The terminal.
Every release.

Improved

• Memory can be turned off for one project or for all: /memory off --project, /memory off.
• Agents started during a conversation see what it has already written to memory.
• A memory row names the note by its place in the memory folder, not by its full path, and an agent's work on notes shows as memory too.
• /memory tells you where everything from the old memory was saved: ~/.nilux/memory-export.md.

Fixed

• A search in the memory folder that runs without asking leaves out hidden files, such as .env and .git.
• Memory rules also apply when the memory folder is reached through a link.
• A project's instruction file can no longer pass itself off as your own instructions.
• An agent started by another agent no longer takes its task for your words: it does not send your project to an address only the task names, or change settings, hooks or instruction files because the task asks.
• The check that keeps secrets out of memory notes and instruction files also recognizes DeepSeek and Tavily keys and a password in a PostgreSQL URL.
• Two projects with the same folder name no longer get each other's old notes: those stay in ~/.nilux/memory-export.md.
• Stopping an agent with Esc no longer leaves its row saying "Working".
• An agent that finished just before you stopped the turn keeps its answer, so its work is not run again.
• In Auto mode, an answer retried after a model outage replaces what the failed attempt showed instead of following it.
• Reopening a workflow step shows the attempt that answered it, not one that failed before it.
• After a reconnect or /resume, the model still knows what a picture link in your message showed.
• Resuming a conversation from before 1.0.32 no longer sends the model memory recalled by the old memory system.

Added

• Project memory lives in plain files you can read and edit, with MEMORY.md as its index.
• /memory shows what a project remembers and where, and turns memory on or off for it.
• MEMORY.md and instruction files can include other files with @ lines.

Improved

• Notes from the old memory move into the new memory folders; the rest is saved to ~/.nilux/memory-export.md.
• A dropped connection to the model no longer stops your turn or your agents.
• In Auto mode Nilux waits out a model outage and carries on by itself.
• A rewind also undoes what the conversation wrote to memory.
• The agent notices when an instruction file changes during a conversation.

Fixed

• A slash command typed after a space or on a new line runs as a command.
• An agent you stop shows as interrupted, and its card closes.
• Continuing a workflow reuses what already finished.
• A retried answer leaves no half-written copy in the terminal.
• A picture link in your message is processed once, not on every step.
• Calls to MCP tools no longer come back as plain text with some local models.
• A ~/ path means the same file in diffs, reads and writes.
• An MCP server that closes mid-request no longer leaves it hanging or crashes Nilux.
• The usage page names each charge by what it was for.

Improved

• Outside Auto mode, deleting a file always asks first: "Always allow" is not offered for it, and rules saved earlier do not cover it.
• /help lists every command.

Fixed

• Node.js 20 is the stated minimum, the version Nilux needs to start.
• /init writes nilux.md, so its instructions load on Linux too.
• Disconnecting an MCP server leaves it stopped, not failed.
• A compacted conversation shows its context size as soon as it is resumed.
• A background command printing without line breaks no longer slows the session.
• A background command whose log cannot be saved no longer closes Nilux.
• On Windows, stopping a command also stops a server it started through a script or a pipe.
• nilux mcp add --help shows an example that works.
• /config, /model, /effort and the sign-in screen describe what they do.

Improved

• A session starts through a flaky network: it keeps trying instead of failing at the first dropped connection.
• A connection that stops answering is replaced on its own, and reconnecting during a reply adds no line to the conversation.
• A reconnect in the middle of a long turn picks up without sending finished work again.
• /compact names the background work still running, and the agent keeps knowing about it after the compaction.
• With a context limit set, /context shows each part as a share of your limit, and what lies past it on one line.

Fixed

• "Always allow" on a command that starts with a variable, like NODE_ENV=test npm test, saves the command itself.
• A deletion behind sudo, env, nohup or a variable is recognised as a deletion.
• Two sessions saving a setting or a permission rule at the same moment no longer lose one of them.
• A background command or picture started by an agent reports back to that agent, not to your conversation.
• A background task's output is no longer reported as a change you made.
• Agents started right after an automatic compaction report back as usual.
• A custom agent saved with no tools gets none, instead of all of them.
• A custom agent's description with --- in it, or on several lines, reads back as written.
• On a Mac without developer tools, Nilux no longer brings up the prompt to install them.
• A file that ends with a newline reports the right number of lines.

Added

• /context shows what fills the context window.
• /add-dir lets the agent work in a folder outside the project; /add-dir alone lists them.
• Sessions are named automatically (switch it in /config).
• A page the agent makes for you to look at — a mockup, a chart, a report — gets an "open in browser" link.

