Skip to content

Changelog

This page tracks user-facing changes to Cosine CLI.

Changes since 2.0.12 (published 2026-04-22).

Shared tool-call UI across CLI, Desktop, and dashboard

Section titled “Shared tool-call UI across CLI, Desktop, and dashboard”

The timeline’s tool-call rendering has been unified across the CLI, Desktop app, and dashboard. Tool chains now look and behave identically everywhere, including hydrated artifact and deployment cards.

When you drill into a subagent, the Tasks section in the sidebar stays visible for the primary task so you can click straight back to the top-level session instead of having to back out manually.

  • Hydrated artifacts now render correctly in the timeline — opening a past session no longer leaves blank spots where artifact cards should be.
  • ggrep now caps match counts and directory depth on very broad searches so a runaway grep can’t hang the tool loop or blow up memory. Over-broad searches surface a clear truncation signal to the agent.
  • Fixed Claude Max sessions not picking up your Claude CLI credentials because SettingSources wasn’t being set — local Claude auth now feeds through correctly.
  • Fixed stale model IDs in the resilient model fallback chain so retries no longer hit missing-model errors after a provider rename.
  • Logging out now clears your team slug but preserves a custom API base URL, so self-hosted and on-prem setups stay configured after a sign-out.
  • Reverted an in-flight GPT 5.5 configuration change that was causing model-selection regressions.

Changes since 2.0.11 (published 2026-04-21).

  • Subagents and the main task are now shown separately in the sidebar, making it easier to see at a glance what the top-level agent is doing versus what its children are doing in Swarm or Multiplan mode.
  • A large set of Desktop app polish improvements landed alongside this CLI release, including portal-rendered composer menus, a cleaner ModeDropdown layering, human-readable wait-tool durations (e.g. 1m 30s instead of 90000ms), and a live countdown on active wait tool calls.
  • Fixed the option selection in confirmation and questionnaire panels jumping when background agent events fired during decision time.
  • Fixed remote runtime branch naming so generated branches are always unique and safely scoped, preventing collisions when resuming remote tasks.
  • Fixed Claude Max sessions exposing the internal load_tools helper as an MCP tool — it is now correctly excluded from the tool list.
  • Fixed attachments on the prompt sometimes not being sent with the intended message; attachments are now tracked on the message they belong to.
  • Fixed Windows ggrep builds by adding the required build prerequisites and linker configuration.
  • Reverted an in-flight GPT 5.5 configuration change that was causing model selection regressions.

Changes since 2.0.10 (published 2026-04-20).

Cloud sessions opened from the dashboard now hydrate the full session UI — timeline, tool calls, artifact cards, and panel state — rather than just replaying text. Revisiting a past remote task in the dashboard now looks identical to the live view.

Added support for Moonshot’s Kimi K2.6 model (via Fireworks and self-hosted local Kimi K2.6), including routing, pricing, and tokenizer settings. You can pick it from the model selector alongside the existing Kimi variants.

The headless runner now supports resuming an existing session ID, making it easier to drive scripted, long-running agent flows without losing history.

  • Fixed attachments attached to the initial prompt being dropped when a session was launched remotely.
  • Hid the slim base image from the environment wizard so users only see the supported image choices.

Changes since 2.0.9 (published 2026-04-17).

Plan mode no longer auto-triggers an exit when you edit the plan. Instead, the agent is explicitly reminded to call exit_plan_mode once the plan is ready to review, and feedback you give during planning keeps the agent in plan mode instead of kicking it out prematurely.

  • Fixed a set of cross-cutting Desktop and CLI issues: assistant layouts in split-view panes, queued-message timing bugs, plan-mode transitions, and a handful of chonkyllm proxy edge cases.

Changes since 2.0.8 (published 2026-04-17).

  • Improved auto-update robustness. The updater now verifies downloaded binaries before swapping them in, so a partial or corrupted download can no longer overwrite a working install. Update failures roll back cleanly instead of leaving the CLI in a broken state.

Changes since 2.0.7 (published 2026-04-16).

Extra shortcut flags have been added for launching into specific modes and configurations, documented in the Flags reference. Mode and reasoning settings are persisted across launches.

  • Remote runtime sessions now always run on a dedicated branch, so the agent’s commits never land on your working branch by mistake.
  • Cleaned up the reasoning UI in the timeline and composer.

Changes since 2.0.6 (published 2026-04-16).

Added support for Anthropic’s Claude Opus 4.7, both through the Claude Max integration and direct API. The model picker surfaces it alongside the existing Opus and Sonnet options.

Added a new “Max” reasoning level for supported models, giving you access to the highest-effort reasoning configuration directly from the reasoning picker.

  • Fixed an Opus 4.7 regression where the model was being ordered incorrectly in the picker and occasionally resolved to the wrong underlying entry.

Changes since 2.0.5 (published 2026-04-16).

Auto-generated branch names for remote sessions

Section titled “Auto-generated branch names for remote sessions”

When you start a remote task, Cosine now derives a descriptive branch name from your initial prompt instead of using a random or numeric branch name. Makes it much easier to find what the agent was working on later.

Changes since 2.0.4 (published 2026-04-14).

Exported artifact transport for remote runs

Section titled “Exported artifact transport for remote runs”

Remote sessions can now stream exported artifacts (PDFs, plan .md files, and other generated documents) back to the dashboard. The Overview panel surfaces them inline so you can download or preview them without leaving the task view.

Nudges and queued messages on remote sessions

Section titled “Nudges and queued messages on remote sessions”

You can now nudge and queue follow-up messages on environment-backed remote sessions the same way you can locally, with stacked items visible in the timeline until they run.

The status bar animation can be coloured per running CLI instance, making it easier to tell multiple cos2 windows apart at a glance. A mode-based palette (Plan, Auto, Manual, Swarm) is also available.

The side panel now has a properly designed empty state instead of a blank pane when no preview or browser session is active.

