OpenCoWork: baixar versions anteriors

Added
- Sign in to a RoutIn desktop account from the sidebar. Tokens stay encrypted in Main (safeStorage); the window only sees the profile, never the access or refresh token.
- Phone remote control from the account menu: enable, scan a QR, and pair a phone over SignalR. Remote terminal execution and git commits from the phone are off by default.
- Added Claude Fable 5.1, Gemini 3.8 Flash (official thinking_level LOW/MEDIUM/HIGH), Gemini 3.7/3.6 Flash, and Gemini 2.5 Pro/Flash/Flash-Lite to the Routin AI catalog.
- Anthropic models now expose a thinking-budget slider on the model switcher (minimum 1024 tokens, default 10K, capped by the model's output limit).
Changed
- A quoted prompt typed while a run is in flight is injected into that run (agent/append-messages) instead of waiting for a brand-new turn after the current one finishes.
- New sessions pick a model in this order: the composer selection, the project binding, a fixed new-session default, then the global active model.
- Native Worker outbound HTTP now uses the same OS/system proxy Electron resolved (Clash, Surge, a corporate PAC). Provider traffic no longer bypasses the proxy the renderer already honors.
- Channel auto-reply and cron background runs now persist late-disclosed thinking in front of the answer, matching the desktop transcript.
Fixed
- Fixed Anthropic interleaved thinking signatures overwriting an earlier block. Each signature stays on the thinking block it belongs to, so the next user turn is no longer rejected by the Messages API.
- Fixed OpenAI Responses reasoning item ids being dropped from persisted thinking blocks, which broke cross-turn reasoning replay on endpoints that never return encrypted_content.
- Fixed Copilot-only local model fields leaking onto other providers when a Copilot row was copied or restored.

Changed
- Message-list tail follow now stops writing scrollTop after the stream ends. A short settle window still pins Thought collapse and widget paint; idle transcripts no longer keep chasing the bottom.
- Hosted sessions reopen when the API key, endpoint, thinking toggle, or reasoning effort changes, instead of reusing a session-open template that still had the old provider pinned.
Fixed
- Fixed inline widgets going blank or shrinking when a later streamed snapshot was shorter or truncated. The live merge keeps the longest complete widget_code.
- Fixed Gemini chat/completions using the catalog's static thinking_level instead of the selected reasoning effort, and still emitting reasoning_effort after thinking was turned off.

Changed
- 1M-capable models (GPT, Gemini, Claude, Qwen, and other chat models with a ≥1M window) now default to a 360K standard window instead of 272K. Turning on 1M context uses the model's native long window — including 2M — instead of always falling back to GPT's 1.048M.
- App updates moved from Data Management into System Settings. The Data page is now backups, migration, and reset only.
Fixed
- Fixed a quoted prompt vanishing from the live transcript before SQLite had the row. The optimistic bubble stays in the resident window, and the list pins it into view immediately instead of waiting for a session reload.

Changed
- Consecutive thoughts and ordinary tools now share one Exploring run, so completed thinking folds into the collapsed header instead of sitting as a separate block.
- The live thinking header shows the latest cleaned fragment and only ticks when a new sentence starts.
- Write can create or overwrite a file without a prior Read. Edit and NotebookEdit still require a matching last-read snapshot.
Fixed
- Fixed Read/Grep and other tool cards disappearing when tool_call_start arrived before tool_use_generated, or when the live agent store already tracked the call. The stream now always upserts the assistant tool_use block.
- Restored missing tool cards from the tool-result journal and request-debug payload when a result exists without a matching tool_use in the message.

Added
- Added GLM-5.3-Flash to the official BigModel catalog (1M context, vision, and thinking).
- Added Hunyuan Hy4 Preview (1M context; thinking defaults to high, and off maps to the official no_think effort).
- Added Qwen3.8 Max, Qwen3.8 Flash, and Qwen3.8 27B to the official Qwen catalog (1M context, hybrid thinking, and vision).
- Added Hy4 Preview, Qwen3.8 Max, and Qwen3.8 Flash to the Routin AI catalog with matching context, pricing, and thinking settings.
Changed
- Retired Hunyuan Hy3 Preview after TokenHub routes that traffic to Hy3 on 2026-08-31; saved Hy3 Preview rows are removed on the next preset restore.
- Qwen3.8-Max-Preview on Routin AI now matches production Max (128K output, vision, official pricing, and low/medium/xhigh effort) and is marked deprecated so saved preview rows can move to the production model.
Fixed
- Fixed the session change-review panel hitting "Maximum update depth exceeded" when a folder had no git branch list. The Zustand selector now returns a stable empty array instead of allocating a new one on every snapshot.

What's Changed
- Fixed consecutive Read tool cards staying on "receiving parameters" or looking hung. Parallel tool events are now committed in sequence order, the durable outbox no longer skips a hole, and the desktop stream holds later envelopes until the missing sequence arrives.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/v1.3.20...v1.3.21

What's Changed
Added
- Added a Windows x64 compatibility installer (*-win-x64-compat). Its Native Worker publishes as .NET 8 AOT with the baseline instruction set, so machines without SSE4.2 can start without lowering the official .NET 11 SKU. Compat builds do not use the official auto-update channel; download the matching installer from the GitHub Release.
Changed
- Consecutive read-only tools (file reads, directory listings, searches, memory lookups, web fetches, and status checks) now run concurrently up to the configured parallel-tool limit, instead of strictly one at a time. Writes, edits, and shell still run in the order listed.
- Tool results now keep up to 64K characters to match the advertised bounded-preview-64k policy (was 16K). Read stops on a line boundary and tells the model the next offset, instead of cutting the middle of the file.
- Session task tools (TaskCreate / TaskGet / TaskUpdate / TaskList) share one schema across desktop, hosted sessions, and CLI, so hosted runs can express dependencies and the blocked / in_review statuses the task board already shows.
- Slow Worker RPC such as image generation now goes through the durable job client instead of an inline request that could time out.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/1.3.19...1.3.20

What's Changed
Added
- Added a live FPS, CPU, and memory overlay when Developer Mode is on. It samples the window frame rate and system load about once a second, pauses while the window is hidden, and appears in the main app plus detached session and SSH windows.
Changed
- Dragging the workbench right-panel splitter now updates width on animation frames and freezes the split surfaces for the drag, so the chat and review panes do not reflow on every mouse move.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/1.3.18...1.3.19

What's Changed
Added
- Added OpenCowork Import Protocol v2. Desktop deep links (opencowork://import/provider#…) and the CLI device-login callback can now upsert built-in or custom channels with models and parameters; large catalogs can arrive via configRef. Legacy Routin-key (v1) imports still work.
- Added first-party presets for Hunyuan, StepFun (and the StepFun plan channel), Mistral, Meta (Muse Spark), Groq, Vertex AI, LM Studio, NVIDIA NIM, Cerebras, Together AI, Fireworks, ModelScope, PPIO, Novita, Infini, and Hugging Face.
- CLI assistant markdown now shares one layout and height pass (tables, fences, headings, lists) so the transcript viewport measures the same tokens it paints.
Changed
- Refreshed built-in catalogs: OpenAI GPT-5.4/5.5 Pro, Anthropic Claude Opus 5, Google Gemini 3.6/3.7 Flash, Azure GPT-5.4 Pro, plus newer models on OpenRouter, SiliconFlow, Baidu, Gitee AI, OpenCode, Copilot, Codex, and Volcengine.
- Codex and Copilot OAuth now share one client-identity helper (user-agent and editor headers). ChatGPT Codex backend URLs get client_version, and /backend-api/codex/* paths are recognized.
- Custom providers can pick Vertex AI as the request type. Routin device-login opens with protocol=2 so the web page can return a v2 payload.
- The import toast reports how many channels were written and skipped.
Fixed
- Fixed CLI markdown replies under-measuring their height and hard-clearing the terminal on every frame.
- Fixed root npm ci / prepare failing on cli:build because the repo-root hoist of wrap-ansi was the untyped 7.x used by yargs, not the typed 9.x the CLI imports.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/1.3.17...1.3.18

What's Changed
- Feat/http by @239573049 in https://github.com/AIDotNet/OpenCowork/pull/161
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/1.3.16...1.3.17

What's Changed
- fix(pricing): DeepSeek peak windows only apply on weekdays by @xyzs996 in https://github.com/AIDotNet/OpenCowork/pull/160
New Contributors
- @xyzs996 made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/160
-
Added
- Added DeepSeek V4 Flash Vision Exp (deepseek-v4-flash-vision-exp) to the official DeepSeek and Routin AI catalogs, with vision enabled and the same context, thinking, and pricing as V4 Flash.
Changed
- Ox Alpha on Routin AI no longer allows turning thinking off; reasoning_effort is now low/high/max (default max), matching GLM-5.3.
Fixed
- Fixed a healthy provider stream dying mid-read (connection reset or truncated SSE body) aborting the turn: the Worker now retries that class of fault when nothing visible has been emitted yet, so a headers-only drop no longer wastes a long wait.
- Fixed live tool cards staying on “receiving args” after the model finished the call: tool_use_generated now marks the call running, and a persisted tool_result wins over a stale overlay streaming status.
- Fixed compaction cuts storing the Worker’s run-scoped asst:<runId> handle, which never matches a transcript row: the host now spares the streaming desktop message, and the in-chat divider falls back to the last assistant turn before the summary instead of disappearing.

Added
- Added a dismissible /init banner when the selected project folder has no AGENTS.md, so Cowork, Code, and ACP sessions can seed workspace protocol without hunting for the command.
Changed
- Removed Auto model routing and the session “follow global model” option; sessions now bind a concrete model, and new sessions snapshot the current global, project, or configured default.
- Refreshed the model switcher and Model settings panel around search results and capability chips so the catalog’s context, thinking, and tool support is visible while picking a model.
Fixed
- Fixed transient provider failures aborting a turn that had not yet produced output: OpenAI Responses stream errors (overloaded_error, rate_limit_error, server_error, and the other retryable codes), response-header timeouts, and HTTP 408/409/425 now retry; a rejected previous_response_id on the WebSocket path falls back to HTTP SSE instead of failing the turn.
- Fixed hosted-session open or send failures stranding the turn: Main now falls back to an idempotent legacy agent/run with the same run id, restores persisted sessions after a worker restart, and forwards the raw error detail through IPC so the chat card shows the actual cause instead of a generic unknown.
- Fixed compaction dividers vanishing into the collapsed tool-process section and landing on the wrong assistant turn: the recorded cut now keeps the summary above the spared streaming row.
- Fixed Stop/terminate hooks never reaching the host after the user cancelled a run, because the reverse request inherited the already-cancelled run token.

Added
- Added live context-compression progress in chat: the summarizer's draft streams into a status card with retry attempts, and long compressions no longer trip the IPC wait timeout because progress frames are live-only while the durable outbox records only completed compression events.
- Added read-only orchestration tools to the ACP lead agent — Skill, MemoryList/MemoryRead/MemorySearch, codegraph_explore, WebSearch, WebFetch, TeamCreate/SendMessage/TeamStatus/TeamDelete, CronList, and Notify — so the parent can plan and delegate without gaining Write, Edit, Bash, or CronAdd.
- Added attached user images to Gemini generateContent, Gemini Interactions, and OpenAI chat-completions by normalizing wire image sources into inline bytes or URLs instead of dropping them.
- Added a live thinking preview: the collapsed reasoning header shows the latest thinking line while the model is still reasoning.
Changed
- New user turns now pin to the top of the viewport via a bottom spacer, so a long reply scrolls beneath the question instead of pushing it off screen.
- Split the 4,000-line SettingsPage into focused panels (General, Model, Runtime, Data, Memory, Analytics, System, About, plus the provider and AI-coding workbenches) behind a shared nav and primitives module, and dropped the unused SettingsDialog.
Fixed
- Fixed Fast mode being forced on for every hosted-session, cron, and channel run: the main-process provider config copied the model catalog's priority service tier straight onto the request, so fast-capable models like GPT-5.6 Sol/Terra/Luna were billed at the priority tier even with the Fast toggle off. Only Codex OAuth still ignores the toggle, because its plans are always priority.
- Fixed auto-compression silently never running for models without a configured context length: an unknown window produced a null compression config while the input-area gauge kept measuring against the 200K default, so both the hosted-session assembler and the interactive run path now fall back to that same default.
- Fixed a session staying permanently busy after a send failed during preflight (#159): the streaming pointer, abort controller, and sidecar run id leaked, which disabled the plan review actions and queued every later message instead of dispatching it, leaving Continue as the only path that still worked. Such a failure now also reports an error card instead of the turn silently disappearing.
- Fixed EnterPlanMode forking a second plan row and .plan file when a run died after ExitPlanMode (#159), which left the review card pointing at the abandoned plan; awaiting-review and approved plans are now reopened in place and moved back to drafting while they are rewritten.

