Changelog

Patch Changes

• 2e0bd8a: Fix extension update resilience (spaces and MCP refresh), tool state UI, and tool healing validation

What's Changed

• Fix lifecycle updates, tool healing, and MCP refresh by @aschung01 in #110
• chore(release): version packages by @github-actions[bot] in #111

Minor Changes

• ab6a7ec: Add a /compact slash command to the chat composer. Typing / now lists built-in commands (under a "Commands" group) alongside skills; selecting /compact manually compacts the conversation — summarizing the full history and sending only that summary to the model on the next turn, while the UI keeps showing every original message. Supports "compact-then-send" (/compact <text> compacts, then sends the remaining text) and surfaces a toast for every outcome (compacted, too short, or failure).

  Also fixes the compaction model resolution so it correctly handles the stored provider:model key (previously it failed to find a provider, which silently broke both manual and automatic compaction).

What's Changed

• chore: refresh models.dev snapshot by @github-actions[bot] in #106
• feat(extension): add /compact slash command by @aschung01 in #107
• chore(release): version packages by @github-actions[bot] in #108

Minor Changes

• 4630aed: Add a context-usage indicator to the chat header: a circular progress ring that shows tokens, usage %, and cost on hover, and a detailed breakdown (provider/model, context limit, token split, total cost, timestamps) on click.
• 67c8be2: Add scheduled tasks: recurring, cron-like agent runs that execute as full background agent sessions (DOM + chrome.debugger, system prompt, compaction). Driven by a bundled /schedule skill and create/list/update_scheduled_task tools, with a Scheduled dashboard, create/edit dialog, status badges, and a per-task auto-approve toggle.

Patch Changes

• 2965aed: Fix chat input lag in long conversations. The message list is now memoized and extracted from the input's render path, so typing no longer re-renders every message (markdown + syntax highlighting) on each keystroke.

What's Changed

• perf(extension): memoize chat message list to fix input lag by @aschung01 in #101
• chore: refresh models.dev snapshot by @github-actions[bot] in #103
• feat(extension): add context usage indicator to chat header by @aschung01 in #104
• feat(scheduled-tasks): recurring background agent runs by @aschung01 in #105
• chore(release): version packages by @github-actions[bot] in #102

Patch Changes

• 6f47a91: Show subagent-opened tabs in the Context card with a "Subagent" badge and grouped cleanup, sort sidebar chats by creation time, and open new agent tabs in the conversation's own window instead of whatever window Chrome has focused. Also fixes the Stripe MCP connector endpoint (404).

What's Changed

• fix(extension): hide subagent chats from chat lists by @aschung01 in #96
• fix: subagent tabs in context card, Stripe MCP URL, sidebar sort, new-tab window by @aschung01 in #98
• chore(release): version packages by @github-actions[bot] in #99

Patch Changes

• 278b9e0: Tune the completion check to reduce false rejections and latency: the evaluator now runs as a single fast pass (no tool calls), accepts a reasonable interpretation of ambiguous requests instead of looping the agent, and no longer rejects page-grounded facts based on its own stale knowledge (e.g. "that batch doesn't exist yet").

What's Changed

• fix(extension): reduce completion-check false rejections and latency by @aschung01 in #94
• chore(release): version packages by @github-actions[bot] in #95

Patch Changes

• cf239e1: Fix agent completion/approval notification clicks to open the conversation in the tab and window where the agent actually ran, instead of the last-active window (which sometimes spawned a new window).

What's Changed

• fix: route agent notification clicks to the originating tab/window by @aschung01 in #86
• chore: refresh models.dev snapshot by @github-actions[bot] in #91
• chore(actions): bump softprops/action-gh-release from 2.6.2 to 3.0.0 by @dependabot[bot] in #90
• chore(actions): bump pnpm/action-setup from 4.1.0 to 6.0.8 by @dependabot[bot] in #89
• chore(actions): bump actions/labeler from 5.0.0 to 6.1.0 by @dependabot[bot] in #88
• chore(actions): bump the actions-minor-and-patch group with 2 updates by @dependabot[bot] in #87
• chore(actions): bump the actions-minor-and-patch group across 1 directory with 3 updates by @dependabot[bot] in #93
• chore(release): version packages by @github-actions[bot] in #92

Minor Changes

• 832e867: Agent can now clean up the tabs it opened: a closeTabs tool closes the conversation's tab group (or specific tabs you opened) with a reversible Undo toast, plus manual control over the workspace Context card. Adds a bundled writing-skills skill that guides the agent through authoring a new skill and installing it via create_skill. Also fixes closeTabs being rejected by the Anthropic API (its tool input schema now serializes to a top-level object), which previously broke agent turns on Anthropic models.

