Release notes
Versions roll out to Positron and RStudio separately, so a change may arrive on one before the other. Releases before 1.10 list Positron changes only.
1.12.0 — 2026-11-12
Added
- New conversations in trusted workspaces start in Auto approval mode, unless you or an administrator have chosen another mode. Untrusted workspaces still start in Restricted
- The assistant can find and read your past conversations in the current workspace with the new
recalltool and/recallskill. TUI:pa recalllists, searches, reads, and exports saved conversations from the command line - Posit Assistant loads
AGENTS.mdfiles in subdirectories of your workspace when the assistant reads files there - Custom header values in
providers.jsoncan read environment variables with{env:NAME} - macOS: an opt-in
"cobox"sandbox backend for shell commands - Positron: “Open in New Window” in the sidebar moves a conversation into its own window in one step
- Positron:
/settings-reportshows Positron’s own AI settings, such aspositron.assistant.*andchat.disableAIFeatures, alongside Posit Assistant’s - The model picker has a “Manage Posit AI Pass account” item
- Support for Kimi K3.1 Pro on Posit AI Pass
- Code blocks have a “Wrap lines” toggle for long lines in diffs and file reads
- A new
ui.reduceMotionsetting (auto,on, oroff) controls animations. Activity indicators now pulse instead of freezing when reduced motion is on - Experimental:
bash,powershell, andexecuteCodecan run in the background, and a task indicator shows running tasks. A shell command that runs longer than 30 seconds moves to the background automatically
Changed
- The auto mode classifier uses Gemini Flash 3.8 or later for Gemini, retries once after a malformed response, and is less strict about actions you asked for or approved
- Declining a tool call also skips the other undecided calls from the same response
- YOLO mode and
PreToolUsehooks that allow a call also waive the extra confirmation the assistant can request for a risky call - Ask mode runs Positron commands that Positron marks read-only
- The sandbox shield explains why an enabled sandbox is unavailable, and
/sandboxtoggles sandboxing even when it is - The
@file picker ranks files as you type and can reach files deeper in the workspace - Posit AI Pass shows only the newest Claude Haiku, Sonnet, and Opus in the model picker
- Copy Selection and Copy Response copy only the label of a file link
- Positron and RStudio: conversation tabs line up with the surrounding editor chrome, and the tab strip no longer overlaps the sidebar
- A Posit AI Pass sign-out now reads “Signed out, sign in again”, with a Sign in button in Positron and “Enter to sign in” in the TUI
Fixed
- Claude no longer fails with a 400 “The block is bound to a different conversation” error on Amazon Bedrock, on newer Claude models, or after you edit an earlier message
- Positron: Posit AI Pass no longer signs you out when Positron starts, and windows no longer reuse a refresh token another window replaced
- Positron: a turn no longer hangs when an R session crashes while the assistant runs code. It ends with the crash error
- A running reply stays intact when you switch away from its conversation and back, or reload the view
- Amazon Bedrock: an expired AWS SSO session is reported when only an AWS profile is set, and OpenAI models appear right after you sign in again
- A blank
providers.jsonno longer causes an error - Shell commands using zsh-only syntax, such as glob qualifiers and
=(...), ask for approval instead of passing as read-only - Path restrictions apply across case and Unicode variants of a path, such as on macOS
- Files pinned by an administrator’s
askrule no longer appear in search results without approval, and headless runs decline all pinned asks - Codex engine: the hook is found on Windows
- TUI: WebP images decode without replacing the global
URL, and the TUI no longer prints aMaxListenersExceededWarningon launch
1.11.1 — 2026-11-05
Fixed
- Positron: On Posit Workbench, Microsoft Foundry no longer shows “Credential pending” with no models when Posit Assistant starts before the Workbench extension
1.11.0 — 2026-11-04
Added
- Databricks: Unity AI Gateway routes are supported alongside classic serving endpoints, so newer gateway-only models work
- In Posit Workbench with Posit Connect identity federation, the Posit Connect provider lists Amazon Bedrock models from Connect AWS integrations and sends them through the Connect LLM Gateway
- Support for Gemini 4 Argon, with pricing
- Tool calls decided by the auto mode classifier show a shield icon, and expanding one shows the classifier’s reason
- TUI: Shift+Enter inserts a line break in the chat input
- TUI: Ctrl-Z suspends to the shell, like other full-screen terminal programs
Changed
- Auto mode reviews tool calls with Claude Haiku (5.5 or later) instead of Claude Sonnet wherever your provider offers it. In our evals it was just as accurate, and approvals are faster and cheaper
- The auto mode classifier sees the assistant’s AskUser questions and your answers
- Restricted mode respects
allowrules from admin, global, and project settings, and never asks approval for a tool that has been explicitly allowed - Shell commands that exit with a non-zero code show the exit code instead of being marked as failed. Commands that time out, are canceled, or fail to start are still shown as errors, even if they exit with code 0
- Copying a chat selection that includes prose always copies Markdown, whether you use Cmd+C, right-click Copy, or the toolbar’s Copy. A selection inside a single code block copies the raw code
- The GitHub Copilot engine runs shell commands with Posit Assistant’s shell tool, so they follow the same permission rules, sandbox, and environment as native chats
- The AI Providers panel lists providers as they respond instead of waiting for the slowest one
- Built-in AI providers are listed alphabetically, with Posit AI Pass first
- Positron: On Posit Workbench, the Workbench-managed Microsoft Foundry token is sent only to the Foundry endpoint Workbench configured. A Foundry base URL pointing elsewhere shows Foundry as unavailable, with an explanation
Fixed
- With an agent engine enabled but no AI provider set up, the setup prompt no longer covers the chat input
- Google Vertex AI: a missing
GOOGLE_APPLICATION_CREDENTIALSkey file is named on the provider row and in the Edit dialog, instead of being reported as a network error - Signing in to a server that issues short-lived OAuth tokens no longer triggers bursts of repeated token refreshes
- Closing a conversation tab or panel frees its memory
- Typing very fast in the chat input no longer causes a “Maximum update depth exceeded” error
- Accessibility: the chat input has an accessible name, AI Providers dialog buttons name their provider, the font size stepper is reachable with the keyboard, message actions and the history More button appear on keyboard focus, and a hidden tip’s Dismiss button no longer takes focus
- Positron: Posit Assistant’s memory use in the extension host no longer grows with every message, and the pane no longer goes blank after recovering from a long freeze
- Positron: On Posit Workbench with Microsoft Foundry set up by an administrator, existing Positron AI provider settings, including model overrides, are kept
- Positron: File tool errors show file paths instead of
file://URIs - TUI: Built-in skills are found in binary-only installs
- TUI: Conversations with many tool calls no longer print a