Added
- Added a Native Worker per-tool durable journal (runtime_tool_results, 3-day retention) written on the emit path the instant a tool finishes, exposed as agent/tool-results-lookup, so a tool_result that never reached the messages table is recovered from disk.
Changed
- Renderer persists the assistant row and its tool_result at every tool boundary instead of waiting for the 30s streaming checkpoint or run end, writing the assistant row first so an interrupted write leaves a healable dangling tool_use rather than an orphan tool_result.
- Hosted sessions snapshot their canonical history after each fully answered tool batch instead of only at loop end, so an interrupted turn does not roll back past tools that already ran.
- Dangling-tool_use healing and Continue consult the Worker journal before falling back to the in-memory tool-call cache, which is empty after a restart.
Fixed
- Fixed a finished tool being silently re-executed on the next turn after a crash, app kill, or worker recycle mid-turn: its unpaired tool_use was stripped from provider input, so the model forgot the call happened and re-ran it — worst case replaying a long sub-agent Task from scratch.
- Fixed sub-agent and context-compression child runs overwriting the parent hosted session's history, which they could do because they borrow the parent sessionId while owning an unrelated conversation.

Added
- Added GLM-5.3 on BigModel (official and coding-plan presets) and Routin AI as the flagship model, with 1M context, thinking that cannot be disabled, and reasoning_effort limited to low / high / max.
Changed
- Applied the shared compact request view in Main session assembly and Native Worker so hosted session-open sends only the compact fence (boundary + summary + preserved tail) instead of reloading every pre-compression turn, and reopens the hosted session when that fence changes.
Fixed
- Fixed CI npm ci failing after 1.3.9 by syncing the lockfile for the new shiki dependency.
- Fixed typecheck failures from an unused cn import and a null vs undefined thinking-config mismatch.

Added
- Added inline file diffs for Write/Edit tool results in chat, including snapshot-backed previews and syntax-highlighted mutation cards.
- Added live sub-agent grouping in the CLI transcript (queued/running/completed, activity, tokens, model, elapsed) and matching desktop presentation phases.
- Added the OpenCode Zen provider preset (opencode.ai/zen) with GPT 5.6 Terra/Luna, GPT 5.5, Claude Opus/Sonnet 5, Grok 4.6, Qwen3.7 Max, DeepSeek V4, GLM 5.2, MiniMax M3, and Kimi models.
- Added shared ACP session-mode tool catalogs so the lead stays orchestration-only while sub-agents inherit the full implementation tool set and run in code mode.
- Added Native Worker AgentRuntimeToolError so recoverable tool failures become model-safe <system-remind> text instead of AOT resource keys such as IO_PathNotFound_Path.
- Added chat home insight copy, a hover-card titlebar mode switch with keyboard shortcuts, and motion helpers for popovers and preview rails.
Changed
- Refreshed chat presentation (tool cards, thinking blocks, skills menu, input area, workspace sidebar) around disclosure, diffs, and plan review.
- Bounded provider "context window exceeded" recovery to two overflow compressions per run so a misconfigured window cannot loop forever.
- Treated Home as project-scoped when a project is already selected, so new sessions from the home picker inherit the project.
- Removed the unused renderer prompt-optimizer module.
Fixed
- Fixed Windows ARM CI npm ci failing on hung Electron downloads by caching the binary, retrying install, and falling back to the npmmirror already used by electron-builder.
- Fixed Native Worker tool failures leaking framework resource keys to the model after AOT publish (UseSystemResourceKeys).

Added
- Hosted interactive agent sessions in Main: tool-capable chat now starts turns via startRun on the generated runtime protocol, with Main assembling session context and journaling projection while Native Worker remains the loop authority.
- Added a CLI localhost callback for Routin device login so CLI-only installs (no opencowork:// handler) can finish browser authorization, plus manual login triggers (cowork login, /login, provider L).
- Added CLI /init to inspect the workspace read-only, preview a drafted root AGENTS.md through AskUserQuestion, and write only after an explicit create/update choice.
- Unified CLI thinking intensity (/effort, /think, /config, /model) into one model-aware axis (off / on / auto / named levels) instead of a separate Thinking toggle plus effort row.
- Added channel reply reliability: chunk long messages to platform limits, cache inbound message-to-chat routing, send DingTalk P2P via oToMessages, and reply to QQ DMs as well as groups.
- Added Grok 4.6 (grok / grok-4.6, including xhigh reasoning) on xAI and Routin AI, defaulted xAI to Grok 4.6, and added Grok Imagine Image 2.0 on Routin.
- Added architecture-boundary ADR/docs and CI runtime-baseline / contracts / typecheck gates on the release build workflow.
Changed
- Pinned session system-prompt construction and streamlined tool catalogs so hosted runs reuse a stable prompt/tool surface instead of rebuilding it per turn in the renderer.
- Refreshed English and Chinese READMEs around the Worker-hosted runtime, CLI, and channels, and shipped matching architecture screenshots.
Fixed
- Fixed CLI dropping back to the shell after a successful cowork login; the process now enters the interactive TUI the same way as a bare cowork start.
- Fixed resumed-session CLI copy showing a replacement character by unifying the i18n key and middle-dot separator.
- Removed an accidental duplicate graph-generation module from the draw canvas path.

Fixed
- Fixed desktop agent runs that were accepted by Native Worker but never streamed to the renderer: Main now acknowledges durable agent/stream envelopes after journaling/delivery, falls back to a live window when routing is missing, and proactively recovers a wedged outbox instead of waiting for a false 45s sidecar_unavailable timeout.
- Fixed CLI-only first installs failing with SQLite Error 1: 'no such table: sessions' by calling db/initialize during Native Worker handshake and ensuring the shared ~/.open-cowork/data.db gets the full app schema (not Job tables only).
Changed
- Replaced the hard 45s first-progress failure with recover-and-wait while a Job is still queued/running: desktop now calls agent:recover-stream (subscribe/replay + journal push) and only fails when the Worker is fatal/restarting or a terminal Job still has nothing to deliver.
- Nested project sessions under the project row in the workspace sidebar so expand/collapse can animate height instead of remounting flat list siblings.

Added
- Added non-interactive CLI print mode (cowork -p / --print) with text, json, and stream-json output, optional --timeout / --max-turns, and stdin prompt support for CI pipelines.
- Added CLI session continue/resume flags (-c / --continue, --resume <session-id>) so a folder’s stored CLI session can be restored before entering the UI or print mode.
- Added cowork update --repair to reinstall the platform Native Worker binary without a full package reinstall.
- Added CLI /mcp and /skills commands, plus a CLI-side MCP host that syncs ~/.open-cowork/mcp-servers.json and serves Worker reverse requests (mcp:call-tool, mcp:read-resource).
- Added Native Worker session-host routes (agent/session-open, agent/session-send, agent/session-close) so a hosted session keeps the run template and canonical conversation inside the Worker.
- Added shared Native Worker wire-protocol helpers (src/shared/native-worker-protocol.ts) and cli/scripts/sync-shared.mjs to vendor those contracts into the CLI package before typecheck/build.
- Added CLI host-adapter registry for reverse-request surfaces (permissions, AskUserQuestion, plan mode, skills, MCP) so unregistered methods fail explicitly instead of hanging the Worker.
- Added fullscreen transcript windowing by estimated message height, SGR mouse wheel/click scrolling, bracketed paste, and a user-overridable keymap (~/.open-cowork/cli-keymap.json).
- Added PTY golden snapshot tests (cli/test/) with a deterministic fixture runtime (OPENCOWORK_CLI_FIXTURE=1) across the documented width matrix, plus a session-host smoke script.
- Added a first-run provider onboarding flow that recommends Routin AI, opens the API-key page in the system browser when available, and keeps a manual-link fallback for SSH/containers (OPENCOWORK_CLI_NO_BROWSER=1).
Changed
- Desktop and CLI Native Worker clients now share the same frame decoder, heartbeat constants, and protocol framing instead of maintaining duplicate socket parsers.
- Expanded the provider setup panel with grouped catalogs (configured / presets / custom), multi-step progress, and clearer review/save messaging.
- Wired npm run cli:sync-shared into CLI typecheck/build, and added npm run cli:test for the golden suite.
Fixed
- Hardened doctor output around MCP config parse failures and missing custom skills directories so readiness checks report actionable status instead of silent gaps.

Added
- Added a full CLI Markdown renderer with headings, lists, tables, block quotes, links, task items, terminal-safe wrapping, and syntax-highlighted code blocks.
Changed
- Improved desktop streaming responsiveness by bounding live Markdown rendering, avoiding redundant scroll corrections, and stopping idle typewriter animation frames.
- Unified provider-facing tool-schema sanitization across Gemini request paths while preserving local validation semantics for unsupported provider keywords.
- Improved CLI runtime coordination and Plan-mode transitions, including cancellation of an active Worker turn when leaving Plan mode.
Fixed
- Hardened Anthropic and OpenAI tool-call payload generation and validation for provider-specific schema and argument shapes.

Added
- Added an SQLite-backed Native Worker Job runtime with idempotent submission, status/result queries, cancellation and command APIs, bounded concurrency, lane-aware scheduling, and durable event storage.
- Added separate Control and Event IPC transports for the desktop app and CLI, with independent Event reconnection plus acknowledgement and replay of unconsumed agent-stream envelopes.
- Added CLI workspace-file references through @ search, including a structured reference bar, prompt-history and rewind restoration, workspace-bound path validation, and bounded text-context injection.
- Added a live CLI turn-status row for requesting, thinking, responding, and tool-use phases, showing elapsed time, transfer direction, token estimates, reasoning effort, and animated activity text.
- Added compact CLI diffs for successful Edit tool calls, including file names, addition/deletion totals, contextual lines, and truncation for large replacements.
- Added masked terminal provider setup through cowork config, /provider, and /config, with common quick presets, custom compatible endpoints, and atomic writes to the provider store shared with the desktop app.
Changed
- Bumped the Native Worker protocol to version 2 and extended route discovery with execution mode, result mode, and lane-policy descriptors so hosts can dispatch eligible methods through the durable Job runtime.
- Changed Job wait semantics so a foreground timeout leaves committed work queued or running for later status checks, while an explicit abort issues a best-effort cancellation even when it races Job submission.
- Made queued Jobs survive Worker restarts, marked interrupted in-flight work with an explicit failure state, and propagated cancellation through Agent providers and tool execution.
- Expanded Native Worker verification around durable Jobs, split IPC, message windowing, and sub-agent history persistence.

