What's New

- Workspace UI now shows Auggie credit usage stats, with per-agent breakdowns on hover.
- Home view gets workspace status and hover cards so you can see what's happening across your spaces at a glance.
- Codex model picker now uses your installed codex CLI when available, so newer models like GPT-5.5 show up automatically.
- Links in notes are now underlined so they’re easier to spot.
- Workspace creation messaging is clearer when running without a worktree. - - Security: tightened the HTTP MCP bridge so it no longer exposes itself broadly on the network.
- Fixes include idle agents no longer appearing as "responding" in the overview, animation glitches and noisy queued-stream logs, home workspace column widths, .app files now detected by content instead of extension, and stale agent replies no longer overwriting newer user messages.

• Model picker now surfaces the actual provider error (e.g. invalid key, network failure) instead of a generic "Failed to load models" message.

• Fixes: duplicate assistant messages appearing in chat, double initial agent stream when opening a workspace.

+ 0.3.8 is out.

• Diff viewer performance improvements.

• Copy the current browser panel URL with a keyboard shortcut.

• Model picker no longer shows stale or cached values from previous sessions, and unknown model names now fuzzy-match to the closest available model instead of being silently dropped.

• Fixes: ACP tool-call updates not matching correct agents, agent wake notifications occasionally creating duplicate messages, deleted workspaces reappearing when multiple were removed in quick succession.

+ 0.3.7 is out.

• Changes panel loads noticeably faster on workspaces with many agents, and the file explorer no longer re-renders the entire tree every time a file is saved.

• Agent state is preserved correctly when switching between workspaces instead of momentarily showing the wrong agent.

• Fixes: duplicate message flashes during streaming, Emacs-style Ctrl key shortcuts in the editor firing on Windows/Linux where they shouldn't.

New Features

- Transcript Persistence: New `keepTranscriptOnNewSession` setting preserves conversation history when starting a new TUI session.

- Non-Interactive Login: Added `--no-tui` flag to `auggie login` for headless and scripted environments.

Bug Fixes

- Fixed TUI flickering during indexing, typing, and sub-agent work.

- Fixed stale agent state persisting after starting a new session with `/new`.

- Fixed `rebuild` command failing when script file paths were passed instead of file contents.

- Fixed queued request IDs not being preserved across retries.

Today we're shipping Prism: a new option in the Augment model picker that efficiently routes each turn to the model that fits the work.

On our internal multi-turn coding benchmark, Prism matches the best individual model on quality at 20–30% lower cost per task than frontier models.

Teams sending 10,000 user messages a month can expect to save $20,000 on their token spend, at similar or better quality.

We know that developers and teams have strong preferences for different model families: with Prism, you can stay in the model family you like, at lower cost.
→ Prism (GPT + Kimi) targets GPT 5.5
→ Prism (Claude + Gemini) targets Opus 4.7

Prism's job is to switch only when the expected win from a different model exceeds the cost of the cache eviction.

Augment can do this because we are model agnostic. No single model wins every task, so we give our customers access to all the industry leaders. That's the foundation Prism is built on: a pool of frontier models, with the routing decision made per turn rather than at the start of a session.

Prism is in the picker today: VS Code, JetBrains, CLI (/model), and web.
Billing rolls up under a single Prism line item. The underlying model that handled any given turn isn't surfaced.

Improvements:

- Model Defaults: Default models updated to Claude 4.5/4.6.

- Image Paste: Cmd+V is now accepted as an image paste shortcut in addition to the existing binding.

- Shell Support: Shell tools now support `sh`, and the TUI uses your configured shell for async commands.

- Billing Summary: Account display now shows a billing summary with credits and cost breakdown including sub-agent rollup.

- MCP Environment Variables: MCP servers support environment variable expansion and working directory injection.

- Indexing Visibility: Improved status visibility when codebase indexing gets stuck.

- Session List Pagination: Large session lists are now paginated to avoid timeouts.

Bug Fixes:

- Fixed ESC key behavior in Enhance mode and slash command menus no longer interrupting the agent.

- Fixed image attachments not passing through the TUI message queue.

- Fixed tool-call blocks disappearing in cloud-connect TUI history.

- Fixed `session delete` failing when workspace ID is missing.

- Fixed duplicate rows appearing in the CLI transcript queue.

- Fixed API retry error messages appearing raw in TUI and headless output.

- Fixed session errors not being surfaced to users.

- Removed spurious "Logging configured" message on CLI startup.

- Fixed `/stats` message count accuracy by excluding tool-result continuations.

- Fixed `/web` fallback link formatting.

- Fixed `view_range` display for end-of-file ranges in tool call titles.

• Model picker no longer swaps your selected model when a provider is slow to load — auto-fallback now waits for the user's provider to actually settle instead of reacting to transient empty results.

• Cmd+F in chat now reliably scrolls to every match, including those inside virtualized turns that were previously skipped.

• Agent renames and deletes propagate to other windows within seconds instead of minutes.

• Fixes: Stop button clicks during agent creation are honored instead of being silently lost; delegation-group race that left coordinators stuck waiting on already-finished sub-agents; custom agent names preserved across subsequent renames; infinite render loop (effect_update_depth_exceeded) when opening some workspaces; hardened PR discovery and assorted changes panel bugs from the underlying git state refactor.

