[3.96.0]

• ✨ New chat settings and diagnostics: Show message timestamps with configurable time and date formats, tune the prompt-cache TTL for Anthropic profiles, and generate privacy-safe diagnostics reports.
• 💸 More accurate provider costs and caching: Keep Anthropic tool results intact, adopt cumulative output tokens in Anthropic and MiniMax costs, read DeepSeek prompt-cache counters, target the xAI Responses cache, and bill Mistral cache reads at 0.1x.
• 🧭 More reliable and faster chats: Resume completed tasks after a restart, restore checkpoints when the saved commit is missing, auto-reload a dead webview renderer, throttle streaming updates, and load locales on demand.

New Chat Settings and Diagnostics

• Add time and date format settings and message timestamps (#2221 by @KarthickMahalingam, PR #2223 by @KarthickMahalingam)
• Add a user-facing prompt-cache TTL for Anthropic profiles (PR #2271 by @simurg79)
• Add privacy-safe diagnostics reports (PR #2292 by @app/zoomote2)

More Accurate Provider Costs and Caching

• Keep tool_result blocks contiguous in Anthropic requests by grouping image blocks separately (#1307 by @alaskanloosemoose-dotcom, PR #2243 by @myk1yt)
• Adopt cumulative message_delta output tokens in Anthropic and MiniMax cost (#1727 by @myk1yt, PR #2242 by @myk1yt)
• Read documented top-level prompt-cache counters for DeepSeek (PR #2252 by @myk1yt)
• Send body prompt_cache_key for xAI Responses cache affinity (PR #2256 by @myk1yt)
• Bill Mistral cache reads at 0.1x instead of the full input price (PR #2257 by @myk1yt)
• Map xAI tool_choice to the Responses API shape with a typed request body (PR #2036 by @easonLiangWorldedtech)
• Add MiMo V2.6 Flash and Pro model configurations for OpenCode Go (#1997 by @WebMad, PR #2246 by @myk1yt)

More Reliable and Faster Chats

• Resume completed tasks after a restart when a message write is interrupted (#2130 by @edelauna, PR #2297 by @app/zoomote2)
• Restore checkpoints when the saved commit is missing (PR #2269 by @simurg79)
• Detect a dead webview renderer via heartbeat and auto-reload (PR #1715 by @myk1yt)
• Report emergency condense like the automatic path (PR #1838 by @PierrunoYT)
• Summarize only the API-visible history when condensing (#1342 by @RemainIndoors1, PR #2300 by @RytisLT)
• Propagate beyond-EOF read-file errors instead of reporting the file as empty (PR #1734 by @myk1yt)
• Throttle partial message updates while streaming (PR #2120 by @PierrunoYT)
• Replace axios with fetch in key-info hooks (PR #2261 by @myk1yt)
• Load locales on demand instead of eager-globbing all 18 (PR #2267 by @myk1yt)
• Replace the debounce package with a local leading-edge helper (PR #2268 by @myk1yt)

New Features and Integrations

• Add a dependency-free BM25 ranker tool (PR #1207 by @edelauna)
• Add abort-signal wiring for lm-studio and qwen-code (PR #1309 by @easonLiangWorldedtech)
• Add Keenable to the MCP marketplace (#2312 by @ilya-bogin-keenable, PR #2313 by @ilya-bogin-keenable)

Testing, CI, and Repository Automation

• Cancel obsolete PR CI runs and remove duplicate builds (#2289 by @edelauna, PR #2294 by @app/zoomote2)
• Post surviving mutants as a PR comment (#2287 by @edelauna, PR #2293 by @app/zoomote2)
• Run webview CI tests with four workers (#2288 by @edelauna, PR #2308 by @app/zoomote2)
• Use plugin-side related-test discovery for the extension package mutation gate (PR #2211 by @easonLiangWorldedtech)
• Add test-utils guardrails and convert remaining stragglers (PR #2279 by @app/zoomote2)
• Drop test-only tool-use interfaces from shared/tools (PR #2258 by @myk1yt)
• Show the Zoo Code panel in the Solheim smoke video and bound the apt step (#2158 by @edelauna, PR #2195 by @app/zoomote2)

Refactoring and Code Health

• Move default provider ID functionality into the provider definition (PR #2233 by @p12tic)
• Resolve profile model IDs through provider definitions (PR #2239 by @p12tic)
• Remove no-explicit-any suppressions in utils (PR #2281 by @app/zoomote2)
• Remove no-explicit-any suppressions in small directories (PR #2286 by @app/zoomote2)
• Close lint coverage and knip hygiene gaps (PR #2280 by @app/zoomote2)
• Sweep dead locale keys behind an AST gate (PR #2263 by @myk1yt)
• Remove dead CSS classes from the global stylesheet (PR #2251 by @myk1yt)

Dependencies and Maintenance

• Update @tanstack/react-query to v5.105.0 (PR #2180 by @app/renovate)
• Update ignore to v7.0.14 (PR #2181 by @app/renovate)
• Update posthog-js to v1.440.0 (PR #2182 by @app/renovate)
• Update posthog-node to v5.57.2 (PR #2183 by @app/renovate)
• Update GitHub Actions (PR #2186 by @app/renovate)
• Update qs to v6.16.1 [security] (PR #2214 by @app/renovate)