Added
- Added CLI clipboard image paste via Ctrl+V for vision-capable models (PNG/JPEG/GIF/WebP, 20 MB each, up to 10 images), with macOS pasteboard, Windows STA Clipboard, and Linux wl-paste/xclip backends sending structured image/base64 content blocks to Native Worker.
- Added a model-aware /effort panel that only offers reasoning-effort levels supported by the active model, and surfaces thinking support in the CLI status line.
- Added Native Worker and sidecar diagnostics snapshots (startup phase, last exit, stderr tail, active runs, last initialize error) and attach them to desktop chat runtime failures for easier troubleshooting.
- Added ordered assistant content segments in the CLI so thinking and answer text keep the real stream order instead of collapsing thinking to the end of the reply.
Changed
- Expanded OpenAI Chat Completions reasoning extraction in Native Worker through a shared ReadReasoningText path that accepts multiple provider aliases and nested text shapes without duplicating or stringifying opaque payloads.
- Improved CLI transcript and status projection: provider retries update a single bottom activity row, collapsed thinking shows duration with ctrl+o expand, context/cache metrics avoid noisy sub-1% rounding, and terminal resize/Ctrl+L fully redraw the frame.
- Hardened main-process Native Worker lifecycle tracking with richer session diagnostics and clearer initialization failure capture in NativeAgentRuntimeManager.
Fixed
- Fixed markdown local-path links (file:// and Windows drive-letter paths) being stripped by react-markdown, so chat answers, thinking blocks, previews, changelog, and release notes can open those targets again.

Added
- Added the `/rewind` command and Rewind panel: every user message becomes a checkpoint, with conversation restore that copies only the canonical prefix into a new session row, plus code restore that replays tracked local changes in reverse order through Native Worker and skips paths whose current hash no longer matches the tracked post-edit state.
- Added the `/config` command with a searchable terminal editor for the shared settings that affect the CLI host (active model, thinking, automatic context compression and threshold, dedicated compression model, provider timeout, tool and sub-agent concurrency, CodeGraph surface); updates go through Native Worker, with a compatibility path for older installed workers.
- Added the `/compact [focus]` command that compresses canonical context through `agent/compress-context`, replacing and persisting the Worker message history on success while keeping the original messages on summarizer failure.
- Added the `/context`, `/cost`, and `/doctor` commands for canonical context usage and compact trigger reporting, recorded usage cost aggregation with shared per-million-token prices, and the same read-only Worker transport probe as `--doctor`.
- Extended the model picker with a capability-driven second step mirroring the desktop model switcher: reasoning effort, Anthropic `budget_tokens` and cache TTL, priority service tier, built-in search, Responses WebSocket, and image generation, plus read-only protocol, context, output, and price metadata.
- Added a persistent two-row status projection with canonical context usage, cache read hit rate, and price-based cost, refreshed on canonical runtime transitions.
- Added `settings/patch-persisted-store` to Native Worker for atomic nested persisted-store patches.
Changed
- Restricted the command registry to commands with a complete CLI host implementation; unadvertised names (`add-dir`, `background`, `branch`, `diff`, `init`, `mcp`, `memory`, `resume`, `theme`) are rejected locally instead of appearing in the menu.
- Made `/clear` clear the runtime's canonical messages and the session's persisted message rows through Native Worker.
- Persist model selection to the shared provider-store index and write model/settings changes to the same split store used by the desktop app.
Fixed
- Hardened tracked-change rollback in Native Worker: symbolic-link paths are skipped to protect their targets, and the current file state is validated against the tracked post-edit hash so external modifications are never overwritten.
- Detected OpenAI Responses WebSocket transport-unavailable errors and surfaced them as explicit exceptions instead of silent stream failure.

Added
- Added Agent Runtime v2 capability negotiation with versioned tool manifests,
session and project context binding, and runtime feature discovery.
- Added agent-definition schema v2 parsing with explicit tool allowlists,
bounded turn limits, and compatibility warnings for legacy definitions.
- Added dependency-free verification for compact request-view pairing and
truncation behavior.
Changed
- Hardened sidecar tool authorization with manifest validation, provider-visible
tool reconciliation, input-schema checks, and project-scoped extension tools.
- Preserved raw provider tool arguments and rejected malformed or hook-mutated
tool input before execution with structured error results.
- Changed context compression to use the generated summary as the complete
model-visible history by default, while pairing boundaries and summaries by
id and keeping their ordering stable.
Fixed
- Fixed compact summaries being missed when message sorting moved them before
their associated compression boundary.
- Added session and canvas project checks to prevent capability snapshots from
being reused with the wrong agent run context.

Added
- Added leaf-worker guardrails for sub-agents and teammates so they cannot spawn nested agents.
Changed
- Updated sub-agent tool inheritance and prompts to retain the parent's tools except Task, with clearer delegation guidance.
Fixed
- Prevented recursive sub-agent creation by removing the Task delegation tool from resolved sub-agent tool sets.

Added
- Added a full spreadsheet preview and editing workspace for XLSX, XLSM, CSV, and TSV files, including formulas, formatting, multi-sheet navigation, undo/redo, save support, and an ExcelJS round-trip verification script.
- Added an app-wide keyboard shortcut layer and a title-bar mode switch shared by the main workspace and detached session windows.
- Added file-type-aware icons and richer file references across the composer, inline file tags, and file tree.
- Added an Office Suite architecture plan covering plugin contracts, tool families, milestones, and delivery risks.
Changed
- Migrated Gemini providers from the retired generateContent flow to the Interactions API, including native-worker support and legacy provider-configuration migration.
- Hardened native-worker HTTP handling with typed DNS, connection, TLS, proxy, reset, and incomplete-stream classification; configurable retry budgets; jittered backoff; connect deadlines; TCP keepalive; and clearer user-facing network errors.
- Reworked context compression to preserve the 12 most recent messages, allow longer summaries, isolate circuit breakers per provider, and preserve the original conversation when summarization fails instead of silently truncating it.
- Collapsed successful tool-work details into a compact process summary while keeping failed, cancelled, approval-gated, and visual-output runs expanded.
- Made subagent_type optional with custom as the default, increased the default sub-agent turn budget, refreshed OpenAI model pricing, and consolidated layout and shortcut behavior between regular and detached sessions.
Fixed
- Fixed tool-result continuation events being dropped when a previous message ID was reused.
- Improved streaming auto-scroll release behavior and message-rail ordering and follow state.

Added
- Added message windowing for long conversations, including targeted verification scripts for windowing behavior and pure message-window data handling.
- Added native-worker request timeout handling and resilient message/session persistence paths for long-running agent tasks.
Changed
- Updated native-worker context compression, video tool integrations, Seedance IPC handling, provider settings, and related renderer UI.
- Enabled recursive Git submodule checkout in the native-worker build workflow.
- Improved collapsible tool-output panels by observing dynamic content size changes.
Fixed
- Flattened top-level oneOf, anyOf, and allOf tool input schemas before Anthropic Messages API requests, preventing HTTP 400 validation failures.
- Fixed stale animation-completion state in collapsible panels.

Added
- Added an openai-video provider type with a dedicated native-worker module and Sora 2 preset model, alongside the existing Seedance and xAI video paths.
- Added Seedance 2.0, 2.0 Fast, and 2.0 mini video models to the Volcengine and Routin AI presets, and Gemini 3 Pro Preview plus Gemini 3.6 Flash to the Google preset.
- Added custom provider icon upload (PNG, JPG, WebP, GIF, SVG up to 1 MB) with validation, surfaced through ProviderIcon for user-added providers.
- Expanded the draw-canvas assistant into a full agentic tool loop with 23 canvas tools — read/create/update/delete/duplicate/connect/move/resize/select nodes, run/retry/cancel generations, image and video generation, image editing, trigger management, and canvas import/clear — plus approval prompts for destructive actions.
- Added assistant conversation controls: message edit-and-regenerate, regenerate, delete with side-effect warnings, pinned node context, and image attachments for vision-capable models.
- Added a per-canvas Agent workspace with ensure/rename/trash IPC handlers; deleting a canvas now moves its workspace to the system trash.
- Added node event triggers so one node can run another when it changes state (queued, started, progress, succeeded, failed, cancelled, interrupted, updated), with cycle and duplicate detection.
- Added dedicated image, video, and text generation node types with quality and size controls, including Sora reference-image size validation and single-reference limits.
- Added a node error banner that exposes the complete provider response with copy-to-clipboard for troubleshooting.
- Added HTTP Range support to the local media protocol so video seeking works and MP4 files with trailing metadata no longer stall on playback.
- Added a disconnect action with loading state to the SSH connected workspace and connection detail panel.
Changed
- Migrated CodeGraph.Core, CodeGraph.Tests, and Worker.Runtime into the external AIDotNet/CodeGraph submodule, removing 200+ locally vendored C# sources. Cloning now requires git submodule update --init --recursive; predev.mjs and publish-native-worker.mjs fail early with an explicit message when the submodule is missing.
- Extracted MCP tool and resource definition builders (buildMcpToolDefinitions, buildMcpResourceDefinitions) so the canvas agent can reuse the same schemas as the chat runtime.
- Reworked canvas graph persistence, node factory, and generation hooks around the new node types and trigger graph.
Fixed
- Restored the 1.0.1 through 1.2.4 changelog history that was accidentally dropped from CHANGELOG.md.

Added
- Added a compact active-additions count to the chat composer and refreshed the skills menu with clearer sections, descriptions, and plugin/skill indicators.
- Added chat actions and localization support for the updated composer menu, including plan mode, goal mode, and media attachment entries.
Changed
- Improved chat action handling and transcript metadata for the updated composer interactions.
- Expanded local media access to support persisted images and videos from both application storage locations.

Added
- Added a background/attached browser webview mode so agent-driven navigation runs off-panel without stealing the right panel; added a runtime reattach path that rehydrates in-flight agent state after HMR or window reload.
- Added a Vietnamese translation pass covering AI Coding, CodeGraph, SSH, pet, chat, layout, and settings copy across all bundled locales.
- Added CodeGraph grammar manifest tooling with validation and a predev script, and expanded PE export name parsing plus grammar asset status/UI messaging.
- Added a dedicated Canvas Assistant model picker so the draw-canvas assistant no longer piggybacks on the global chat model.
Changed
- Split the monolithic SSH store into modular slices and removed the accidentally checked-in * 2.* file duplicates left over from an earlier merge.
- Reworked the settings page context-compression controls (moved into Settings, cleared from the model switcher and provider panel) and unified compression across renderer, native worker, and automatic flows.
- Refined native-worker publish, sidecar management, and CodeGraph asset resolution paths.
Fixed
- Fixed CodeGraph import resolver and extraction/resolution test coverage for expanded grammar cases.
- Fixed a class of provider retry edge cases via a shared retry-policy helper across Anthropic, Gemini, OpenAI Chat, and OpenAI Responses transports.

Added
- Added a dedicated compression model setting, allowing context summaries to use a model separate from the active chat model.
- Added a global context compression ratio setting and migrated legacy per-model compression ratios to the global value.
- Added Vietnamese translations for AI Coding and CodeGraph features.
Changed
- Unified context compression behavior across the renderer, native worker, and automatic compression flows.
- Enhanced SSH settings with additional options and clearer connection status messages.
- Added provider request retry telemetry to show the retry attempt, HTTP status, and next delay in the chat UI.
Fixed
- HTTP 500 and other transient provider failures now retry up to 10 times with progressively increasing delays before stopping.
- Removed obsolete per-model compression ratio controls and safely cleaned up legacy persisted settings.

Added
- Added Kimi K3 model support with parameter carry toggles and quota display for Kimi provider.
- Enhanced CodeGraph RPC daemon with code visualization, indexing UI, and worker runtime support.
- Enhanced canvas assistant with local tool loop, draggable panel, context chips, and per-project session persistence.
Changed
- Refactored SSH module with complete lifecycle rewrite (connection handling, IPC handlers, terminal components, and store management).
- Removed legacy sidecar module and cleaned up unused email-drafter and post-to-x skills.
- Enhanced MessagePack IPC serialization and shared type definitions for improved data fidelity.
- Improved renderer stores, tools, settings UI, and internationalization across multiple components.
- Updated native worker with CodeGraph executor and enhanced tool support.
- Refined sidecar management and file handling in main IPC layer.
Fixed
- Improved SSH terminal session management and JSON parsing robustness.

