Flint v0.4.1
PROJECT › Changelog

Changelog

Notable changes in each release. The measurements behind a change are in experiments.

0.4.1 - 2026-10-02#

Added#

  • Settings → About: the version and build you're running, links to the docs, changelog, repository and issues, and the commands to update. Nothing is checked online.

0.4.0 - 2026-10-02#

Added#

  • Edit before writing: Edit on a diff card opens the model's new lines for you to correct, and Write my version writes yours instead. The model is told exactly what was written.
  • Hunk-by-hunk accept: a diff with more than one change has a checkbox on each, and only the checked changes are written. The model is told which ones weren't.
  • Cloud models: one review for a turn's edits. Their edits are staged, not written, and the next read or edit in the same turn sees them; after the reply, one card shows every changed file, with Write all and Discard all. A shell command while edits are staged isn't run, since it would see the old files, and a new message waits until the review is decided (E32).
  • Cloud models can change several places in a file in one edit_file call (edits).

Changed#

  • Diffs show each change as its own hunk instead of one block from the first change to the last.
  • An edit that removes a name the file still uses says on which lines, so a rename doesn't stop at the import (E32).
  • An edit whose old text isn't in the file is pointed at the one line that contains every part of it, and told to send that whole line, then the whole changed line. A small model had otherwise repeated a garbled version of the line until the step cap ended the turn (E30).

0.3.0 - 2026-10-02#

Added#

  • File tools in folder chats. The model reads files in the attached folder on its own (with a cloud model, only after you approve each read, since the file goes to its host), and edits or creates them through a red/green diff you approve before anything is written. Paths can't leave the folder, the shield's secret-file rule is checked on the real path after symlinks and .. are resolved, and an edit to a file that changed since it was proposed is refused (E30).
  • The folder anchor lists the folder's real file paths, not just its top level, so the model names files that exist instead of guessing: the right file in 29 of 30 answers, up from 3 (E29).
  • In a folder chat, file names the model mentions that aren't in the folder are listed under its reply, so an invented path stands out (E27). The next request tells the model too, so it checks before building on a name it made up (E28).
  • A message too big for the chat's context window is refused before it is saved, instead of failing at Ollama after the wait. The limit counts only what compaction can't remove, so older history doesn't shrink it. The composer warns when a message will push older messages out of verbatim context.
  • The agent plan card has a small flint whose face shows the run's state: working, waiting for your approval, done, or stopped.

Changed#

  • The step cap counts only steps that asked for your approval (3); reads and refused proposals are bounded by a total of 8.
  • A started agent plan card folds into one compact row.

Fixed#

  • The shield's SSH-key rule missed key names with digits, so cat ~/.ssh/id_ed25519, today's default key type, was offered for approval instead of being blocked.

0.2.1 - 2026-09-27#

Changed#

  • An agent that looks through the folder is told to search file contents with grep -rni and one short word instead of guessing file names. A command with no output says (no output), and a grep that matched nothing says so and suggests a shorter word (E26).
  • The edit button is hidden on @agent messages, which can't be edited.

0.2.0 - 2026-09-27#

Added#

  • @agent <task> splits a task over the attached folder into subtasks, each run by its own agent with a clean context window, and writes one answer from their results.
  • You approve the plan first. One model call proposes the subtasks and their files, checked in code: only files directly in the folder, never a web search from the model. The plan appears as a card where you edit the instructions, add or remove files, add a web search yourself, remove subtasks, then Run or Discard. A task that doesn't split is answered as a normal chat.
  • Agents read what they're given, or look for it. An agent with files works from them (big files keep their start and end, sized to fit its window). An agent with no files, on a model with tool support, looks through the folder with shell commands that appear in the main chat as "Agent 2 wants to run", each waiting for your Approve, Deny or Reply instead, with the same safety checks as the chat's own commands. Models without tool support, like phi3, work from their inputs only.
  • Bounded, and visible. At most "max agents" run at once (2 locally, 10 for cloud models, both changeable in Settings), an agent asks for at most "commands per agent" commands (8 by default) and stops after 5 failures in a row, and each has 5 minutes of model time (waiting for you doesn't count). Stop cancels everything. An Agents label after the context meter opens a window with each agent's task, inputs, status and full transcript, live while it runs.
  • The answer comes back to the chat. It streams in like any reply, under a "From 3 agents" block listing what each agent found, and says plainly when a part is missing. Your task, the agents' short results and the answer join the chat's history; the transcripts never do. An agent that read nothing can't give an answer, since a small model otherwise invents one.
  • Typing @ at the start of a message lists the commands (@agent, @web, @memory, @memory save, @compact) with what each does.
  • Switch an empty chat's model from the header; it locks once the chat has its first message.

Changed#

  • The context meter shows from the start of a chat, not only after the first reply.

Fixed#

  • In Docker, attaching a folder from your machine failed with "folder does not exist", since the container only saw its own files. The compose files now mount your home folder at the same path, so the folder browser starts there and attached paths are the real ones. Set FLINT_FOLDERS to one folder to expose only that.

0.1.4 - 2026-09-26#

Added#

  • The docs are published at flint.lusansapkota.com.np, from the docs/ folder, in Flint's own dark colors with a grouped page menu: a home page with a quick start (Docker first), screenshots, the changelog, and the reasoning behind the design (where the ideas come from, what was deliberately not built).
  • Set your own context window in Settings → Connection, one for local models (for a GPU with room for more than the defaults) and one for cloud models, with common sizes (32k = 32768, 1M = 1048576, ...) to pick from. Cloud models get 32k by default instead of the 4k/8k meant for small local ones, and no model is given more than it supports.
  • Ollama cloud models (:cloud) are tagged "cloud" in the model picker, Settings and the chat, with a line saying the chat is sent to ollama.com. Not being signed in to Ollama now says to run ollama signin.
  • Change a chat's attached folder with the Change link on the folder chip. The new folder replaces the old one as the chat's standing context.
  • A "Saved to memory" card where you saved a memory from a chat, which expands to show what was saved and stays after a reload.
  • Settings → Memories shows which chat each memory was saved from, with a link back to it.
  • Reply instead on a waiting shell command: tell the model what to do rather than just denying it. Your reply is an ordinary message, so @web and @memory work in it.
  • Log out from Settings too, at the bottom of its sidebar.
  • Edit your name and email, and change your password, in Settings → Account. A new email or password needs your current password, and a password change logs out your other devices.