MaxListenersExceededWarningabout resize listeners over the screen
1.10.0 — 2026-10-29
Added
- Positron: OpenRouter is now available as a built-in AI provider on Positron 2026.11 and later
env.filterin your global settings keeps API keys and other inherited variables out of the processes Posit Assistant starts: shell commands, hooks, local MCP servers, and its own R and Python processes. Posit Assistant’s own configuration and authentication variables are now withheld automatically; see Command Environmentenv.setin your global settings sets variables for those same processes, with{env:NAME}to insert a variable from Posit Assistant’s environment; see Setting variables for commands- Local MCP servers accept
environmentas an alias forenv; a server that sets both is skipped with an error - Positron: The About dialog shows Positron’s build number after its version, such as “2026.11.0 build 223”, and includes it when you copy the details
- Positron: The assistant can see every open console and notebook session and choose which one runs code, instead of waiting on a busy foreground console
/settings-reportshows where each setting and provider credential comes from, with secrets masked, and can be exported as Markdown. In Positron, “AI: Create Diagnostic Report” includes it- Agent engines (Claude Agent, Codex, GitHub Copilot, and ACP agents) no longer require
experimentalFeatures, and each can be configured from the AI Providers dialog. In Positron this needs 2026.11.0-133 or later - Positron: A Sign in to GitHub button on the GitHub Copilot engine row in the AI Providers dialog
- The thinking-effort control in the model picker works for GitHub Copilot engine models
- Support for GPT-6.1 Terra on OpenAI and Amazon Bedrock, with pricing
- Support for MCP servers that use the 2026-07-28 protocol. Saved remote-server sign-ins are tied to the authorization server that issued them
Changed
- A provider row shows “API key rejected” when the provider rejects the saved key while listing models, instead of “Connected”; its Edit dialog says which key was rejected
- The API-key provider configure dialog shows where the key comes from, and when a saved key overrides an environment variable, Details explains it and offers removing only the saved key. Provider rows no longer show “via ” or the long clear label, and Anthropic’s key link points to platform.claude.com
{env:NAME}in MCP server configuration reads the variables a local server would inherit afterenv.filterandenv.set, instead of Posit Assistant’s whole environment. A reference to an unset or withheld variable now stops the server with an error naming the variable instead of expanding to an empty string; see MCP server environment variables- Fuzzy matches in the slash-command and file pickers and in conversation search rank higher when the typed letters start words, sit close together, or come early in the name, and the highlight shows those letters
- Positron: The assistant searches file contents in local workspaces with the
greptool, backed by Positron’s bundled ripgrep, instead of thesearchtool. To switch back, turn offassistant.fastSearchand reload the window. Remote-only workspaces keepsearch - Positron: Unlike
search,grepsearches only the working folder, reads files from disk rather than unsaved editor changes, and matches line by line. By default it skips.gitignored files and dot-directories other than.github,.gitlab, and.circleci. To find files by name, the assistant uses the shell - Positron: Command rows and approval prompts show a Positron command’s name, such as “Select Interpreter Session”, instead of only its id, and label each argument by name
- Positron: Ask mode runs Positron commands that Positron marks read-only, such as reading the Viewer, and still refuses every other Positron command. Which commands are marked depends on the Positron build
- RStudio: The
greptool uses ripgrep or the systemgrepby default, without experimental features. To switch back, turn offfeatures.fastSearch; see Search - RStudio: RStudio keyboard shortcuts, such as F6 to move between panes and Ctrl+2 to focus the console, work while focus is in Posit Assistant (requires a recent RStudio build). Posit Assistant no longer uses Ctrl+Shift+1 through 9, leaving them to RStudio’s pane zoom
env.filterandenv.setalso apply to the Claude Agent, Codex, GitHub Copilot, and ACP agent engine processes- Conversation tabs are shown by default in Standalone, RStudio, and the Positron sidebar
- The overflow menu is regrouped, with import and export items in a submenu, and the Thinking effort submenu lists levels from lowest to highest
- Hovering a top conversation tab swaps its status icon for the close button in place instead of widening the tab
- Shell command rows show the command’s description instead of the raw command
- Shell output over 20,000 characters keeps the start and end and names the omitted lines, instead of cutting off the end
- The Snowflake model picker lists the models your account can use, with refreshed pricing
- An expired AWS sign-in prompts you to sign in when you press Send, instead of failing partway through the reply
Fixed
- Databricks OAuth sign-in uses the correct discovery address and avoids repeated delays when discovery is unavailable
- A clear message when a Bedrock model isn’t enabled in your AWS account, instead of a raw AWS Marketplace error
- Error notifications from a subagent say they came from a subagent and name its model
- Rate-limit and payment error notifications include the provider’s explanation, such as which limit was hit and when to retry
- Starting a new conversation or opening one from history replaces the current tab’s title in place instead of closing the tab and opening a new one
- Opening a conversation with unread activity no longer briefly shows the unread dot in its tab
- Dragging a conversation tab to a new position no longer makes it or a neighboring tab briefly collapse and grow back, or settle in the wrong place
- Positron: File excludes (
positron.assistant.aiExcludes) enforced by a Workbench administrator now block Posit Assistant’s read and edit tools on matching files, instead of asking for permission - Positron: Disconnect on the GitHub Copilot row hides Copilot models again, and setting
providers.copilot.enabledtofalsehides them too - Positron: The AI Providers dialog’s bottom corners now follow the dialog’s rounded frame instead of poking out square
- With
sandbox.enabled, shell commands are refused with an explanation when the sandbox is unavailable (such asbwrapmissing on Linux), instead of running unconfined - PowerShell commands that end in a statement such as
foreachorifno longer fail with an error about2>&1 - R plot errors raised at print time no longer garble the output
- Dark mode: native scrollbars follow the dark theme, and the Plugins, Settings Report, and AI Providers panels in Positron render in dark colors
- Links and quotes in your own messages are readable in the default dark theme
- Positron: Cmd-Z in the chat input undoes one step at a time instead of removing the wrong text
- Positron: Posit Assistant keeps the active tab when it runs code
- Positron: A tab hidden for several minutes in Chrome no longer shows a false “Request timeout” error when you return
- TUI:
-mselects the model in the interactive TUI. Headless runs honor the sandbox setting, and long conversations no longer triggerMaxListenersExceededWarning
Releases before 1.10 list Positron changes only.