Housekeeping

• Merge the previous v3.94.0 release preparation after its publication; this is prior-release housekeeping, not a v3.96.0 feature (PR #2272 by @app/zoomote2)

[3.94.0]

• ✨ Broader model and provider support: Use Claude Haiku 5.5 through Anthropic, send images to compatible VS Code LM models, tune temperature for OpenCode Go Responses models, and cancel requests cleanly across more providers.
• 🔐 More flexible MCP OAuth: Configure static OAuth client settings, use custom callback paths, discover authentication details from WWW-Authenticate, and see callback failures before retrying.
• 🧭 Smoother chats and long sessions: Rename tasks in history, drag files into chat without holding Shift, reduce memory retained during long sessions, and keep subtask links and replaced webviews stable.

Broader Model and Provider Support

• Add Claude Haiku 5.5 to the Anthropic provider (#2208 by @app/zoomote2, PR #2212 by @app/zoomote2)
• Add runtime image input support to VS Code language models (PR #2159 by @simurg79)
• Add image input conversion and shim data parts for VS Code language models (PR #2196 by @simurg79)
• Update VS Code language model metadata and regression tests (PR #2255 by @simurg79)
• Add opt-in temperature support for OpenCode Go Responses models (#2055 by @AntzCode, PR #2160 by @AntzCode)
• Add abort-signal support for Anthropic, Anthropic Vertex, xAI, and MiniMax (PR #1293 by @easonLiangWorldedtech)
• Add abort-signal support for VS Code language models (PR #1300 by @easonLiangWorldedtech)
• Add abort-signal support for Gemini, Mistral, and LiteLLM (PR #1303 by @easonLiangWorldedtech)
• Add ensureModelFetched to the six catalog providers (PR #2191 by @p12tic)
• Resolve the CLI model ID through provider definitions (PR #2192 by @p12tic)
• Count compatible cache reads for OpenAI (#1591 by @GeenccMustafa, PR #1662 by @PierrunoYT)

More Flexible MCP OAuth

• Add static OAuth client settings for MCP servers (#2096 by @SoccerField24x7, PR #2129 by @SoccerField24x7)
• Support custom MCP OAuth callback paths and WWW-Authenticate discovery (#2218 by @sdbrother0, PR #2219 by @sdbrother0)
• Report MCP OAuth callback failures before retrying (#2187 by @SoccerField24x7, PR #2238 by @SoccerField24x7)

Smoother Chats and Long Sessions

• Rename tasks in history (#2112 by @designPrj, PR #2122 by @designPrj)
• Allow dragging files into the chat input without holding Shift (#2139 by @juneleung, PR #2184 by @app/zoomote2)
• Reduce chat memory retained during long sessions (PR #2217 by @app/zoomote2)
• Preserve subtask link identities (PR #2228 by @simurg79)
• Dispose view listeners and ignore stale views when the webview is replaced (#2205 by @edelauna, PR #2213 by @app/zoomote2)
• Replace dispatch lock booleans with DispatchState (#374 by @edelauna, PR #2198 by @edelauna)
• Treat apiConfiguration as always defined in the webview (PR #2234 by @p12tic)

Testing and Repository Automation

• Use the real VS Code toolkit in Playwright visual baselines (#2105 by @edelauna, PR #2163 by @app/zoomote2)
• Fix five webview visual tests that failed on every pull request (#2206 by @edelauna, PR #2209 by @app/zoomote2)
• Link review gate Details to the workflow run (PR #2204 by @app/zoomote2)
• Move the release kickoff to 12pm EST (PR #2190 by @app/zoomote2)

Dependencies and Maintenance

• Update build, lint, and test tooling (PR #2101 by @app/renovate)
• Update AI SDK core and providers (PR #2103 by @app/renovate)
• Update @modelcontextprotocol/sdk to v1.32.0 [security] (PR #2131 by @app/renovate)
• Update @next/eslint-plugin-next to v15.5.28 (PR #2132 by @app/renovate)
• Update @types/node (PR #2135 by @app/renovate)
• Update esbuild-wasm to v0.28.2 (PR #2137 by @app/renovate)
• Update undici to v6.29.1 [security] (PR #2169 by @app/renovate)

Housekeeping

• Merge the previous v3.92.0 release preparation after its publication; this is prior-release housekeeping, not a v3.94.0 feature (PR #2166 by @taltas)

[3.92.0]

• ✨ Faster model choice and stronger model support: Switch models from the chat composer. Use Claude Haiku 5.5 on Amazon Bedrock, get more accurate Kimi K3 capabilities on OpenRouter, and use corrected Gemini 4 Argon limits on Vercel AI Gateway.
• 🧭 More reliable terminal, editing, chat, and remote workflows: Collapse command output by default when you choose, get clearer editing failures, keep reasoning and API rows visible, and avoid startup and proxy issues in remote workspaces.
• 🔌 Expanded MCP marketplace: Connect Zoo Code to AnySearch, Mailtea, and Tradehand through the MCP marketplace.