  • Fixed a hard crash that could occur when the Swarm QA engine processed certain tool-approval sequences.
  • Fixed tool approvals staying in a “pending” state after a questionnaire was answered, and fixed custom answer input not accepting keystrokes until the pane was re-rendered.
  • Fixed a “Rejection:” prefix being stripped from the timeline when a nudged or queued message was displayed after a rejection.
  • Improved the reliability of portr tunnel port scanning — remote preview URLs come up faster and retry automatically when the pod’s port isn’t immediately ready.
  • Cleaned up the project settings panel layout.
  • Tightened interrupt handling so stopping a run feels snappier.
  • Moved the multiplan configuration model to GPT 5.4 1M for higher-quality multi-plan generation.

Changes since 2.0.3 (published 2026-04-10).

The desktop and CLI model selectors now group models by provider and capability into clear segments (e.g. Anthropic, OpenAI, Gemini, local), making it much faster to find the model you want in a long list.

A full environment lifecycle has been introduced on the backend and wired through the dashboard and CLI. Remote environment sessions now have explicit lifecycle states (creating, ready, paused, failed, etc.) that surface in the connection indicator and session shell, so you always know why an environment isn’t responding.

Added Lumen Outpost as a first-class model option with its own tokenizer configuration.

  • Fixed context-window token accounting so in-flight turn tokens are included in the progress bar, eliminating the “above the limit but UI says fine” discrepancy.
  • Fixed remote gitops flows when running against environments built from a custom image.
  • Updated several Desktop keyboard shortcuts — Cmd+N now opens a new task, Ctrl+T opens a new split-view pane, and Ctrl+ ` toggles the terminal. The old Cmd+T terminal shortcut has moved to Ctrl+ ` to free up Cmd+T for new tabs.
  • Fixed stale terminal panes in Desktop so opening the terminal panel reflects the current session state immediately instead of after a roundtrip.
  • Fixed experience_mode being dropped from detached thread and quick-prompt launches, so your workflow mode (Auto/Plan/Swarm/Manual) carries over when you spawn a new task from those entry points.
  • Aligned workflow-mode labels across onboarding and the task picker to the canonical Manual / Auto / Plan / Swarm names.

Changes since v2.0.2 (published 2026-04-07).

Custom environments are now built on Cosine’s updated base images, which include a faster tool-search runtime, a more reliable cx_auto_generate_image flow, and the latest dependency set. Existing environments rebuild on demand into the new images.

The Desktop app received a large batch of UI fixes: the subagent/Swarm panel auto-dismisses when all subagents finish (and re-opens automatically if new ones spawn), onboarding text and backgrounds in dark mode have better contrast, and ghost-text in the prompt box is correctly aligned with the input typography.

  • Fixed fresh Windows installs failing to launch because of missing MinGW and libgcc DLLs. libgcc is now statically linked into the Windows build, so the CLI runs out-of-the-box without installing a separate runtime.
  • Fixed Gemini reasoning: reasoning-content payloads are now shaped correctly for Gemini models in the chonkyllm proxy, resolving the “Gemini won’t reason” regression.
  • Fixed the prompt queue so queued follow-up prompts lay out correctly and are drained in the right order.
  • The dashboard PR review panel no longer depends on the editor websocket being connected, so it works even for archived or non-live sessions.

Changes since v2.0.1 (deployed 2026-04-01).

You can now attach files directly in the chat prompt, making it easier to share context with the agent without relying on @mentions or copy-paste.

The CLI now supports desktop automation on supported platforms, enabling the agent to interact with GUI applications. Headless Linux environments (musl builds) gracefully fall back to no-op stubs so the feature does not affect existing workflows.

  • The browser pane now opens with a smooth reveal animation and no longer collapses the sidebar.
  • Window zoom via double-click on macOS works correctly without the yoyo resize effect.
  • Model selection via the + button and Cmd+K now updates the UI immediately in split-view layouts.
  • The Cosine role icon is now visible in dark mode.
  • Dock icon notification badges are no longer clipped.

The status bar animation can now display a unique color per running instance, making it easy to distinguish between multiple cos2 processes. A mode-based color scheme is also available (purple for Plan, green for Auto, etc.), configurable from settings.

Errors returned by the backend are now displayed as clearly formatted red message cards instead of raw JSON or stack traces. Request IDs are preserved for debugging.

Cross-Origin Navigation in Dashboard Browser

Section titled “Cross-Origin Navigation in Dashboard Browser”

The integrated browser in the dashboard no longer restricts navigation to same-origin URLs, so you can follow links freely during agent-driven browsing sessions.

  • Fixed resumed sessions incorrectly showing MCP servers as not loaded — the tool inventory is now resynced on startup and resume.
  • Fixed startup command execution in remote environments.
  • Fixed Gemini model compatibility — encrypted reasoning content is no longer sent to Gemini models, and a retry mechanism handles upstream rejections gracefully.
  • Fixed the native tool-search feature sending an empty tools.tool_search call when no deferred tools were available, which caused request rejections on the OpenAI Responses API.
  • Fixed alpha desktop builds incorrectly receiving stable-channel updates.
  • Improved streaming error recovery — the classifier now recognises Anthropic tool-mismatch errors in streamed response.failed events and attempts automatic replay.

Changes since v2.0.0-beta21 (published 2026-03-12).

As you type, the prompt input now shows inline ghost-text completions. A predicted continuation appears after your cursor and can be accepted with Tab or the right arrow key. The overlay tracks your cursor position precisely as you move through text.

Sessions now persist between runs. When you launch the CLI, your last active session is automatically resumed so you can continue exactly where you left off. You can also explicitly resume a specific session by ID. The resume flow correctly prioritises remote-hosted sessions when running in cloud environments.

Remote environments can now run and stream a VNC display, enabling the agent to operate GUI applications visually. This unlocks browser automation, desktop apps, and any workflow that requires a visible screen.