What's Changed

• fix(landing): resolve mobile layout overflow and animate FAQs by @aschung01 in #81
• feat(agent): agent tab cleanup with reversible undo and manual Context card control by @aschung01 in #80
• feat(skills): bundled writing-skills skill + fix closeTabs Anthropic schema by @aschung01 in #82
• chore: refresh models.dev snapshot by @github-actions[bot] in #83
• chore(release): add changeset for #80 + #82 by @aschung01 in #84
• chore(release): version packages by @github-actions[bot] in #85

Patch Changes

• 7c36ce6: Favorite tabs behave Arc-style: a favorite is recognized by hostname (adopts the first matching open tab, stays recognized while you navigate within the same site, and survives service-worker restarts); reordering favorites in the overlay now moves the real Chrome tabs and keeps them ordered between pinned and regular tabs (with bounce-back on manual reorder, returning a favorite to its own slot). Spaces now reliably reattach to their windows across browser restarts (via a durable home-tab anchor with pinned-tab fallback) instead of losing favorites and spawning new empty spaces; pinned tabs are persisted per space and reopened when a space's window is recreated (favorites are no longer auto-opened). Also: reuse an existing Settings tab instead of opening duplicates; Settings logo no longer navigates; remove Esc-closes-Settings; Models search supports "/" focus and Esc-to-clear with hints; chat-delete dialog keycap styling + optimistic removal; "Send now" on the next queued message; "Continue" action on the chat error banner; overlay footer logo respects dark mode; fix overlay logo/Actions menu close-then-reopen flash; scope the "working on this tab" blocker to the agent's actual tab.

What's Changed

• feat(docs): add changelog page by @aschung01 in #72
• feat(bench): declarative harness configs, headless subagents, and Kernel browser pools by @aschung01 in #73
• feat(extension): add Context card to workspace side panel by @aschung01 in #76
• chore: refresh models.dev snapshot by @github-actions[bot] in #75
• fix(extension): tab ordering, settings tab reuse, overlay menu flash, and assorted UX fixes by @aschung01 in #74
• chore: refresh models.dev snapshot by @github-actions[bot] in #77
• chore(release): version packages by @github-actions[bot] in #78

Patch Changes

• 610c535: Aux model picker shows all configured models; token-AND model search; fix approval-interrupt bug; expose executePython; improved MCP approval + skill-install UX.

What's Changed

• feat(extension): configurable aux model picker, token-AND search, and approval fixes by @aschung01 in #70
• chore(release): version packages by @github-actions[bot] in #71

Patch Changes

• 6010425: Fix tool-approval "Always allow", add other-tab awareness, and assorted abort/UX fixes:

  • Respect "Always allow" by calling needsApproval with the AI SDK's positional input argument (previously input was always undefined, so the same-origin allowlist was never consulted and approval was always required).
  • Surface the user's other open tabs to the agent as an awareness-only ## Other open tabs block. Titles/URLs are sanitized (newlines/control chars collapsed, length-capped) before reaching the system prompt to prevent prompt injection, and only http(s) URLs are exposed.
  • Bind the shared active tab on the first message of a new conversation so the legend marks it [active] immediately; the bind now honors a { ok: false } background response instead of pinning an unbound tab.
  • Skip the completion check when the user aborts generation (the SDK emits an abort chunk and closes cleanly, which previously caused the check to grade an abandoned draft).
  • Stop TipTap from nesting pasted markdown links on each copy/paste/resend cycle.
  • Disable per-word stagger in the markdown renderer to fix concurrent/parallel reveal of streamed sections.

What's Changed

• fix(ci): resolve release tag from input, not event_name by @aschung01 in #65
• fix: always-allow approval, tab awareness, and abort/UX fixes by @aschung01 in #66
• chore: refresh models.dev snapshot by @github-actions[bot] in #69
• chore(release): version packages by @github-actions[bot] in #67

Minor Changes

• dbbda3c: Subagents: the chat agent can now delegate focused tasks to specialized subagents that run with fresh context, their own tool allowlist, and isolated tab/window state.