Faster Model Choice and Stronger Model Support

• Add an inline model selector to the chat composer (#1502 by @daewoongoh, PR #2034 by @daewoongoh)
• Add Claude Haiku 5.5 to the Amazon Bedrock model catalog (PR #2100 by @app/zoomote)
• Correct Kimi K3 capabilities and reasoning effort on OpenRouter (#1316 by @CodeCruncher88, PR #1325 by @easonLiangWorldedtech)
• Clamp Gemini 4 Argon to Google's limits on Vercel AI Gateway (PR #2073 by @taltas)

Terminal, Editing, Chat, and Remote Reliability

• Add abort-signal support to the native Ollama provider (#404 by @Fuzzillogic, PR #1299 by @easonLiangWorldedtech)
• Fix Destructive Command Guard startup errors in remote workspaces (#1603 by @atoomnetmarc, #1770 by @DaubnerF, PR #1639 by @PierrunoYT)
• Route the Bedrock code-index embedder through the system proxy (#1772 by @LouisClt, PR #1773 by @LouisClt)
• Require HTTPS for custom Google Gemini base URLs while allowing loopback HTTP (PR #2033 by @easonLiangWorldedtech)
• Add a setting to collapse terminal command output by default (#605 by @AmitTzah, PR #2042 by @bigyao25)
• Report all apply-diff failures with clearer diagnostics (PR #2053 by @simurg79)
• Keep reasoning and API request rows visible after batched reads (#2061 by @danGO-79, PR #2126 by @PierrunoYT)

Expanded MCP Marketplace

• Add AnySearch to the MCP marketplace (#1956 by @hieutran0413, PR #1955 by @hieutran0413)
• Add Mailtea to the MCP marketplace (#2140 by @davepoon, PR #1869 by @davepoon)
• Add Tradehand to the MCP marketplace (#2121 by @Blakefolgado, PR #2123 by @Blakefolgado)

Task Lifecycle Reliability

• Add run-state divergence comparison and metrics (RSK-05) (#1795 by @edelauna, PR #2047 by @edelauna)
• Add the run-state turn sub-machine (RSK-09) (#1800 by @edelauna, PR #2127 by @edelauna)
• Route task latch and initialization writes through run-state events (RSK-06) (#1796 by @edelauna, PR #2128 by @edelauna)
• Make Task.create share its startup path and promise with Task.run (#2145 by @edelauna, PR #2146 by @app/zoomote)

Testing, Performance, and Repository Automation

• Enable Knip production analysis and remove dead code (#519 by @edelauna, PR #527 by @edelauna)
• Add a trusted evaluation fixture runner and execution traces (#1896 by @edelauna, PR #1909 by @edelauna)
• Keep the has-conflicts label while GitHub recomputes mergeability (#2045 by @edelauna, PR #2048 by @app/zoomote)
• Refresh fork PR review-state labels after reviews (#2046 by @edelauna, PR #2049 by @app/zoomote)
• Add weekly release kickoff automation with maintainer rotation (PR #2075 by @app/zoomote)
• Base stale PR timing on author activity instead of bot activity (#2076 by @edelauna, PR #2077 by @app/zoomote)
• Stabilize restart-persistence E2E coverage and improve task-message read errors (#2078 by @edelauna, PR #2080 by @app/zoomote)
• Prevent the review-state label workflow from listening to itself (#2086 by @edelauna, PR #2088 by @app/zoomote)
• Freeze review-state labels while GitHub computes mergeability (#2089 by @edelauna, PR #2091 by @app/zoomote)
• Speed up lint and type checks with caching and staged-only pre-commit lint (#2090 by @edelauna, PR #2092 by @app/zoomote)
• Keep conflict status separate from review-state turn labels (part of #2089 by @edelauna, PR #2102 by @app/zoomote)
• Dispatch fork review follow-up using the correct repository metadata (#2133 by @edelauna, PR #2134 by @app/zoomote)

Dependencies and Maintenance

• Update build, lint, and test tooling (PR #804 by @app/renovate)
• Update KaTeX to v0.18 (PR #1975 by @app/renovate)
• Update simple-git to v4 (PR #1976 by @app/renovate)
• Update Zod to v4 (PR #2029 by @app/renovate)
• Pin dependencies (PR #2035 by @app/renovate)
• Update AI SDK core and providers (PR #2071 by @app/renovate)
• Update @changesets/cli to v2.31.1 (PR #2081 by @app/renovate)
• Pin dependencies after provider updates (PR #2087 by @app/renovate)
• Update @eslint/js to v9.39.5 (PR #2094 by @app/renovate)
• Refresh the lockfile (PR #2097 by @app/renovate)

Housekeeping

• Merge the previous v3.90.0 release preparation after its publication; this is prior-release housekeeping, not a v3.92.0 feature (PR #2074 by @JamesRobert20)

[3.90.0]

• ✨ Providers and models: GitHub Copilot is now a provider: sign in and Zoo Code loads your Copilot models. IO Intelligence is a provider with its own API key. Gemini 4 Argon is on Gemini and Vertex. GPT-6 Luna is on OpenCode Go. Gemini 4 Argon and GPT-6 Luna are on Zoo Gateway as well.
• 🧭 Tool execution: Repeating the same tool call gets a warning, then Zoo Code stops that loop. Disabling attempt_completion no longer leaves a task unable to finish. Queued messages run the same command-protection checks as commands you send directly.
• ⚡ Reasoning and streaming: Anthropic adaptive thinking shows in chat, and thinking tokens are counted. The R1 toggle keeps reasoning from llama.cpp, Ollama, and LM Studio. Codex keeps encrypted reasoning with the summary you see. VS Code language-model streams use guarded recovery when a stream fails.

