Release Notes

What changed in each release of Codumentor, newest first.

0.5.0 — 2026-10-04

Changes since the 0.4.0 release candidates. Entries marked Enterprise describe features the Community edition does not contain.

Added

Choose the model and how hard it thinks. The conversation composer now has separate Model and Effort controls instead of Fast/Slow/Custom mode cards. Settings → Models brings together connections, your default for new conversations, background work and a read-only explanation of which model does each job. The terminal client has the same model-and-effort picker; /model and print-mode --model also select models.
Bring your own model connections. Add a personal endpoint or sign in with your ChatGPT subscription from Models settings, where deployment policy permits it. Background helpers can use a lighter model, but the privacy rule prevents automatically moving work to a less private connection.
A conversation list that tells you what is happening. Rows show the current request, live activity and a checked outcome in the conversation's language, rather than just a title. Sticky headlines, changed-file counts, running-agent counts, elapsed time, sharing markers and a recent-headline trail make active work easier to follow. A user-scoped live feed updates the list without waiting for the next routine refresh; pinned rows get useful summaries too.
Manage your conversations by asking. The assistant can find your own conversations by workspace, date, pinned/shared status or unread state, and preview a batch operation before asking approval to pin, unpin, rename or mark them read/unread. Discovery results include how recently each conversation was active.
Live voice conversations. An optional realtime speech-to-speech engine sits alongside Classic voice mode. Switch between Classic / Live when both are configured: Live supports natural turn-taking, interruptions and steering while the agent works, and talks about the finished answer while the full text stays on screen. It requires a configured OpenAI Realtime connection; spoken tool approvals remain opt-in.
Choose your dictation language. Select a language or leave it on automatic detection, with the detected language shown in the voice interface.
Choose the working folder in the browser. Community user-home installs can pick a project from the header or welcome screen, remember recent folders, or chat with No folder. Switching restarts the app after confirmation if work is running; conversations from another folder remain readable but must be continued in their own folder. Project configuration asks for trust in the browser before the restart.
Restart from the Admin page. Admin → Server, and Restart now after a configuration reload, let administrators restart without shell access. Confirmation lists running work: agent turns resume, approval waits survive, and in-flight ingestion and scheduled runs do not resume. Connected clients reconnect automatically.
Choose databases per conversation. Tick the connections the assistant may use on the welcome screen, or choose none to turn database access off. The selection is enforced on database tool calls as well as in the prompt; a banner shows when access is narrowed.
Test approval rules before relying on them. The rule dialog explains what the server actually allows, shows uncovered parts of a command and saves rules atomically before rechecking the request. Web settings gains Test a command and diagnostics for conflicting or ineffective rules; the terminal approval editor also explains coverage and rule warnings.
Know exactly what you are running. Settings → About shows the version, build and flavor, links to the bilingual release notes and copies build information for support. The menu footer shows the edition and version; --version and build-info identify the actual build, including release-candidate and development status.
Enterprise — Save branch selections. Name, restore, rename and delete multi-repository branch-selection presets in the branch picker.
Enterprise — Slack context, not just the last message. The assistant receives the surrounding thread or direct-message history and can read more channel or thread context on demand, subject to access checks and bounded history.

