Glyph 1.1.14: 155 QA reports, reviewed one by one

This release came out of a full pass over every screen, panel and dialog in Glyph. It is almost entirely fixes, and they share a theme: things that failed quietly. A permission you thought you had granted, a save that was still running after you closed the form, a filter that emptied a column and then said there was nothing there. Each one now says what actually happened. Requests that hang no longer strand the screen you are on Reading your saved credentials could block unrelated requests, leaving settings, model loading and other ordinary actions unfinished with no explanation. Credential reads are now isolated and bounded, so a slow keychain no longer stops the rest of the app. Saving a project, an employee profile, a schedule, a ticket, a squad, an MCP server, a voice clone, a skill or a model choice can now be cancelled, and each one gives up on its own instead of spinning forever. Where a request may have reached the server before it timed out, Glyph says so rather than guessing. Closing a form now actually abandons the request behind it. Previously a cancelled save kept running, and reopening the form could show a permanently disabled button. Actions that failed silently now say so Approving or denying a tool the agent wants to run could fail to send. The prompt disabled itself anyway, so the agent stayed blocked on a permission you believed you had given. The prompt now reports the failure and lets you answer again. Cancelling a running background agent could fail without a word, leaving the agent running and holding one of your agent slots. Turning a connector on or off in the composer's + menu could revert with no message, which looked like the switch had ignored the click. Deleting a skill, an MCP server, a memory, a chat, an image or an audio file now reports a failure instead of appearing to succeed. Notes in a project workspace are stored on your computer; when that write failed, the note stayed on screen and then disappeared on the next reload. Glyph now tells you. Sign in and connectors Connector sign in could stop listening for your approval as soon as anything else on the page updated, so a sign in you completed in the browser never finished. It now keeps listening, and you can stop waiting on a browser window you closed. Signing in with Hugging Face or Google no longer starts a second attempt on top of the first. Resetting a password now asks for the reset before confirming your email, rather than the other way around. Everything reachable by keyboard The knowledge context picker, the chat chapter list, the chat history filters, the ticket filters, the agent mode menu, the image preview, the knowledge document preview and the paywall, gallery, download, GlyphNet and voice dialogs can all now be opened, moved through and dismissed from the keyboard. Escape now closes one level at a time and returns focus where you left it, instead of dropping you at the top of the page. Choosing a filter value or closing a preview no longer loses your place. Messages that were telling you the wrong thing A knowledge stack you had deleted showed as "No Context" while it was still attached to your next message. It now says the context is unavailable so you can clear it deliberately. A file whose diff could not be loaded was reported as a file with no changes. A graph node whose content failed to load showed the failure as if it were the node's own text. A column emptied by a filter said there were no issues at all. The Files tab said only "File browsing is unavailable" and stayed blank until something else happened to refresh it. It now explains why and offers to try again. Cloud usage totals now line up with the records they are drawn from. The account screen shows the version you are actually running. Typing the wrong thing no longer costs you the message Pressing Enter in the tool picker pinned whichever tool was highlighted — except when it did not, because the list you saw and the list Enter used were filtered differently. The agent browser silently ignored an address it could not open; it now tells you what went wrong. Standard text shortcuts work again in the image editor prompt. Updates and installation A failed update check now offers to try again instead of ending at an OK button, and a check started from the menu can no longer fail with no message at all. If Glyph cannot move itself to your Applications folder, it now says so instead of quietly continuing to run from the installer disk image, where ejecting the image breaks startup. Performance and packaging Replies begin appearing sooner: safe text is streamed immediately instead of being held back while Glyph waits to see whether it was the start of a stop marker. Model warmup no longer competes with the chat you are actually using. Inference processes Glyph started are now shut down when it quits. Apple Silicon builds now include the MLX native helpers and Metal kernels they need. Glyph 1.1.14 is available from fortaify.com, and installed copies update themselves.