1.9.1 — 2026-10-23
Fixed
- GitHub Copilot models missing after updating to 1.9.0; after Positron starts, they can take a few seconds to appear
1.9.0 — 2026-10-21
Added
- Chat font size control
- Option to sort conversation history by creation time or by last activity
- Quarto documents in the visual editor are shared as file context
- The assistant can work in Quarto documents that show output inline
- On Posit Workbench, Assistant gateway requests identify the Workbench session and project
Changed
- Posit Assistant now handles sign-in for AI providers, instead of Positron’s authentication extension
- Positron commands and Positron-provided skills are on by default (
assistant.advancedPositronIntegration) - New conversations start with the model of the conversation you are looking at; the
preferredModelsetting, when set, always wins - Removed the experimental direct GitHub Copilot provider and its
assistant.enableCopilotSdksetting - Provider connection URLs and Vertex settings are saved in
providers.json - Admin provider rules from environment variables are kept when they contain unknown settings, which now produce warnings
- The Workbench gateway metadata header is now
Gateway-Metadata, set byPWB_GATEWAY_METADATA_HEADER - Posit AI Pass offers DeepSeek V4.1 Flash in place of V4 Flash
- OpenAI models on Amazon Bedrock show detailed reasoning summaries
- Conversation cost includes edited-away and deleted messages
- Notebook context uses fewer tokens
- The date is sent with each turn instead of in the system prompt, so the prompt cache survives a change of day
- New speech-bubble icon for new conversations
- Send, attach, and web search buttons have accessible names for screen readers
Fixed
- Configure AI providers freezing Positron 2026.06 and earlier
- Open conversation tabs showing an outdated model list after the model catalog changed
- Exported conversations revealed in the wrong folder on Windows and Positron Server
- HTTP 400 errors on Bedrock GPT models after tool calls with empty output
- Auto-mode classifier responses with commentary before the JSON block being treated as malformed
1.8.0 — 2026-10-13
Added
- Claude Haiku 5.5 support, with pricing for its higher rate on prompts over 100K tokens; its Posit AI Pass cost multiplier shows the range across both rates
!command mode in the chat input:! <command>runs a shell command and!r/!pythonrun code, without a model request- Provider-native web search for OpenAI, Amazon Bedrock, and Gemini models
- The assistant can see what you are viewing in the Data Explorer
apiKeyHelperinproviders.jsongets a provider’s API key from a command, for rotating or vault-issued keys- Portkey key-type option, for connecting through a Portkey gateway
- Clear notice when an AWS guardrail blocks a Bedrock response
- Bedrock access denials and expired AWS logins are explained in chat
- The MCP section reports servers that were skipped and why
- Guidance for deploying Python content to Posit Connect Cloud
Changed
- Low-difficulty subagents, web search, and summaries use Claude Haiku 5.5 when available
- Claude Sonnet 5.5 cache reads are priced at the new $0.10/MTok rate, including in costs of earlier conversations; Posit AI Pass cost multipliers are relative to the cheaper Sonnet 5.5
- Web fetch answers its prompt with a low-tier model reading the page, using fewer tokens; its usage appears in the Session HUD
- Subagents get their own system prompts and the project’s memory files
- Simpler subagent tool and shorter system prompt
- The active file is described only when it changes
- Long user messages scroll instead of expanding the chat
- MCP servers with a
urland nocommandare treated as remote - The AI Providers panel explains why a save failed when
providers.jsonis invalid - Requests to Bedrock, Vertex, Ollama, and LM Studio identify Posit Assistant
- Removed the todo list tool
Fixed
- “Off” thinking effort now turns thinking off on Claude models that think by default, such as Claude Opus 5 and Haiku 5.5
- Costs are now reported for Claude models on Google Vertex, including earlier usage
- Claude 400 errors when earlier thinking no longer matches the prompt
- GPT-6.1 Sol on Bedrock failing when thinking is “Off”
- Windows file paths outside the workspace opening without their drive letter
1.7.0 — 2026-10-07
Added
- Sign in to Posit Connect with a server URL and device-code OAuth
- The agent can see and use apps in the Viewer pane through Positron commands
- Faster
greptool, now available in Positron, using ripgrep - Terms of service notice when signing in to third-party providers
- Warning when Posit AI Pass setup opens on the wrong posit.ai account
- GPT-6.1 Sol pricing, and the model is promoted in model pickers
- Clickable user-message previews in the Session HUD
Changed
- Conversations open in new editor tabs instead of nested conversation tabs
- Long conversation names are truncated in editor tab titles
- Built-in model tiers prefer the standard model over Pro variants
- The editor selection is shared as a position rather than its text
- Session variables are sent once, then only what changed, to reduce token usage
- Shorter system prompt and tool descriptions, and cheaper message and conversation summaries
- Write and edit tools can write to the OS temp directory without prompting
- Yolo mode auto-approves shell commands that can’t be split into subcommands
- Token counts under 1K are shown as fractional K
- Tool file access is more strictly checked
- Rate-limit error message nudges toward Continue
- Copilot models get more time to load on remote and server hosts
Fixed
- Plan mode failing to exit on OpenAI-compatible providers
- Unix socket failures in the assistant’s shell on macOS, and writes to
/dev/nulland/dev/fdblocked in the macOS sandbox - 400 errors after Claude runs a
web_searchalongside other tools - Duplicate ggplot output when running R code
- Misplaced trailing quote in tool call captions
- Code blocks in chat keep the previous theme’s syntax colors after switching between light and dark themes
- Low-contrast title on the active conversation tab in dark themes
1.6.1 — 2026-10-01
Fixed