New Features

- Added `/settings` built-in slash command to quickly open extension settings from chat

- Added connectivity check to Augment servers before completing login to surface network issues early

- Added recovery of forked conversations so the parent conversation's exchanges leading up to the fork point are restored alongside the forked exchanges

Bug Fixes

- Fixed auth header lookup being incorrectly attempted for stdio MCP servers

- Fixed MCP OAuth and env-var secret key casing to use uppercase consistently

- Fixed token counting failures with gpt-tokenizer by adding character estimation fallback

- Fixed prompt enhancer not receiving conversation context in webview

- Fixed plan mode tools incorrectly appearing in cloud agent sessions

- Fixed dropped code selection when sending a chat message after clicking into the chat input

- Fixed settings drawer missing padding on narrow viewports

- Fixed first line of chat code blocks being clipped under the sticky header

- Fixed pinned open files not being included in retrieval context or surfaced to the agent

- Fixed update notifications pointing pre-release users at stable versions instead of newer pre-release builds

- Fixed repository allowlist rejecting Bitbucket Server URLs with ports, bare hostnames, or `/scm/` / `/git/` path prefixes

- Fixed MCP OAuth refresh failing when the authorization server omitted `refresh_token` from a refresh response

- Fixed MCP server headers being lost when exporting servers via "Copy JSON"

- Fixed checkpoints not being created in some conversations after conversation-ID migration

- Fixed checkpoint display index being off-by-one (first checkpoint showed as "Checkpoint 0")

- Fixed the Edits tab dropping updates during bursts of refresh dispatches

- Fixed summarized text being shown as the user message when recovering summarized conversations

- Fixed stale local git repository info (branch, remote, root) being injected into the GitHub API tool description and cached for an hour

- Fixed noisy "Failed to load hooks configuration" toast appearing on webview open for users who had never configured hooks

Improvements

- Added 'service unavailable' to retryable gRPC error patterns for improved connection resilience

- Improved accessibility across common UI components (buttons, navigation, dropdowns, text areas)

- Improved error messages by surfacing fetch error cause details (code, message) in logs and user-facing errors

- Improved data table column sizing and layout

- Show a recovery banner when no chat history is found in the backend for the current conversation

- Show the prompt enhancer (rewrite prompt) button while an exchange is in progress

- Surface hooks load errors inline on the Hooks settings page

- Sort built-in commands alphabetically in Settings → Commands

- Make the stream interruption error message generic instead of incorrectly attributing interruptions to token limits

New Features

- Added built-in slash commands, including /fork and /ask.

- Added a Hooks settings view in the IDE for managing lifecycle hooks.

- Restored Figma OAuth support in IDE settings.

Performance

- Improved thread panel and chat history performance for large numbers of conversations.

- Reduced settings UI reflow and sidecar process overhead.

Bug Fixes

- Improved automatic retry behavior for mid-stream network failures and temporary service unavailability.

- Fixed project and source root detection in multi-project IntelliJ workspaces.

- Fixed MCP startup, OAuth secret, environment secret, and encoded secret handling.

- Fixed Ask User tool completion, error, persistence, input, and rejected-action states.

- Fixed sub-agent tools being marked canceled early and duplicate sub-agent entries in the panel.

- Fixed terminal and shell tools appearing stuck when commands were waiting for input.

- Fixed PowerShell quoting for Windows command execution.

- Fixed chat recovery for first chats, long histories, recovered tool states, and SSH sleep/wake interruptions.

- Fixed recovered chats showing incorrect timestamps.

- Fixed edited messages using the wrong draft model.

- Fixed message editing while messages are queued.

- Fixed Cmd+Backspace deleting more than the current input line.

- Fixed native shortcuts like Cmd+V and Cmd+A in thread search and rename inputs.

- Fixed thread sorting, age buckets, pin state, deletion navigation, and unread badge counts.

- Fixed empty new conversations showing an error overlay.

- Fixed rules, guidelines, and hierarchical rule path handling.

- Fixed save-file operations adding an extra newline at end of file.

- Fixed Figma detection false positives.

- Fixed chat UI rendering errors that could leave the webview blank or broken.

Other Improvements

- Improved accessibility across chat, settings, navigation, tabs, drawers, and icon buttons.

- Improved readability of tool call displays in chat.

This is the big refactor we've been working on for the past 2 weeks, plus a new setup flow for new users.

Changes:

- Setup flow redesigned for a smoother first-run experience, with fixes for opencode-only setups.

- Agents are more resilient under memory pressure — they are no longer killed mid-response, even in background workspace tabs.

- Terminal titles are now sanitized to prevent credentials from leaking into the window title.

- Agent suggested prompts are back.

Fixes:

- Repeated identical messages are no longer silently dropped.

- User messages are preserved during concurrent saves.

- In-flight message queueing is more reliable.

- Chat messages merge correctly with on-disk content when reopening a chat.

- Delegated agent wake-up and subscription reliability improved.

- Spec and note updates now reactively reflect task changes.

- Dropdown crash from duplicate options resolved.

- Custom specialists stranded during a prior migration now recover correctly.

- Fixed a streaming hang when the "done" notification arrives before the stream closes.

We’ve switched Augment’s default model to Opus 4.7. And it’s now 50% off until April 30th. Try it in Intent, the Auggie CLI, or our IDE extensions. Existing threads won’t change unless you switch models!

Built for async workflows, CI/CD, and automation

Most models can generate code quickly. But in real workflows, things break down - plans drift, tool state gets lost, and work restarts halfway through.

In our testing, Opus 4.7 is noticeably better at staying on plan over longer horizons, adapting to changes, and finishing with verification instead of restarting mid-run.

The result:

• fewer tool-state failures
• fewer “re-derive the approach” turns
• more reliable execution in CI/CD and long-running tasks

It’s also a strong everyday model, with a more opinionated default—more likely to challenge assumptions and suggest better approaches.

You’ll see the biggest impact in Intent workflows, where agents execute work end-to-end. Opus 4.7 is 50% off until April 30.