Safari is now available alongside Chrome and Firefox as a browser option for agent-driven web tasks.

The CLI now checks for updates in the background and prompts you to restart when a new version is ready. Homebrew-installed builds are updated automatically — no manual brew upgrade needed.

Pass a prompt from a file when starting the agent:

cos2 start --prompt-file ./task.md

Useful for longer or templated prompts and scripted agent invocations.

Set the agent mode and reasoning level directly at startup without navigating the settings menu:

cos2 start --mode plan --reasoning high

Shorthand mode flags are also available: --am (auto), --mm (manual), --pm (plan), --sm (swarm):

cos2 start --am # Start in auto mode
cos2 start --pm --reasoning high # Plan mode with high reasoning

Flags passed to cos2 start (like --mode and --reasoning) are now saved and restored across sessions, so your preferred configuration carries over automatically. Switching modes via Shift+Tab or the settings dialog also persists the selection.

A keyboard shortcut has been added to open the commands panel directly, without navigating through submenus.

Press Ctrl+\ at any point during agent reasoning to interrupt and get the agent’s current answer immediately, without waiting for the full response cycle to complete.

When an MCP server connects, the TUI now shows exactly how many tools it loaded (e.g. Vercel MCP (14 tools)), making it easy to verify integrations are healthy at a glance.

The interface for discovering and installing MCP servers from within the CLI has been redesigned for clarity.

The AI’s native tool-search capability is now enabled by default for everyone and persisted across restarts. The --enable-tool-search flag has been removed since the behaviour is now the default.

Collapsible Changes View in the Overview Tab

Section titled “Collapsible Changes View in the Overview Tab”

The file-change summary in the Overview tab can now be collapsed and expanded, keeping the UI manageable when the agent has touched many files.

The transcript view now opens anchored to the bottom of the conversation and preserves scroll position when you reopen it.

Auto-Generate Dockerfiles for Environments

Section titled “Auto-Generate Dockerfiles for Environments”

When creating a new environment, Cosine can automatically generate a suitable Dockerfile based on your project’s language and tooling, rather than requiring you to write one from scratch.

Each cloud environment can now be configured with its own compute size independently, so demanding environments get more resources without over-provisioning everything.

Remote environments now support full interactive terminal sessions with proper PTY allocation, enabling interactive programs (REPLs, TUIs, prompts) to run correctly in the agent’s terminal.

Auto Git Configuration in Remote Environments

Section titled “Auto Git Configuration in Remote Environments”

Git identity and authentication are automatically configured when spinning up a remote runtime session, so the agent can commit, push, and open PRs without requiring manual credential setup.

The Environments submenu has been reorganised under the Workspace section in the navigation, grouping project-level settings in one place.

You can now log out of your Cosine account directly from within the CLI without needing to touch configuration files.

  • Fixed Windows builds depending on a separate GCC runtime DLL, so fresh Windows installs can launch the CLI without missing-runtime errors.
  • Fixed large paste handling — pasting a large block of text into the prompt no longer hangs or drops characters.
  • Fixed intermittent @filename mention search failures; file search now also runs asynchronously while you type so it no longer blocks input.
  • Fixed TUI layout corruption caused by ANSI escape sequences in tool output.
  • Fixed sidebar keyboard selection drift when the list updates while navigating.
  • Fixed stale Anthropic todo items without IDs failing to clear.
  • Fixed multiple scroll position bugs in the chat timeline that caused the view to jump unexpectedly when new content arrived.
  • Fixed thinking block markdown rendering styles in the timeline.
  • Fixed over-counted context token usage that was causing premature compaction.
  • Fixed Plan mode context being rebuilt too eagerly after tool calls, causing unnecessary latency.
  • Fixed repeated streamed tool-call argument chunks being incorrectly stripped, which produced truncated tool inputs.
  • Fixed the agent to handle malformed tool-call payloads gracefully rather than entering an error state.
  • Fixed MCP OAuth flow bugs in the connection flow and TUI OAuth UI components.
  • Fixed the code edit tool to normalise indentation when matching, so edits succeed across codebases with mixed tabs and spaces.
  • Fixed the Anthropic-provider edit tool to enforce exact string matching and prevent incorrect partial matches.
  • Fixed tool registration failures in Plan mode that caused certain tools to be unavailable during planning.
  • Fixed the remote gitops push-to-branch and PR creation flow, including correct git identity and auth handling.
  • Fixed remote task history resume not correctly restoring conversation state in the Platform view.
  • Fixed a trailing blank row appearing at the bottom of the TUI under certain terminal sizes.
  • Fixed remote todo list display and subagent token graphs not rendering correctly in remote sessions.
  • Updated all remaining user-facing “checkpoint” references to “commit” for consistency.
  • Reduced environment boot time through a refactored bootstrap sequence that parallelises more initialisation work.
  • The grep tool now streams results in chunks, preventing memory pressure and crashes when searching large codebases.
  • @mention file lookup now runs asynchronously so typing in the prompt is never blocked while results load.
  • Provider-side context compaction is now available, reducing context window pressure in long sessions.
  • Improved scrolling performance through long conversations, particularly in sessions with many tool calls or large outputs.

Changes since v2.0.0-beta20 (published 2026-03-10).

  • Added GitHub Copilot sign-in with cos login-with-copilot, so you can connect your Copilot account from the CLI and use Copilot-backed models directly in Cosine.
  • You can now queue follow-up prompts while a run is still active. Nudges and queued prompts are shown separately, and queued items stay visible in the timeline until they run.
  • Added /link-project so you can attach another project to the current session, with an option to save that linked project for future sessions in the same workspace.
  • Tool approvals are now more informative: edit-like tool calls can show a diff preview before approval, and approval cards use clearer user-facing descriptions.
  • Expanded the model picker with updated model options, including GPT 5.4, GPT 5.4 1M, and GitHub Copilot-backed choices, while making model selection behave more consistently across settings and task prompts.
  • Fixed image path detection so pasted paths still work when they are quoted or followed by trailing punctuation.
  • Fixed remote runtime host override handling so custom hosts are respected more consistently.
  • Fixed remote runtime configuration selection and write-location issues.
  • Fixed queued prompt handling so stacked items lay out correctly and stale completion events do not accidentally drain the queue.
  • Fixed interrupted-run UI so cancellations are shown more cleanly.

