What's New

Reliability & Performance

• Chat history now keeps saving when some conversations repeatedly fail to save. Newer conversations are no longer blocked behind the failing ones, and conversations are not marked as saved until a write actually succeeds.
• A large backlog of unsaved conversations now drains on its own instead of waiting for your next chat action, and transient save failures are retried automatically with backoff.
• Very large chat histories no longer abort saving or loading with a "Maximum call stack size exceeded" error.
• Conversation titles are generated once per conversation. Overlapping triggers no longer send duplicate title requests.

Bug Fixes

• Agent file edits and diagnostics lookups no longer fail when a language extension reports a diagnostic without an error code.
• Augment Migration commands no longer appear in the Command Palette when migration is not available to you.

Connect Azure DevOps

Azure DevOps is now available to every organization. Connect a bot account with a personal access token for a specific organization, or with OAuth to have Cosmos find every Azure DevOps organization that account can reach. Agents get access to those repositories and can respond to pushes and pull request activity through webhooks Cosmos sets up for you.

New features

• Service accounts are now available on Enterprise and team plans. Admins can create service accounts, issue API tokens for them, and see what each account can access.
• GitHub Enterprise Server is now in the Connectors catalog, with a setup guide.

Improvements

• Fleet Management can update the model for many Experts at once, using either a model you choose or each Expert's recommended model.
• Agents no longer merge pull requests.
• The Linear connector walks you through setting up Ticket Manager.
• Cosmos asks admins to reconnect Linear when its permissions are out of date.
• The session sidebar shows when the session owner's connection to an integration needs attention.
• Groups and members synced from your identity provider are read-only in Cosmos.
• Double-click inline code to copy it.

Bug fixes

• Sessions no longer show an empty transcript when you move between them.
• The Stop button no longer gets stuck.
• The right pane stays open when the left sidebar collapses.
• The Space switcher no longer lists Default twice.
• Environments are no longer wrongly labeled invalid.
• The Files panel refreshes after each agent turn.
• Choosing an Expert after picking a model now uses the Expert's model.

Changes below are new since the last stable IntelliJ plugin, v0.488.0-stable.

Improvements

• The agent now asks for your approval before reading a file that looks sensitive — .env files, private keys and certificates, and credential- or secret-named files — as well as any file excluded by your .augmentignore. Directory listings are unaffected, and .gitignore is still not treated as a read boundary.
• MCP servers reached over HTTP, SSE, or Streamable HTTP now go through your configured HTTP proxy, so MCP tools work on machines that require a proxy for outbound traffic.
• Skills with invalid SKILL.md frontmatter are now shown in Settings with an actionable error instead of being silently dropped. An invalid skill also no longer hides a valid skill of the same name.

Reliability & Performance

• The plugin's sidecar process can now use up to 4096 MB of heap, instead of Node's default ceiling of roughly 1.7 GB, so loading a very large conversation no longer exhausts memory. If it still runs out, you now get an explicit out-of-memory message instead of the sidecar disappearing silently. The ceiling can be overridden with the -Daugment.sidecar.maxOldSpaceSizeMb IDE system property.
• Chat now automatically retries a response that is interrupted part-way through the stream, instead of ending the turn with an error you have to resend.
• Opening Find Action no longer freezes the UI while the BYOK actions read your stored credentials; that lookup now runs off the UI thread.
• After a restart, a turn that was interrupted now stays finished. Previously late startup loading could revive it, leaving the conversation appearing to still generate with tool calls stuck unresolved.

Bug Fixes

• Fixed remote tool integrations (GitHub, Linear, Notion, and others) disappearing from Settings when the first tool fetch happened before sign-in completed. The list now heals as soon as you sign in, instead of staying empty for up to an hour or until you restart the IDE.
• Fixed the agent being unable to read a file by absolute path from outside the project's content roots, such as a plugin-installed skill under ~/.augment/skills/.
• Fixed a response that hits the model's output limit with no tool call ending as a silently truncated success. It is now surfaced as an interrupted response.
• Fixed agent checkpoints failing to save for files in deeply nested workspaces, where the generated checkpoint filename exceeded the filesystem's limit. Existing checkpoint filenames are preserved.
• Fixed a git checkout that deletes and immediately recreates a file racing the file watcher into deleting the freshly recreated file.
• Fixed a chat message whose Markdown could not be parsed breaking the rendered response. It now shows an inline "Failed to render Markdown." notice, and rapid streaming updates always settle on the latest content.
• Fixed a failed MCP Tool Search setting write in Settings being reported as a load failure, which made the whole MCP server list look broken.

New Features

• Chat submit is now a rebindable VS Code keyboard shortcut. "Submit Chat Message" (Enter) and "Submit Chat Message Immediately" (Ctrl+Enter / Cmd+Enter) appear in Keyboard Shortcuts and can be remapped.