Improved

Configuration changes need one clear confirmation. When you ask for a change, the configuration assistant proposes it and uses the approval dialog as the single confirmation; ambiguous requests, warnings and edits to shared bases still prompt a discussion. It can identify which loaded file sets a value, edit an extends: base, patch individual named list entries and remove entries without replacing the whole list. Diffs mask secrets, masked values survive edits, old key spellings are updated per file, and the proposal correctly says when a restart is required, including plugin changes.
A smoother Community first start. Bare Community launches open the web app; setup offers a provider menu, defaults to DeepSeek V4.1 Flash and creates a persistent signing key and a copyable administrator password for user-home installs. The English/Hungarian personal welcome page appears before login. Personal defaults avoid automatic vector indexing on first start and enable deferred tool discovery so a larger toolkit need not fill every prompt.
Voice follows the conversation better. Classic mode can answer spoken follow-up questions about the agent's reply instead of repeating it, resumes promptly after an interruption and filters recogniser hallucinations and playback echoes. Revised answers are not read from the beginning again, and failed utterances, rejected language hints and missing-model errors no longer silently strand the voice session.
Navigate and resume more comfortably. Keyboard navigation and Ctrl/Cmd+Shift+←/→ start or switch conversations; swiping past the last conversation starts a new one. Reopened chats land on unread content when there is any, otherwise restore your position. Sidebar titles are available on hover, collapsed thinking shows the latest signpost, and completed subagent cards show wall-clock duration. Dialogs keep focus rather than losing it to the composer.
Clearer progress and partial results. Tool groups distinguish mixed success from complete success, count their outcomes and surface the first failed call when collapsed. Completed file-write cards no longer look pending; the terminal client distinguishes a budget-pressured stop from an ordinary finish.
The assistant understands its environment better. File tools accept the sandbox paths used by the shell and explain boundary failures; shell failures give relevant network/interpreter hints. Goal workers receive the tools they were granted, subagents retain their own tools, and text-only or unattended channels no longer receive browser-only rendering instructions. Saved documents report whether a download was shown or a file was saved.
More complete bilingual help. Guides now cover mature features including exports, sharing, attachments, mentions, spreadsheets, presentations, scheduled runs and the development server.
Enterprise — A more informative push workflow. The dialog explains diverged-remote choices, shows in-flight operations and changed-line counts, and marks when the remote branch has moved — even when there is nothing to push. It offers workspace synchronisation, prevents pushing during a sync, keeps edited-field focus across refreshes and places a push marker in the conversation, visible in focus mode. Generated session-branch names are shorter while existing branches remain reusable.

Fixed

Dropped model streams recover without replaying tools. Model transport failures retry the model call rather than already completed work. Codex handles dropped or stalled streams, incomplete streamed tool arguments, usage-cap errors and provider-required effort corrections more reliably; unusable calls report an error rather than disappearing. Subscription calls keep their prompt-cache affinity.
Conversation status stays truthful. Approval-stopped, cancelled and failed work no longer gets a success outcome; outcome claims are checked against execution and test evidence. Finishing or cancelling a parent closes its subagent turns. Event sequence assignment and visibility are atomic, and a completed answer wins over an older partial buffer when reconnecting mid-stream.
Security boundaries hold beyond tool calls. Sessions are tied to the configured instance issuer so a sibling instance's session cannot be reused. External-agent bridge directories are isolated per operating-system user, and file-boundary checks close symlink escapes. Enterprise — restricted users no longer receive hidden repository skills, instructions, permalinks or organisation-memory content through plugin hooks or the memory interface; memory storage respects the same access policy.
Ingestion remains responsive and reports empty results honestly. Re-ingestion's CPU-heavy work runs in a child process, default code filters include .tsx and .jsx, and an empty ingest is reported as a warning rather than plain success.
Community and Windows startup are more reliable. Model profiles and Office editors load in Community without the sandbox plugin, packaged Word/PowerPoint dependencies are available, and Windows sandbox-path handling preserves absolute paths and separators. A taken listen port produces a useful explanation, Windows keeps startup errors visible, and supervised restarts preserve the exact launch command without false port-conflict reports.
Goals stay in durable storage. Goal setup and import use the same persistent task database as the scheduler, and refuse an ephemeral store instead of accepting a goal that disappears after shutdown.
Built-in plugins load across source and packaged installs. Configuration can use either supported module-path spelling without silently losing plugins when moving from a checkout to a binary.
Configuration and interface details. config and status can display a configuration that the current edition cannot start; About build information appears after a fresh login; recalled-conversation cards show their snippets and use a preview when the title is missing; folder trust prompts stay above the folder browser; and model/effort selections remain visibly selected on hover.
Enterprise — Branch and push correctness. Pending changes and push targets are anchored to the branch's fork point, the default branch is never stored as a conversation choice or checked out by the branch-switching plugin, and subagents cannot half-switch their parent's workspace. Remote-only branches resolve from the remote reference, working diffs show this conversation's changes rather than unrelated checkout dirt, and oversized diffs return bounded file summaries instead of freezing the browser. The working-diff handler and push plugin load correctly again.