  • explore — read-only research subagent. Use for background investigation that should not mutate state.
  • general — read/write subagent for general-purpose delegation when the parent's tools fit but the work would bloat the parent's context with verbose output.

  Each delegation appears inline as a collapsible trace block with a live transcript, a step counter, and a phase title the subagent updates as it works. Subagents run with isolation: peer puts the child in its own tab group within the same window (default), and incognito opens a fresh incognito window with no shared cookies/auth/storage that auto-closes when done. Stop now correctly cancels in-flight subagents along with the parent.

What's Changed

• feat(extension): subagents - explore and general agents by @aschung01 in #62
• chore(release): version packages by @github-actions[bot] in #63

Patch Changes

• 402b8d0: Add Attio as a built-in MCP connector (new "CRM" category) with OAuth, all 37 tools, and light/dark icons. Also fixes MCP reliability:

  • OAuth flow now sends a state parameter, required by strict providers like Attio (previously failed with "Authorization page could not be loaded").
  • Connected MCP servers are no longer torn down and reconnected on every home.html load — this eliminated a multi-second window where the agent saw zero MCP tools.
  • An active chat conversation now rebinds to the latest transport when MCP tools finish loading, so newly connected tools become available without a refresh.
  • Settings → Connectors left panel border now spans the full height.

Patch Changes

• b5717e9: Fix three small chat issues:

  • Completion check no longer runs while a tool call is paused on human approval. The drafted text at that point is mid-narration ("I'll now run X to do Y") and isn't a final response — the gate now waits for the next iteration after approval, when the tool actually has output.
  • Long tool error logs collapse to ~10 visual lines with an inline expand toggle. Single-line errors that wrapped to dozens of visual lines now collapse correctly (the previous clamp only counted \n-delimited lines). Applied to executeCode/executeOnPage/executePython error output, the skill tool's error block, and the top-level chat error banner.
  • "Always allow on " now reliably persists before the agent resumes. The previous implementation fired the storage write and the approval synchronously, so the next tool call's needsApproval check could race the write and re-prompt — most reproducible on home.html with back-to-back executeOnPage calls.

What's Changed

• fix: completion-check pending tools, error log clamp, always-allow race by @aschung01 in #55
• ci(release): build extension zip from changesets workflow by @aschung01 in #53
• chore(release): version packages by @github-actions[bot] in #56

Minor Changes

• 2604829: Agent: a separate skeptical evaluator now reviews each drafted final response before it reaches you, and asks the agent to revise if the response is incomplete or unsupported by what was actually observed. Mid-loop refinements show as a compact "Refining answer" pill; unresolved concerns surface as a soft warning. Choose a cheaper or faster evaluator model in Settings → General, or leave it on the default. Chat exports and the per-message Copy button now include the rejection block as part of the audit trail.

Patch Changes

• 2604829: Agent reliability fixes:

  • Tab handles now persist across mid-stream conversation switches, so in-flight tool calls in the previous chat don't lose the tab they were targeting.
  • Stranded tool calls left over from interrupted streams heal cleanly on edit/retry/regenerate instead of breaking the conversation on resume.
  • Approving a tool call while other tools are still running in the same step no longer drops the auto-resume — the agent now picks up the approved call once the rest of the step completes.
  • Editing a user message in chat-db now logs a warning when the target id can't be found, surfacing the historical "stale tail after edit" failure mode at first repro instead of silently months later.
  • The completion-check evaluator now pins to its transport instance, so multiple concurrent agent windows can't drift across each other's models mid-stream.
  • Evaluator-error fallback messages no longer expose internal error details to the chat or markdown export; detailed errors stay in the developer console.