Added
- Rewrote the Draw page as a node-graph canvas (text/image/config nodes with connections, upstream/downstream generation, crop/mask/outpaint/upscale/split dialogs, an asset library, a prompt library, and a canvas assistant panel), replacing the previous form-based layout.
- Added Seedance video generation support via a Volcengine Ark provider, including a new native-worker module for the async Ark video protocol, main-process IPC handlers, and a "Seedance Video" provider option with a video model category in provider settings.
- Added inpaint/outpaint mask editing for the native OpenAI Images worker (action: 'edit' with a mask part) and a corresponding renderer-side mask utility.
Changed
- Moved sidebar navigation so the primary rail now links directly to the Draw page, with Plugins moved into the overflow menu.
- Simplified the SSH page toolbar by removing the redundant Hosts/SFTP pill switcher and unused tone-border helpers.
- Restyled the SubAgent hover card and icons to use theme-aware tokens instead of hardcoded dark colors, and removed unused dot-matrix and order-label helpers.
- Trimmed duplicated logic out of ImageEditDialog and ImagePreview in favor of shared mask/graph utilities.

What's Changed
- feat(build): add Windows no-install zip distribution by @cipherTing in https://github.com/AIDotNet/OpenCowork/pull/125
Added
- Added a SubAgents workspace and runtime management panel for viewing active, completed, and historical sub-agent runs with their details and context.
- Added Windows no-install ZIP distribution support for users who prefer portable builds.
Changed
- Improved SubAgents navigation, history loading, concurrency state, permission context, and localization across the workspace.
- Updated update handling so Windows no-install builds direct users to manual release downloads instead of using the installer auto-update flow.
Fixed
- Added a safe one-time retry for OpenAI Responses requests interrupted before the first streaming event, avoiding duplicate streamed output or tool execution.
- Prevented Windows no-install packages from contaminating installer update metadata.

What's Changed
- feat: render Bash-execution artifacts as a card with Reveal/Open actions by @warischa in https://github.com/AIDotNet/OpenCowork/pull/126
- fix(chat): 优化折叠动画并修复虚拟列表展开滚动跳动 by @cipherTing in https://github.com/AIDotNet/OpenCowork/pull/130
- chore(dev): silence common npm run dev startup warnings by @cipherTing in https://github.com/AIDotNet/OpenCowork/pull/129
New Contributors
- @warischa made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/126
Added
- Added support for background proxy task generation and processing, enabling sub-agents to run background tasks independently.
- Enhanced session change review mechanism with improved detection and handling of session transitions.
- Improved background sub-agent handling with better lifecycle management and state synchronization.
Changed
- Optimized team tool usage logic for better performance and reliability in multi-agent scenarios.
- Refined session-change detection and background task execution workflows for improved stability.

Added
- Added comprehensive permission management system with tool whitelist and command rules, enabling fine-grained control over which tools can be executed and under what conditions.
- Added Permission Panel in settings for managing per-tool allowlists and command execution rules.
- Added PermissionDialog UI component for real-time permission grant/deny approvals during tool execution.
- Introduced AgentRuntimePermissionPolicy in native worker for enforcing permission rules on the C# sidecar.
Changed
- Enhanced SkillsPanel to display and respect permission configurations for skill execution.
- Improved ModelSwitcher with better provider selection feedback.
- Refined native worker communication with updated permission policy enforcement across streaming and tool-execution flows.

Added
- Added split provider persistence under ~/.open-cowork/ai-provider/, with atomic per-provider JSON writes, automatic migration from the legacy config.json entry, and native-worker sync coverage.
- Added versioned built-in provider presets and a "Restore defaults" action that refreshes built-in model definitions while preserving credentials, enabled states, and custom providers.
Changed
- Reduced main-process startup work by deferring terminal, SSH, image/GIF, migration, channel-provider, updater, MCP, and database initialization until needed or after the first window begins loading.
- Split renderer window surfaces and lazy-loaded secondary pages, update release-note rendering, Mermaid support, tokenizer data, and per-language locale bundles to improve first-paint time and reduce memory use.
- Delayed the first Native Worker spawn until shell-environment initialization completes and moved high-volume startup diagnostics to deferred file writes.
- Extracted SSH connection and proxy-jump payload resolution from the full SSH IPC handler so Git, cron, and sidecar consumers can load it independently.
Fixed
- Prevented a rejected second app instance from accessing Electron session APIs or starting the Native Worker before app readiness, eliminating Session can only be received when app is ready during shutdown.
- Temporarily stopped sending the unsupported "mode":"pro" parameter for GPT-5.6 Sol/Terra ultra requests while retaining their medium-effort fallback.

Added
- Added model-aware prompt cache controls for OpenAI-compatible Responses and Chat APIs, including stable cache keys, retention settings, explicit breakpoints, and cache write/read token accounting.
- Added native file-tool recovery for operating-system permission failures, allowing users to grant scoped folder access and retry read, write, edit, notebook, list, glob, and grep operations.
- Added the ultra reasoning-effort level and refreshed saved provider model capabilities when built-in reasoning presets gain new levels.
Changed
- Unified the composer send and stop actions so the primary button stops an active run and returns to sending when idle.
- Improved reasoning-effort presentation with clearer maximum-level effects and provider-specific defaults.
- Normalized prompt-cache usage and billing calculations across chat context totals, analytics, and exported conversations.
- Simplified home and project composers by hiding the duplicate working-folder picker while retaining folder selection in the surrounding page.
Fixed
- Persisted completed or interrupted results for dangling tool calls after stops, errors, or worker crashes so continuation does not silently replay long-running tools or sub-agents.
- Improved native sub-agent and tool-call continuation handling to preserve streamed arguments, results, and terminal state across Responses and Chat API runs.
- Filtered invalid Responses input parts and improved reasoning and cache-usage event parsing for OpenAI-compatible providers.

Added
- AI Coding support for ClaudeCode and Codex, including dedicated settings panels, configuration storage, localized UI strings, and launch helpers.
- Project terminal actions for starting ClaudeCode or Codex sessions from the bottom dock with configured commands, permissions, and environment overrides.
- Right-panel terminal tabs so local and SSH terminal sessions can move between the bottom dock and the right-side workspace.
- Fullscreen support for the bottom terminal dock, allowing the dock to take over the conversation area and restore the previous height when exiting.
- Introduced the Creative Production extension with MCP assets, enabling material import and step-by-step creation workflows including mood board, shot/style intake, ad/scene/offer/logo exploration, and generative polish across 9 skill packs.
- Added local prototype setup and user context management to the product-design extension, supporting prototype bootstrap scripts, user context initialization, communication protocols, and key overrides.
- Collapsible tool-call execution runs in chat, allowing tool-call output to be collapsed for a cleaner conversation view.
- Web search block component for displaying server-side web search results with expandable source list
- Native worker auto-restart with exponential backoff and heartbeat monitoring
- Lifecycle events (onReconnect, onDisconnect) for native worker management
- Support for Anthropic and OpenAI built-in web search tools
- Permission management UI translations and full-access mode confirmation
- Content blocks utility for upserting web search blocks during streaming
- Builtin search enabled flag in provider store configuration
Changed
- Reworked the title-bar folder action to open the existing right-side Files tab instead of a separate working-folder panel.
- Restored the closed right panel to a zero-width state without a persistent border or icon rail.
- Improved bottom terminal sizing and xterm layout so maximized or resized terminals keep their bottom row visible.
- Refactored code structure for improved readability and maintainability.
- Renamed product design references and removed related dead code.
- Refactored AssistantMessage to simplify debug tool call handling
- Improved native worker resilience with crash logging and stderr capture
- Enhanced provider store with built-in search capability detection
Technical
- Extended terminal session state with surface tracking and environment overrides for AI Coding launches.
- Added shared ClaudeCode/Codex launch resolution utilities and settings routes.
- Creative Production extension ships with a standalone MCP server, widget assets, and Python scripts.
- Product Design extension includes prototype templates, bootstrap scripts, and user context management tooling.
Fixed
- Native worker now properly restarts on unexpected crashes with exponential backoff
- Web search blocks update in place instead of stacking duplicates

What's Changed
- feat(updater): confirm install after update download by @cipherTing in https://github.com/AIDotNet/OpenCowork/pull/119
New Contributors
- @cipherTing made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/119
Added
- Added an install-confirmation step to the auto-updater: downloaded updates now wait for the user to choose "Update now" (restart and install) or "Update later" instead of restarting automatically.
- Introduced update:status and update:install IPC handlers so the renderer can query whether a downloaded update is ready and trigger installation on demand.
- Added a "ready to install" update state in the title bar and settings, with matching UI strings across all supported locales.
Changed
- Deferred quitAndInstall until the user confirms; install failures now surface an update:error back to the window instead of force-quitting the app.
- Clarified the auto-update setting description to reflect that updates download automatically but ask before restarting to install.
- Refined the context-memory compression description to reflect model-aware compression thresholds and summary-based history preservation.

Added
- Implemented comprehensive capybara pet companion system with personality-driven interaction model
- Added pet experience progression system with level-based milestones and stat tracking
- Implemented pet skin management system with customizable appearance options
- Added pet agent runtime for proactive companion behavior and voice integration
- Implemented pet memory system for contextual awareness and personality persistence
- Added pet pose/animation system with 13 different animation states (idle, sleep, eat, play, swim, bathe, etc.)
- Created pet studio for pose generation, expression customization, and skin preview
- Added native worker OpenAI Audio module with configurable audio processing and streaming support
Changed
- Removed documentation website from main repository (consolidate in separate docs project)
- Enhanced provider stores (Routin AI, Xiaomi) with new capability management
- Improved IPC messagepack channel routing with additional serialization support
- Refined usage analytics tracking with enhanced provider metrics
- Improved C# native worker OpenAI Audio Models with robust deserialization and error handling
Technical
- Integrated multi-modal pet agent with adaptive response generation
- Implemented pet persistence layer with Zustand-based state management
- Added pet-specific IPC handlers for lifecycle and data synchronization
- Extended localization system with pet UI and messaging (en/zh)

Added
- Added Anthropic Claude API integration to the native worker with proper message validation and handling for trailing user turns (required by Claude Opus 4.6+, Sonnet 5, and Fable 5).
- Implemented cache creation token tracking split by time windows (5-minute and 1-hour) for improved cache cost analysis and display.
Changed
- Removed plan executor from native worker runtime; plan execution now handled through the unified agent loop.
- Enhanced token usage metrics display with detailed cache creation breakdown and improved cost calculations.
- Improved SSH workspace connection handling and status monitoring with better error recovery.
- Refined permission dialog presentation and model switcher UX for better clarity.

Fixed
- Ensured first-run global memory directory initialization and onboarding profile writes can create missing USER.md without surfacing startup errors.
- Broadened missing-file detection for Windows and native-worker messages such as Could not find file, allowing onboarding, settings, and memory fallback paths to create files instead of failing.
- Made local MemoryRead create missing memory files from templates before reading them.
- Treated missing OpenAI Responses previous_response_id replay errors as recoverable and retried with full sanitized input.
Security
- Upgraded the native worker SQLitePCLRaw bundle to 3.0.3, pulling SourceGear.sqlite3 3.50.4.5 and clearing the SQLitePCLRaw.lib.e_sqlite3 vulnerability warning.

Fixed
- Removed BOM character from package.json for better compatibility with various tools and parsers.
- Added pointer-events-none to user message locator container to prevent it from intercepting clicks on underlying UI elements.
Changed
- Optimized streaming and debug overhead by implementing on-demand caching and batch refresh instead of full recalculation on every increment, reducing O(n²) complexity in main and sidecar processes.
- Enhanced debug request body handling by upgrading from in-memory truncation to persistent disk storage for better readability of long request bodies.
- Added constraints for sub-agent and plan mode operations: sub-agents can no longer directly create/execute/approve plans, and plan mode tools are disabled by default in sub-agent contexts.
- Implemented context injection for plan mode, plan revision, plan execution, plugin channels, system commands, and slash commands to ensure model input consistency with current runtime state.
- Added prompt cache key truncation and derived sub-agent cache keys for OpenAI Responses API.
- Added duplicate task deduplication during runtime to prevent identical sub-agent prompts from being executed multiple times.
Fixed
- Fixed sidecar manager initialization and communication flow.
- Improved session image export functionality.
- Updated routing AI provider configuration.