Improved

• Auto mode never stops for a dialog: each action is allowed or refused with its reason.
• In auto mode, a small deletion in the project runs once a copy is kept for /rewind.
• In auto mode, the agent can undo its own edits with git.
• The low-balance warning comes at the threshold you set on the billing page.
• Each session keeps its own model across /resume and reconnects.
• Pictures stay with the conversation and come back after a reconnect or /resume.
• A queued message keeps its pictures.
• Workflow runs and their notes survive a restart and /resume.
• The model list marks a model that is not answering.
• Work sent from another session runs under a limit, and its approvals say whose request it serves.
• A folder's own permission rules apply only as you accepted them.
• "Always allow" is offered only where it saves a rule.
• The package lists the licences of everything it is built from.

Fixed

• The status line keeps the last digit of the session cost visible.
• Finished background results are no longer sent again after a reconnect.
• /compact after /resume keeps the conversation's memory and agents.
• A message sent while connecting or during a server restart is no longer lost.
• You stay signed in when several Nilux sessions run at once, and a network hiccup no longer signs you out.
• A git command or a tool call that never answers no longer holds the turn.
• /resume says "just now" for a session under a minute old.
• File tools no longer follow a link out of the scratchpad.

Added

• /workflow: a team of agents works through one task in phases, after you approve the plan.
• Plan a workflow with the agent in an ordinary conversation before starting it.
• A board for every run: agents, phases, live activity and transcripts.
• A Browser plugin in /plugins: the agent opens pages, clicks, fills forms and reads the screen.
• "My Chrome" mode drives your own Chrome through the Playwright Extension.

Improved

• Slow MCP calls move to the background, like Bash commands.
• Bash rows can show diffs of the files a command changed.
• Subagents and workflow agents stream their answers, so long replies no longer break off.
• A subagent's web searches and page fetches run in parallel.
• Each agent gets its own session on MCP servers that support it.

Fixed

• Esc stops only the current turn: background agents and tasks keep running and still report back.
• Rows no longer shift or overlap on Windows after a full redraw.
• Reasoning before a tool call is restored in order after /resume.
• The live token counter no longer double-counts.
• No more empty rows between tool calls.
• Search over past sessions no longer returns messages you cancelled.

Added

• Hooks: run your own commands, HTTP calls or MCP tools on agent events. /hooks shows what is loaded.
• /plugins installs TypeScript and Python language servers: every edit is checked locally, and the agent sees the errors it introduced.
• Response streaming modes and a context budget setting in /config.

Improved

• Auto-compact is on by default for new installations.
• After a reconnect or an update, running work resumes and an idle conversation stays idle.
• Subagents run until the task is done, with no fixed turn or time limit.
• A more accurate cache countdown in the status bar.
• Agent lists stay readable in small terminals.
• The response language setting is followed more reliably.

Fixed

• A finished background command is reported once.
• Prompt suggestions no longer show up as errors.
• Messaging a finished subagent resumes it instead of starting a new one.
• Shift+Tab no longer desyncs the shown and the active permission mode.
• An interrupted turn keeps the output already produced.
• Repetitive reasoning is stopped early.
• MCP tools with unusual schemas no longer block agents.
• Bash on Windows no longer creates a nul file.

Added

• Lower Nilux Standard prices, about a third less on a typical conversation.
• Give a running session a task from the terminal.
• Updates are offered at startup instead of installing silently.
• Search reports skipped directories, and Bash reports how much output was cut.

Improved

• Delegated agents can use every tool they are given.
• Messages from another local session can be rewound to.
• Fewer early stops while the context still has room.
• The tool being written is shown while the assistant writes it.
• Nested shell commands no longer inherit the outer command's allow rule.

Fixed

• Interrupting a tool batch keeps every result paired with its call.
• Truncated tool results read the same live and after /resume.
• Failed searches and truncated results are reported as such.
• Read and Edit agree on line numbers.
• A failed tool's output survives /resume.
• Forked conversations keep their checkpoints.
• Failed or stopped delegates are reported correctly and return their partial work.
• Image analysis is billed to the right turn.
• Connection retries no longer repeat the reasoning.
• Cancelling a Skill restores the full tool set.

Added

• An agent list under the prompt: Alt+A to focus, Enter to open, Ctrl+X to stop.
• Open an agent to read its work as a chat and message it, running or finished.
• Agents survive a server update.
• The cursor dims when the terminal loses focus.

Improved

• Finished agents leave the list after a minute.
• Background agent reports arrive in full.
• A finished agent opens on its report.
• Several agents finishing together no longer trigger extra turns.

Fixed

• Approval dialogs work with the agent list open.
• A resumed agent keeps its type and tool count.
• Leaving the list no longer drops your last message into the prompt.
• Background agents no longer stay on "Initializing".
• Messaging a running agent no longer makes it re-read its whole run.