Improvements

• Workspace-defined Augment hooks now run only in trusted workspaces. Command hooks supplied by a workspace stay disabled until you grant workspace trust, and are armed as soon as trust is granted.
• The agent now asks for approval before viewing sensitive files — .env files, private keys and certificates, credential- and secret-named files, and paths excluded by .augmentignore.
• Skills in .augment that fail to parse are now listed in Settings with the reason they were rejected, instead of disappearing silently.
• MCP Tool Search in Settings now distinguishes a failure to save your changes from a failure to load the current configuration.
• Remote MCP servers (HTTP and SSE) now honor the HTTP_PROXY and HTTPS_PROXY environment variables.

Reliability & Performance

• Chat responses now recover from mid-stream interruptions automatically. Recovery is suppressed after a tool call has already completed, so tools are never re-executed by a retry.
• A stream that ends abnormally now reports a neutral, retriable "response was interrupted" message instead of guessing that your network connection failed.
• Recovered chat turns settle into a finished state. Tools that were running when a conversation was interrupted no longer appear to run forever after it is restored.
• The conversation you are working in is preserved when the extension host restarts or reactivates.
• Remote tool integrations no longer disappear when sign-in finishes while the extension is still starting up.

Bug Fixes

• Responses cut short by the model's output limit are no longer reported as complete. Chat now shows an explicit truncation failure.
• Markdown that fails to parse now renders an inline fallback message instead of leaving the response blank.
• Agent checkpoints in deeply nested workspaces no longer fail to restore because of over-long checkpoint filenames.
• Files that Git recreates during a checkout are no longer deleted again by the agent's checkpoint file watcher.

Control which models your team can use

Admins can now choose which models people in their organization can use. Set a default on the new Model Access page and override it for individual models. Fleet Management flags Experts that use a model your organization no longer allows.

New features

• Admins can manage every secret in the organization from Secrets, including its owner and the Spaces that can use it, and delete secrets they no longer need.
• Admins can create user groups and assign them to resources to manage access.
• Browser push notifications for agent updates are available in beta. Turn them on in Settings.

Improvements

• Sessions can clone repositories from more than one GitHub organization.
• Incident Investigator can investigate incidents you describe directly in Cosmos, not only from alert threads.
• Agents can create private Slack channels and start group DMs. A Slack workspace admin needs to reconnect Slack in Settings > Connectors to turn this on.
• Runlayer is now in the MCP catalog.
• You can filter connectors that need attention, and MCP errors explain how to fix them.
• Environments can repair their repositories, and failed refreshes tell you what to do next.
• Sessions stay active while you use a tunnel.
• Session snapshots are now kept for 14 days.

Bug fixes

• You can save changes to a shared Expert, and saving an Expert no longer changes its hidden triggers.
• Organization members can edit organization files.
• Notifications you've read are now marked as read in all your open tabs.
• File previews update when someone else edits the file.
• Logging out returns you to Cosmos.
• Fixed the Settings menu shortcut and added compatibility for Dvorak keyboards.
• DevRev webhooks no longer fail signature checks.

Tell us how your agent did

Rate any completed response with a thumbs up or thumbs down. If a response missed the mark, choose a reason such as "Didn't follow my instructions" or "Too slow or stuck." This helps us improve where agents fall short.

Improvements

• Teammates joining an existing workspace get a shorter onboarding that focuses on connecting their own accounts.
• 28 more vendor-hosted MCP servers are available, including Mixpanel, Pendo, BugSnag, Swagger, and Zephyr.
• Retired models no longer appear in model pickers, and Experts using a retired model switch to its replacement automatically.
• Cloud sessions start and resume faster.
• The Expert integrations picker shows a Connect button for Jira and Confluence when they aren't connected.
• File comments now open in a panel on the right.
• "Back to app" is now at the top of the Settings sidebar.

Bug fixes

• Resumed conversations keep their extended reasoning instead of losing it.
• Sent messages no longer stay stuck in the queue.
• File attachment details no longer appear inside your messages while they stream.
• Diffs now load for sessions that don't clone a repository.
• MCP reconnects now save correctly.
• Workers started by an Expert stay in the same Space and use the model the Expert specifies.
• Built-in Expert templates now scope Bitbucket triggers to the intended events; reapply existing Bitbucket Experts to stop unwanted sessions.
• Onboarding no longer shows a blank page after checkout.
• EU links no longer get stuck in a redirect loop.
• GitHub links on iOS no longer open an empty browser tab.

New Features

• Reasoning Effort Flag: Added a flag to control the model's reasoning effort from the CLI.

Bug Fixes

• Fixed a crash that could occur when replaying a session summary that included unsupported thinking content.
• Removed the `secret get --reveal` flag since it did not work.
• The daemon now validates the host's Git version at startup and reports a clear error if it's too old.
• Fixed the agent failing to parse file view ranges from malformed tool calls.
• Improved error messages when an MCP server's header secret configuration fails.
• Fixed prompt module output being cut off before completion.
• Fixed duplicate error notifications when a VFS sync failed.
• Extended the knowledgebase download timeout to prevent failures on large codebases.
• Patched security vulnerabilities identified in a routine dependency scan.
• Fixed several issues where VFS syncing could destroy or resurrect deleted files.

