[shakibbinkabir]

Eight things I ran into using Coucou on Windows every day, one commit each. This folds in #318 and #321, which I am closing, so everything is in one place.

Take any subset. Five commits stand alone, and the other three each need only the one before them. I applied every row of the table to main separately and ran the tests on each.

#	Commit	What	Kind	Needs
1	df9b63a	Auto-close test no longer races its own timers	Test only	—
2	157f37c	A hidden island stops painting	Bug (0 % CPU rule)	—
3	4f82cbc	Room for the chat's model list	Bug in the port	—
4	f6798e8	A provider's error is one sentence, and a note stays in its card	Bug	—
5	d92b538	A chat error's note can be closed, and closes by itself	Small new behaviour	4
6	bcae11b	The picker marks a model its provider refused	Small new behaviour	—
7	f5dd758	A refused model stays marked between launches	Small new behaviour	6
8	89d2b70	The picker shows the providers that are set up	Proposal, departs from the Mac	—

They are ordered from "plain fix" to "your call". Commit 1 is first so that this PR's own Build check stops failing at random.

What changes on screen

Shipped views stay as they are. These are the only visible differences, so you can judge them against the no-restyle rule:

• 3: while the chat's model picker is open, the island is 312 px tall instead of 240. Nothing inside it changes.
• 4: a note longer than four lines ends with an ellipsis instead of filling its card.
• 5: the note shown after a chat error gets the finished card's "OK" button.
• 6: a model the provider refused is greyed in the list with "Not available" (a string the app already has).
• 8: a cloud provider without a key is not offered in the picker.

Commits 1, 2 and 7 change nothing you can see.

The repo's rules

• Secrets: the picker only asks the credential store whether a key is there, as before. No key leaves it.
• Network: no new request anywhere. The "Not available" marks come from the answers to real chat turns; nothing is probed.
• 0 % CPU when hidden: commit 2 is about exactly that. The only timer added is the note's single 10 s timeout, which exists only while a note is up.
• Disk: one new file, refused-models.json next to the log, with model names only. The uninstaller removes it with the log. settings.json gets no new field.
• Mac and iPhone: no app code is touched. Commit 1 changes one Swift test file.
• Pill IDs: untouched.

The commits

1. Auto-close test no longer races its own timers (was #321)

"Test auto-close timing" fails now and then on PRs that do not touch the state machine: line 27 on #307, line 54 on #320 and on this PR. Two cases sleep for a fixed time and then assert that a pending timer has not fired, with 150 ms and 100 ms to spare. Both timers are now a full second out. What each case checks is unchanged, and the test takes about 1.6 s longer.

On a macos-latest runner on my fork, compiled the way scripts/test-auto-close.sh does:

	Old test	New test
Idle, 10 runs	0 failed	0 failed
9 busy loops on 3 cores, 40 runs	0 failed	0 failed
Process frozen 200 ms out of every 500 ms, 30 runs	21 failed (16 at line 27, 5 at line 54)	0 failed

Plain CPU load did not reproduce it. Freezing the process did, at the two lines CI fails on. The freeze is my model of a stalled runner, not something I observed on the failing runs.

2. A hidden island stops painting (was #318)

style.css pauses infinite animations only in views that are not .on. The view that is on stays on while the island is compact or hidden, so the ticker's shimmer kept repainting at 60 fps inside an island of zero height. #content now carries .live while it is visible, and the existing rule also covers #content:not(.live) *.

CPU time of coucou.exe plus its WebView2 processes over 30 s, starting 3 s after the window shrinks to the wake strip, after a Claude Code session:

	Total	Renderer	GPU process
0.2.0 as released (two runs)	2,297 ms and 2,297 ms	1,344 ms	797 ms
With this commit	656 ms	500 ms	under 15 ms

About 0.5 s per 30 s remains, and this commit does not address it. It is not the page: a CDP profile of the hidden island shows 0 ms of script, layout and style over 20 s, and the AudioContext is suspended. It sits on Chromium's thread-pool threads and I could not attribute it to app code.