- Providers sometimes showed no models at startup when credential checks are slow, until “Reload model list” was used
1.6.0 — 2026-09-29
Added
- Claude Sonnet 5.5 support; it is now the default model
- GPT models get an
apply_patchtool for multi-file edits, and file names in its tool calls are clickable features.submitKeysetting to choose the key that submits chat input- “Reload model list” item in the model menu, which now stays pinned while scrolling
- The agent knows about Posit Package Manager
- The
posit_assistantR option is set while code runs in the R session
Changed
- Chat input is plain text: no auto-formatting, and newlines are preserved
- Markup typed in chat messages is shown as literal text
- Quieter thinking row with tighter spacing
- Tool icons align with the text, and running tools use a shimmer animation instead of a spinner
- Improved subagent rows, with a Task panel and streaming launch rows
- Abnormal finish reasons (such as truncated or filtered responses) are now shown in chat and in subagent results
- Slimmer system prompt
Fixed
- Setup prompt flashing on Positron startup
- Timeouts when opening very large conversations
- Model lists and message sends staying consistent when provider settings change
- Agent Layout opening duplicate Assistant tabs
- Migration banner reappearing in new panels
- Multi-line instructions after slash commands like
/reportbeing dropped - GPT OSS tool calls on Databricks
1.5.0 — 2026-09-23
Added
- Claude Opus 5.5 and GPT-6 (Astra, Sol, Luna) support
- OpenCode as a built-in provider
- Estimated costs for Snowflake Cortex models
assistant.previewFeaturessetting; Positron IDE command support now requires Preview features instead of Experimental features- Explore subagents can run shell commands, and canned subagents can be granted additional tools
- The approval dialog explains when auto mode couldn’t evaluate a tool call
Changed
- Turn costs are shown by default, and are labeled as estimates
- Model picker is ordered by capability
- Removed the fast/balanced/smartest labels for Posit AI Pass models
- The agent uses Positron’s package install commands instead of naming a shell install command for the Python environment
- Lighter streaming spinner; top tabs no longer show a hover actions button (right-click menu is unchanged)
- Reset buttons use an undo icon
Fixed
- Sidebar failing to start for workspaces on Windows SMB shares
- Model-discovery hangs and duplicate provider fetches
- Request failures on Bedrock models without cache-point support, such as Kimi K3
- OpenAI prompt cache misses after turns with tool calls, and Claude idle cache keepalives
/notebooknow creates notebooks whose cells run- Skills in symlinked directories (such as
.claude/skills→.agents/skills) are now discovered, and skill rediscovery no longer races - MCP OAuth on Posit Workbench, and MCP servers no longer need re-authorization after a restart
- Search tool calls that pass a single glob pattern as a string
- Accurate per-turn token growth; the tab strip no longer shifts while closing tabs
1.4.1 — 2026-09-10
Fixed
- Bedrock requests for GPT 5.6-series models failing with
Unsupported parameter: 'reasoning.summary'
1.4.0 — 2026-09-10
Added
- Per-turn footer in the chat with cost, token counts, timing, a copy button, and a usage graph; cost is also shown live while streaming
- Subagent task descriptions are shown in a collapsed expando instead of a tooltip
- Gemini 3.8 Flash support
- Support for auth-less custom Anthropic providers behind SSO proxies
Changed
- Modes are ordered least-to-most permissive, with Ask first
- Agent shell commands run with the user’s login shell, captured at launch even when a workspace
.envchanges it - Faster auto-mode permission decisions through parallel classification
- Icon buttons have press feedback; edit, delete, resend, and sibling-navigation buttons are disabled while streaming
- Long file paths in tool calls are truncated on the left
Fixed
- Stuck composer when a prompt is rejected before a turn starts
/compactfailing with “Invalid signature” on Snowflake Cortex, and refusals with attached documents or unpriced models- Read tool output is capped to keep long sessions within context
- Users stay signed in through transient OAuth refresh failures
- Fixed GPT-5.6 context budgets and recovery for oversized chats
- Model tier preferences now reliably select the intended model
- Tool calls now settle correctly when cancellation interrupts setup
- Pending tool-call placeholders are retired once their rows render, including stray placeholders caused by thinking blocks
- Auto-mode salvages decisions from length-truncated classifier responses
1.3.1 — 2026-09-03
Fixed
- Removed plot message about
chartLabel.
1.3.0 — 2026-09-02
Added
- Microsoft Entra ID authentication for Microsoft Foundry
- In-app AWS SSO sign-in for Bedrock
- Conversation cost display and token usage preferences, with a per-turn cost strip and richer chart tooltips
- Live subagent token usage in the cost tooltip
- Thinking-effort control for subagents
- Up-arrow history recall in the chat input
- Editor validation for
settings.jsonandproviders.json, from a hosted JSON schema - Administrator controls for MCP servers and lifecycle hooks
- Plugin management is now available by default
- Assistant can read Positron’s documentation bundled on disk, version-matched to the running build, when answering questions about Positron
- GLM 5.3 and DeepSeek V4 Flash models in Posit AI Pass
- Request log inspection for developers: a command to open the latest request log, plus a browser-based log viewer
Changed
- The Posit AI subscription service is now called Posit AI Pass
- Thinking effort has moved into the model selector and is shown as text
- A single unified subagent row, with effort labels, cost, raw tool data, and live status
- The assistant now prefers Positron’s own commands for environment tasks such as inspecting interpreters, packages, and layout
- Skill reference links resolve against the skill’s own directory
- Refreshed plugins dialog
assistant.posit.cois trusted for JSON schema downloads by default- The
assistant.*plugin and marketplace settings have been removed. Plugin state now lives in~/.posit/assistant/settings.json.