Changed
- Improved native worker development selection and publish flow so ready debug builds are preferred and failed native publishes do not wipe bundled resources.
- Updated Electron Builder metadata by removing invalid and deprecated Windows build version fields.
Fixed
- Ensured AI API requests always carry a non-empty versioned User-Agent header such as OpenCowork/1.0.2 across renderer, main proxy, cron, sidecar, native worker, image, audio, skills, souls, and WebSocket request paths.
- Normalized legacy default User-Agent placeholders like OpenCowork to the versioned app header while preserving provider-specific custom values.

Added
- Added native MessagePack sidecar transport and native worker runtime coverage for agent tools, database, file, Git, shell, SSH, terminal, sync, extension, memory, goal, task, cron, and provider execution paths.
- Added request stop support for agent runs, including shared run identifiers and stop-aware provider/runtime models.
- Added message windowing, stable long-context handling, and richer export output for chat transcripts.
- Added Luckin Coffee extension resources and bundled native worker publishing for platform-specific release artifacts.
Changed
- Improved live tool-call rendering with streaming tool argument updates, foreground/background session synchronization, and session change summaries.
- Updated release CI to publish the native worker per target runtime and include .NET setup in Windows, Linux, and macOS packaging jobs.
- Refined the user-message locator into a compact right-edge marker with hover previews for faster navigation in long conversations.
- Default command-tool live previews to the session working folder when no explicit cwd is provided.
Fixed
- Forwarded native and SSH shell output chunks through shell:output so terminal/tool cards update while commands are still running.
- Reduced stale shell-output listener buildup by using a single native worker forwarding listener.
What's Changed
- fix: UI Optimize by @Austin-Patrician in https://github.com/AIDotNet/OpenCowork/pull/113
New Contributors
- @Austin-Patrician made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/113
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/1.0.0...1.0.1

Added
- Added per-session model selection mode (inherit / auto / manual) so a session can follow the global model, route automatically, or pin a specific provider+model independently of the global active model.
- Added model_selection_mode column to sessions (additive migration; existing provider+model rows backfilled to 'manual').
- Added centralized session model resolution in session-model-resolution.ts.
- Added "Follow global model" option in ModelSwitcher with split setSessionModelManual/Auto/Inherit actions on the chat store.
- Added thinking/reasoning support for MiniMax and Kimi (Moonshot) providers.
- Added live SSH process monitor panel (SshProcessMonitor) for inspecting remote host processes.
Changed
- Stopped mutating the global provider store on session switch/sync; model state is now session-scoped.
- Refactored Anthropic API integration for cleaner cache control and response handling.
- Reworked ModelSwitcher and InputArea to support the new per-session model mode.
- Improved SessionListPanel, PreviewPanel, and SSH terminal status panel layouts.
- Updated multiple provider presets (Baidu, Gitee AI, MiniMax, Moonshot, OpenRouter, Qwen, SiliconFlow, Routin AI) with refreshed model lists, thinking support, and pricing.
- Threaded model selection mode through channel auto-reply, channel handlers, and DAOs.
Fixed
- Kept the first tool-call id when providers re-emit ids per streaming delta, preventing tool-call matching failures.
- Various UI consistency and component rendering improvements.
Removed
- Removed the legacy project wiki subsystem (wiki-dao, wiki-handlers, ProjectWikiPage, wiki-tool, wiki-generator) in favor of the memory/AGENTS.md workspace protocol.
- Removed unused architecture diagram SVG assets.

Added
- Added request timing metrics (TTFT, TPS) display in InputArea for real-time performance monitoring.
- Added cache shape debugging information in AssistantMessage showing system hash, tools hash, message prefix hash, tool count, and cache read ratio.
- Added new models to Routin AI provider: Kimi K2.7 Code HighSpeed, MiMo V2.5 Pro, MiMo V2.5 with thinking support.
- Added getCacheCreationTokens and getUsageCacheHitRate utility functions for improved token usage tracking.
- Added calculateCacheReadRatio function for cache analytics and debugging.
- Added RequestTiming type support for detailed request performance metrics.
- Added Volcengine (火山引擎) provider preset with Doubao Seed 2.1 Pro/Turbo models supporting vision, function calling, and thinking.
- Added Doubao Seed Evolving, Seed 2.1 Pro (260628), and Seed 2.1 Turbo (260628) models to Routin AI provider.
- Added sub-agent workspace protocol injection — sub-agents now automatically load project-level AGENTS.md as authoritative workspace context.
- Added parallel tool calls prompt for sub-agents to maximize concurrent independent tool invocations.
- Added sub-agent runtime cache policy with prompt cache key generation for OpenAI Responses WebSocket backend.
- Added turn-context injection for plan mode sub-agent runs.
- Added stable tool definition sorting (by name → description → schema) for deterministic cache keys.
- Added cache-shape module with stable hash/serialize utilities for prompt cache debugging.
- Added SmoothTokenNumber component with animated token counter transitions.
- Added thinking content extraction and model avatar with breathing dots animation for streaming placeholder state.
- Added model info hover card panel in model switcher showing provider, capabilities, and auto-routing target model.
- Added toolUseOrderById index for sub-agent inline rendering decisions.
- Added canRenderInlineSubAgentRun heuristic for multi-sub-agent message layout decisions.
- Added resolveSubAgentProviderConfig with session-scoped provider resolution.
Changed
- Refactored chat streaming placeholder from static "生成回复" to "Thinking.../正在思考中……" with per-model avatar and animated breathing dots.
- Replaced model switcher plain label with hover card showing model name, provider, and capability tags for auto-routing clarity.
- Deferred large Read tool result text rendering to lazy mount for performance.
- Memoized MCP active tool computation to reduce store selector re-derivation.
- Set pending assistant message timestamp to current time instead of epoch 0 for correct sort order.
- Hidden file diff dialog close button to match interaction design.
- Enhanced chat components (AssistantMessage, ToolCallCard, InputArea) with improved UI and performance.
- Improved settings pages and analytics overview with better data visualization.
- Updated Anthropic API integration with optimized cache control logic and cache target selection.
- Improved MCP tools and agent runtime with better error handling and stability.
- Enhanced internationalization (en/zh) with updated translations.
- Updated Routin AI provider models with new pricing and thinking support configurations.
- Various performance and stability improvements across the application.
Fixed
- Fixed widget rendering logic in ToolCallCard with simplified state management.
- Fixed cache hit rate calculations and token usage tracking.
- Fixed various UI inconsistencies and improved component rendering.
Removed
- Removed unused imports and properties (Loader2, Wrench icons) for cleaner codebase.
- Removed redundant model thinking indicator properties for simplified component API.
Fixed
- fix(mcp): expose connected MCP tools in chat for MCP-enabled sessions.
- Fixed sub-agent inline rendering: background sub-agents now properly hidden and synchronous sub-agents rendered inline when no visible content sits between them.
Documentation
- Added comprehensive Repository Guidelines (AGENTS.md) covering project structure, build commands, codi

Added
- Added Code workspace mode with Explorer and Source Control sidebars, central preview/editor tabs, Monaco diff editing for Git and agent changes, and project terminal docking.
- Added SCM diff plumbing for reading file content at Git refs, editable unstaged/untracked diffs, and split/inline diff mode switching in previews.
- Added configurable concurrent sub-agent limits in Settings plus queued/dequeued stream events and clearer approval/retry status updates for teammates and Task runs.
- Added automatic recent visual artifact reuse for image-to-code and other visual follow-up prompts, including image reference preservation for OpenAI-compatible backends.
- Added the built-in create-extension skill and new provider presets for glm-5.2 and kimi-k2.7-code.
Changed
- Refreshed the chat tool-call rendering stack with inline context-compression summaries, completion metrics, richer compact headers, and improved image/tool cards.
- Replaced the right-panel Context entry with a Review-first workflow and refreshed the Settings navigation layout.
- Updated app-plugin resolution so project settings inherit global plugin defaults more consistently.
- Switched Electron download mirrors to npmmirror.com and excluded html-to-image from renderer dependency prebundling.
Fixed
- Fixed context compression so repeated summarizer failures now fall back to local truncation instead of letting context grow unbounded.
- Fixed Git diff preload/cache behavior and preview save refresh flow for edited diffs in the workspace.
- Fixed Windows native rebuild/package flow by skipping prebuilt module rebuilds in postinstall and disabling duplicate electron-builder rebuilds.

Added
- Added shell environment variable management in Settings with localized labels and IPC support.
Changed
- Scoped session change summary cards to the current assistant message and its tool calls so earlier assistant messages can show their own file-change summaries.
- Improved terminal environment handling so configured shell variables are applied when launching terminal sessions.

Added
- Added heuristic Auto model routing signals for task type, tool intent, complexity, risk, and routing reasons.
- Added Auto model routing metadata display in the model switcher, including complexity and risk labels.
- Added localized Auto routing fallback, complexity, and risk labels in English and Chinese.
Changed
- Reworked Auto model selection to combine heuristic signals, sidecar classifier output, low-confidence reuse, and policy overrides before choosing main or fast models.
- Limited sidecar prompt tool context to requests that actually need tools, reducing unnecessary channel and MCP context for simple prompts.
- Simplified page transitions to avoid extra animated wrapper behavior.
Fixed
- Fixed unsafe fast-model routing for complex, high-risk, tool-required, or workspace-aware requests by forcing main-model routing when policy gates require it.
Added
- Added Runtime Status Panel with Git integration: branch info, ahead/behind status, changed file count, and diff line summaries for the working directory.
- Added SSH connection context display in the runtime status panel showing connection name and working folder.
- Added source files panel in runtime status showing selected files from input drafts.
- Added model management defaults system: creating provider models auto-applies managed model parameters when the model ID matches.
- Added preset vs custom provider categorization in model management ("预置" / "自定义").
- Added responsesImageGeneration.partialImages config option for OpenAI Responses models.
- Added runtimeStatusPanelOpen state to UI store with toggle/set and persistence support.
- Added compact request view logic (applyLatestCompactRequestView) to filter UI-only messages and show only the latest compaction boundary in sent requests.
Changed
- Rewrote RuntimeStatusPanel: switched from hover-triggered narrow panel to a toggleable panel with dedicated open/close button in the title bar.
- Redesigned FileChangeCard layout: replaced inline CompactEditDiff with NewFileContent component for new file previews.
- Improved Write/Edit tool live streaming previews: show tail content previews instead of hiding until complete (content_preview, old_string_preview, new_string_preview).
- Enhanced GoalRuntimeService with status change detection (completion/blocked transitions) and more robust lifecycle management.
- Improved model management search placeholder to "搜索模型/服务商……" for broader search scope.
- Updated alert-dialog action/cancel buttons with min-w-0 max-w-full for better text overflow handling.
Fixed
- Fixed goal continuation dispatch to properly fire on goal status transitions in chat actions.
- Fixed pending write preview truncation: no longer appends trailing ellipsis when content already starts with '…'.
- Fixed real-time Write FileChangeCard defaulting to collapsed state.
Removed
- Removed background session toast warning from ask-user-tool.

Added
- Implemented WebDAV sync functionality with full UI integration for managing sync settings.
- Added SyncPage component for configuring and monitoring WebDAV sync operations.
- Introduced IPC channels for sync configuration, status queries, and conflict resolution.
- Created sync type definitions for managing sync configurations and operations.
- Enhanced SshTerminal to handle terminal resizing and notify remote sessions.
- Added UI store management for opening and closing the sync page.
- Implemented conflict resolution handling in the sync process.
Changed
- Updated localization files for sync-related strings in both English and Chinese.