Changes since v2.0.0-beta19 (published 2026-03-07).

  • Checkpoint detail cards now show the provider model ID for the model that produced the change, making it easier to attribute a checkpoint to a specific model when you’re switching between providers.
  • Terminal tool calls now emit stall nudges and support interactive streaming so you can see progress (and re-engage the agent) on long-running shell commands instead of the UI appearing to freeze.
  • Context window token totals are now emitted for the timeline HUD, giving you a live view of context pressure.
  • First backend work landed for the new environment management API that later powers the dashboard environments UI.
  • Fixed the desktop streaming indicator flickering during long generations.
  • Fixed the /theme command not honouring user theme selection because the theme change message wasn’t being handled.
  • Fixed Swarm QA being off by default; the toggle now initialises to on, and the enabled/disabled labels in the settings UI are clearer.
  • Fixed small-model and micro-model support in the agent-to-LSP bridge, alongside assorted toast notification and Sentry wiring fixes.
  • Cleaned up a stray Ctrl+A agents label in the footer.
  • Raised the default TTFT timeout to 45 seconds so slower providers no longer time out before the first token.
  • Hardened the chat-completions proxy against provider rate-limit errors; the CLI now surfaces rate limits instead of hanging.

Changes since v2.0.0-beta18 (published 2026-03-07).

  • Added a diff viewer to the Git Session Manager so you can inspect checkpoint changes without leaving the CLI.
  • Swarm mode now lets you turn QA verification agents on or off with /swarm-qa or from the command menu.
  • Subagents now share MCP and LSP context more reliably with their parent session, reducing setup friction in Swarm workflows.
  • Added initial support for sandbox-backed remote environments in managed runtime flows.
  • The project initialization dialog now includes a “No, not this time” option so you can skip setup and continue immediately.
  • Fixed ANSI terminal rendering issues.
  • Fixed shifted number and symbol keys in VS Code terminal input.
  • Fixed nested subagents appearing under the wrong parent in the Agents panel.
  • Fixed UI state bleed-through between active and background agents in multi-agent sessions.
  • Fixed a background-task scheduling issue that could cause Swarm runs to return early.

Changes since v2.0.0-beta17 (published 2026-03-05).

  • Added Lua plugin support with cos plugin install, cos plugin list, and cos plugin remove, plus a plugin runtime that can subscribe to agent events, queue nudges, make HTTP requests, and perform sandboxed file operations.
  • Added plugin documentation covering installation, the Lua API, lifecycle, and example plugins.
  • Added xhigh reasoning support for GPT-5.4 in the CLI reasoning-effort controls.
  • Swarm can now nudge active work and surface semi-async subagent progress so long-running tasks feel more responsive.
  • cos update now automatically runs the installer for non-Homebrew installs, reducing the manual steps after an upgrade.
  • The slash-command menu has been cleaned up for faster navigation and clearer command discovery.
  • Plugin logs are now surfaced in the CLI panels so plugin activity is easier to inspect while debugging.
  • Lumen and Devstral have been removed from the CLI model picker.
  • Resolved a subagent hang that could affect VS Code-hosted CLI sessions backed by CLI2.

Changes since v2.0.0-beta16 (published 2026-03-04).

  • Rolled up a batch of miscellaneous CLI2 fixes covering tool handling, streaming, and TUI polish.

Changes since v2.0.0-beta15 (published 2026-03-04).

  • Internal version bump with no user-visible changes.

Changes since v2.0.0-beta14 (published 2026-03-04).

  • Internal release-tagging bump with no user-visible changes.

Changes since v2.0.0-beta13 (published 2026-03-04).

  • Added automatic task name derivation from user prompts in the Agents sidepanel to make multi-agent sessions easier to scan.
  • Expanded CLI2 event coverage to improve telemetry and product insights.
  • Fixed additional Swarm mode issues to improve stability in parallel subagent workflows.

Changes since v2.0.0-beta12 (published 2026-03-03).

  • Fixed update check functionality to properly detect and prompt for new CLI versions.
  • Fixed Swarm mode bugs including subagent checkpointing issues that could cause state inconsistencies.
  • Fixed terminal display issues where the terminal could appear white or unresponsive.
  • Fixed backspace behavior in filter dialogs — Ctrl+H now works correctly as backspace.
  • Fixed user question handling on subagents so prompts are properly routed when using parallel agent workflows.
  • Improved Homebrew install detection on Windows CI environments.

Changes since v2.0.0-beta10 (published 2026-02-25).

Swarm mode enables parallel agent workflows by allowing the primary agent to spawn and orchestrate multiple subagents simultaneously. When activated, the agent can delegate independent tasks to subagents that run in parallel, significantly speeding up complex multi-file operations. Subagents appear in the Agents panel (Ctrl+3 / Alt+3) indented under their parent, where you can monitor progress, view diffs, or kill individual agents. This is particularly useful for large refactoring tasks, cross-module changes, or when you need multiple independent investigations running concurrently.

The CLI now supports a complete light color theme alongside the existing dark theme. Light Mode provides better visibility in bright environments and reduces eye strain during daytime usage. The theme automatically applies to all UI components including the timeline, side panels, dialogs, and terminal views. You can switch between themes through the settings or use the quick toggle in the commands menu.

Rebase Quick Action in Git Session Manager

Section titled “Rebase Quick Action in Git Session Manager”

A new Rebase action has been added to the Git Session Manager, allowing you to quickly rebase your current branch without leaving the CLI interface. This streamlines common git workflows when working with agent-generated branches, making it easier to keep feature branches up to date with main before merging or continuing work.