Fixed
- Syntax highlighting now follows the theme under
window.autoDetectColorScheme - The caret stays in view on Shift+Enter and when pasting plain text
- The cursor position is included in open-file context when nothing is selected
- The last conversation tab can now be closed
- Shell tool child processes are given
/dev/nullstdin instead of an open pipe, so they no longer hang waiting for input - Reconnect now refreshes expired AWS SSO and Google ADC credentials
positronCommandresult truncation is structure-aware- Root-targeted compaction links
- Seeded thinking effort and web search are preserved when the default draft model is selected
/dev-updatereports errors and no longer stalls when the update prompt is dismissed
1.2.0 — 2026-08-24
Added
- Ask mode is now generally available
- Search conversation history by name or message content
- User-level memory file (
~/.posit/assistant/AGENTS.md) for cross-project preferences - Menu item for delete current conversation action
posit-assistant.newChatcan be invoked from the command palette- The R/Python session context toggle is remembered per conversation
Changed
- Built-in Bedrock AWS region/profile configuration moved into
providers.json - Faster conversation switching by removing startup network calls
- Faster streaming rendering for tool-heavy responses and live thinking blocks
- Prompt caching for the auto-mode permission classifier
Fixed
- GPT-5.6 cache reads stalling during tool loops
- Posit AI sign-in failures are now actionable and recoverable
- Simplified startup and hardened model discovery
- Models are kept when a credential check times out
- Stray trailing bullet/asterisk artifacts in assistant messages
- Usability papercuts across chat, history, and the composer
- The credential-loss toast no longer appears on provider sign-out
- Updated GPT 5.6 Sol pricing
1.1.0 — 2026-08-18
Added
- Gemini API models: Gemma and Gemini 3.6/3.7 Flash, with cached-token accounting
- Update per-model output token limits for Posit AI models
Changed
- The
/and@suggestion popup is restyled as a sectioned panel anchored to the composer bar - File attachment chips are unified on a single design
Fixed
- Editor webviews now recover correctly after a reload in Positron
- Webviews are resilient to WebKit message-transfer failures
- Shell tool processes are interrupted when a turn is cancelled
- The cache keepalive widget is hidden when the selected model has no keepalive policy
- Stuck client-state mutex and corrupted-file startup failures
- Slash-command double-Enter and focus loss after
/clear executeCodesupport for non-R/Python Positron runtimes- Markdown rendering performance during streaming
- Ollama and LM Studio model registrations now declare their catalog
1.0.1 — 2026-08-13
Changed
- Databricks models now route over native vendor APIs
- The Posit products skill now covers Publisher troubleshooting commands and updated Connect deployment guidance
1.0.0 — 2026-08-13
Added
- Retry, Resend, and Continue actions for failed or cancelled turns
- Host-application skill source, so Positron can ship its own skills
- Support for Posit Package Manager plugin marketplaces (HTTP and non-side-banded git servers)
- CSV export and request timing in the token usage panel
- Experimental LiteLLM provider (built-in and custom) and experimental Portkey gateway provider
- Experimental Bedrock GovCloud support, plus support for all commercial AWS Bedrock regions
/dev-updatenow shows the build time of the available dev build
Changed
- Auto mode, default model, and persona are now global (user-level) settings
- Exploration is merged into
launchSubagentas a subagent type - Provider configuration issues are surfaced in the UI
- Skill directory reads and Posit documentation fetches are pre-approved
- Model name display is more compact, and the sidebar icon appears immediately at startup
- Updated Sonnet 5 pricing and the Kimi K3 context limit; refreshed the Snowflake Cortex model catalog
- The diagnostics bundle command is folded into the Positron diagnostics command
- The Positron command list is no longer included in the system prompt
- The dev build command is renamed to “Install latest dev build”
- The legacy Positron settings reader is version-gated to pre-migration Positron
- The troubleshooting note is hidden on Positron 2026.07+
Fixed
- Model discovery and session startup no longer hang
- “Model with id X not found” error when switching models
- Session creation on Windows by keeping workspace paths real
- Queued messages are restored after stopping a turn
- Duplicate user message after editing on hub-backed surfaces
- Conversation search rendering and focus in Positron
- Notification banner transparency with an opaque underlay
- Run in Console language errors no longer appear in the UI error banner
- Unknown errors now produce normalized diagnostics
- Platform skill file access is scoped to loaded skills
0.10.0 — 2026-08-03
Added
- Snowflake Cortex credential discovery from
connections.toml - Bedrock Mantle support for gpt-oss and GPT-5.x models
getConsoleHistorytool for reading recent Positron console history- User-configurable lifecycle hooks via the
hookskey insettings.json - Plugin and marketplace support for installing bundled MCP servers and skills (experimental)
- Tiered model selection for subagents, so simple subagent work can run on a cheaper model
- Explicit prompt caching for GPT-5.6
/dev-updatecommand to check for and install newer dev builds
Changed
- Token usage panel redesigned with cost-first columns, request counts, and a grid layout
- Model picker reordered: GPT-5.6 moved up, older Claude models moved down
launchSubagentis now always offered, with nesting capped and a shorter label- Images in the
readtool are resized to fit Anthropic’s many-image cap - Remote MCP transports consolidated, with hardened OAuth handling
- Model transport errors now include better diagnostics
- Streaming output has a minimum flow rate for smoother rendering
- Updated GPT-5.6 pricing to reflect OpenAI’s rate reduction
- Posit product knowledge moved into a dedicated skill
- The
positronCommandtool and prompt list are gated behind a feature flag
Fixed
- Cache keepalive timer no longer shows a shorter time when a ping is in flight, and uses the correct color
- Starting a new chat no longer fails on the first invocation after a fresh start
- Subagent content no longer leaks out of collapsed subagent blocks
- File paths in user messages are no longer rendered as slash commands
- Subagent tool allowlists are validated before launch
0.9.1 — 2026-07-28
Added
- Pricing and model metadata for Kimi K3
Fixed
- All questions in a parallel Ask User batch are now shown, instead of only the last one
- Renaming the active conversation no longer silently hangs
- Long file and tool names no longer overlap in the Session panel
- Provider HTML is no longer included in API error messages
- Enforced Positron settings now cover provider enablement and model pinning, not just connections
- Fixed a schema error in the
positronCommandtool
0.9.0 — 2026-07-28
Added
- Multi-conversation substrate with concurrent conversation tabs and background execution
positronCommandtool for running Positron IDE commands- Databricks as a language model provider