- Added context compression status card to surface current compression state more clearly in the UI.
- Improved message merge logic so history is preserved when consolidating messages.

Added
- Enhanced ModelSwitcher component with Popover dropdown for better model selection UX.
- Added workspace directory change detection with automatic refresh and provider-based model grouping.
- Implemented message write tracking and flush mechanism for chat sessions.
- Added browser goal runtime support for agent sessions.
Fixed
- Fixed sub-agent event handling for background sessions with session-scoped state lookup.
- Improved code structure for better readability and maintainability.

Added
- Unlocked Chat-mode tool selection so models can now use file, terminal, MCP, and other tools when available, while preserving approval flow constraints.
- Added automatic default working directory creation for Chat sessions without a project context, preventing ambiguity in file-based tasks.
- Improved manual context compression with more aggressive shrinking based on actual compressible message count and recent input token estimation.
- Enhanced model switcher to display models grouped by provider for better readability and navigation in long lists.
Fixed
- Fixed sub-agent event handling for background sessions by removing foreground session checks and using session-scoped state lookup.
- Delegated sub-agent cleanup to session-aware state synchronization for more reliable background session management.

Added
- Added goal runtime service support for blocked and usage-limited goal states so active goals can persist, resume, or pause based on blocker and usage conditions.
- Added dedicated browser tool cards and browser tool name grouping so browser tool calls render consistently in assistant messages.
- Added a Xiaomi coding preset and fast-route model selection support for separate main and fast model workflows.
Changed
- Reworked goal persistence, schema migration, and event emission to support objective replacement, blocked audits, usage limits, and live elapsed-time tracking.
- Improved browser handling with shared built-in storage cleanup, more resilient webview JSON parsing, and screenshot encoding that preserves the native image format.
- Refined the chat composer, working-folder picker, sidebar and project navigation, and task calendar so project-scoped sessions and queued messages behave more predictably.
- Updated OpenAI Responses websocket session scoping in the cron background runtime so agent runs use distinct connection keys.
Removed
- Removed placeholder code-compatible LSP and worktree tool stubs from the code-compatible tool catalog.

Added
- Added Soul marketplace page for discovering and browsing community souls.
- Finalized marketplace install flow with end-to-end soul installation from marketplace.
Changed
- Refactored working directory selector dialog with improved UX and streamlined layout.
- Unified file and tool icon styles across chat cards and session panels for visual consistency.
Fixed
- Added Stage1BuildResult wrapper in memory pipeline to preserve filter reasons and original content, ensuring filtered entries are recorded with accurate metadata instead of fallback values.
Removed
- Removed WeChat UI send skill and its automation scripts.

Added
- Added automatic memory summarization system with backend pipeline, frontend panel, and dynamic context-injection into agent prompts.
Added
- Added automatic sort_order normalization before message reads so sessions with dirty sequence data are repaired in place only when anomalies are detected.
Fixed
- Fixed message query ordering drift caused by gaps, duplicates, or out-of-order sort_order values in messages rows within the same session.
- Added created_at ASC as a secondary ordering key across message queries and stopped upsertMessage conflict updates from overwriting recovered sort_order values.

Changed
- Reworked the draw page image-generation flow with richer prompt optimization, style blending, and image-quality controls.
- Improved OpenAI image provider routing and Responses compatibility so image requests carry the right generation settings across providers.
- Expanded draw history, error reporting, and chat-mode prompt context so failed runs and optimized prompts surface more useful diagnostics.
Added
- Added prompt-style presets and a user core suggestion field to help shape image prompts before generation.
- Added richer error details for image-generation failures, including provider and request metadata when no image output is returned.
Fixed
- Fixed image-generation handling for transparent-background and no-output cases across the draw workflow.

Changed
- Docs IA restructure: Reorganized documentation from flat "Getting Started / Core Concepts / Features / Plugins / Providers / Architecture / Development" into task-oriented navigation: Start, Install, Channels, Agents, Capabilities, Skills, Models, Platforms, Ops, Reference, and Help. Added index pages and meta.json for each section, with redirects preserving all legacy URLs.
- Rewrote docs landing page with streamlined layout, new DocsComponents and CopyCommandButton components, and improved Mermaid diagram rendering.
- Refactored chat-store (441 lines changed) with enhanced session management and state flow.
- Expanded fs-tool (402 lines changed) and bash-tool (128 lines changed) with richer file system and shell execution capabilities.
- Enhanced fs-handlers (279 lines), db-handlers (102 lines), ssh-handlers (81 lines), and shell-handlers (28 lines) in the main process for more robust IPC.
- Refactored messages-dao data access layer for improved message persistence.
- Redesigned GoalSessionControls component with better goal session UX.
- Enhanced InputArea with improved input handling and SkillsMenu with richer skill browsing.
Added
- Code-compatible tool (code-compatible-tool.ts, 321 lines): New tool providing code-agent-compatible aliases for OpenCowork's tool system, enabling seamless interoperability.
- Tool input sanitizer (tool-input-sanitizer.ts, 72 lines): Input validation and sanitization layer for all tool calls.
- New channels.ts IPC channel definitions for extended renderer-main communication.
- Enhanced cron-tool, search-tool, and plan-tool with additional capabilities.
- Expanded tool-types with new tool interfaces and type definitions.
- Docs application screenshot added to public assets.
Fixed
- Updated i18n strings in English and Chinese chat locales for consistency.
- Improved dynamic-context and memory-files agent runtime modules.
- Enhanced use-chat-actions and use-plugin-auto-reply hooks for edge cases.

Fixed
- Isolated built-in browser session storage to prevent contamination of the user's native browser profiles.
- fix(cron): replaced string-concatenated output with chunk-based buffer decoding to avoid encoding truncation.
- fix(weixin): added i18n error keys for QR code and login failure scenarios.
- fix(todo): respected teamToolsEnabled setting in hasActiveTeam guard.
Added
- feat(images): enhanced GIF fallback when no structured frames are available, with improved error branch handling.
- feat(teams): added filtered task definition and disabled-state filtering for team tools.
- feat(plugin): rewrote auto-reply flow with streaming text append, error tracing, full persistence, and replay support.
- feat(chat): added session deduplication by ID (dedupeSessionsById) and image preview in InputArea.

Fixed
- Protected real Chrome/Edge/Brave/Chromium profiles by keeping Electron's writable browser session data inside OpenCowork's isolated storage.
- Clarified browser settings copy so selected browser profiles are used for identity emulation, not direct writable storage reuse.

What's Changed
- fix(style): 修复 SSH 中代码片段页面的样式不统一问题 by @alwaystom in https://github.com/AIDotNet/OpenCowork/pull/98
New Contributors
- @alwaystom made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/98
Changed
- Refreshed the SSH and theme settings UI, including SSH connection management, terminal status presentation, and theme preset handling.
- Improved the chat change review flow with a cleaner review card layout and updated transcript rendering.
Fixed
- Fixed inconsistent code-block styling in SSH support workspaces.

Added
- Added a runtime status panel in the session view that surfaces sub-agent execution summaries from the title-bar control without occupying the input area.
- Added a standalone runtime todo list with collapse/expand handling and earlier-task indicators for long task histories.
Changed
- Unified sub-agent run data aggregation for the runtime panel and sub-agent list so both views share filtering, summaries, failure details, and tool-state mapping.
- Moved in-progress todo presentation out of embedded tool-call rendering and replaced the ping animation with a rotating progress indicator for calmer task updates.
- Removed duplicated layout-side session workspace derivation so panel state is computed from the shared runtime data source.

Changed
- Unified Markdown rendering across changelog, preview/detail panels, system command cards, team notifications, plan reviews, context compression summaries, AskUserQuestion previews, sub-agent reports, and thinking blocks by reusing the shared preview Markdown plugin set.
- Moved the session goal bar below the chat input and made it collapsible, with localized show/hide controls to keep the input area calmer while preserving quick goal access.
Fixed
- Fixed completed team workflows leaving an empty assistant placeholder stuck at the bottom as "thinking" after all team tasks were marked complete.

What's Changed
- feat(i18n): translate Chinese to English + add zoom + optimize dev by @IMMENSE-BUILDER in https://github.com/AIDotNet/OpenCowork/pull/95
New Contributors
- @IMMENSE-BUILDER made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/95
Added
- Added persistent session goals with database storage, sync events, and goal-aware runtime tools for get_goal, create_goal, and update_goal.
- Added a built-in /goal slash command plus context-panel controls to view, edit, pause, resume, and clear the active session goal.
- Added goal usage accounting and auto-continue support so active goals can carry across turns with token/time tracking and budget limits.
- Added task-page controls to abort active runs and render cron plans with second-level precision in the calendar view.
Changed
- Improved cron schedule validation to use node-cron, normalize cron expressions and time zones, and reject invalid schedules before enabling them.
- Reduced task-run loading to the visible calendar window so the task page scales better with larger run histories.
- Hardened OpenAI-compatible chat streaming with a pre-stream OAuth refresh retry for 401/403 failures and better handling for providers that delay terminal SSE chunks.
Fixed
- Fixed Claude base64 image payloads on the cron Anthropic path to send media_type instead of mediaType.
- Fixed scheduled-state propagation for cron run completion payloads and broadened OAuth expiry parsing for providers that return nonstandard expires_at fields.

Added
- Added assistant-message branching so a new session can be forked directly from a previous reply.
- Added Anthropic tool replay normalization to keep tool_use and tool_result history aligned when restoring forked or background sessions.
Changed
- Scoped right-panel, terminal, browser, SSH preview, and related UI state by both session and project to prevent cross-session leakage.
- Hardened the dev startup flow by clearing the Vite cache and pinning the renderer port before launching.
- Refined Codex OAuth header handling to strip session_id and conversation_id outside supported chatgpt.com/backend-api/codex flows.
- Refreshed the packaged desktop icons.
Fixed
- Prevented stale or misaligned Anthropic tool history when replaying forked sessions.

Added
- Refactored backend tools and frontend panels with full search/grep/cache and rich preview capabilities.
- Added new renderer components for rich content preview caching and search result display.
- Enhanced IPC tool channel to support grep search, tool cache, and content preview.
Changed
- Restructured backend tool registration and frontend panel layout for better maintainability.
- Improved tool execution pipeline with caching layer and optimized data flow.
Fixed
- N/A

Added
- Added reasoning mode support for Anthropic/OpenAI with thinking/reasoning parameter passthrough, cache control, and prompt caching markers.
- Added browser plugin capability with IPC for cookie cleanup and tool re-registration on project switch.
- Added new DAO interfaces for querying user messages only and for reverse-lookup run changes by sessionId and toolUseIds.
- Added reasoning effort mapping directly supporting xhigh without client-side normalization.
Changed
- Refactored streaming chat and tool chain to be runtime-state-driven: removed legacy long_running_mode field, now driven by current runtime state and configuration.
- Narrowed theme presets to the default only; removed global theme panel, SSH terminal theme panel, and redundant session title display. Settings migration falls back to default theme on old versions.
- Simplified message list to always load all session messages at once; removed "load earlier messages" button, auto-fill, and scroll anchor recovery. Added session-level deduplication to prevent duplicate tail tool restores.
- Completed Anthropic SSE/usage handling: unified message_start/message_delta/message_stop and data.type, aggregated input/output/cache read/cache creation/reasoning token stats, with cache writes billed per 5m/1h buckets. Tool call end events flush at stream end; message_end acts as fallback.
- Rewrote Clarify mode prompt as a strict "clarify first, then plan" flow with enforced AskUserQuestion/EnterPlanMode/ExitPlanMode closure.
- Enhanced file edit tool to preserve original line-ending style (CRLF/LF), avoiding mixed line endings.
- Tool output with structured errors is now recognized as failure instead of success.
- Run change queries expanded from exact runId match to also support sessionId and toolUseIds reverse-lookup.
- Improved stream rendering with new typing render pool, finer-grained animation classes, and progressive Markdown/table/component reveal.
- AssistantMessage now binds run changes precisely via tool_use ids, filtering out failed file tool results.
- Cron recovery marks still-running background runs as aborted on app restart to prevent hanging states.
- Enhanced request header forwarding security to avoid duplicating body-managed headers.
Fixed
- Fixed multi-line code block and local path recognition in Markdown rendering.
- Stopped duplicate tail tool restoration when resuming sessions.