Measure what your agents deliver with Outcome Analytics

Outcome Analytics is out of beta and open to everyone on your team. Results on the Outcomes tab are grouped by category. You can add your own metrics as columns and select a single outcome to see its trend over time. Search, filter, and export results to CSV, then drill into the sessions behind each outcome. Session costs remain visible only to admins.

New features

• Choose a reasoning effort level when selecting a model for sessions and Experts.
• Connect self-hosted GitLab instances, Apollo, and more hosted OAuth providers.
• Experts can now be set to use every integration you have connected.

Improvements

• Secrets can no longer use names reserved by the platform, and you get a warning before a name overrides a built-in credential.
• Automations show who created each trigger and when it was last updated, and run history shows who owns each session.
• Every webhook trigger type now uses the same setup flow.
• Deleting an Expert also removes its Automations.
• Deep Code Reviewer now runs on GPT 5.6 Sol.
• Press Escape to leave focus view in Files.

Bug fixes

• Chat messages are no longer dropped or repeated while a response streams.
• Leaving Settings returns you to the page you came from.
• Sessions that need an isolated workspace are no longer sent to machines that can't create one.
• You can remove unavailable repositories from an environment.
• MCP servers keep their scope and Space settings when you add or reconnect them.
• Sessions can no longer be added to folders in a different Space.
• Slack event details with non-English text are no longer cut off.

Changes below are new since the last stable IntelliJ plugin, v0.482.3-stable.

Improvements

• Command hooks supplied by a repository's own `.augment/settings.json` now run only in projects you have marked as trusted. In an untrusted project those hooks are blocked; hooks you configure yourself at the user, system, or plugin level continue to run as before. Changing a project's trust state takes effect on the next hook event, with no restart needed.
• MCP servers that announce a changed tool list during a session are now picked up automatically. Previously the tool list was read once at startup and could go stale for the rest of the session.
• MCP servers are now verified with a real round-trip at startup, so a server that completes its handshake but cannot actually service calls no longer shows as connected.
• Removing an MCP server now sticks. Previously a removed server could reappear when the set of active servers was rebuilt.

Groundwork

• Included early groundwork for a future guided migration experience. It is not enabled in this release and has no effect today.

Reliability & Performance

• Chat now recovers quickly when a response stream drops mid-way. Previously an interrupted request could keep spinning until it eventually surfaced a "Request appears to have stalled" message; it now surfaces a retriable failure promptly so you can try again.
• Your most recent conversations are no longer lost on restart after a long chat history grows past the webview's storage limit. Previously the conversation list could revert to an earlier point.
• Saving a conversation that times out or fails is now retried automatically instead of being marked saved, and concurrent saves of the same conversation no longer pile up.
• Long-running MCP tool calls now respect the configured timeout, and a tool result that does not match its declared schema is passed through instead of failing the whole call.
• Long conversations that cross a summarization boundary no longer fail with a provider error on models that return reasoning content.
• Chat startup is more tolerant of a slow local connection between the chat panel and the plugin, and retries only the cases that can succeed on a retry.
• Refreshing a project's source roots now tolerates files that have gone away mid-refresh instead of failing the refresh.

Bug Fixes

• Fixed the chat panel missing indexing progress and the initial codebase summary when the panel was opened after indexing had already started. The current indexing state is now replayed once the panel is ready, including after a panel reload.
• Fixed the Context Usage popover overstating how full the context window was by counting MCP tool definitions twice. Tool definitions are now counted once, and the breakdown reconciles with the reported total.
• Fixed a "Failed to render" error in chat caused by a code block whose language tag was unrecognized. Unknown languages now fall back to plain text.
• Fixed a "Failed to render" error in chat that could appear for conversations saved by an older version of the plugin.

Compatibility

• Added support for JetBrains 2026.2 IDEs (build 262). The plugin was previously capped at 2026.1.
• The plugin now requires JetBrains IDEs 2025.2 (build 252) or newer. It previously installed on 2024.2 and later. If you are on an IDE older than 2025.2, this update is not offered and your currently installed version keeps working; update your IDE to 2025.2 or newer to receive further plugin updates.

New Features

• Daemon Pool Bundles: Added declarative configuration for daemon pools.

Improvements

• Environment Rebuilds: Environment rebuild operations now run asynchronously.
• Plan-Mode Editors: Plan-mode editor paths now expand `~` to the home directory.
• MCP Scopes: Private and shared MCP scope changes now work when organization-level access controls are disabled.
• Local Daemons: Prevented duplicate local daemons from starting.
• Explore Agents: Corrected the tool allowlist for explore sub-agents.

Bug Fixes

• OAuth Recovery: Improved recovery when hosted MCP OAuth sessions expire.
• Azure DevOps Projects: Preserved Azure DevOps repository project information.
• Authentication Security: Auth session tokens are now written with restricted file permissions.