- Command to focus the Posit Assistant chat input
- Option to clear all conversations at once
Changed
- Some Posit AI models now use a blended per-token cost multiplier
- Streaming re-renders and Shiki highlighting are throttled to reduce power usage
- Chat input now pastes plain text instead of formatted content
- Single-tilde strikethrough disabled in Markdown rendering
Fixed
- Deleting the current conversation no longer shows an error message
- Late-credential recovery gap on Positron startup
- Startup stall and slow conversation switches from absent auth providers
- Delay when a provider is not present
- HTML conversation import and export/import format issues
- Tool-call header filename misalignment
- Notebook tools failing to match the active notebook on Windows
- Subagent prompt no longer shown in token usage tooltip
0.8.0 — 2026-07-17
Added
- Auto-select the best available model when no named default matches the provider catalog
- Token usage is now recorded for subagents and the auto-mode classifier
- Positron IDE skill for answering questions about Positron features and workflows
Changed
- Automatic micro-compaction removed; compaction is now triggered explicitly via a command
- Code execution uses
NonInteractivemode to avoid combining Assistant input with Console input - Credential-loss notification reworded to match its action
Fixed
- Model selection no longer substitutes a recently-used model when that model is unavailable
- Text file attachments that browsers report no MIME type for (
.R,.qmd, etc.) are now accepted - Provider error reasons are surfaced to users instead of a generic “Error making request” message
- Long URLs in the chat composer now wrap instead of overflowing
- Sign-out no longer resurrects credentials
- Status bar stability, workspace switching, and tool display polish
- Extension enabled context is registered earlier, preventing brief flickering
- The edit tool now shows the searched-for pattern when it can’t find the target text
0.7.0 — 2026-07-10
Added
- EDA log feature
- Local models can now be registered with core, enabling local provider support
- “Refresh Models” menu item
- Cache keepalive status widget, showing when the prompt cache is being kept warm
- Estimated token usage is now shown for Copilot models
- Mid-stream approval-mode switching between Normal and Auto
- Workspace git branch is now recorded in conversation metadata
- Withheld plots now show a caption (and a broken-image fallback) for non-vision models
Changed
- Provider settings moved from
settings.jsontoproviders.json - Mode and thinking pickers consolidated into a single control in the prompt input
- Cache-keepalive pings are now enabled by default
- Local providers are now labeled experimental
- Various tool-call display and status bar polish: consistent spacing, a connecting line across consecutive tool calls, and a collapsible status bar
Fixed
- Malformed tool-call inputs are now repaired before they can poison a conversation
- Incorrect Positron
baseUrlsettings are now fixed at the source - The legacy providers block in
settings.jsonis no longer deleted - Notebook tools are hardened against
getContext()rejection - GovCloud Bedrock model ID prefix issue (via an
ai-libupdate) - A catalog-default region is no longer treated as Bedrock authentication
/reportQuarto frontmatter is kept to title-only
0.6.5 — 2026-07-06
Fixed
assistant.preferredModel.providernow accepts the camelCase Positron settings name (e.g.amazonBedrock) in addition to the canonical provider ID, so preferred-model selection works as configured- Thinking/reasoning display is shown again for Gemini models using the Interactions API
- Haiku 4.5 tool calls on Snowflake no longer fail (eager tool streaming is disabled for that combination)
0.6.4 — 2026-07-03
Added
- Claude Sonnet 5 support
Changed
- Cache keepalive pings now also run during long tool calls, not just between turns, keeping the prompt cache warm through slow tool execution
Fixed
- Tool-result images are now hidden from models that don’t support vision, instead of being sent anyway
- Content after a subagent call no longer disappears when a message is reloaded or edited
uvvirtual environments are now detected even whenVIRTUAL_ENVis unset (e.g. when Positron launches the Python kernel directly by interpreter path), anduv addis recommended overuv pip installwhen apyproject.tomloruv.lockis present- Haiku 4.5 tool calls on Posit Workbench (Bedrock) no longer fail with an HTTP 400 error
0.6.3 — 2026-06-29
Changed
- Posit Assistant once again installs and works on older Positron versions (before 2026.07), restoring support for the legacy
positron.assistant.enablesetting on those versions - Connect Cloud deploy skill no longer assumes R is installed: it selects the deploy route based on the active session language and supports deploying from a GitHub repository
- Ask User dialog: multi-part questions are now collected and submitted together via a reworked tabbed question interface
- Removed “currently available in public preview” text from the splash banner
Fixed
- Plan files saved to the home directory (
~/.posit/assistant/plans/) are no longer blocked on Posit Workbench when the open workspace has no.posit/assistant/directory
0.6.2 — 2026-06-25
Added
- Resend button on user messages to re-submit the previous turn
- Gemini models now use the stateful Interactions API
Changed
edittool now matches despite whitespace differences (indentation, blank lines, trailing spaces)assistant.enablednow defaults totrue. The sidebar icon and editor panel commands are hidden when disabled. Theassistant.sidebarViewsetting has been removed (sidebar visibility is now controlled byassistant.enabled).- Breaking: the legacy
positron.assistant.enablesetting is no longer honored. Admins and users who relied on it must switch toassistant.enabled(or use Positron admin policy) to disable Posit Assistant. Existing users with only the legacy setting set tofalsewill see the assistant become enabled after this update. - Posit Assistant is now gated on Positron’s
ai.enabledswitch — toggling it enables or disables the assistant - Posit AI: Opus 4.8 and Opus 4.6 label updates; Opus 4.7 hidden
- Posit AI: summarized thinking now available for all models
Fixed
- Token usage display: implicitly cached tokens shown correctly; Gemini 2.5 Flash pricing added
0.6.1 — 2026-06-23
Added
- Shell tool now inherits Positron’s terminal environment contributions (e.g.
quarto,airbinaries added by extensions) /diagnosticscommand to generate a diagnostics bundle for troubleshooting (Posit Assistant: Generate Diagnostics Bundle)- Prompt-cache keepalive pings, with new settings to enable and to control the number of pings
- Summarized thinking blocks for newer Anthropic models (Claude Opus 4.7+, Fable 5, Mythos 5)
- Usage tips displayed in the chat interface
- Skills: support for non-standard metadata fields in
SKILL.mdfrontmatter - Conversation ID sent on every Posit AI request for better attribution
Changed
- Session HUD restructured: token usage panel promoted to top-level, tooltips consolidated, chart improved
- Streaming re-render performance improved by ~25%
- Slash commands (e.g.