What's Changed
- feat: 引入 markstream-react 实现 llm 返回 md 时候 流式增量渲染（只渲染新的部分消耗低） by @NOBB2333 in https://github.com/AIDotNet/OpenCowork/pull/90
Added
- Added a new sidebar entry for drawing, with menu highlighting integrated so the feature is discoverable from the main navigation.
- Added streaming markdown incremental rendering support via markstream-react so LLM responses render only newly arrived content.
- Added clarify-prompt and AskUserQuestion flow improvements to make interactive follow-up questions more reliable.
- Added guarded session-clearing actions in the sidebar to reduce accidental destructive operations.
Changed
- Aligned SSH workspace chrome with theme tokens for more consistent visual integration.
- Stabilized provider transport and image persistence in the main process to improve reliability during content handling.
- Improved chat prompt handling and refined the main user-interaction flow.
Fixed
- Prevented the message list from auto-scrolling while AskUserQuestion is pending.

Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.9.85...0.9.86
Added
- Added OpenAI image part support utilities and request_debug event type for richer streaming observability.
- Added model context length and max output token parsing so discovered model capabilities are reflected in provider settings.
- Added request_debug event emission in cron execution, image content filtering, and a 20-result cap on search tool output for consistency across environments.
Changed
- Improved OpenAI chat provider with structured token usage tracking and image part support for more accurate streaming metadata.
- Normalized search result limits across SSH, local, and cron tool execution paths to cap at 20 results uniformly.
Fixed
- Stopped auto-scroll when AskUserQuestion is pending, preventing the message list from jumping during user input prompts.

Added
- Improved wiki generation to collect project source files locally and over SSH with ignore-aware filtering.
Changed
- Synced teammate runtime state so running tasks and member status stay aligned during team execution.

Changed
- Refined chat context compression to use context tokens in usage accounting.
- Preserved image requests during provider routing so image-related turns remain intact.

Changed
- Unified terminal output with the thinking stream display and added a session clear entry point.
Fixed
- Completed terminal session cleanup immediately when a terminal session has already ended, avoiding duplicate pending states.

Added
- Added image-read gating so send actions are blocked until images finish loading.
- Added session message backfill logic to reload the message window after resolving the real total count when the view is empty.
Changed
- Improved shell result handling so terminal output can connect directly to the session flow.
- Updated the model panel interaction model and kept the inline terminal experience in the tool panel.
- Persisted sidebar and panel state, and refreshed debugger and model default values.

Added
- Added inline interactive terminal session rendering in the tool panel for shell execution workflows.
Changed
- Switched shell execution to terminal-backed sessions for a more consistent interactive command experience.
- Persisted sidebar and panel state, and refreshed debugger and model default values.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.9.74...0.9.80

Added
- Added GPT 5.5 model to built-in RoutinAI provider presets.
Fixed
- Expanded streaming message context window and fixed Q&A auto-accept and preview styling issues.

Added
- Added interaction tooling to extend the built-in browser with automation-oriented workflow support.
Changed
- Updated package metadata for the new patch release.
Fixed
- Synchronized lockfile root metadata version with the package manifest version.

Changed
- Tightened the title-bar project action group styling and switched to a lighter-weight layout.
- Unified the workspace panel and image preview flow, added context statistics, and avoided loading full message history unnecessarily.

Added
- Added Responses image generation support and connected it to the streaming preview/config flow.
- Added richer SSH workstation capabilities, including cross-host transport, split-window sync, and theme synchronization.
Changed
- Refactored the SSH and terminal systems to support more unified workspace behavior across local and remote sessions.
- Improved SSH connection inspection, file browsing, SFTP workspace flows, terminal status panels, and related i18n coverage.
- Updated theme presets and runtime theme sync so UI appearance stays consistent across layout and settings surfaces.
- Reworked provider, settings, and chat-side flows to align with the new image-generation and SSH experience.
Fixed
- Fixed several SSH, terminal, and session-store edge cases during navigation, switching, and remote workspace usage.
- Fixed response-stream integration details so image-generation and related protocol changes remain compatible.

Added
- Added background task scheduling support across the main process, IPC layer, and cowork UI.
- Added Responses WebSocket protocol support for provider routing and chat streaming flows.
- Added richer chat streaming error cards with localized coverage for transport, auth, rate-limit, timeout, cancellation, and parameter failures.
Changed
- Synchronized session runtime state across windows so streaming progress, tool activity, and detached-session focus stay consistent.
- Refactored the working-folder browser into a dedicated drawer, persisted its width, and improved remote workspace navigation behavior.
- Optimized right-panel, message-list, and layout performance to reduce unnecessary rendering and keep panel interactions smoother.
- Refined chat components and tool rendering for WebSocket streaming, including preserving visualize widget tool inputs for downstream rendering.
Fixed
- Fixed background and provider streaming retries to better handle upstream transport failures, early disconnects, and short-lived error bursts.
- Fixed session creation and persistence ordering so initial messages do not race ahead of newly created sessions.
- Fixed database migration behavior to skip remapping chat sessions into project-scoped records.

Added
- Added session isolation improvements to keep conversation-scoped tool sessions separated across windows.
- Added default project/workspace recovery behavior for unbound sessions when an existing project is available.
Changed
- Updated chat session and sidebar flows to better align with multi-window interactions.
- Refined websocket and streaming handling to improve turn consistency and context delivery.
- Continued improving context compression, message metadata, and related data-flow paths.
- Improved SSH remote search handling and error resilience.
- Kept analytics overview loading lazy for better settings performance.
- Updated README and documentation with clearer usage guidance.
Fixed
- Fixed stream delta flushing behavior to avoid redundant animation-frame scheduling.
- Fixed queued websocket runs to honor turn boundaries more reliably.
- Fixed remote search to skip ignored directories consistently.

Added
- Added websocket timeout handling improvements with better context-token-aware streaming behavior.
- Added plan execution gating and persisted context compaction metadata to prevent state drift in planned workflows.
Changed
- Improved sidecar runtime resiliency by adding circuit-break behavior for repeated streaming transport failures.
- Refactored context compression and added message metadata support, including database schema and data-access updates.
- Improved chat assistant message summarization to categorize file-change results more precisely.
- Refactored chat and sidebar interactions to isolate tool sessions by conversation scope and redesigned home input handling.
- Optimized SSH handler behavior with better search error handling.
- Updated README and project documentation to include special thanks and better usage guidance.
- Improved settings performance by lazily loading analytics overview data.
- Improved chat tool activity and diff presentation in review tooling.
Fixed
- Fixed SSH remote search behavior to skip ignored directories.
- Fixed queued websocket runs to correctly honor turn boundaries.
- Fixed tool behavior around skipped directories and search edge cases in remote SSH sessions.

Changed
- Replaced the direct electron-builder install-app-deps postinstall step with scripts/postinstall.mjs to rebuild native Electron dependencies with version detection and platform-aware module skipping.

Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.9.1...0.9.2

Added
- Added dedicated sub-agent limits to cap tool use, read scope, and runtime behavior for safer agent execution.
- Added change review sheet and file-change utility helpers to support richer file diff review flows in the chat UI.
Changed
- Refined sub-agent creation, resolution, runner, and default prompt flows to better enforce tool availability and execution constraints.
- Updated filesystem and search tool handling for sub-agents and teammate runners to align with the new execution limits.
- Improved skills and steps side panels, plus chat review card interactions, for clearer review workflows.
- Standardized English and Chinese chat locale copy for the updated sub-agent and review experience.
Fixed
- Fixed sub-agent and runtime protocol behavior in the .NET sidecar to keep agent execution consistent.
- Fixed streaming and review card state handling in the renderer when file changes transition across statuses.

Added
- Added WebSocket session transport support to improve stability and responsiveness for real-time streaming messages.
Changed
- Updated tool card and thinking block expand/collapse behavior during streaming sessions to keep UI state consistent.
- Improved WebSocket channel status handling with reconnect fallback guidance when connection failures occur.
Fixed
- Fixed tool call and file-change cards not properly resetting collapsed state when transitioning from streaming to completed status.
- Fixed message list auto-scroll behavior during long streaming output to reduce jitter and false scroll triggers.

Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.8.2...0.8.4

Added
- Added new VSCode tasks for building, publishing, and watching the .NET project.
- Added SSH window management and IPC handlers for improved user experience.
Changed
- Enhanced AgentLoop to summarize tool input for streaming.
- Improved message handling in OpenAiResponsesProvider.
- Updated TypeScript build info files for web and node configurations.
Fixed
- Removed unused PNG file to reduce bundle size.
Notes
- This release focuses on agent streaming improvements, SSH window management, and build configuration enhancements.

Added
- Added a dedicated in-app changelog dialog with optional AI translation support.
- Added direct PluginSendMessage and PluginReplyMessage tool support for CronAgent channel delivery flows.
- Added window-region capture plus clipboard image helpers to improve image export and copy flows.
- Added support for opening local file paths directly from markdown links and inline code references.
Changed
- Refined assistant workspace rendering with collapsible tool execution summaries and first-class widget output blocks.
- Improved file change cards to better distinguish created versus modified writes and to show richer inline previews.
- Optimized session lookup, background-session buffering, sidebar subscriptions, and session switch remount behavior for heavier multi-session workloads.
- Tightened cron agent delivery instructions and plugin tool registration so plugin-bound sessions keep the expected messaging tools available.
- Bumped the app version to v0.8.1.
Fixed
- Reduced widget iframe sizing and resize jitter issues in chat output.
- Prevented background-session updates from being dropped or applied stale when switching back to an active session.
- Improved copied chat image reliability by capturing the window region directly and normalizing clipboard image writes.
Notes
- This release focuses on chat workspace polish, plugin-channel delivery, session-state reliability, local file navigation, and export/copy stability.

Added
- Enhanced file system operations and tool integration across the agent runtime and UI surfaces.
- Improved Task tool and sub-agent execution flow.
Changed
- Reworked agent/runtime IPC and sidecar protocol plumbing to keep renderer and backend paths aligned.
- Refined file change and tool call card rendering for the new tooling flow.
- Updated the steps panel, chat store, and supporting agent/tool handlers to match the revised file-system workflow.
- Updated contributor guidance, build/development commands, README links, and repository housekeeping.
- Bumped the app version to v0.8.0.
Fixed
- Added reduced-memory recovery after OOM.
- Added GIF background color handling to improve export consistency.
- Stabilized file edit operations when source strings use different quote characters.
Notes
- This release focuses on file-system tooling, task/sub-agent workflow improvements, chat UI polish, runtime stability, and developer tooling/documentation updates.

What's Changed
- 在 README.zh.md 添加 Codex-Manager 友链与搭配使用说明 by @qxcnm in https://github.com/AIDotNet/OpenCowork/pull/71
New Contributors
- @qxcnm made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/71
Added
- Enhanced the Task tool and sub-agent execution flow.
Changed
- Reworked agent/runtime IPC and sidecar protocol plumbing to keep renderer and backend paths aligned.
- Refreshed orchestration and sub-agent card UI.
- Optimized transcript row metadata and lookup.
- Updated contributor guidance, build/development commands, README links, and repository housekeeping.
- Bumped the app version to v0.7.18.
Fixed
- Added reduced-memory recovery after OOM.
- Added GIF background color handling to improve export consistency.
Notes
- This release focuses on task/sub-agent workflow improvements, chat UI polish, runtime stability, and developer tooling/documentation updates.