Plan mode now supports auto-approve, which automatically approves the generated plan without requiring manual confirmation for each step. This is ideal when you want the agent to create and execute a plan in one flow while still benefiting from the structured planning phase. The setting can be toggled per-session or configured globally for users who prefer the plan-then-execute workflow without interruptions.

Plan files are now injected into the agent’s context more consistently and reliably. When a plan is created, its content is properly formatted and made available to the agent throughout the session. This ensures the agent always has access to the full plan context when making edits, reducing drift between the intended plan and actual execution.

Open Terminal Action in the Terminal User Interface (TUI)

Section titled “Open Terminal Action in the Terminal User Interface (TUI)”

You can now open an interactive terminal session directly from the TUI using the new Open Terminal action. Accessible from the commands menu (/ or Ctrl+P), this creates a new shell session without leaving your current conversation. The terminal appears in the Terminals panel (Ctrl+2 / Alt+2) where you can switch between it and your chat session seamlessly.

Terminal output can now be sent directly back to the model as context. After running commands in an interactive terminal, you can capture the output and include it in your conversation with the agent. This eliminates manual copy-pasting when sharing command results, error messages, or logs with Cosine.

When the agent creates a plan, it now appears inline in the chat timeline as a structured implementation plan block. These plans are collapsible, show completion status, and can be referenced throughout the conversation. This makes it easier to track what the agent intends to do and check off items as they’re completed.

The CLI now supports native streaming responses from Google’s Gemini models. This provides real-time token-by-token output for Gemini-backed conversations, reducing perceived latency and allowing you to see responses as they’re generated rather than waiting for complete responses.

Text selection in the timeline has been significantly improved. You can now select and copy text from assistant responses more reliably, with better visual feedback during selection. The copy UI has been updated to provide clearer confirmation when text is copied to the clipboard, making it easier to extract code snippets, explanations, or other content from agent responses.

  • Fixed hangs that could occur during websocket inference.
  • Fixed rendering issues where output_text assistant messages could display incorrectly.
  • Fixed message ordering in resumed sessions so tool execution blocks and assistant output are interleaved correctly.
  • Fixed a ggrep mutex issue that could cause instability.
  • Removed stray debug logs from the terminal popout dialog.
  • Fixed stale tool definitions after model switches by invalidating cache correctly.
  • Fixed terminal overlay responsiveness issues.
  • Fixed dialog search behavior so Ctrl+H works as backspace.
  • Fixed Linux CGO build/linking reliability by switching to external linker mode.
  • Improved CLI2 CGO build speed by roughly 3x.
  • Added improved generation/stream telemetry (including trace, latency, and TTFT capture) to improve performance observability.

Changes since v2.0.0-beta10 (published 2026-02-25).

  • Added an Open Terminal TUI action that lets you spawn an interactive shell from inside the CLI and attach it to the Terminals panel.
  • Terminal output can now be sent back to the agent as context with a single action — no more copy-pasting command results manually.
  • Implementation plans now appear as structured, collapsible blocks in the chat timeline as the agent produces them.
  • Native Gemini streaming is now supported, so Gemini responses render token-by-token instead of arriving in one chunk.
  • Improved text selection and copy UI in the timeline — multi-line selection is more accurate and copy confirmation is clearer.
  • Timeline todo items now wrap correctly instead of being truncated at the edge of the pane.
  • AI generation telemetry now captures trace context and latency, and records TTFT on the first streamed chunk to power more accurate performance monitoring.
  • Fixed Ctrl+H not working as backspace in dialog search fields.
  • Fixed tool definitions going stale after a model switch; the definitions cache is now invalidated correctly.
  • Fixed terminal overlay responsiveness during heavy streaming.

Changes since v2.0.0-beta9 (published 2026-02-24).

  • Context window normalization is now strictly validated before turns, preventing the agent from silently operating with a truncated or mis-sized window.
  • Added a runtime-configurable flag to disable the TTFT watchdog in Harbor, useful when evaluating very slow providers.
  • Fixed a batch of bug fixes rolled up as “yet more bug fixes”.
  • Fixed the agents list wrapping incorrectly in the agents panel.
  • Fixed CSV and .ipynb files being incorrectly excluded from indexing and search.
  • Fixed desktop-side build errors and restored a backend security flag that had been dropped.
  • Renamed the internal Lumen prompt and prompt-RFT behaviours for clarity.
  • The debug tab in the TUI now only shows in debug builds.
  • Updated the Codex model ID to 5.3 to match the provider’s current naming.

Changes since v2.0.0-beta8 (published 2026-02-23).

Added a /feedback slash command to send feedback directly from the CLI, plus an in-app tip for internal users highlighting the command.

Agent memories are now a first-class feature — stored, surfaced, and updated through the agent directly rather than through a separate side flow. Memory content is captured into the agent’s working context on each turn.

When you create a new remote task, Cosine now generates a descriptive branch name from your prompt in the task dialog rather than requiring you to invent one.

New shortcuts let you jump between the main tasks in your session without using the sidebar.

  • Fixed compaction behaviour that could leave the agent confused about which parts of the conversation were summarised.
  • Fixed local tasks failing to run when no project was associated with the session.
  • Fixed missing context-window entries for claude-sonnet-4-6 variants, which caused those models to be rejected.
  • Fixed Windows checkpoint commands emitting CRLF warnings by disabling core.autocrlf on those calls.
  • Fixed key bindings inside plans being swallowed by the outer timeline.
  • Fixed ggrep spawning rg for literal patterns by adding a fixed-string mode.
  • Fixed auth clients not being reloaded after a fresh ChatGPT login.
  • Reworked remote task init so staging pods come up reliably.
  • Reduced CPU and memory overhead across the agent, TUI, and per-session state — longer sessions now feel noticeably lighter.