/auto) can now be used while a response is streaming - Increased max output tokens for Claude models
- Input editor preserves bold, italic, inline code, and strikethrough formatting when sending a message
Fixed
- Plan-file writes now respect user/project permissions (
settings.jsoneditrules were being shadowed) - Errors from
vscode.lm.selectChatModels()are caught @-mention file scanning slowness and popup lifecycle bugs@fileand/commandcompletions: fuzzy matching restored and various popup issues fixed- Tool working directory and system prompt working directory now align with the first local workspace folder
- Project
settings.jsonpartial entries no longer cause Zod defaults to silently override global settings - Prompt-cache breakpoint added at the previous user turn (fixes cache collapse at turn boundaries)
AskUserdialog and other session tools now route to the correct panel when multiple Posit Assistant panels are open
0.6.0 — 2026-06-11
Added
- Run in Console action for code blocks and inline selections — R and Python
- Provider icons next to the model name in the top bar and model menu
- Slash command completions for all skills (built-in, project, and user-scoped)
- Prompt to migrate workspace
.positaidirectory to.posit/assistant - Credential-loss notifications when provider sessions expire
- Support for Fable 5 and Mythos 5 models
- Line counts in
editandreadtool summaries - Verification step after
/reportfor Quarto documents - Publish
posit-assistant.hasChatModelscontext key so Positron can gate actions on whether a usable model is configured
Changed
- Configuration directory changed from
.positaito.posit/assistant, both globally (~/.positai→~/.posit/assistant) and per-workspace (<workspace>/.positai→<workspace>/.posit/assistant) - Reintroduced web fetch in sandbox mode
- Format R code in system prompt with Air
- Set
POSIT_ASSISTANT=1environment variable in shell sessions - Stack action buttons vertically when the panel is narrow
- Show eye icon for R/Python chip in compact mode
- Discourage assigning to intermediate objects in generated code
Fixed
- Drag-and-drop file attach in Positron webview
- Invisible “Allow” button on some themes and leaked Markdown escapes in tool paths
- AWS Bedrock credential refresh in Positron
- Insert at Cursor and Insert into New File hanging for 30 seconds
0.5.11 — 2026-06-05
Added
- Tooltips on the New Conversation, History, and More Actions buttons in the top bar
Changed
- Web fetch now returns an error in sandbox mode
- Styled tooltips, popovers, and hover cards to match Positron’s native editor hover widgets across light, dark, and high-contrast themes
- Updated the feedback URL to point to the assistant-feedback repository
Fixed
- Anthropic, OpenAI, and Gemini API keys with a version-less base URL (ai-provider-bridge v0.0.8)
- Honor each model’s declared output token limit (e.g. 128k for Opus 4.8)
- Restore the active mode when switching back to a conversation
0.5.10 — 2026-06-03
Added
- Experimental DeepSeek provider
- Auto mode
- Text selection toolbar with Copy, Quote, and Memory actions
- Support for Opus 4.8
Changed
- Rebuilt the chat input and refreshed the visual treatment for chat references
- Evaluate
executeCodecalls expression-by-expression, interleaving code and output - Bash-aware line wrapping for code blocks
- Render mode banners directly above the chat input
/clearnow preserves the current model settings, differentiating it from/new- Automatically approve safe bash and PowerShell commands, plus within-workspace
cd - Auto-allow builtin skill invocations
- Use PowerShell’s built-in parser for command permission resolution
- Never automatically approve network commands
- Avoid notebook tools on
.qmd/.rmdfiles - Avoid overriding
reposininstall.packages()
Fixed
- Interim fix to sync
assistant.enabledwithpositron.assistant.enablefor existing users preferredModel.idignored when set in.positai/settings.json- Sandbox fallback for Windows bash
- Google Vertex / Positron authentication path
- Plan-mode tooltip path now readable in the Positron dark theme
- Added timeouts to Positron runtime context gathering
0.5.9 — 2026-05-26
Fixed
- Command handling for configuring providers across Positron versions
0.5.8 — 2026-05-25
Added
- Google Vertex as a provider with Application Default Credentials (ADC) authentication
0.5.7 — 2026-05-24
Added
- Button to move conversation to editor panel
assistant.enabledsetting to control extension visibility- Support for
authentication.configureProviderscommand - Cost view in token usage chart
- Support for custom HTTP headers per provider
Changed
- Updated extension icon
- Improved prefix context and ephemeral context
- Use structured Bash command parsing with unbash
Fixed
- Invalidate stale model cache when
assistant.enabledtransitions to true - ConfigurationOverlay sizing
- Stop awaiting for session creation
- Responsive layout and spacing for setup overlay at narrow widths
- Script tag escaping
0.5.6 — 2026-05-20
Added
- Concurrent tool call execution
- Self-knowledge skill for answering questions about Posit Assistant
- Capture stderr output in bash tool
- Line numbers in read tool UI when file links are used
- Re-encode images for improved size and compression
Changed
- Show bash tool calls collapsed by default
- More compact open files context format
- Improved tool call visual feedback during streaming
- Smoother streaming with minimum speed and immediate buffer flushing
- Set max height for chat input
- Trim model list for Snowflake Cortex
- Required Positron version bumped to 2026.04.0
Fixed
- Filter
<|channel>token leaks with Gemma models - Use
vscode.openfor non-text files in link handler - Script tag handling in streamdown
0.5.5 — 2026-05-14
Added
- Attachment button in chat input
- Support for
skills.pathssetting to configure custom skill locations
Changed
- Updated toolbar and menu styling
Fixed
- Graceful recovery from corrupted conversation index
- Handle missing language model messages
- Menu width sizing
0.5.4 — 2026-05-12
Added
- Math and Mermaid diagram rendering in Markdown
- Notebook active-cell context injection
- Platform information in About dialog
- Preview URL support for web applications
- Combine Posit AI and VS Code configured MCP servers
Changed
- Removed
<topic>tags from rendered assistant messages
Fixed
- Text wrapping in chat messages
- Streaming smoothness for subagent and tool-execution panels
- UI performance with subagent panel rendering
- Strip
file://URIs from model-facing workspace context - Normalize Windows
file:///URIs in file links - Quote filenames with spaces in completions
- MCP configuration in Positron
- Agreement-pending recovery on all platforms