What's Changed

• chore: refresh models.dev snapshot by @github-actions[bot] in #48
• feat(bench): optional Cloudflare R2 upload for run artifacts by @aschung01 in #49
• feat: completion check — review and refine drafted responses before they ship by @aschung01 in #50
• chore(release): version packages by @github-actions[bot] in #52

Patch Changes

• 98f6004: Agent: tab arguments on browser tools are now explicit (tab: "t1") instead of relying on an implicit "active tab". Fixes the "Always allow on this domain" approval button not appearing when the agent acted from the home tab. Conversation tab handles persist across service worker restarts. (#39)

What's Changed

• chore: OSS foundation by @aschung01 in #26
• chore(actions): bump marocchino/sticky-pull-request-comment from 2.9.4 to 3.0.4 by @dependabot[bot] in #27
• chore(actions): bump actions/checkout from 4.2.2 to 6.0.2 by @dependabot[bot] in #29
• chore(actions): bump actions/setup-node from 4.4.0 to 6.4.0 by @dependabot[bot] in #30
• chore(actions): bump actions/upload-artifact from 4.6.2 to 7.0.1 by @dependabot[bot] in #31
• chore(actions): bump peter-evans/create-pull-request from 6.1.0 to 8.1.1 by @dependabot[bot] in #28
• chore(deps)(deps): bump dotenv from 16.6.1 to 17.4.2 by @dependabot[bot] in #34
• chore(deps)(deps-dev): bump @types/node from 22.19.19 to 25.9.1 by @dependabot[bot] in #37
• chore(deps)(deps-dev): bump vitest from 2.1.9 to 4.1.7 by @dependabot[bot] in #35
• chore(deps)(deps): bump @openrouter/ai-sdk-provider from 1.5.4 to 2.9.0 by @dependabot[bot] in #33
• fix(bench): drop deprecated baseUrl from tsconfig by @aschung01 in #38
• chore(deps)(deps-dev): bump typescript from 5.9.3 to 6.0.3 by @dependabot[bot] in #36
• feat(agent): explicit tab args + fix Always-allow-on-domain approval button by @aschung01 in #39
• chore(release): make package.json the version source of truth by @aschung01 in #40
• chore: add changeset for v0.3.0 by @aschung01 in #41
• Revert "chore(release): make package.json the version source of truth (#40)" by @aschung01 in #42
• chore(release): make package.json the version source of truth by @aschung01 in #43
• fix(ci): use pnpm run version in Changesets workflow by @aschung01 in #44
• chore(release): version packages by @github-actions[bot] in #45
• chore(release): adopt Changesets workspace tag format by @aschung01 in #47

New Contributors

• @dependabot[bot] made their first contribution in #27
• @github-actions[bot] made their first contribution in #45

What's Changed

• fix(extension): rail layout mount and agent-transport compound keys by @aschung01 in #24
• chore(release): bump extension to v0.2.1 by @aschung01 in #25

What's Changed

• feat(extension): add webFetch tool + provider/model UX fixes by @aschung01 in #17
• feat: per-conversation file uploads to OPFS workspace by @aschung01 in #18
• feat(extension): Working folder viewers, resizable side-panel, and OPFS uploads by @aschung01 in #21
• feat: OpenBrowse offline evaluation harness & runner hardening by @aschung01 in #20
• feat(extension): message queueing while agent is streaming by @aschung01 in #22
• chore(release): bump extension to v0.2.0 by @aschung01 in #23

What's Changed

• feat(landing): mono headline, refined widths, manual-install CTA by @aschung01 in #15
• fix(extension): heal broken chat sessions and stranded tool calls by @aschung01 in #16

What's Changed

• feat: Add agent task planning UI and todowrite tool by @aschung01 in #1
• feat(overlay): unified ranked search with personalized shortcuts by @aschung01 in #2
• feat: detachable side panel popover with per-tab scoping by @aschung01 in #3
• docs: restructure site (sidebar, sections, new pages) by @aschung01 in #4
• feat: landing page redesign and self-hosted typography by @aschung01 in #5
• feat: add Option+Space global chat popup with conversation/draft persistence by @aschung01 in #6
• fix(agent): message-based compaction + reduce mid-task reluctance by @aschung01 in #7
• fix(overlay/search): rank open tabs above frecent history by @aschung01 in #8
• feat: agent skill system, virtual workspace, and OPFS filesystem tools by @aschung01 in #9
• refactor: convert to pnpm monorepo (apps/* + packages/*) by @aschung01 in #10
• feat: add Python execution sandbox via Pyodide and UI improvements by @aschung01 in #11
• feat(registry): dynamic provider registry powered by models.dev by @aschung01 in #12
• fix: auto-update bundled skills when extension is upgraded by @aschung01 in #13
• Docs update + manual install + release & CI workflows by @aschung01 in #14

New Contributors

• @aschung01 made their first contribution in #1

Full Changelog: https://github.com/openbrowse-ai/openbrowse/commits/v0.1.0