Windows AI that stops answering: say so sooner, log it; Phi Silica's line

[kellylford]

On 17 Oct, after Windows' AI components (WindowsWorkload.*) updated without a restart, Quick Describe returned no descriptions: the helper said Ready, loading the models worked, but describing and reading text both hung until the PC was restarted. The log showed nothing, and a request could wait minutes before saying anything.

Changes

• Giving up sooner. A request to the helper times out after 90 seconds (was 3 minutes; a description takes about 3 seconds, 6 to 18 the first time), counted from when it reaches the helper, not while it waits its turn. It says: "Windows' AI didn't answer. If this keeps happening, restarting the PC may help."
• A request waiting behind one that timed out gives up at once, rather than waiting as long again, so describing's text comes from the built-in OCR straight away. This includes a hung warm-up at start (each warm-up request has 60 seconds). The next request after that tries a fresh helper. A request given up by Escape or a newer one doesn't count, and neither does a warm-up that only waited behind the person's own slow picture.
• No model downloads inside a request's time limit. At start, Quick Describe gets the text recognizer ready too, if Windows says it needs it, by having a separate helper read a blank picture (up to 15 minutes, like --prepare). Reads use the built-in OCR meanwhile. A description isn't asked for while its model isn't ready; the result says "Windows is still getting its image description model ready. The first time, this can take a few minutes; try again then." To know the recognizer's state, the readiness check now uses the helper's --details (it was --check).
• Logging. A warm-up that hangs is now logged ("Warming up didn't work: ... timeout"). Describing logs when Windows AI can't be used, once per state rather than with every picture. Getting the text recognizer ready is logged.
• Check Windows AI's Phi Silica line. Its ready state only says the model is on the PC; using it is refused (Access is denied, tried on this PC). It now reads "Phi Silica, Windows' language model, is on this PC. It isn't supported in Quick Describe yet." when Ready, and "Phi Silica, Windows' language model, isn't supported in Quick Describe yet." otherwise.

An independent review found a flaky test, two holes in the fast-fail around the warm-up, and the download risk; all are fixed in the second commit.

Tests

dotnet test --solution QuickDescribe.sln: 320 passed, 2 skipped (the screen trials). HelperConnectionTests alone passed 15 runs out of 15, and each timing test passed 10 of 10 run on its own. New tests: a request behind a timed-out one, or behind a hung warm-up, gives up at once without starting a helper; a warm-up that waits its turn past its limit isn't a timeout; a request behind one given up still goes ahead; a model still getting ready isn't asked to describe and the text still comes; when the recognizer counts as needing getting ready; the Phi Silica lines.

Checked on Kelly's PC: the installed helper's --read on a blank picture, as used to get the recognizer ready, exits at once with no text. Getting a model ready couldn't be tried here, as both are already ready.

Screen reader check

Only text changes; nothing about focus or announcements. With JAWS, NVDA and Narrator:

1. Help > Check Windows AI. The Result box gets focus; arrow down to the Phi Silica line and check it reads "is on this PC. It isn't supported in Quick Describe yet."

🤖 Generated with Claude Code