Providers and Models

• Add a GitHub Copilot provider with sign-in and live model capabilities (#1971 by @KarthickMahalingam, PR #1969 by @KarthickMahalingam)
• Add the IO Intelligence provider (PR #1758 by @smiraldr)
• Add Gemini 4 Argon to the Gemini and Vertex AI providers (PR #2043 by @app/zoomote)
• Add GPT-6 Luna support for OpenCode Go (#1979 by @AntzCode, PR #2003 by @AntzCode)
• Port six Copilot model entries to the VS Code language model provider (PR #2032 by @simurg79)
• Complete the Copilot model port with safe input ceilings (PR #2052 by @simurg79)
• Always send the OpenCode Go session header from a per-instance session id (#1511 by @404CN, PR #1710 by @myk1yt)
• Show sane defaults for custom OpenAI-compatible models (#1866 by @p12tic, PR #1848 by @p12tic)
• Fetch model lists through the ApiHandler interface (PR #1906 by @p12tic)

Safer Tool Execution

• Add tiered tool-repetition detection with a soft warning and a hard stop (#771 by @p12tic, PR #1829 by @p12tic)
• Validate tools against the request's policy snapshot instead of live settings (#1642 by @DaubnerF, PR #1749 by @DaubnerF)
• Keep attempt completion available when it is listed in disabled tools (#1640 by @DaubnerF, PR #1751 by @DaubnerF)
• Apply command protection checks when draining queued messages (#1170 by @edelauna, PR #1699 by @myk1yt)
• Scope tool state to each tool call (#373 by @edelauna, PR #2015 by @edelauna)
• Warn that apply_diff search blocks must be whole lines (#868 by @lowlight-ua, PR #1712 by @myk1yt)
• Discover skills inside a symlinked container directory (#1842 by @p12tic, PR #1843 by @p12tic)

Reasoning and Streaming

• Surface adaptive thinking display and thinking tokens for Anthropic models (#1326 by @easonLiangWorldedtech, PR #1327 by @easonLiangWorldedtech)
• Preserve reasoning for local OpenAI-compatible models with the R1 toggle (#1118 by @ch405canova-sudo, PR #1119 by @ch405canova-sudo)
• Preserve encrypted Codex reasoning alongside visible summaries (PR #1966 by @WebMad)
• Integrate guarded recovery into VS Code language-model streaming (PR #1608 by @simurg79)
• Abort in-flight OpenAI Codex requests (PR #1290 by @easonLiangWorldedtech)
• Resolve shared AI SDK tool-call deduplication and error handling gaps (PR #799 by @easonLiangWorldedtech)

Task Lifecycle and Maintenance

• Add the run-state field, write helpers, and projection (RSK-04) (#1794 by @edelauna, PR #1985 by @edelauna)
• Decode documents from required buffers (PR #2002 by @WebMad)
• Remove redundant captureException calls that double-count errors (PR #1702 by @myk1yt)
• Write contributor docs and templates in Simplified Technical English (#1986 by @zoomote, PR #1987 by @app/zoomote)
• Add a typed-test checklist to the ESLint suppressions guidance (#1419 by @coderabbitai, PR #1420 by @easonLiangWorldedtech)
• Key the label-review workflow concurrency group per pull request (#1707 by @zoomote, PR #1739 by @app/zoomote)
• Keep a pull request awaiting the author when a third party re-requests review (#2026 by @edelauna, PR #2027 by @app/zoomote)
• Keep the has-conflicts label when mergeability is unknown (#2025 by @edelauna, PR #2028 by @app/zoomote)
• Turn off unused-variable linting in src and remove dead code (PR #2038 by @app/zoomote)

Dependencies

• Update undici to v6.28.1 (PR #1874 by @app/renovate)
• Update axios to v1.20.0 (PR #1875 by @app/renovate)
• Update shell-quote to v1.11.0 (PR #1973 by @app/renovate)
• Update @modelcontextprotocol/sdk to v1.31.0 (PR #1974 by @app/renovate)
• Update @ai-sdk/openai-compatible to v3 (PR #1899 by @app/renovate)
• Update ai to v7 (PR #1907 by @app/renovate)
• Update pretty-bytes to v7.2.0 (PR #1880 by @app/renovate)
• Update react-i18next to v15.7.4 (PR #1887 by @app/renovate)

Housekeeping

• Merge the previous v3.88.0 release preparation after its publication; this is prior-release housekeeping, not a v3.90.0 feature (PR #1980 by @edelauna)

[3.88.0]

• ✨ Provider and model support: Use Grok Build and improved custom-model support. This release also fixes Bedrock model limits and sends the Codex verbosity setting.
• 🧭 Task reliability: This release fixes cancellation handling, tool-turn continuation, and task-message saving.
• ⚡ Faster editing and chat: File writes no longer use artificial delays by default. Chat messages carry less file data, and code actions reach the last active chat.

