FLASH CLI v0.4.4

What's Changed

• Add an interactive model picker: bare /model opens an arrow-key list of the models on this machine, typing filters it, and a name that matches nothing is offered as a download with a progress bar. /model <name> still switches straight to one by @Natuworkguy
• Add @ file mentions: type @ anywhere in a message to pick a file from a dropdown, use / to walk into a directory, and the model reads whatever you pointed at. Dot-entries stay hidden until you type the leading dot by @Natuworkguy
• Add a fetch tool that reads a web page as text with no browser and no vision model. It returns the final URL, title, description, and body, capped at 20,000 characters, and hands JSON, plain text, and CSV back untouched. A page drawn by JavaScript comes back with a note that tells the model to screenshot it instead by @Natuworkguy
• Teach the model when to read and when to look: fetch for what a page says, screenshot for how it looks. Search results are now fetched before being relied on, rather than answered from a snippet by @Natuworkguy
• Add SHOW_STATS: a dim line under every reply with the tokens the turn used, how long it took, the generation rate, and how full the context got. Tool rounds are summed into one line, and SHOW_STATS=0 hides it by @Natuworkguy
• Add NUM_CTX: set a token count, or max for the largest window the model's architecture supports. Flash says once when the setting did nothing and why, and once per model when no window is pinned and Ollama's default applies, with the ceiling and how to raise it by @Natuworkguy
• Fix voice mode speaking replies to typed messages. Only a turn you spoke gets a spoken reply; a typed turn is answered in writing and hands the prompt back by @Natuworkguy
• Name the file path in the write tool's result line by @Natuworkguy
• Add httpx to the requirements. install.sh and install.ps1 pull it in; a manual install needs pip install -r requirements.txt again by @Natuworkguy

Flash Onyx 2.4

• Put the steps before a computed answer: a number or verdict it still has to work out comes after the arithmetic and a one-line check, instead of being guessed first and justified after by @Natuworkguy
• Stop fabricated tool claims: "I ran", "I checked", and "the output shows" are only said when a tool call actually happened this session, and the honest form otherwise is "from memory" by @Natuworkguy
• Keep shell calls clean: a tool call contains only the command, never a comment or a narration by @Natuworkguy
• Every DONE line in a multi-part report carries the thing that proves it, quoted: the test output, the log line, the grep match by @Natuworkguy
• Cut the filler harder: no acknowledgement openers like "Perfect!" or "Got it!", no scare-quoting the user's words back at them, no ** bold in a chat reply by @Natuworkguy
• Sound more like a person across a whole session: vary the shape of replies from turn to turn, let paragraphs be lopsided, and repeat a word rather than reaching for a synonym by @Natuworkguy
• Write files to disk through the write tool and hand back the path, never a page pasted into chat as a code block by @Natuworkguy

Tip

Use Natuworkguy/flash-onyx-2.4:31b-cloudbase to try the cloud build, or flash-onyx-2.4:12b on consumer hardware.

Full Changelog: v0.4.3...v0.4.4