Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
Python agent-orchestrationai-agentai-coding-assistantclaude-codecoding-agentcomputer-usedesktop-applocal-firstmcppersonal-jarvisvoice-assistantvoice-coding
fix(desktop): text is selectable again — every window ran on pywebview's kiosk default
rubenluetke10-beepc39accc081fe
diff --git a/docs/BUGS.md b/docs/BUGS.mdindex af6fd60f93..ecc5de2f67 100644--- a/docs/BUGS.md+++ b/docs/BUGS.md@@ -12147,6 +12147,71 @@ dev GET unchanged).why the recovery still needs one restart: an app that booted with `style=none`has no Tk root, so the bar cannot be swapped in live.+---++## BUG-173: the chat says a turn cost "10 tokens" when it cost fifty thousand — the live counter reads the placeholder and skips the cache (MEDIUM, FIXED 2026-08-24)++**Symptom (maintainer, 2026-08-23).** The live status line under a running+agent-chat turn read `Noodling… (31s · ↓10 tokens)` after half a minute of work+that included two tool calls and a thinking block. "I did not use just 8+tokens."++**Evidence.** One real Claude Code turn, `--output-format stream-json`, the+`assistant` line's own usage object:++```json+{"input_tokens": 2, "cache_creation_input_tokens": 30420,+ "cache_read_input_tokens": 21355, "output_tokens": 1}+```++and the `result` line that closes the same turn:++```json+{"input_tokens": 2, "cache_creation_input_tokens": 30420,+ "cache_read_input_tokens": 21355, "output_tokens": 4}+```++**Root cause — two mistakes stacked.**++1. `translate_claude_line` counted `input_tokens` and `output_tokens` only, so+ the 51 775 cached tokens that carry the turn were dropped. The input side+ read as **2** instead of **51 777** — the CLI reports the cached part BESIDE+ `input_tokens`, never inside it.+2. `output_tokens` in an `assistant` line is the `message_start` snapshot: a+ placeholder that only becomes true in the `result` line. Summing one+ placeholder per message is a message counter wearing a token counter's+ clothes — ten messages read as "10 tokens".++The finished-turn receipt was right all along: the `result` branch of the same+translator already read all four fields. Only the live line was blind, which is+why the number looked plausible until someone compared it with a bill.++**Fix.**++* `jarvis/agent_chat/runner_cli.py` — `_CLAUDE_USAGE_KEYS` counts all four+ fields, and a message's usage is merged with `max()` so a placeholder can+ never walk a real count back down.+* `jarvis/ui/web/frontend/src/components/agentchat/toolView.ts` — `inputTokens`+ adds `cache_read_input_tokens` and `cache_creation_input_tokens`. The OpenAI+ shape (`cached_input_tokens`) is deliberately NOT added: there the cached+ tokens are already inside `prompt_tokens`, and adding them would double-count.+* `AgentTimeline.tsx` — the live line no longer shows a token count at all.+ There is no honest one to show: measured against a real multi-step turn, every+ `assistant` line reports `output_tokens: 2` and only the closing `result` line+ carries the true 981, while the input side legitimately re-counts the whole+ conversation at every step (21 355 → 51 902 → 54 094 cached tokens read). The+ sum is arithmetically right — it matches the CLI's own `result` to the token —+ but "↑283.2k" under a one-sentence question reads as a bug, and a live figure+ that is either far too small or unbelievably large is worse than none. The+ receipt, which runs on the authoritative `result` line, keeps ↑ in / ↓ out /+ cost. If a CLI ever streams true usage, the line can come back.++**Lesson.** Two readers of one stream drift apart. The receipt path and the live+path sat thirty lines from each other in the same file; one was extended for+cache tokens, the other was not. A number that crosses layers needs one reader+and a parity test (AP-4) — a second, hand-rolled reader is a bug with a delay+fuse.+## BUG-174: right after a restart, voice on local models is unusably slow or returns nothing — the Whisper engine was built ON the event loop, and a slow build was answered by starting it again from zero (HIGH, FIXED 2026-08-24)**Symptom (maintainer, 2026-08-24 morning).** "I used it after a restart with@@ -12223,6 +12288,7 @@ supervisor and the supervised were on the same thread. And "it is taking toolong" is not a diagnosis: replacing a worker only helps when the worker isstuck, never when it is merely loading. Patience needs evidence, so the enginehad to say which one it was.+## BUG-175: "ask him what Jarvis version 2 can do" answers nothing — the model was handed a wiki tool whose mode is switched off, and the wiki that WAS working sat unused beside it (HIGH, FIXED 2026-08-24)**Symptom (maintainer, 2026-08-24).** A question typed into the chat, sixteen@@ -12542,3 +12608,39 @@ so — and never let a fallback fire on the absence of evidence when thefallback's failure mode is doing the thing twice.---++## BUG-179: nothing in the app can be selected or copied — every window was created on pywebview's kiosk default (HIGH, FIXED 2026-08-24)++**Symptom.** Drag the mouse across an agent's answer in the Agentic IDE chat+and nothing highlights. Same on the timeline, the tool rows, the receipts,+every view in the app, on every platform. Copying a path, an error message or+a block of generated code out of Jarvis is impossible with the mouse. Typing+still works, which is why it reads as a quirk of one view rather than a window+setting.++**Root cause.