Provider and Model Support

• Add Grok Build model metadata for xAI (#407 by @edelauna, PR #1687 by @myk1yt)
• Honor custom model IDs for OpenAI-compatible providers (#1845 by @p12tic, PR #1846 by @p12tic)
• Update OpenAI sane-default parameter values for custom models (#1845 by @p12tic, PR #1847 by @p12tic)
• Add an openAiStrictToolSchemas setting for OpenAI-compatible gateways (#1827 by @easonLiangWorldedtech, PR #1828 by @easonLiangWorldedtech)
• Send the Codex verbosity setting in requests (#1958 by @WebMad, PR #1977 by @tosinxt)
• Expose the Opus 5.5 output token limit on Bedrock (PR #1903 by @PierrunoYT)
• Resolve custom ARN capabilities from the underlying Bedrock model (#1968 by @kw-bivanov, PR #1970 by @kw-bivanov)

Task Reliability

• Recover completed tool-turn continuation (#1883 by @jaszhix, PR #1884 by @jaszhix)
• Make task-message saving stage-independent and finalize open partial tool asks (#1933 by @easonLiangWorldedtech, PR #1927 by @easonLiangWorldedtech)
• Check abort before an ask row posts (#1803 by @edelauna, PR #1895 by @edelauna)
• Add shared abort-signal cancellation-scope helpers (PR #1651 by @easonLiangWorldedtech)
• Forward the task abort signal in OpenAI and LM Studio handlers (#1450 by @allo-, PR #1686 by @myk1yt)
• Fix permanent DeepSeek and OpenAI request failures from lone UTF-16 surrogates in task history (#461 by @yoloseeyou, PR #1831 by @app/zoomote)
• Reject unsupported MIME types in image blocks (PR #1961 by @WebMad)

Faster Editing and Chat

• Remove artificial write delays by default (part of #1375) (PR #1381 by @easonLiangWorldedtech)
• Omit originalContent from webview messages and fetch it on demand (#1885 by @daewoongoh, PR #1943 by @daewoongoh)
• Route code actions to the last active chat (#1944 by @WebMad, PR #1946 by @WebMad)
• Preserve Unicode when clipping file lines (PR #1960 by @WebMad)
• Keep the streamed read path empty until content is available (PR #1962 by @WebMad)

Safety, Code Index, and Maintenance

• Add blanket auto-deny for unapproved commands in hands-free mode (#1569 by @DaubnerF, PR #1760 by @DaubnerF)
• Remove automatic startup embedder validation (PR #1879 by @WebMad)
• Refresh code-index status routing on workspace-folder changes (#1857 by @WebMad, PR #1957 by @tosinxt)
• Parse MCP tool names with double underscores in server names (#1731 by @myk1yt, PR #1737 by @myk1yt)
• Add the pure run-state kernel module (RSK-01) (#1791 by @edelauna, PR #1942 by @edelauna)
• Remove the dead Task.isPaused field (RSK-12) (#1793 by @edelauna, PR #1894 by @edelauna)
• Disable extensions when launching the Extension Host for debugging (#1859 by @daewoongoh, PR #1860 by @daewoongoh)
• Revert the apply_diff line-number prompt guidance (PR #1908 by @app/zoomote)
• Add packages/types to changed-code mutation testing (#1892 by @edelauna, PR #1893 by @app/zoomote)
• Wait for the resolved context window in the visual chat scenario (PR #1941 by @p12tic)
• Remove the stray tab before the Tool Use Guidelines heading (#1415 by @daewoongoh, PR #1440 by @easonLiangWorldedtech)
• Update posthog-js to v1.435.3 (PR #1586 by @app/renovate)
• Update posthog-node to v5.55.0 (PR #1587 by @app/renovate)
• Merge the previous v3.86.0 release preparation after its publication; this is prior-release housekeeping, not a v3.88.0 feature (PR #1897 by @taltas)

[3.86.0]

• ✨ New model support: Use GPT-6.1 Sol, get the corrected Claude Opus 5.5 output budget, and select DeepSeek V4.1 Flash on OpenCode Go.
• 🧭 Reliable task aborts and interruptions: Preserve the first abort reason, reset stream-abort state for each request, and stop interrupted tasks from replaying rejected subtasks.
• 🛠️ Dependable tools, prompts, and UI streaming: Keep streamed tool arguments intact, enforce MCP tool policy, prevent invented line numbers, and smooth chat updates.

Model Support

• Add GPT-6.1 Sol support for OpenAI (#1863 by @WebMad, PR #1864 by @WebMad)
• Correct the Claude Opus 5.5 output budget (#1853 by @jaszhix, PR #1854 by @jaszhix)
• Add DeepSeek V4.1 Flash metadata for OpenCode Go (#1648 by @carloslato, PR #1659 by @PierrunoYT)

Task Reliability

• Preserve the first abort reason when a task stops (#1802 by @edelauna, PR #1811 by @edelauna)
• Reset stream-abort state for each API request (#1801 by @edelauna, PR #1812 by @edelauna)
• Prevent interrupted tasks from replaying rejected subtasks (#1714 by @huankimtran, PR #1726 by @edelauna)