0.4.0 — 2026-09-21

Entries marked Enterprise describe features the Community edition does not
contain.

Added

Formulas are typeset. LaTeX in an answer now renders as mathematics in the transcript instead of showing up as source.
Pinned conversations have a place of their own. Pin the work you are actually on: pinned conversations get their own section in the sidebar, which you can resize and collapse, and they still appear in the full list so none of them quietly disappears.
Swipe between conversations. On a phone or tablet, swipe left and right to move from one conversation to the next, and swipe a row sideways to uncover its actions.
A conversation list that gets out of the way. On a narrow desktop window it collapses to a strip you can open when you need it. When conversations re-sort they glide rather than jump, so the one you are reading stays findable, and a conversation you come back to opens where you left it — a freshly loaded one at the start of the last answer, not somewhere in the middle of it.
Keep the screen on. A new setting stops your device dimming or sleeping while the assistant is working.
A pause for dictation. Hands-free voice input now has a think-pause toggle for collecting your thoughts mid-sentence, and a visible countdown before the message is sent.
Images with your own subscription. If you connect a ChatGPT/Codex subscription instead of an API key, you can now send images to it.
Choose the shell. Administrators can set which interpreter shell commands run under (tools.shell_interpreter).
A larger Community edition, and a new name for the other one. The free Community build now also ships the developer subagent, goals, database access, Word/Excel/PowerPoint editing and external coding agents. The full build is now called Codumentor Enterprise.
Windows. Plugin interfaces build correctly on Windows, the terminal client's remote-control channel works there, and a batch of portability fixes came with them.
Enterprise — Pushing remembers what happened. The push dialog keeps a history of every push with the commits each one carried, the commit message is editable and survives a reload, and Codumentor now asks the remote what it will accept before pushing — offering a branch that will land instead of failing on a diverged one. When syncing your isolated workspace with the remote hits conflicts, the assistant can resolve them for you.
Enterprise — You can see what memory recalled. Knowledge the assistant pulled in on its own now appears in the transcript as cards, private to you, and you can hide them without switching recall off. A new recall engine asks a model which of your organisation's knowledge is relevant instead of relying on similarity scores.

Improved

Approvals you can actually judge. The prompt you meet most was rebuilt: shell commands are pretty-printed and syntax-highlighted with the risky parts marked, JSON details are highlighted too, changes to configuration or to a ticket show the change itself rather than raw arguments, and a long request scrolls, folds into disclosures, states the current mode and can be stopped.
Faster to load, lighter in the browser. Pages are served pre-compressed, a plugin's interface is fetched only when something needs it, and plugin bundles are now built minified against one shared toolkit.
Turns start sooner. Preparing the workspace for a turn moved off the critical path and now costs a fixed couple of steps however many repositories you have, and re-indexing after a push waits instead of competing with a running turn.
Long conversations hold together for longer. Compaction now remembers what it already summarised and reuses it, so it re-summarises far less of a long conversation.
A calmer, more consistent look. The web interface follows one shared visual language — accent colours that mean something, a single type ramp, status cards outlined rather than filled.
Smaller touches. The final answer's thinking folds into the work chip in focus mode, and focus sits ahead of Share in the header; Auto mode gets more room per turn (up to 100 automatic approvals, a 12-second judge timeout); new scheduled-run counts moved onto the Auto-runs tab; repository-removal cleanup options start ticked; the "needed by plugins" secrets section hides itself once everything is set; and the terminal client's colours improved.
Clearer words for what this is. The welcome screen and positioning copy, in English and Hungarian, now describe Codumentor as the Intelligent Work Environment you orchestrate.
Enterprise — A context budget you can read. The context visualiser now shows one card per turn and one bar you can actually interpret.