Changes since v2.0.0-beta7 (published 2026-02-20).

  • Worktree branch names are now auto-generated and the git context is injected into the agent prompt, so the agent always knows which branch it’s working on.
  • Context compaction now accounts for the overhead of system and instruction messages, making compaction decisions more accurate.
  • Commit namer now generates both a title and a body for agent-authored commits, producing more useful commit history.
  • Added session-affinity headers on non-Codex client requests so repeated turns in the same session land on the same backend pod.
  • Integrated the Mason LSP installer and flattened the LSP package structure, making LSP installation more reliable across environments.
  • Fixed trailing empty assistant messages causing rendering gaps in the timeline.
  • Fixed background-task questions leaking into the foreground session; they now stay pinned to the background session they were raised from.

Changes since v2.0.0-beta6 (published 2026-02-20).

  • Added a TTFT timeout so the CLI can recover from providers that accept a request but never produce a first token.
  • The selection hint in the timeline now only renders for multi-line selections, reducing visual noise during ordinary navigation.
  • Fixed a TUI crash during the login flow and a stale appContext reference that lingered after logging in.
  • Reverted the helptext rendering width calculation to a character count so it renders correctly under every terminal width again.

Changes since v2.0.0-beta5 (published 2026-02-20).

Added a /jira slash command so you can pick Jira tickets directly from the prompt and have the agent pull context from the ticket.

Agents that are done or in an error state now expose a close task action directly from the side panel.

  • Fixed a hang that could occur when entering Plan mode multiple times in one session.
  • Fixed terminal action buttons not appearing on hover in the Terminals panel.
  • Hid the Ultra menu item outside debug builds.
  • Fixed subagent ordering in the Agents side panel flickering between renders.
  • Fixed ggrep not escaping regex metacharacters when UseRegex=false, so literal-looking queries now match literally.
  • Fixed p and y key bindings hijacking text input fields.
  • Fixed the selection highlight appearing for zero- or one-character selections.

Changes since v2.0.0-beta4 (published 2026-02-17).

You can now paste images directly into the prompt with Ctrl+V, alongside the existing drag-and-drop support.

File search is now available in the new-task dialog — @filename autocompletion works from the very first turn, including matches inside .ipynb files.

Added Gemini 3.1 Pro as a selectable model with its own LiteLLM configuration.

claude.md files alongside agents.md are now auto-discovered and merged into the agent’s context on startup.

  • Significantly improved TUI rendering performance during long streams and heavy timelines.
  • Fixed a scroll jump caused by backspace under certain cursor positions.
  • Fixed a text-input regression where a leading t was swallowed in fresh prompts.
  • Fixed several MCP auth and patch-apply display bugs.
  • Fixed an error that could occur immediately after a compaction pass.
  • Fixed model change not rendering — model switches now trigger a full timeline re-render.
  • Fixed session naming to pick more useful titles.
  • Fixed checkpoint creation under tight network conditions by extending the context deadline from 10s to 20s.
  • Restored Claude intermediate messages to context after fixWindowItemsForCompat, so message detection doesn’t miss them.
  • Paused beamline animation when idle and cached frame renders to reduce idle CPU.
  • Added more headless CLI flags for the scaffolding pipeline.

Changes since v2.0.0-beta3 (published 2026-02-17).

A new login flow landed, with a cleaner sequence for OAuth providers and better handling of mid-flow errors.

Telemetry is now enabled by default and cannot be disabled via config. The CLI always initialises the PostHog client and applies relaxed semver handling so new builds don’t get dropped from dashboards.

  • Fixed tool-usage state leaking between sessions when switching between them.
  • Batched a “miscellaneous small bug fixes” sweep across the TUI.
  • Switched to the production PostHog key for official builds.

Changes since v2.0.0-beta2 (published 2026-02-16).

  • Added a Tips surface in the TUI that rotates short product tips while you work.
  • Lazy-loaded startup operations so the CLI reaches an interactive prompt faster on cold boot.
  • Added PostHog tracking for checkpointing activity.
  • Simplified find/replace handling in the edit tool, removing a class of mis-matched indentation failures.
  • Corrected relative-path handling when committing checkpoints so paths outside the repo root no longer cause checkpoint failures.
  • Strengthened streaming stream-closure handling to reduce stuck streams at the end of a response.
  • Aligned the settings nomenclature around “checkpointing” (from the mixed “checkpoints/commit” terminology).

Changes since the initial 2.0 beta opt-in (published 2026-02-16).

This is the first broadly-available v2.0.0-beta build. Highlights brought over from pre-2.0 work and landed for the 2.0 beta rollout:

  • Ultra daemon topology — a long-lived background daemon that owns LSPs, MCP servers, browser agents, and session state, so the CLI TUI starts and restarts instantly without re-bootstrapping each time.
  • Runtime debugger panel — a new TUI panel for inspecting the live agent runtime, including tool state, context size, and in-flight requests.
  • Remote CLI2 execution — the CLI can now delegate a task to a remote environment and follow its output from your local TUI, the foundation for later remote runtime features.
  • VSCode 2 Extension — a new VS Code extension that hosts a CLI2 agent inside the IDE.
  • Browser agent v4 — a rewritten browser automation agent with improved reliability against modern JS-heavy pages.
  • --checkpointing flag — toggle checkpointing on or off when starting a session, including a disable path for repos where checkpointing isn’t wanted.
  • Find-and-replace tool — a new find_replace_in_file tool for in-file regex replacement with robust line-ending normalisation.
  • login-openai command — log in with OpenAI directly from the CLI and centralise the API key via InferenceAPIKey instead of env vars.
  • MCP protocol integration — first-class MCP tool support, an MCP configuration UI, and plan-mode tweaks that respect MCP-provided tools.
  • Kimmie mode — a tuned preset for working with the Kimi family of models.
  • Auto-accept mode — a configurable mode that auto-approves non-destructive tool calls for faster iteration.
  • Link Project — attach another project to the current session from the TUI.
  • Project indexing for the agent — the agent and memory tools now share a project-wide index for faster retrieval.
  • Dynamic suggested tasks — the homepage now suggests tasks based on your recent activity.
  • Rate-limit retry handling — provider rate limits are retried automatically with a visible status indicator in the TUI.
  • Linux AMD64 and ARM64 builds — official Linux builds land alongside macOS and Windows.
  • Session list pagination and terminal titles — session lists paginate cleanly and terminal titles reflect the active task.
  • X-high reasoning — the xhigh reasoning effort is available in the reasoning picker.
  • Session saving improvements — session state persists more reliably across restarts, with stored branches.
  • Anthropic tools and more models — added GPT-5, more Anthropic models, and GLM 5.
  • Side panel show/hide, navigation fixes — the side panel can be toggled, and navigation bugs on its tabs were fixed.
  • Fixed a large-file read regression.
  • Fixed LSP processes dying on use; they now lazy-start and shut down when idle.
  • Reduced LSP memory usage during lookups.
  • Fixed tool-result ordering when many tool calls ran in parallel.
  • Optimised the CLI2 LSP server scan to speed up start-up times.
  • Fixed streaming bugs in the new inference pipeline.
  • Fixed MCP headers not being forwarded on outbound requests.
  • Fixed the homescreen perf regression on large project lists.

The 1.x series shipped under the cos binary from the original Cosine CLI codebase. Cosine CLI 2.0 is a ground-up rewrite and is tracked above. Release notes for 1.x are included here for completeness.

Changes since 1.15.1 (published 2025-11-14).

  • The list_tasks tool now tolerates both raw arrays and wrapped payloads from the API, so task listings never silently come back empty when the backend changes its response shape.

Changes since 1.15.0 (published 2025-11-03).

A new headless mode lets you run Genie non-interactively, suitable for scripted and CI-driven agent runs. Includes support for piping extra prompts, a Terminal output help section, and Home/G and End/G shortcuts for jumping to the top or bottom of the terminal viewer.

A proof-of-concept Azure DevOps integration was added for teams using Azure DevOps repositories.

  • Fixed absolute /tmp paths failing in headless file reads.
  • Fixed headless-terminal-cos so headless runs pick up the correct terminal mode.
  • Fixed git checkout -- . in the CLI so clean-working-tree resets actually run.

Changes since 1.14.5 (published 2025-10-24).

Added ft-gpt-5 model support in the CLI inference selector.

The workspace now opens to a welcome screen with an onboarding animation the first time you enter it.

You can now continue into a session without importing a project, backed by an ephemeral cli project type. The Tasks view hides for CLI-only projects that aren’t linked to a Cosine-hosted project.

The input area now virtualises very large pastes — the full content is preserved and submitted, but the UI no longer blocks while rendering it.

CLI signing preflight now runs at startup, surfacing approval prompts early rather than interrupting the agent mid-task.

  • Stream errors now propagate back up from the agent loop instead of being swallowed silently.
  • Implemented a dry merge step in keef so speculative merges don’t mutate the working tree.
  • Renamed Project.Type to Provider and cleaned up “CLI project” detection to use the new helper.

Changes since 1.14.4 (published 2025-10-23).

  • Fixed glob pathspec handling for git calls so @-mention globs like src/**/*.ts resolve correctly.
  • Fixed a second pathspec regression that affected @filename matches with shell metacharacters.
  • Fixed a login flow bug that could leave the CLI stuck on the provider picker.

Changes since 1.14.3 (published 2025-10-23).

When continuing a previous session, the CLI now attaches to the existing task instead of forking a new one, preserving history and task state.

Added a settings panel directly in the CLI for inspecting and editing CLI configuration without touching config files.

Changes since 1.14.2 (published 2025-10-20).

  • The CLI can now auto-login using saved credentials on start-up, removing the manual step for already-signed-in users.
  • Improved the VS Code “@file” context toggle so it reflects the current pin state correctly.
  • Upgraded to the Go 1.23 toolchain and fixed @path parsing regexes that broke under the new compiler.

Changes since 1.14.1 (published 2025-10-18).

  • Timeline refreshes are now coalesced during streaming coding output, cutting CPU usage during heavy generations.

Changes since 1.14 (published 2025-10-15).

  • Added a dividing line between chat turns in the timeline so scrolling back through a long session is easier to parse.
  • Improved keef startup time on large projects by avoiding repeated repository scans.
  • Rewrote the workspace bridge for cleaner IPC between the CLI and daemon.

Changes since 1.13 (published 2025-10-13).

Initial release of the Cosine Genie VS Code extension, which hosts a CLI agent inside VS Code.

Major refresh of the streaming pipeline for smoother token rendering in the timeline.

  • Remote pod cleanup runs even when init fails, so failed starts don’t leave orphaned pods behind.
  • Fixed daemon write-bytes accounting so large tool outputs no longer get truncated over the daemon bridge.

Changes since 1.12 (published 2025-10-10).

  • Fixed Esc cancelling a nudge instead of just closing the nudge input.

Changes since 1.11 (published 2025-10-02).

Added a profiles system so you can switch between named CLI configurations (credentials, model choice, workspace overrides) with one command.

  • Fixed a cli-timeline-created-at bug where timeline entries were keyed by incorrect timestamps and appeared out of order.
  • Fixed cli-import regressions that affected importing a project from a fresh install.

Changes since 1.1 (published 2025-10-01).

  • Added an overlay for opening a URL in the browser directly from the CLI.
  • Removed the legacy cos serve command, which was superseded by the daemon architecture.

Changes since 1.0.7 (published 2025-09-25).

You can now pick up an existing task on start instead of always creating a new session.

Added CLI support for stopping a running task or nudging it with an additional prompt while it’s working.

Diffs in the timeline now wrap cleanly at the terminal width, and new inline hints guide first-time users through the main CLI actions.