Added

• Two sessions on the same computer can ask each other questions.
• Message a running agent without stopping it.
• Git status in the bottom line: branch, changes, ahead/behind, open pull request.

Improved

• Delegated agents inherit your project instructions.
• Multi-agent runs ask for approval before they start.
• Delegated agents follow your /effort level.
• Background commands keep their output.
• A declined tool is treated as your decision, not an error.
• Long files are written in one step.
• Image reads get a heading of their own.

Fixed

• Crashed or stopped agents are no longer reported as finished.
• Agents launched together no longer swap results.
• Background agent reports lead with their findings.
• Resuming an interrupted agent works.
• An image that cannot be attached is still mentioned.

Added

• Background agents keep all their tools and ask for approval as usual. Ctrl+B also backgrounds Task agents.
• Closing the CLI ends the session.
• Agents ask before using an MCP server, and "always allow" applies to one tool.

Improved

• Command results say clearly what happened.
• Answers follow the language you write in.
• Work continues after a compact instead of starting over.
• The context percentage and the cache warning are shown together.
• Weekends are billed at the standard rate.
• A smoother loading animation.

Fixed

• A path at the start of a message is no longer read as a command, and a pasted /compact works.
• Repeated empty tool calls are stopped and not charged.
• Agents cannot overwrite files they have not read.
• Messages sent in a burst after a reconnect are handled as one turn.

Added

• Approval dialogs explain what a command does, in your language.
• CLI sign-in with Google or GitHub.

Improved

• Pairing a terminal requires signing in again.
• Cheaper web page reading on Pro.
• Images follow the conversation through /compact and /rewind.
• The language setting is validated.

Fixed

• No more calls to a non-existent Suggestion tool.
• Password prompts from commands like psql no longer hide off-screen on Windows.

Added

• The model sees pasted screenshots directly.
• Images are kept with the conversation and restored by /resume.
• Image generation is back, rebuilt.
• Choose the quality before generating, with price and time shown.
• Transparent backgrounds.
• Generate a new image in the style of an existing one.

Improved

• Size follows your request, not the quality level.
• The output format follows the file extension.
• Generation runs in the background.
• /rewind restores images too.

Fixed

• Several images in one message all reach the model.
• Switching models no longer misreports image support.
• Quoted text is readable again.

Fixed

• Tools in a batch no longer stay on "Running...".
• Retry messages show the real reason.
• Sending an image to a text-only model gives a clear error.

Fixed

• /resume and reconnects restore the conversation exactly, without lost context or a full-price re-read.
• Reasoning before a tool call survives a restart.
• Approval dialogs no longer outlive their connection.
• The right model is shown after a reconnect.

Improved

• MCP servers can write to the session scratchpad.
• The peak badge and the cold-cache warning are shown together.
• Safer clipboard cleanup.

Added

• Terminal copies paste as clean text.
• Read opens PDFs, including page images and full-page rendering.
• Read opens Word and Excel files.
• Read opens Jupyter notebooks with their outputs.
• /doctor diagnoses your setup.

Fixed

• Long silent commands no longer freeze the turn.
• A multi-line paste collapses in every input.
• No more repeated question loops.

Added

• /effort sets how deeply the model reasons: low, high or max.
• Usage in provider peak hours bills at 2x, and the status bar shows when a peak window is on or coming.
• An Auto-Compact option in /config, off by default.

Improved

• The per-turn tool limit is removed.

Added

• The model and the running conversation cost in the status bar.
• The cost includes subagents, workflows and tools, and updates in real time.
• The cost is saved with the conversation and restored by /resume.

Fixed

• A dropped connection no longer leaves "Generating…" stuck.
• Esc always returns control.
• A reconnect always rejoins your live session.

Improved

• Temporary files go to a per-session scratchpad outside the project.
• Scratchpad writes show as Scratch and need no approval.
• After a rewind, your message returns to the input box.
• The assistant is told what a rewind restored.

Fixed

• Rewind finishes in seconds instead of timing out.
• Files created after the rewind point are removed, with a recoverable copy.
• Rewind never touches binary files.
• Edit previews no longer flash into an error.
• /rewind during a response asks you to stop it first.

Improved

• The context bar matches the real limit, and a full conversation offers /compact or /clear.
• Faster and cheaper /compact.
• Approval requests wait while you are typing.

Fixed

• Queued messages and drafts are never lost.
• Rewind keeps a recoverable copy of the files it changes.
• Shift+Enter works in VS Code and similar terminals.
• Code blocks in an unknown language render as plain text.
• Generating an agent with /agents works, and the cursor lands after a pasted image.

Improved

• MCP tools fail less often, including after /compact, /resume and rewind.