Tools, Prompts, and Streaming

• Preserve streamed native tool-call arguments across partial chunks (#695 by @awschmeder, PR #700 by @awschmeder)
• Truncate VS Code LM tool results safely within the context window (PR #1606 by @simurg79)
• Validate native MCP tool calls through the shared tool-policy layer (#1646 by @DaubnerF, PR #1724 by @DaubnerF)
• Prevent invented line numbers in apply_diff prompts (PR #641 by @faitno)
• Warn against including line numbers in search_replace, edit_file, and read_file inputs (#716 by @edelauna, PR #1685 by @myk1yt)
• Batch repeated tool preambles in the chat UI (#1656 by @petersjostrom, PR #1658 by @PierrunoYT)
• Avoid flushing throttled UI state for partial say messages (#1851 by @daewoongoh, PR #1852 by @daewoongoh)
• Keep markdown-sourced open-file requests inside the workspace (PR #1762 by @easonLiangWorldedtech)
• Use 127.0.0.1 for the development server to avoid IPv6 localhost resolution issues (#1588 by @hnbdr, PR #1589 by @hnbdr)

Code Index and Maintenance

• Introduce workspace scopes behind the code-index registry (PR #1766 by @WebMad)
• Add code-index feature scopes and centralize status subscriptions (PR #1768 by @WebMad)
• Centralize code-index workspace scopes and enablement (PR #1778 by @WebMad)
• Use immutable code-index configuration snapshots (#1788 by @WebMad, PR #1815 by @WebMad)
• Separate code-index service factories and embedder validation (#1816 by @WebMad, PR #1818 by @WebMad)
• Extract code-index scan execution without changing behavior (#1819 by @WebMad, PR #1834 by @WebMad)
• Extract code-index single-file preparation (PR #1836 by @WebMad)
• Update lru-cache to 11.5.3 (PR #1583 by @app/renovate)
• Update Mocha to 11.8.0 (PR #1585 by @app/renovate)
• Document P1 persisted ownership and the run-state kernel target architecture (PR #1706 by @app/zoomote)
• Add a community-approved label for community code approvals (#1872 by @edelauna, PR #1873 by @app/zoomote)
• Merge the previous v3.84.0 release preparation after its publication; this is prior-release housekeeping, not a v3.86.0 feature (PR #1810 by @taltas)

[3.84.0]

• ✨ New SOTA models added: Use GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 across supported providers.
• 🧭 More reliable tasks and subtasks: Keep delegated modes isolated, preserve subtask links after repeated stops, and protect orchestrator settings when slash commands switch modes.
• 🛠️ More dependable terminal, provider, and code-search behavior: Improve terminal behavior across Windows and non-English environments, strengthen provider responses and cancellation, and make code search use the correct workspace more consistently.

Minor Changes

• Add GPT-6 Sol and GPT-6 Luna to OpenAI model catalogs (PR #1755 by @app/zoomote)
• Add Claude Opus 5.5 across supported model providers (PR #1756 by @app/zoomote)
• Correct the Vertex Claude Opus 5.5 maximum output limit to 128K tokens (PR #1777 by @wrsfmss)
• Report Amazon Bedrock output truncation and expose model token limits (#1668 by @LouisClt, PR #1718 by @PierrunoYT)
• Sanitize lone UTF-16 surrogates in VS Code LM text and tool input (PR #1605 by @simurg79)
• Clear stale native tool-call arguments when finalization fails (#1221 by @canblmz1, PR #1634 by @canblmz1)
• Preserve concurrent MCP settings when the settings file is first created (#1371 by @pajitosingh, PR #1380 by @easonLiangWorldedtech)
• Use the canonical default model for the native OpenAI provider (#992 by @WebMad, PR #1627 by @JunyongParkDev)
• Propagate caller cancellation into dynamic model-catalog fetches (#1615 by @DaubnerF, PR #1683 by @DaubnerF)
• Inherit the host UTF-8 locale in terminals instead of forcing en_US.UTF-8 (#1084 by @scottdk, PR #1713 by @hebulin)
• Prevent Inline Terminal from falling back to cmd.exe incorrectly on Windows (#705 by @myk1yt, PR #1673 by @xcloudx01)
• Report the shell that actually runs commands under Inline Terminal (#1568 by @DaubnerF, PR #1682 by @DaubnerF)
• Search the task workspace instead of the active editor workspace in codebase search (PR #1629 by @WebMad)
• Initialize external code-index task managers before search begins (#1722 by @WebMad, PR #1725 by @WebMad)
• Align codebase-search readiness checks across mode filters (PR #1630 by @WebMad)
• Keep delegated child modes isolated from their parent task (#1632 by @robert-shade, PR #1637 by @PierrunoYT)
• Prevent slash-command mode frontmatter from overwriting the orchestrator task mode (#1776 by @edelauna, PR #1784 by @edelauna)
• Preserve parent and child task links after repeated Stop actions (#1676 by @jaszhix, PR #1678 by @jaszhix)
• Stop task-history updates from writing the entire history to global state (#1542 by @edelauna, PR #1664 by @PierrunoYT)
• Add a guarded VS Code LM tool-call recovery parser and normalized schema conversion (PR #1188 by @simurg79)
• Stabilize restart conversation-history E2E coverage with polling (#1641 by @DaubnerF, PR #1663 by @PierrunoYT)
• Update Vitest to 4.1.11 for security fixes (PR #1582 by @app/renovate)
• Reduce Windows CI cold-start time (PR #1654 by @app/zoomote)
• Mask the context-token counter in Electron sidebar snapshots (#1679 by @edelauna, PR #1680 by @edelauna)
• Add scope-boundary guidance to CodeRabbit global path instructions (PR #1757 by @app/zoomote)
• Allow non-organization members to interact with CodeRabbit chat (PR #1775 by @app/zoomote)
• Isolate workflow configuration tests from the Turbo cache (#1722 by @WebMad, PR #1782 by @edelauna)
• Add the task-lifecycle verification gap report and remediation blocks (PR #1626 by @app/zoomote)
• Add the stable release documentation PR step to the release workflow (PR #1786 by @taltas)
• Update the weekly release-reminder rotation (#1698 by @app/zoomote, PR #1700 by @app/zoomote)
• Merge the v3.82.2 release preparation state into main (PR #1677 by @edelauna)