Fixed

Uploads no longer corrupt binary files, and an attachment may now be up to 50 MB.
A conversation cut off by a restart resumes instead of sitting on a half-finished turn, and opening a conversation restates whether its turn is still running — so a turn that ended while you were looking elsewhere no longer looks stuck.
A subagent waiting for your approval is no longer cancelled by the background sweep that clears abandoned work.
On mobile, the conversation column no longer scrolls sideways, swipe-to-close works again on conversation rows, the "…" menu on a chat row opens, and menus stay on screen instead of running off the edge.
The conversation list no longer flickers when it refreshes.
Tool cards read properly again: they show the friendly tool name, stop wrapping one letter per line, keep a failed shell command on one line when collapsed, and no longer squash a subagent's prompt into a vertical strip.
Diagrams and diffs. Inline Mermaid diagrams render at their natural size instead of being squeezed to the column, each side of a split diff scrolls independently, and a file you edited in the overlay is compared against its own baseline.
Full-screen previews scroll to content that was clipped, and their status layer no longer swallows clicks meant for the app.
Files you attach are read from where the assistant is told they are, dictation no longer leaves its marker inside your question, and @-mentions rank matches that start a word above fragments buried in the middle of one.
Security. A secret stored by one user can no longer reach another user's tool environment, and a configuration file whose authentication block was dropped now says so instead of quietly starting without it. Enterprise — the sandbox no longer hands the agent the host's user list and DNS search domains.
Enterprise — Memory hint cards in the transcript are no longer blank.
And smaller ones. Goals refuse a rotation whose archive destination git ignores, rather than losing the history; a turn that runs away thinking is reported as an error instead of ending in silence; image results no longer fail against a Codex subscription; permalinks stop re-wrapping links that were already links; and the admin feedback summary no longer counts rows in an empty table.

0.3.5 — 2026-09-08

Added

Thinking modes. Choose how hard the model thinks, per conversation: Fast, Slow, or a custom effort, translated to whatever each model actually supports. The composer shows the current setting as a chip, the terminal client has /mode and /effort, and a model that rejects a level is demoted visibly instead of failing the turn.
Auto mode. Pick an approval mode per conversation — on every surface, web and terminal alike. In Auto, a judge model checks each gated call against what you actually asked for before approving it, the transcript records which calls Auto approved, and when Auto interrupts you the dialog explains why. Your permission policy is yours alone: it is per user, and an operator's escalation rules bind Auto everywhere.
MCP without the curl. Connecting a server is a UI action now, each user authorizes a server in their own browser, the tool list is discovered live from the server, and interactive MCP apps can render inside the conversation.
Personal recall. The assistant can search and reread your own past conversations, and with auto-recall on it pulls relevant ones in at the start of a turn — a card in the transcript links back to the conversation it came from.
A community edition. A separate build for personal use, containing a curated set of plugins: the developer subagent, goals, database access, Word/Excel/PowerPoint editing, external coding agents and more, with the personal and minimal flavors and the web interface as its front door.
Instance home and configuration layers. A user-level install has a real home of its own; configuration composes in layers — bundled defaults as the floor, instance and user files on top — with the provenance of every value visible on status.
A sandboxed preview for web apps the assistant builds: run them in an isolated frame, not in your browser's context.
Tasks can fire from events, not only from the clock.
The conversation list order is yours. Sort however you like, with needs-attention the default and a dot preview of what each conversation is waiting on; a three-state filter and a badge in the sidebar find the auto-runs; and "mark as unread" is back in the row menu.
The composer row is open. Plugins can place their own controls in the composer's action row, which now has a state line above the textarea and knows how to overflow gracefully.
The terminal client grows up. It can be driven over an out-of-process control socket, its approvals toggle is always visible (F4), and when it exits it says why.

Improved

Context you can see. A /context command shows what the model is actually carrying, and the compaction that keeps a long conversation alive is now a readable seam in the transcript: a lede you can read without opening it, what each round changed, and which tool results it dropped.
Secrets stay out of the logs. Known secret values are redacted from every log line, and the SVN password no longer appears in the host's process table.
Pushes are authored as the repository's own git identity, not as whoever happens to be running the instance.
Plugins are steadier under load. One instance per process, per-session state where state is per-session, and a disabled plugin writes nothing at startup.

