[HarryCordewener]

Requested by Grave · project thread

Two changes from the same investigation.

Crashes reach client-diagnostics-*.log

A crash in the window left nothing in client-diagnostics-*.log. The only record was in Windows Event Viewer and %LOCALAPPDATA%\CrashDumps. Three gaps are closed, in RunLive (shared by both hosts) and in SharpMUTermApp:

• SharpConsoleUI's own log is now forwarded, Warning and up (ClientDiagnostics.Forward). The framework catches whatever its main loop and queued UI actions (EnqueueOnUIThread, behind our OnUiThread) throw, then logs it to a LogService that writes nowhere unless SHARPCONSOLEUI_DEBUG_LOG is set. When the loop dies it returns 1 and prints the stack to a console, which the window doesn't have.
• Unhandled and unobserved exceptions are recorded while RunLive runs (ClientDiagnostics.RecordUnhandled). This covers AppDomain.UnhandledException, which also catches an exception escaping Avalonia's UI thread, and TaskScheduler.UnobservedTaskException. Neither is marked handled. The file sink is unbuffered, so the line is on disk before the process ends.
• An exception escaping Run is logged before it reaches MainWindow's catch, which only writes to Console.Error. A non-zero exit code is logged too.

The crashes that prompted this were native faults on one machine and would not have reached any of these hooks. The same WebP decodes identically 680 times on Linux x64 with the same ImageSharp and runtime, and other programs on that machine crash with the same codes.

Each MXP picture is fetched and decoded once per session

InlineImageHost's cache held finished rows, keyed by presentation, pane width, box and URL. A room picture re-sent on every look was therefore fetched and decoded again after any resize, split or sidebar change. Looks that arrived before the first fetch returned each started their own fetch. Decoding is the expensive half: a desktop-sized WebP costs tens of milliseconds of SIMD work.

• Decoded pictures are now kept per URL, with a 64 MB pixel budget, least recently used evicted first. A picture larger than the whole budget is used once and not kept.
• A fetch already under way is shared.
• A failed fetch is not remembered, so the next mention tries again.
• The row cache stays, so a repeat at the same size is still not rescaled or re-transmitted under Kitty.

Tests

• CrashLoggingTests (Tui): framework errors are forwarded with their exception, framework debug/info output is not, nothing is forwarded after disposal, and a live app forwards its framework's log.
• InlineImageCacheTests (Tui): a second width doesn't refetch, three looks during one fetch share it, a failure is retried, and the budget keeps or drops a picture correctly.
• Tui (1965) and Avalonia (191) suites are green. The solution builds with no warnings.

🤖 Generated with Claude Code

https://claude.ai/code/session_011SqUn5K4cB6cU9zp3HmVmF

[coderabbitai]

Caution

Review failed

The pull request is closed.

Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}