Changed#

  • @web results carry their publish dates, in the sources card and for the model, when Brave has one for every result. The model then answers "latest version" questions with its source's date instead of presenting a month-old page as current (E21).
  • @web picks more relevant results: its re-ranking now sends nomic-embed-text the search prefixes it was trained with (E20).
  • Logging out asks "Are you sure you want to log out?" first.
  • The @web sources card spans the chat's full width too, and no longer jumps wider as it opens.
  • Shell command cards span the chat's full width, so command output like df -h shows every column instead of scrolling sideways.
  • Deleting a chat asks Yes / No instead of making you type DELETE. Typing DELETE stays for deleting a model, a memory or your account.
  • In a chat without a folder, asking the model to run or check something gets the command to run yourself and a note that attaching a folder lets it run commands, instead of made-up output (E19).

Fixed#

  • A chat whose first reply failed no longer tries to generate a title anyway, which cost a model call (quota, on a cloud model) and kept the input blocked while it ran.
  • The "waiting for the reply" dots no longer stay on forever after an instant response, such as the @compact "Nothing to compact" notice or an error Ollama returns right away.
  • A cloud model's details in Settings → Models show its context length.
  • qwen3.5 no longer fails with "System message must be at the beginning" after @memory, a later @web, or once a long chat is summarized. Its chat template allows one system message, first; Flint now resends such messages as user messages for models that refuse them (E22).
  • nomic-embed-text unloads as soon as an @web search has re-ranked its results, instead of staying in memory for Ollama's default 5 minutes beside the chat model.
  • When Ollama isn't reachable, Flint says so plainly (where it looked and what to do) instead of showing Go's raw connection error, and says it once per page instead of in every section.
  • A saved Brave Search key is no longer sent back to the browser. It was in the Settings page and in /api/me, where the eye button or the page source showed it in full. Settings now shows only its last four characters, with Replace and Remove.
  • After a reload, messages keep their @web / @memory tags as typed, and recalled memories show as "Recalled memories for …" after the message that asked, instead of the instruction text meant for the model. Collapsible headers keep their arrow on the same line.
  • A denied command card shows just "Denied", not the instructions Flint writes for the model.
  • Collapsible headers (sources, system messages) and link hovers no longer flash Pico's default blue.
  • The "Attached folder" card no longer jumps sideways while it opens and closes.

0.1.3 - 2026-09-26#

Added#

  • Delete your account from Settings. It asks for your password and, if you've set security questions, their answers, then a typed DELETE confirmation. Everything in the account is deleted for good, including attachment files, and every session is signed out.
  • Settings says when nomic-embed-text is missing and offers to pull it. It's optional: @web works without it, using Brave's ranking, and it only loads while a search runs. The Installed list marks it, and deleting it says what changes.
  • Load and unload models from Settings → Models. Loading one unloads whatever else is running first.
  • Settings → Connection lists your @web searches: the exact query sent to Brave, when, which chat, and how many this month.

Changed#

  • Settings is laid out like the chat page: sections in a sidebar (Connection, Models, Memories, Account), one at a time, instead of one long scrolling column. On phones they become tabs.
  • Thinking, sources, system messages and model details open and close with a short transition instead of snapping.

Fixed#

  • Display math no longer shows a small vertical scrollbar.
  • A model's details button shows a readable summary (family, parameters, quantization, context length, capabilities) instead of raw JSON or an empty box.

0.1.2 - 2026-09-26#

Changed#

  • A denied command no longer makes the model claim it lacks permissions. It now hears that you chose not to run that one command, and tries another way or asks you instead of guessing the output (E18).
  • The context meter shows how many older messages have been condensed into a summary, for example "100 / 4.1k · 3 condensed".
  • Hovering a chat in the sidebar shows its full title.

Fixed#

  • Clicking the chat you're already in no longer reloads the page.

0.1.1 - 2026-09-26#

Changed#

  • The tool-reasoning nudge is left off when your message says not to run commands ("without running", "don't run", ...). On the benchmark's restraint tasks this went from 0/9 to 7/9, with no loss in tool use (E17).
  • During an @web reply, the search query and its sources are shown as it happens, in a collapsible sources card.
  • The Docker quick start prints Flint's address once it's running.
  • The latest image tag only moves for normal releases, not pre-releases.

Added#

  • A contributing guide, the repository rules, and how to report a vulnerability privately.
  • How to update a Docker install to a new release.

0.1.0 - 2026-09-26#

First release.

  • Chat with local Ollama models, with accounts, SQLite persistence, markdown and offline math rendering, image attachments, a thinking toggle for models that support it, and search across chats.
  • Folder attachment, with a shell tool that only runs a command after you approve it, a shield against catastrophic commands, and precondition checks before a command is offered.
  • Context management for small windows: an explicit token budget, layered background summaries, and recovery when a prompt overflows.
  • @web search (Brave Search, your own key), @memory for facts that carry across chats, and @compact to condense a chat on demand.
  • A context meter, a scored benchmark with switchable scaffolding, and project documentation.
  • A Docker image for amd64 and arm64 on ghcr.io, run with Docker Compose against the Ollama already on your machine.