3. Room for the chat's model list

On macOS the picker is a popover with 260 pt for the list. In the port it sits inside the island, in whatever the chat leaves: 23 px for rows 28 px tall. While the picker is open the island now takes all the height the panel has (PANEL_H - 8), the way the question view already grows for its options.

Before	After

	Island	List	Models fully visible (of 14)
Before	240 px	23 px	0
After	312 px	95 px	3

It does not reach the Mac's 260 pt. That needs a taller window, and PANEL_H is also the desktop Mochi's home zone and the Linux geometry, so I left it alone.

4. A provider's error is one sentence, and a note stays in its card

A chat error is shown as a note, in a card made for one short line. The cloud providers' errors carried the provider's whole message after our own sentence. Gemini's 404 for a model new accounts can no longer use adds three sentences of developer advice and a link: 341 characters over five lines, filling 100 of the card's 108 px.

• net::error_detail keeps the provider's first sentence. Its comment already said "cut down for a message", but only plain-text bodies were cut.
• The note is held to four lines with an ellipsis, with the whole text in its tooltip.

Before	After

5. A chat error's note can be closed, and closes by itself (needs 4)

After a failed chat turn the note stayed up with nothing to close it, one tab away from the chat its own advice points to ("pick another one above the chat box").

• The note gets the finished card's "OK" button, which goes back to the chat.
• Left alone for ten seconds, it folds the island. It does not go back to the chat by itself: the chat takes the keyboard, and nobody asked for that.
• A note with its own timing (a failed drop, gone after 2.4 s) is unchanged: no button, no second timer.

On the real app: OK lands in the chat. With the island's own auto-close set to 120 s, so that only the note's timer could act, the note was still up after 6.5 s and the island was folded after 12 s.

6. The picker marks a model its provider refused

A provider's list offers models a key cannot use. Gemini lists gemini-2.5-pro to accounts it then answers 404 to, and afterwards the list still showed it like any other.

The chat now remembers the models a cloud provider answered 404 for, and the picker shows them greyed with "Not available". They can still be picked, and one that answers again loses the mark. Only a 404 marks a model: quota, rate-limit and server errors can pass, so they do not. It covers the providers in openai_compat.rs; Anthropic and the model servers are untouched.

7. A refused model stays marked between launches (needs 6)

With commit 6 alone the marks are gone after a restart. They are now kept in refused-models.json, next to the log. The file is written only when a mark is added or removed, never on an ordinary answer, and past 200 marks the oldest go.

On the real app: three Gemini models that answered 404 were still reported as refused by a fresh launch.

8. The picker shows the providers that are set up

Every cloud provider is listed, key or not, and picking one without a key leads to "No API key". Model servers already only show once connected. This treats cloud providers the same way: one shows once its key is stored, and the active one always shows, so the picker never hides where the chat goes. With nothing set up, the active provider still says its key is missing and links to Settings.

On the Mac, ModelPickerView always shows the cloud providers, so this departs from it. Drop this commit if you want the two alike. If you want it on the Mac too, I can add the same filter there.

Before (OpenAI has no key)	After

Tests

What was applied to main	npm test	tsc --noEmit
All eight	335 pass	clean
2 alone	324 pass	clean
3 alone	326 pass	clean
4 alone	324 pass	clean
4 + 5	331 pass	clean
6 alone	325 pass	clean
6 + 7	325 pass	clean
8 alone	325 pass	clean

• cargo test --workspace: 200 + 30 pass, run on the full series only.
• Commit 1 is Swift: compiled and run on a macOS runner on my fork (table above).
• By hand on Windows 11 with a release build, including the real Gemini 404 for commits 4 to 7. Every screenshot is the real app, captured the same way before and after.

Not tested

• Linux: not run. It shares the Windows code path, and CI builds it.
• Mac: no Mac app code changes, so there is nothing to run there beyond commit 1's test.
• Commit 3's animation: I measured the island's height before and after and looked at stills. I did not judge the grow and shrink motion by eye.

🤖 Generated with Claude Code