Added
- Improved GIF slicing by trimming transparent margins first and falling back to balanced grid slicing when transparent padding is insufficient.
Changed
- Made file edit matching more robust when quote styles differ, with clearer handling for missing and non-unique patterns.
- Kept sidecar build output compatible with legacy and external paths by copying an additional unpacked binary bundle.
- Added a “Last 3 days” time filter and restored panel transition animation in the interface.
Fixed
- Reduced GIF cutline misalignment and white-edge artifacts.
- Stabilized file edit operations when source strings use different quote characters.
Notes
- This release focuses on image-processing reliability and editor/tooling robustness.

Added
- Added LongCat AI built-in provider preset with 6 models including vision capabilities (Flash Chat, Flash Thinking series, Flash Lite, Flash Omni with image support).
- Added Claude Code CLI source code analysis project for learning and research purposes with TypeScript implementation and MCP protocol support.
Changed
- Bumped the app version to v0.7.15.
- Updated the docs homepage release badge from v0.7.14 to v0.7.15.
Fixed
- No fixes in this release.
Notes
- This release enhances the platform's AI provider ecosystem and adds valuable learning resources for developers interested in high-quality CLI tool implementations.

Added
- Added a clearer comparison-style preview flow for AskUserQuestion, including safer single-select preview handling and improved option presentation across chat surfaces.
- Added stronger OpenAI chat provider payload typing and indexing support in the .NET sidecar to better preserve structured tool and streaming data.
Changed
- Bumped the app version to v0.7.14.
- Updated the docs homepage release badge from v0.7.13 to v0.7.14.
- Refined task, teammate, and follow-up interaction flow across renderer and main-process team runtime paths.
Fixed
- Improved ask-user interaction stability in chat, settings, and workspace flows to reduce edge-case state mismatches.
- Improved tool-call handling and replay consistency in the agent runtime and OpenAI provider integration.
Notes
- This release focuses on agent interaction flow polish, ask-user UX reliability, and provider/runtime consistency updates.

Added
- Added support for the latest application update to include additional runtime safety checks before release checks.
- Added improvements to updater behavior under mixed network conditions.
Changed
- Bumped the app version to v0.7.13.
- Updated the docs homepage release badge from v0.7.12 to v0.7.13.
Fixed
- Kept existing agent and message handling stable while refining update-related flows.
- Improved update package delivery fallback behavior.
Notes
- This release focuses on update pipeline stability and minor UX/behavior consistency updates around release checks.

Added
- Added persisted team runtime state so team manifests, messages, and member sync survive across the lead and worker IPC flows.
- Added isolated renderer-backed team workers so teammate runs can spawn outside the main agent loop when needed.
Changed
- Bumped the app version to v0.7.12.
- Updated the docs homepage release badge from v0.7.11 to v0.7.12.
Fixed
- Restored the Node agent-loop fallback when the sidecar cannot be used so runs continue instead of failing immediately.
- Kept assistant tool activity visible in chat and ensured the model switcher stays focused on the current active model.
Notes
- This release hardens the team runtime path by persisting runtime state in the main process and adding the isolated worker bridge for renderer-side teammate execution.

Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.7.10...0.7.11

What's Changed
- 修复删除会话后错误切换项目的问题 by @guoxdxdxd in https://github.com/AIDotNet/OpenCowork/pull/52
- feat(ssh): 支持单跳跳板机自动跳转与 OpenSSH 主机别名解析 by @guoxdxdxd in https://github.com/AIDotNet/OpenCowork/pull/51
- 添加内置 Git 面板并支持本地/SSH 仓库操作 by @guoxdxdxd in https://github.com/AIDotNet/OpenCowork/pull/53
New Contributors
- @guoxdxdxd made their first contribution in https://github.com/AIDotNet/OpenCowork/pull/52
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.7.9...0.7.10
Added
- Expanded the .NET sidecar role so more provider and message-processing work now runs outside the Electron layer.
- Added shared normalization and replay handling in the provider pipeline to support continued runs and backend-side message shaping.
- Added a reusable code diff viewer in the renderer so file change cards and tool call cards share the same diff presentation.
Changed
- Moved provider formatting, tool output shaping, and stream replay responsibilities into the .NET backend path.
- Consolidated renderer diff rendering around the new shared viewer to reduce duplication between chat cards.
- Bumped the app version to v0.7.10.
- Updated the docs homepage release badge from v0.7.9 to v0.7.10.
Fixed
- Improved OpenAI streaming replay so continued runs preserve tool calls, assistant messages, and formatting more reliably.
- Tightened provider message formatting for tool results and multimodal payloads to reduce replay mismatches and compatibility issues.
- Reduced divergence between the renderer and backend by normalizing more of the message shape in the .NET provider layer.
Notes
- This release continues the backend migration work: more logic has moved from the renderer-facing path into the .NET sidecar to make provider handling more consistent and easier to evolve.

What's Changed
- 修复了 github action 没有正确发布 macos 更新 yml 文件的问题 by @NOBB2333 in https://github.com/AIDotNet/OpenCowork/pull/47
Added
- Added an OpenCode migration workflow in Settings with preview, conflict handling, backups, and selective import for providers, model selections, commands, agents, MCP servers, and instructions.
Changed
- Improved chat and sub-agent runtime handling with better persisted execution history, richer detail rendering, clearer usage/debug traces, and more resilient provider request bookkeeping.
- Hardened main-process IPC delivery and background event forwarding across shell, SSH, cron, API proxy, and Weixin integrations to better tolerate disposed windows and oversized live output.
- Updated the docs homepage release badge from v0.7.8 to v0.7.9.
Fixed
- Reduced updater failures on unsupported ad-hoc signed macOS builds and ignored transient update errors more safely during install checks.

Added
- Added a long-running mode for autonomous sessions so the agent can persist this preference per session, auto-answer AskUserQuestion, and continue iterating until the work is verified complete.
Fixed
- Forwarded refs in the shared Select primitives to improve Radix integration reliability across renderer UI surfaces.
- Improved the app plugin settings panel so unconfigured plugins resolve state safely instead of failing on missing setup data.
- Preserved nested preview panel height constraints to avoid layout regressions in the renderer preview area.
- Sent Chatgpt-Account-Id for OpenAI Responses account-backed requests and included account_id in exported provider account JSON.
Changed
- Bumped the app version to v0.7.8.
- Updated the docs homepage release badge from v0.7.7 to v0.7.8.

Fixed
- Stopped excluding partial-json from the packaged app so the main process can resolve it at startup; fixes Cannot find module 'partial-json' after install on Windows and other platforms.
Changed
- Updated the docs homepage release badge from v0.7.4 to v0.7.5.
Full Changelog: https://github.com/AIDotNet/OpenCowork/compare/0.7.3...0.7.5

Changed
- Improved chat composer and settings layouts with better input height calculation, file editor spacing, provider organization, and clearer SubAgent error presentation.
- Synced the Bun lockfile so virtualization-related dependencies stay aligned with the declared package set.
- Updated the docs homepage release badge from v0.7.2 to v0.7.3.
Fixed
- Prevented IPC broadcasts from failing when renderer windows or frames are already disposed by centralizing safe window send helpers.
- Recovered Weixin message polling after remote session timeouts by resetting the polling cursor and retry cadence automatically.
- Kept embedded SubAgent execution details in a single-column layout for more stable rendering.

Added
- Added a reusable project working folder selector that supports both local desktop folders and SSH targets from chat home, project home, and the workspace sidebar.
- Added persisted SubAgent history snapshots so detail views can continue to show transcript and report context after execution completes.
Changed
- Improved the SubAgents experience with richer detail rendering, transcript-specific tool presentation, grouped history display, and clearer report status feedback.
- Updated the docs homepage release badge from v0.7.1 to v0.7.2.
Fixed
- Fixed usage analytics model and provider resolution by carrying request debug metadata through usage recording and falling back to session context when needed.
Refactored
- Removed the legacy renderer wiki navigation route and obsolete wiki-related UI state wiring.

Added
- Added a dedicated SubAgents detail panel with transcript rendering, execution progress, task input context, and report states for teammate runs.
- Added ACP-specific empty-state hints and homepage copy so users can understand empty sessions more quickly.
- Added a dedicated Routin AI（套餐） built-in provider preset to expose the https://cn.routin.ai/plan/v1 model lineup.
Changed
- Expanded the workspace experience with richer side-panel behavior and improved SubAgents panel layout, navigation, and localization copy.
- Updated Anthropic model capability metadata and reasoning effort labels, including support for the max effort level where applicable.
- Updated the docs homepage release badge from v0.6.6 to v0.7.1.
Fixed
- Improved ACP chat empty-state handling so guidance stays consistent across chat home and message list views.
- Improved wiki document access by returning tree metadata and preserving leaf-level source file references for tool and page consumers.
Refactored
- Removed the legacy OpenAI Responses websocket transport preference across providers, channels, and related settings.
- Refactored project wiki generation toward a tree-based document structure with leaf-node generation flow and sidebar browsing support.

Added
- Added SSH configuration import support for both OpenCoWork exports and OpenSSH config files, including conflict preview, selective import actions, and automatic connection list refresh.
- Added an application-level auto-update toggle in Settings so automatic update checks can be enabled or disabled persistently.
Changed
- Improved project memory resolution to prefer workspace-local .agents files with fallback to legacy root memory files, and applied project AGENTS.md content to prompt recommendations for local sessions.
- Simplified skills market access by allowing users to browse and test marketplace availability without requiring an API key upfront.
- Refined chat and workspace UI behavior across project home/archive pages, sidebar layout, selected-file handling, and session presentation.
- Updated the documentation homepage release badge from v0.6.5 to v0.6.6.
Fixed
- Kept the chat composer state in sync more reliably after document updates and initialization.
- Cleaned up pending session state more thoroughly when deleting sessions or projects.
- Closed the draw page when entering sessions, opened markdown preview links externally, and preserved terminal/tool state consistency after aborted runs.

Added
- Added project workspace navigation with dedicated project home/archive pages, a workspace sidebar, and project-bound channel settings.
- Added a built-in /plan command for entering Plan Mode directly from chat.
- Added personal Weixin media support for sending images/files and downloading inbound image messages for multimodal processing.
- Added an open source agent SDK survey covering Python, TypeScript, C#, and Java options.
Changed
- Improved the chat composer and message actions with file-aware drafting, queued draft editing, and attachment-aware layout behavior.
- Expanded the onboarding tour with dedicated Clarify, Cowork, and Code mode guidance plus updated English and Chinese copy.
- Updated the docs Docker build and CI workflow to use safer memory settings and Node.js 22.
Fixed
- Flushed completed tool events before ending aborted agent runs so terminal tool states stay consistent.
- Fixed the home composer sizing when attachments are present.
- Hardened the macOS unsigned build and release signing flow with stronger validation, ad-hoc signing support, and library validation adjustments.
Refactored
- Simplified the channel settings panel layout by removing redundant project-name prop threading.
- Removed redundant hover tooltip content from the right panel rail tabs.

Added
- New RightPanelRail component for enhanced layout sidebar navigation
Changed
- Improved chat UI components: FileChangeCard, InputArea, LazySyntaxHighlighter, MessageList, and ToolCallCard for better functionality and visual appearance
- Updated layout components: TitleBar and TopBar with refined UI interactions
- Refreshed UI dialog components: alert-dialog and dialog for consistent styling
- Added missing localization strings for layout-related UI in English (en/layout.json) and Chinese (zh/layout.json) locales

Added
- Gemini provider support for the vertex-ai model.
- Data URL to Blob conversion in ImagePreview for more reliable image downloads.
Changed
- Refactored Gemini provider API types and provider store to improve model management.
- Updated CommandPalette model presets to reflect the latest provider options.
- Refined ImagePreview download flow to reuse Blob conversion for data URL sources.

Changed
- Desktop input handlers now lazy-load @jitsi/robotjs and gracefully degrade when unavailable (e.g. unsupported platforms).
- OpenAI and Codex OAuth models now default to preferResponsesWebSocket for the Responses API.
Fixed
- Fixed desktop click/type/scroll failing at startup when robotjs cannot be loaded on the current platform.