** `webview.create_window()` takes `text_select=False` by default,+and on that default pywebview injects `body {-webkit-user-select: none;+user-select: none; cursor: default}` into every page it loads+(`webview/js/customize.js`), on Windows, macOS and Linux alike. No window in+this repo ever passed the flag, so the whole app ran as a kiosk. `user-select`+blocks only mouse SELECTION, not keyboard input — hence the app feeling 95 %+normal and the missing 5 % being exactly copy and paste.++**Fix.**+* `jarvis/ui/desktop_app.py` — `TEXT_SELECTABLE`, spread into the kwargs of+ every `create_window` (main window, tray-reopened main window, detached+ view) — and `jarvis/ui/shell/shell.py` pass `text_select=True`.+* `src/index.css` states the app's own policy at `html body`. Specificity, not+ order, is what makes that stick: the injected `<style>` lands LAST in the+ head, so an equally specific `body` rule would lose to it, and a shell that+ was not updated (an older install, a different frontend host) cannot+ re-impose the kiosk. Chrome that is pressed rather than read — buttons,+ tabs, switches, menu items — opts back out, so dragging across a toolbar+ does not paint it blue.++**Lesson.** A library default that silences an interaction is invisible in+review: nothing in the diff is wrong, the flag was simply never named. When a+window frame can change what the page is allowed to do, the page states its own+policy as well, so the app does not depend on every shell agreeing with it.++---diff --git a/jarvis/ui/desktop_app.py b/jarvis/ui/desktop_app.pyindex 125c732446..2d4c538644 100644--- a/jarvis/ui/desktop_app.py+++ b/jarvis/ui/desktop_app.py@@ -69,6 +69,23 @@ if sys.platform == "win32":LOCK_FILE_PATH = DATA_DIR / "jarvis.lock"+# Every window that shows the app must let people select text with the mouse.+#+# This is not a preference — it repairs a default. pywebview's ``create_window``+# takes ``text_select=False`` unless told otherwise, and on that default it+# injects ``body {user-select: none; cursor: default}`` into every page it+# loads, on all three platforms. The whole app then behaves like a kiosk: no+# dragging over an answer to copy it, no I-beam cursor, nothing selectable+# anywhere. Typing still works, which is why it reads as a mysterious quirk of+# one view rather than a window flag (maintainer, 2026-08-24: "man kann keine+# Sachen kopieren").+#+# Spread into the kwargs of every window created here so no future window can+# quietly inherit the kiosk again. The page still decides where selection is+# unwanted (chrome, buttons) in its own CSS — that belongs to the UI, not to+# the window frame.+TEXT_SELECTABLE: dict[str, Any] = {"text_select": True}+# Views the desktop shell can split off into their own solo window, mapped to# the window-title label. Keys mirror the frontend ``SECTION_IDS``# (``store/events.ts``); labels are deliberately the stable English section@@ -4506,6 +4523,7 @@ class DesktopApp:resizable=True,confirm_close=False,background_color=self._window_background(),+ **TEXT_SELECTABLE,)except Exception as exc: # noqa: BLE001# WebViewException on a backend that cannot create runtime windows@@ -4668,6 +4686,7 @@ class DesktopApp:"resizable": True,"confirm_close": False,"background_color": self._window_background(),+ **TEXT_SELECTABLE,}def _hook_main_window_lifecycle(self) -> None:diff --git a/jarvis/ui/shell/shell.py b/jarvis/ui/shell/shell.pyindex f4212f9e72..3c1fcf0d2f 100644--- a/jarvis/ui/shell/shell.py+++ b/jarvis/ui/shell/shell.py@@ -86,6 +86,10 @@ class JarvisShell:frameless=self._cfg.frameless,easy_drag=self._cfg.easy_drag,confirm_close=self._cfg.confirm_close,+ # Same reason as jarvis/ui/desktop_app.py's TEXT_SELECTABLE: left at+ # pywebview's default this window injects `user-select: none` into+ # the page and nothing in the app can be selected or copied.+ text_select=True,)# Close button = hide (minimize-to-tray behavior)diff --git a/jarvis/ui/web/frontend/src/index.css b/jarvis/ui/web/frontend/src/index.cssindex 1f9af2f3f3..c02e80ca11 100644--- a/jarvis/ui/web/frontend/src/index.css+++ b/jarvis/ui/web/frontend/src/index.css@@ -17,6 +17,38 @@scrollbar-color: hsl(var(--border)) transparent;}+/* Selecting text is the app's decision, not the window frame's.+ PLAIN CSS and `html body` rather than `body`, both deliberately:++ pywebview — every desktop window the app runs in — injects+ `body {user-select: none; cursor: default}` at load time unless it was+ created with `text_select=True` (jarvis/ui/desktop_app.py fixes that at the+ source). That injected <style> lands LAST in the head, so an equally+ specific `body` rule here would lose to it. `html body` outranks it, which+ makes this stylesheet the authority whatever shell, version or window the+ page ends up in — including one that was never updated.++ Everything readable is therefore selectable. Chrome that is pressed rather+ than read opts out below, so dragging across a toolbar does not paint it+ blue; a component that needs an exception uses Tailwind's `select-none`,+ whose class beats both rules. */+html body {+ -webkit-user-select: text;+ user-select: text;+ cursor: auto;+}++button,+label,+[role="button"],+[role="tab"],+[role="switch"],+[role="menuitem"],+[role="option"] {+ -webkit-user-select: none;+ user-select: none;+}+@layer components {/** Voice-orb state language. The procedural material stays inside the sphere;