Fixed

Your draft is saved as you type, not only when you switch away — a crash or a refresh no longer eats what you were writing.
The file browser reopens where the agent last wrote, and committed files stay visible after the workspace overlay is swept.
Conversation search follows the list order you chose, not a raw relevance rank.
A read-only database connection is actually read-only, and a rejected read-only session fails closed instead of limping along.
The file watcher watches the repositories, not your whole home directory — and an exhausted watch budget skips instead of failing to start a conversation.
MCP tool definitions reach the model once, not twice.
A shell command timeout can actually fire on the long-running commands it exists to stop.
The terminal client honors the card-visibility toggles, and a stray Ctrl+C or Enter no longer kills it or arms its quit dialog.
A failed model call is reported as a failed run, not dressed up as an answer.
Compaction leaves headroom and never cuts a tool result loose mid-stream.

0.3.4 — 2026-08-18

Added

Manage repositories from the admin panel. Add a repository, edit its settings, remove it and its leftovers, reorder the list, flip SVN and share switches, and see a health readout — all from the UI, without editing YAML or restarting.
PowerPoint editing grows up. Deck creation is unblocked end to end: content and style tools, tolerant arguments, and a save path that refuses to lose your work.
A Minimal flavor. A stripped-down configuration with four tools and a one-line prompt, for people who want the assistant without the apparatus.
Finer-grained plugin permissions. A plugin can say which of its tools write, a permission gate can name who is left out (not only who is let in) and can target a single subagent, and subagents may read a ticket or a pull request without being able to write one.
Local MCP servers run in their own sandbox, one per server.
Calendar event reminders notify you from your connected calendar.

Improved

The push dialog is readable. No JSON dumps, no hunting for a close button — the diff gets the room it needs, and highlights survive scrolling.
Better Windows support. Document workers start reliably, bundled configurations load as UTF-8, and office paths are normalized.

Fixed

An aborted turn no longer leaves a tool card spinning forever in the transcript.
Overlay reads cannot escape to the host when workspace isolation is on.
Subscription authentication reaches a sandboxed external coding agent instead of dying at the sandbox boundary.
The settings screen reads back what you saved — the auto-approve-all toggle no longer lies on reload.
A broken static-key configuration closes the API, not the whole instance.
Ingestion respects the subtrees a repository declares as its contribution.
A turn cannot end on text that is not an answer.

0.3.3 — 2026-08-01

Added

Voice. Push-to-talk dictation, spoken replies, and a full voice conversation mode: the assistant listens while you talk, steers mid-sentence, filters its own echo, and can take approvals by voice.
Notifications. The browser tells you when a turn finishes, even in another tab: a turn-completion notification, and an unread badge on the tab and favicon for conversations running in the background.
Attachments that stick around. Files you attach are staged on disk where the assistant can save them, they outlive the turn they arrived in, and the staging area sweeps itself by age and per-conversation caps.
Enterprise — Access profiles. An administrator can give each user their own ceiling on what the assistant may do: chat-only profiles, per-user subsets of the repositories, and — for the restricted ones — a mediated question flow, where the assistant's answer is parked for a reviewer and released only after approval.
Sharper memory. Recall now measures relevance against the corpus the agent actually reads, ranks term hints by specificity, and caps how much it injects — and the memory curator runs in its own sandbox, so it cannot act on the conversation transcript. Retired notes are pull-only: they are remembered, never pushed into a turn.

Improved

Security hardening. The instance authenticates by default and refuses to serve open on the network; login attempts are throttled and sessions are revocable; the refresh token no longer sits in the browser's local storage; the HTTP surface is safe without a hardened proxy in front; and the logout endpoint can no longer be abused by anyone who can reach the port.
The push changes-dialog is scoped to the conversation, so it shows the changes this conversation made, not everything in the workspace.