Fixed

• A message sent right after sleep keeps the conversation and the model.
• Long commands and URLs wrap correctly.
• Lines ending in CJK characters or emoji no longer overflow.
• Calls to unknown tools are reported correctly.

Improved

• A multi-line paste collapses into one marker.
• Esc twice clears the prompt.

Fixed

• Typed characters no longer swap on a busy machine.
• A held backspace deletes properly.
• Arrow and function keys no longer print raw codes.
• Pastes arriving in pieces stay intact.
• A new code block no longer scrolls you to the top.
• Brief connection drops mid-reply are retried.
• No startup warning on newer Node versions.

Improved

• Context survives reconnects, sleep and server updates.
• Reconnects cost a fraction of what they did.
• Faster and cheaper /compact.
• /resume and rewind restore conversations exactly.
• Web lookups run in parallel and no longer block the turn.
• A clearer web fetch display.
• The dashboard shows the cache percentage and daily spend.

Fixed

• /resume right after launch keeps your history.
• /resume lists only conversations that can be restored.
• A new session appears in the list immediately.
• A resumed conversation keeps its message count and title.
• /rewind lists only your own messages.

Improved

• MCP tool parameter types are always resolved.

Fixed

• Closing or replacing a memory no longer hangs.
• Unsupported actions fail fast instead of hanging.

Added

• Search and read past sessions of the current project.

Improved

• Many connected MCP servers no longer flood the context: tool details load on demand.

Added

• Four permission modes, switched with Shift+Tab or /mode.
• Plan mode: a read-only plan before any change.
• An optional separate planning model in /config.
• A notification bell.
• A referral program.
• Email verification and password reset.
• Google and GitHub account linking.

Improved

• The dashboard usage chart shows real spend.
• A low-balance banner in the CLI.
• Inline command hints.
• More reliable Shift+Tab on Windows.
• Longer timeouts for heavy MCP tools.

Fixed

• Idle flicker and the approval dialog timer.

Added

• Sign in with Google and GitHub.
• The CLI signs in through your browser.
• The model notices date changes and new commits during a session.

Improved

• Stronger account and session protection.
• Changing your password signs out every device.
• A minimum password length of 8 characters.
• Cheaper long sessions through better caching.
• Faster reconnects.
• Long tool-heavy turns compact instead of failing.
• More reliable top-ups.
• A new loading spinner.
• Colored line numbers in diffs.

Fixed

• Per-turn and image costs are shown correctly.
• Small amounts no longer round to $0.00.
• Edit no longer corrupts text containing $.
• Edit requires a unique match or "replace all".
• Writing a file with no content returns an error.
• Rewind works on case-sensitive file systems.
• Messages sent during a response are queued.
• Output lost in a reconnect is reported.
• Duplicate messages during fast streaming.
• Smoother terminal rendering.

Added

• Automatic reconnect, and the running task continues.
• A 10-minute reconnect window.

Fixed

• No lost or duplicated output after a reconnect.
• Approvals survive a disconnect.
• Tool batches no longer hang after a drop.
• A tool never runs twice after a reconnect.
• Skills accept inline arguments.

New

• Built-in Memory — a global, cross-project memory. The model remembers decisions, bugs, patterns, conventions and context across sessions, and can recall knowledge from any other project while working in the current one.
• A full memory toolset — recall, search and link memories, track goals, and store reusable patterns.
• Cloud Sync — two-way memory sync between your devices: automatic download on startup, upload after each change.
• Stored locally on your machine. Works on any Node.js 18+ with no extra dependencies.

Fixed

• MCP disable now persists to config (not reset after restart).
• MCP disconnect no longer auto-reconnects.
• Improved sync stability.

New

• Background Workflow (Ctrl+B) — send any workflow to background and continue working.
• Live status updates in real time as a workflow runs; the full output is saved to ~/.nilux/bg-tasks/<id>.output.
• Session recovery — /resume restores full workflow state after a terminal crash.
• 3-role workflow: Planner → Coder → Reviewer. Plan shown before execution for approval.

Changed

• Workflow now uses 3 roles instead of 4 — the Tester role was merged into Reviewer.

Fixed

• Faster workflow start — removed a startup delay.
• ESC correctly stops the workflow and returns control to the main agent.
• Approval dialog no longer hangs after ESC.

New

• Prompt suggestions — the model proposes a continuation as ghost text at the end of its reply; press TAB to accept.
• Automatic WebSocket reconnect when you send a message after the connection drops.
• Faster detection of dropped connections.

Fixed

• Emoji now render at the correct width inside tables.

Improved

• Cleaner table rendering.
• Arrow-down navigation goes through chat history first, then the status bar.

Fixed

• Parallel sub-agents no longer hang during long runs.
• Improved overall connection stability.