[3.82.2]

Patch Changes

• Prevent unavailable tools from appearing in system prompts (#505 by @DScoNOIZ, #1240 by @JunyongParkDev, PR #1505 by @DaubnerF)
• Fix DeepSeek Flash image input by adding the new deepseek-flash model ID (PR #1618 by @app/zoomote)
• Fix token usage tracking for Grok and xAI-compatible endpoints whose domains contain "x.ai" (#1483 by @BambinoSK, PR #1484 by @BambinoSK)
• Apply the configured reasoning effort consistently across OpenAI-compatible requests (#993 by @Gringo675, PR #1604 by @JunyongParkDev)
• Preserve the configured LiteLLM model ID in the model picker (#1367 by @easonLiangWorldedtech, PR #1368 by @easonLiangWorldedtech)
• Fix delegated subtasks reading the parent mode in environment details and tool validation (#1623 by @edelauna, PR #1625 by @edelauna)
• Add a file version token to the guarded-write path to prevent stale overwrites (PR #1383 by @easonLiangWorldedtech)
• Extract the code-index manager registry for clearer ownership (PR #1622 by @WebMad)
• Route Roomote pull requests through the CodeRabbit review path (PR #1598 by @app/zoomote)
• Make mutation-testing findings advisory instead of blocking (PR #1610 by @app/zoomote)
• Group mutation warnings by source location to remove duplicate warnings (PR #1619 by @app/zoomote)
• Skip mutation testing while pull requests are in draft (PR #1645 by @app/zoomote)
• Scope the mutation diff against the exact merge base so unrelated changes on main stop inflating the scope (PR #1655 by @app/zoomote)
• Model test bundle dependencies in Turbo so caching stays correct (#114 by @edelauna, PR #1611 by @app/zoomote)
• Separate extension unit tests from bundle smoke tests (PR #1614 by @app/zoomote)
• Move extension source coverage to cacheable test lanes (#118 by @edelauna, PR #1620 by @app/zoomote)
• Cache extension coverage by ownership lanes (#115 by @edelauna, PR #1631 by @app/zoomote)
• Keep coverage caches valid when only verification scripts change (PR #1649 by @app/zoomote)
• Union ownership-lane coverage reports before uploading to Codecov (#1647 by @DaubnerF, PR #1650 by @app/zoomote)
• Validate coverage lanes dynamically in the merge queue (PR #1644 by @app/zoomote)
• Stabilize the accessibility contrast audit during theme changes (#1612 by @edelauna, PR #1613 by @app/zoomote)
• Make CodeRabbit completeness checks advisory (PR #1621 by @app/zoomote)

[3.82.1]

Patch Changes

• Fix task history entries disappearing when users reopen a task (#1279 by @jiangbing321, PR #1319 by @app/zoomote)
• Fix task history entries disappearing when users restart after a task completes (#1453 by @zoomote, PR #1452 by @app/zoomote)
• Fix tasks stalling when interrupted subtasks resume (#1468 by @zoomote, PR #1470 by @app/zoomote)
• Fix nested subtask tool calls stalling mid-task (#921 by @easonLiangWorldedtech, PR #1494 by @app/zoomote)
• Fix reasoning models stopping their thinking after a model selection change (#1348 by @zoomote, PR #1349 by @app/zoomote)
• Fix provider settings requests contacting model services that the user did not select (#1424 by @zoomote, PR #1425 by @app/zoomote)
• Fix MCP OAuth registration failing for servers that advertise unsupported grant types (#1531 by @edelauna, PR #1532 by @app/zoomote)
• Fix commits failing after users interrupt mutation tests (#1514 by @easonLiangWorldedtech, PR #1525 by @app/zoomote)
• Fix the OpenCode Go context meter showing incorrect limits (PR #1428 by @app/zoomote)
• Fix NanoGPT requests dropping optional tool parameters (PR #1590 by @huggix)
• Ignore blank or missing follow-up suggestion answers in the webview (#1226 by @andreluko, PR #1286 by @easonLiangWorldedtech)
• Finish the canonical provider identifier audit across providers (#944 by @WebMad, PR #1493 by @WebMad)
• Add test coverage for code-index and tools lines left uncovered by #1297 (PR #1317 by @easonLiangWorldedtech)
• Add an LM Studio reasoning_content E2E guard (#1075 by @marunguy, PR #1322 by @easonLiangWorldedtech)
• Refresh pull-request review labels after base updates (PR #1584 by @app/zoomote)
• Bump GitHub Actions to node24 runtimes to clear Node 20 deprecation warnings (PR #1534 by @app/zoomote)
• Refine CodeRabbit review checks (PR #1571 by @app/zoomote)
• Clarify CodeRabbit approval checks (PR #1577 by @app/zoomote)
• Update mammoth to v1.12.2 (PR #1472 by @app/renovate)
• Update globals to v16.5.0 (PR #1474 by @app/renovate)
• Update i18next to v25.10.10 (PR #1475 by @app/renovate)
• Update ink to v6.8.0 (PR #1477 by @app/renovate)