Fixed

The conversation list holds its state across refreshes — no more dropped, reverted or leaked entries.
Goal-tick commits are recoverable: a crash mid-tick no longer loses the run's output, and a stalled tick flushes what it has.
Spoken language is established once and carried onto short follow-up utterances, so a quick "yes" is not re-detected as a different language.
The unread badge clears when you tab back to the conversation that was running.
Windows line endings no longer break the memory notes' commits.
The browser is no longer started with its own sandbox disabled on a guess.

0.3.2 — 2026-07-23

Added

Enterprise — Pull-request awareness. The assistant can read the pull requests of your Gitea repositories — what changed, what was reviewed, what the comments say — and, when you ask it to, it posts its review straight to the PR.
Goals grow up. Scheduled runs get deterministic titles (no model call per tick); pausing and stopping a goal is enforced by the system rather than hoped for in the prompt; each run can select its own model; new tools let you add or remove a goal's workers by name; a copied goal tree can be re-registered as a live goal; and a goal whose runs keep failing is counted toward auto-disable.
A headless mode for scripts. codumentor -p now behaves like a proper command-line tool: deny-by-default permissions for unattended use, stable session identity in the output, and clean help text.
Much faster startup. First paint dropped from about 2.25 seconds to 0.45, and --help from 1.9 seconds to 0.24.

Fixed

Memory notes are written to the repository you configured, not to a duplicate under the repositories directory.

0.3.1 — 2026-07-17

Fixed

The knowledge-base search tool stays available even when the vector collection is empty, instead of disappearing from the assistant's toolkit.
Plain folders work as repositories, not just git repositories — a folder without version control is accepted the same way.
The isolated workspace self-heals: stale worktree registrations left behind by an interrupted run are cleaned up automatically, so a broken registration no longer blocks the next run.
The configuration assistant handles plugin overrides when it edits your configuration, instead of losing them.

0.3.0 — 2026-07-14

Added

A complete push workflow. The push button now shows what is uncommitted and what is unpushed, you can review the diff of the pending commits before pushing, and Enter in the passphrase field completes the push.
File previews and diffs. Links to files in the transcript open a rendered preview, and changed files show a side-by-side diff by default.
The configuration assistant grows up. It can now show you the plugin catalog, it understands and validates the configuration schemas of the plugins you have installed, and it preserves your YAML comments when it writes.
Browse Drive folders. Drive search accepts a folder, so you can browse a folder instead of searching the whole drive.
A prompt inspector. An administrator can open any model call in the transcript and read the exact prompt that went out, independently scrollable, with the memory-recall decisions traced.
The marketing site is served at /marketing next to the app.

Improved

Prompt assembly was substantially restructured. The system prompt is now a lean core, and the situational guidance — repository inventory, custom instructions, edition-specific notes, log-collection procedures — arrives as separate context notes placed before your question. What the model sees is easier to follow, and the prompt inspector above makes it easy to verify.
Workspace isolation is stronger. On Linux the assistant's file and shell work can run in a sandbox, and it now gains a seccomp-BPF syscall filter on top of the namespace isolation; when the isolation component is installed, subprocesses are sandboxed by default, and network isolation is on by default.
Agentic memory got sharper. The assistant now reviews its own work mid-turn for things worth remembering, ranks candidate memories with multi-probe retrieval, and writes new notes with aliases so similar recollections consolidate instead of piling up.

Fixed

Security. A path in the web interface could be crafted to read arbitrary files from the server; that hole is closed.
A phantom "thinking" block could appear at the end of a conversation when two processes shared a sequence counter.
Permalinks are now corrected per segment, so a single broken link no longer drags the whole turn's links with it.
Scheduled runs no longer crowd out the conversation list — they are hidden from it on the server side.
Focus mode keeps its live panel when a turn resumes after an approval.
A whole-text revision that carries only the final segment is no longer truncated in the snapshot.
The collapsed work chip stays visible when you scroll to the last response.
Packaged one-file builds now ship their prompt templates, so a fresh install no longer misses them.