole fault treatment table (station name, meter, keeps-talking, line text, bar border) per fault. The component tests assert **structure and attributes** — which element carries the drag region, which carry `data-denpa-hit`, the transform and its origin, which strings appear per fault. They assert nothing about appearance. `npm run typecheck` still reports exactly 585 errors, the pre-existing count, and grepping its output for `pal`, `tauri-pal`, `App.tsx` and `theme` returns nothing. Two new errors did appear mid-work in `pal-bar.tsx` and `pal-composer-panel.tsx` (Chakra `Box as="button"` rejects `type`/`disabled`); they were fixed by switching to `chakra('button')` before committing. `npm run lint` is broken on `main` and was not run. **No Rust tests were added or run.** cargo is not part of the gate and building the Tauri tree in a worktree with no target dir of its own was not worth the turn, so the ⌥ hotkey and the new capability entries are unexecuted code. ## Visual claims never observed rendering I never rendered the pal. Every layout claim below rests on reading design 1b/7e, not on seeing it: - the bubble meets her shoulder at `bottom: 296px` - the bar is only as wide as she is - the model's declared hit rect coincides with the Live2D canvas `App.tsx` draws - the composer sits above the bar - the corner grip lands at the top-right of the pal bounds - scaling the pal group and the out-of-subtree canvas through one shared CSS variable keeps them aligned Also unconfirmed at runtime: that the click-through poll feels immediate, and that a bare `AltLeft` registers as a global hotkey at all. ## Deliberately left out The 7e fault treatments render from a table and are tested, but **nothing drives them** — `PalContainer` passes no fault. The two-hop status that names 401 / no-model / upstream-down is not reported to the client until the adapter route lands (#20/#21), and なぜ has no failure screen (5a/6a–6d) to open. Rendering a fault that cannot occur would be a disabled affordance, which the design forbids. 自発 does not reach the invocation line either: `ProactiveSpeakContext` does not carry the trigger yet, per the design's own deferral. The 5b 大きさ slider is not built — different card, different issue. `pal-scale.ts` is the single store both surfaces are meant to share, so the slider is a read/write against it. `InputSubtitle` is no longer rendered, but the file and its hooks were left untouched rather than deleted, since a sibling branch may reference them. ## Assumptions 1. A bare `Code::AltLeft` is registrable as a global shortcut for held-modifier PTT. Registration failure is logged and PTT is simply absent, matching how the two existing hotkeys behave. Unverified on a running app, and a global Alt hold may interfere with other apps' menu activation on Windows. 2. The corner grip must scale the canvas too, but `App.tsx` owns the Live2D container outside `PalContainer`'s subtree — so the pal publishes `--denpa-pal-scale` on the document element and pet mode's canvas style reads it with the same `transform-origin: left bottom`, rather than remounting the Cubism/WebGL canvas inside the pal group. 3. 不通 replaces the station name only for faults that stop the transmission (upstream-down, 401). no-model and mic-lost leave the name and the live meter alone, because 6c says both hops are up and 6d says nothing but the 受信 key changes. 4. mic-lost has no line and no なぜ button — "nothing else changes" taken literally, so the acid 受信 key is itself the way into the 6d screen. 5. Click-through is polled at 60ms and hit-tested in the renderer, applied via the existing `update_component_hover` command (which also restores focusability), rather than a new Rust cursor watcher. 6. A turn is one transmission plus the message that caused it; an agent message with no question in front of it becomes its own turn, which is how an unprompted transmission will appear once it exists. Relay rows never open or extend a turn. 7. Pet mode *is* the pal, so `PalContainer` replaces `InputSubtitle` in `App.tsx`. 8. The three missing tokens belong in `denpa-tokens.ts`, not local to the pal. ## Review findings Two findings, both addressed, none skipped. **Blocker** — `pal-fault.ts`: no-model (6c) set `barBorderAcid: true`. The code contradicted its own adjacent comment ("both hops are up … the carrier is alive"), and design row 6c states "Both hops ✓" with 声だけで続ける as "a real path, not a dead end". The acid bar edge is the signal that the transmission stopped (README:497, under the base fault treatment), so firing it on a live carrier mis-draws the fault. Flipped to `false`. Verified the acid-bordered *bubble* that carries the 姿がない line is computed independently in `pal-bubbles.tsx` and is unaffected — `barBorderAcid` is consumed only at `pal-bar.tsx:122`, for the bar edge colour. The adjacent comment was tightened to say why the bar edge stays cyan. **Major** — `pal-fault.test.ts`: the no-model case asserted station name, meter, keeps-talking and line but not `barBorderAcid`, which is how the above passed. Added `expect(treatment.barBorderAcid).toBe(false);`. Ran it against the old behaviour first and watched it fail for the right reason: ``` FAIL pal-fault.test.ts > palFaultTreatment > keeps her talking with no model, and leaves the meter alive AssertionError: expected true to be false // Object.is equality ❯ src/renderer/src/components/pal/pal-fault.test.ts:53:37 ``` Then it passes with the fix. The 138/138 tail above is post-fix. One process gap to be honest about: the `caveman:cavecrew-reviewer` pass the global instructions require was **not** run — no agent-spawning tool is available to me as a subagent. Treat this branch as having had one pair of eyes, not two.

aiko referenced this issue from a commit

2026-07-30 03:53:53 +00:00

Merge #18's desktop pal into the receiver chrome

aiko commented

2026-07-30 13:32:53 +00:00

Author

Owner

Shipped. Merged to main at d73f4ed (phase B2i — client chrome integration).

Closed as part of reconciling the board with main: the autonomous run's policy was not to auto-close, so shipped work stayed open and the tracker read as though nothing had landed. That policy is now reversed — issues close when they land.

Shipped. Merged to `main` at [`d73f4ed`](https://git.aiko.works/aiko/denpa/commit/d73f4ed) (phase B2i — client chrome integration). Closed as part of reconciling the board with `main`: the autonomous run's policy was not to auto-close, so shipped work stayed open and the tracker read as though nothing had landed. That policy is now reversed — issues close when they land.

aiko closed this issue

2026-07-30 13:32:54 +00:00