[3.82.0]

• 🔑 Use your Zoo Gateway API key anywhere — bring it to any OpenAI-compatible client or workflow: https://zoocode.dev/models
• 🎁 Free model access for a limited time — get free access to MiniMax-M3 through Zoo Gateway.
• ✨ Brand-new models — GPT-6 Astra and Claude Fable 5.1 are now available.

Minor Changes

• Add verified GPT-6 Astra support across providers (PR #1506 by @app/zoomote)
• Add Claude Fable 5.1 support (#1507 by @navedmerchant, PR #1508 by @navedmerchant)
• Add DeepSeek V4 Flash Vision Experimental support to DeepSeek AI (PR #1438 by @PipeYume)
• Complete DeepSeek V4 Flash Vision Experimental support with verified native metadata, Fireworks availability, image preservation, and reasoning controls (PR #1488 by @app/zoomote)
• Let users add custom request fields for OpenAI-compatible providers (#1104 by @grade1000, PR #1350 by @app/zoomote)
• Add abort-signal support to Amazon Bedrock prompt completion and message creation (PR #1292 by @easonLiangWorldedtech)
• Add an API cancellation helper and regression coverage for prompt-completion options (PR #1288 by @easonLiangWorldedtech)
• Route GPT-5.6 Luna through the Responses API for OpenCode Go (#1431 by @AntzCode, PR #1443 by @AntzCode)
• Prevent OpenCode Go requests from failing when a session header is unavailable (PR #1512 by @app/zoomote)
• Make Gemini CLI handler routing explicit and prevent provider misrouting (#1029 by @WebMad, PR #1442 by @enestunc10)
• Preserve provider reasoning order by yielding reasoning chunks before content (#1461 by @daewoongoh, PR #1462 by @daewoongoh)
• Keep approval controls visible while users scroll through chat (#1447 by @app/zoomote, PR #1448 by @app/zoomote)
• Keep the child-task view open during extension-state updates (#1456 by @app/zoomote, PR #1458 by @app/zoomote)
• Prevent Chat and History labels from clipping at narrow widths (#1444 by @app/zoomote, PR #1445 by @app/zoomote)
• Keep provider-profile actions accessible within narrow settings layouts (#1476 by @DiscantX, PR #1515 by @app/zoomote)
• Remove obsolete cloud-account lifecycle listeners (PR #1353 by @app/zoomote)
• Share the canonical provider-identifier ESLint rule across packages (PR #1421 by @WebMad)
• Add Extension Host visual-regression coverage (#1358 by @app/zoomote, PR #1426 by @app/zoomote)
• Merge the v3.80.1 release preparation state into main (PR #1432 by @edelauna)
• Configure CodeRabbit to perform adversarial pull-request reviews (PR #1433 by @edelauna)
• Stabilize restored-subtask approval coverage in Extension Host E2E tests (PR #1434 by @app/zoomote)
• Run CodeRabbit only after required CI checks complete (PR #1437 by @app/zoomote)
• Capture typed extension-host messages in Story Gallery fixtures (#517 by @edelauna, PR #1446 by @app/zoomote)
• Prevent mocked E2E tests from timing out on streaming previews (PR #1449 by @app/zoomote)
• Remove the unused nightly-extension path from CI (PR #1454 by @app/zoomote)
• Allow reviewed pull requests from forks to enter the merge queue (PR #1455 by @app/zoomote)
• Ensure CodeRabbit reviews updated commits on label-opted pull requests (PR #1457 by @app/zoomote)
• Refresh pull-request labels after CI completes (PR #1467 by @app/zoomote)
• Allow Renovate updates to trigger CI validation (PR #1473 by @app/zoomote)
• Add model checking for concurrent task-lifecycle transitions (PR #1478 by @app/zoomote)
• Require changed-code mutation tests before review (PR #1479 by @app/zoomote)
• Ground CodeRabbit reviews with current web-search context (PR #1490 by @app/zoomote)
• Fix the mutation gate for both stale and current pull requests (#1498 by @edelauna, PR #1499 by @app/zoomote)
• Refresh pull-request review labels after automated reviews and base conflicts (PR #1509 by @app/zoomote)
• Refresh review state for pull requests from forks after automated review (PR #1510 by @app/zoomote)
• Prevent task cleanup from producing unit-test teardown errors (#1526 by @app/zoomote, PR #1527 by @app/zoomote)