- Don’t auto-spawn AWS SSO login on startup for Bedrock
- Inform user when edits appear already applied
- Gate topic tag instructions on model thinking capability
0.5.3 — 2026-05-06
Added
- Non-blocking banner when Posit AI agreement is pending
Changed
- Use
authentication.positai.baseUrlfor Posit AI authentication
Fixed
- Thinking duration display for short and long durations
- Prevent
streamdown:incomplete-linkfrom appearing in rendered messages - Prevent extra newlines when copying user message text
- Suppress empty thinking UI when no thinking content is present
- Fall back to legacy command name on older Positron versions
- Stop polling server indefinitely when in agreement_pending state
0.5.2 — 2026-05-06
Changed
- Updated activity bar icon
Fixed
- AGENTS.md resolution failing when workspace folders return file:// URIs
0.5.1 — 2026-05-05
Added
- Topic-aware thinking display with accessible disclosure panel showing thinking topics and elapsed time
- Notebook session handling in code execution — routes code to the correct notebook session when active
- Auto-resize textarea in AskUser UI
Fixed
- File links in Workbench
- Copilot provider errors when
enableCopilotSdkis disabled - Startup flash of configuration overlay in Positron
- Removed unnecessary table styling
0.5.0 — 2026-05-01
Added
- Smooth streaming with Streamdown renderer (replaces React-Markdown)
- Web search tool for non-Anthropic models on Posit AI
- Clickable file links in chat messages
- Subcommand-aware permission patterns and PowerShell compound command support
Changed
- Updated color themes
- Use context lengths from
/modelsendpoint with Posit AI - Transition Posit AI’s OpenAI path to openai-compatible provider
Fixed
- Code block styling and rendering
- High contrast theme support
- Submenus fitting in narrow windows
- Exit plan mode buttons wrapping
- Context length precedence for Posit AI
- Cached token display and pricing for OpenAI and Gemini
- Sanitize content streamed to UI
- Inline CSS to work around Firefox bug
- Hide “Always allow” for unsafe or unparseable shell commands
0.4.2 — 2026-04-29
Added
- Cost column header in model selector for Posit AI models
Fixed
- Honor saved permissions for owner-only (600) files instead of re-prompting
- Show read-denied files in directory listings with
[access denied]annotation - Use SNOWFLAKE_HOST for private-link/RCR Snowflake endpoint construction
0.4.1 — 2026-04-29
Fixed
- Fix model overrides not applying for Bedrock ARN identifiers
- Prompt instead of deny for owner-only (600) files
- Restore cleaning mode prompts dropped during rename
0.4.0 — 2026-04-29
Added
- Data cleaning sub-mode
- Custom instructions migrated to skills and memory
getPlottool for retrieving plots- Support for
positron.assistant.models.overrides.<provider>setting - Support for queuing messages from user while streaming content
Changed
- Renamed “mode” concept to “persona”
- Switched to core tool implementation for
read,write,edittools - Downgraded default
**/.*aiExcludes pattern from deny to ask - Unified code rendering across platforms
- Updated HTML export header appearance
Fixed
- Pass workspace context to subagents so they know their working directory
- Permission handling: normalize Windows relative paths, case-insensitive path comparison, strip leading slash before Windows drive letter
- Fix delay in loading conversation history
- Fix HTML export
- Prevent quarto-authoring skill from triggering for Rmd
0.3.3 — 2026-04-24
Added
- Show token cost multiplier for Posit AI models
- Fallback behavior when no workspace is open
posit-assistant.newChatcommand with support for transientdata:URI attachments
Changed
- Model preferences now persist per-conversation; the
positron.assistant.preferredModelsetting is restructured into dotted sub-keys (.id,.provider,.thinkingEffort,.webSearch) - Improved appearance of image and file attachments, and attachment icons
- System prompt prefers
bash(gh)overwebFetchfor GitHub queries - Updated greeting text
Fixed
- React streaming performance regressions
- Dynamic sidebar/topbar switching for Positron
- Contrast issues in some themes
- Sanitize ANSI control characters in tool output
- Tooltip diamond coloring
0.3.2 — 2026-04-17
Added
- “Configure LLM Providers” item in the menu
- Support for Claude Opus 4.7
- Added support for
xhighandmaxthinking effort levels
Changed
- Will ask to write plan files (
plans/*.md) even ifaiExcludeswould normally block them - System prompt updated to prefer dedicated tools over Bash and to more strongly discourage running R via
Rscript
Fixed
- Fix blank plot captures in
executeCode - Accept file paths without a URI scheme in read/write tools
- Handle CRLF line endings in skill files
- Reduce verbosity of permission debug logging
0.3.1 — 2026-04-14
Added
- Positron-native notebook tools via
positron.notebooksAPI - Drag and drop support for attachments
- Per-conversation model persistence
positron.assistant.models.includesetting to filter available models
Changed
- Migrated chat scroll to use-stick-to-bottom library
- Auto focus on chat input after user actions
Fixed
- Fix model change indicator and selection revert during streaming
- Remove nested scroll containers causing duplicate scrollbars
- Remove leading newlines in Quarto reports
- Add height limit to text preview
0.3.0 — 2026-04-10
Added
- Support for Gemini, OpenAI, AWS Bedrock, Snowflake Cortex, Microsoft Foundry, and OpenAI-compatible providers via new auth system
- Inline expression guidance in
reportskill
0.2.1 — 2026-04-08
Changed
- If Positron trusts the workspace, that trust setting is used automatically
0.2.0 — 2026-04-08
Added
- Image and PDF input support (experimental)
aiExcludessetting to exclude files from AI context- Bash tool sandboxing on macOS and Linux
- Python environment description on conversation start
- Predictive modeling skill for Python
- Python equivalents for R-specific prompt sections
- Independent collapse toggles for tool call sections
- Menu item to copy conversation path and other info
- Preferred/recommended models highlighted in model selector
Changed
- Renamed from Databot to Posit Assistant
- Config keys changed from
posit.assistant.*toassistant.* - Renamed
quarto-reportskill toreport - Unified tool name for running code
- User message background color lightened
- UI text uses
--ui-font-sizesetting - Model selector centers on current model when opened
Fixed
- Subagent continues loop when user declines a tool
- Settings menu and dev mode fixes
- Stopped streaming text/html to chat
- Don’t send tool_result for server-executed tool calls
- Unified tool confirmations on InlineToolConfirmation
0.1.1 — 2026-03-31
- Public release