The GenieProvider now emits timeline events for retry attempts, so you can see when a request is being retried rather than silently hanging.

Expanded the @file token recogniser so more prefixes and path shapes register as mentions, and the mentions dropdown now shows all matches instead of truncating to a small list.

  • Reduced goroutine churn in the CLI websocket handler under heavy streaming.
  • Improved reattach behaviour when the websocket drops mid-task.
  • chore: daemon now writes logs to both stdout and a file for easier debugging.

Changes since 1.0.6 (published 2025-09-23).

  • Fixed a Sentry spam issue where expected websocket closures were being reported as errors.
  • Fixed file navigation so arrow keys wrap around at the first/last entry.

Changes since 1.0.5 (published 2025-09-12).

First release of the long-lived daemon process that backs the CLI, improving start-up time and enabling persistent LSP and session state across invocations.

  • Optimised timeline rendering for large markdown blocks so long assistant answers don’t stall the UI.
  • Disabled the debug panel in non-debug builds.

Changes since 1.0.4 (published 2025-09-12).

  • Removed the legacy terminal viewer that was superseded by the integrated terminal.

Changes since 1.0.3 (published 2025-09-11).

The branch selector now supports fuzzy filtering and a scrollable list, making it far faster to pick a branch in repos with long branch lists.

  • Fixed bullet character rendering in timeline entries on terminals without full Unicode support.
  • Fixed the preset toggle shortcut conflicting with VS Code terminal bindings.
  • Fixed multiple key bindings that were incorrect on macOS terminals.
  • Removed an over-eager full cleanup that was preventing concurrent CLI sessions from coexisting.

Changes since 1.0.2 (published 2025-09-10).

  • The CLI version now appears in the footer for easy reporting.
  • Updated terminology — “promote” is now “publish” and “remote” is now “cloud” throughout the CLI copy.
  • Refreshed login interface with clearer banners and color treatment.
  • autoAccept option is now honoured on CLI send-message calls, enabling scripted approvals where supported.
  • Reworked key bindings for consistency across views.
  • Improved onboarding feedback for project import.
  • Removed the separate workspace option from the top-level menu.
  • Ensured proper cleanup of terminal streams and resources on exit.
  • Fixed markdown rendering clipping text at the terminal edge.
  • Terminated active terminal sessions when the CLI exits so background shells don’t linger.

Changes since 1.0.1 (published 2025-09-10).

Added a command to open the proposed diff in your external editor so you can review changes in a full-size viewer.

CLI commands now honour a unified ignorer so .gitignore and Cosine exclusions are applied consistently across tools.

Changes since 1.0.0 (published 2025-09-10).

  • Build-only release with no user-visible changes.

Changes since 0.2.2 (published 2025-09-09).

A purpose-built side-by-side diff viewer landed for reviewing agent changes, with better wrapping and clearer change markers than the previous unified view. Diffs now render for every action type that produces changes, not just edits.

The workspace now opens with a welcome message and a short animation to introduce first-time users.

You can now interrupt a running task cleanly, with the CLI prompting for confirmation before exiting.

  • Fixed a set of border-rendering issues around panels and dialogs.
  • Ctrl+C now follows a two-step confirm flow before exiting, so accidental presses no longer kill the session.
  • Removed the detach flow from CLI task promotion; remote tasks now follow a simpler publish path.
  • Added a full-screen seeding spinner when a message is being prepared.

Changes since 0.2.1 (published 2025-09-08).

  • Shift+Enter now inserts a newline in the prompt, enabling true multi-line input.
  • Added a “Continue Locally” option for remote tasks so you can take a remote task back to your local machine mid-flight.
  • Expanded the timeline event coverage for agent actions.
  • Cleaned up stale shadow file system folders left behind by aborted sessions.
  • Refactored the promotion flow in the CLI for fewer edge-case hangs.

Changes since 0.2.0 (published 2025-09-01).

  • Added the ability to create remote tasks directly from the CLI.
  • Added a tool_result event type to the timeline so tool outputs are captured and replayable.
  • Added file attachments on task creation.
  • Added remote task viewing in the CLI, alongside a large refactor to task creation and websocket session management.
  • Fixed duplicate timeline events appearing on task resume.
  • Fixed several diff display options for CLI-initiated changes.
  • Improved CLI timeline scrolling to keep the active line in view.
  • Removed auto-pm behaviour in favour of explicit mode switches.

Changes since 0.1.4 (published 2025-08-05).

This release was a significant foundational update to the Cosine CLI with substantial UI and architecture work.

The CLI now has a dedicated Tasks view backed by API data and an explicit project-import flow, replacing the earlier implicit project detection.

A new CLI header surfaces the current project, team, and active model, and the model picker was updated with the current set of supported models.

Presets, accept-with-changes, and connection error rendering

Section titled “Presets, accept-with-changes, and connection error rendering”

Added CLI presets for common configurations, an “accept with changes” action so you can tweak an agent-proposed edit before committing to it, and clearer connection-error rendering.

Added a Simple UI mode that strips the CLI back to a minimal layout for constrained terminals or scripted usage.

Ctrl+C/Ctrl+V copy and paste now works reliably inside the CLI, and the timeline renders messages in arrival order instead of interleaving them incorrectly.

If the websocket drops mid-task, the CLI now reattaches to the task on reconnect instead of starting a fresh session.

  • Refactored the timeline into an MVC pattern so updates are driven by model changes instead of ad-hoc re-renders.
  • Moved markdown and git handling onto the shared pz stack for consistency with the web app.
  • Fixed a scroll-view implementation that was silently non-scrollable.
  • Fixed the tab content arrangement inside the main CLI view.
  • Fixed a shadow-file cleanup regression that could leave workspace state inconsistent.

Release notes for CLI builds earlier than 0.2.0 are intentionally omitted; those were internal and pre-beta rollouts with no general availability.