BrowserOS Updates & Release Notes

• feat(claw): add privacy-safe Rust analytics (#1962) (40d3b67)
• fix(browserclaw): poll live sessions in background (#1965) (30d3d02)
• docs(mcp): clarify host metadata contracts (#1964) (20b3197)
• docs(core): document input coordinate basis (#1963) (caf48b0)
• feat(claw-server): add release-safe version contract (#1959) (269fdf1)
• docs(core): document screenshot capture contract (#1961) (a9e7249)
• docs(core): explain navigation readiness polling (#1960) (9a7b4da)
• docs(core): explain snapshot reference semantics (#1958) (bedab67)
• fix(claw-app): merge replay navigations by tab (#1957) (4872335)
• docs(core): document page identity (#1956) (150e84b)
• chore(claw): rename shared database file (#1955) (3b8d3f5)
• docs(cdp): clarify connection epoch (#1954) (a9ce228)
• fix(claw): tell agents to copy a foreign tab's URL into their own tab (#1953) (4f70bab)
• fix(claw): keep agent tabs open for inspection (#1952) (7fee5e3)
• fix: cap automatic action settling at 50ms (#1951) (9352d73)

Showing the latest 15 of 93 changes. View the full changelog.

fix(app): make custom model IDs discoverable in the provider picker (#…1917)

The Model combobox showed a models.dev snapshot as if it were the provider's real catalog. LM Studio ships three entries there, so users running their own loaded models concluded BrowserOS did not support them (bug report: "when you select LM Studio it doesnt show up with the model list just 3 pre defined ones"). Free-form entry already worked but was invisible: nothing in the trigger, the search placeholder, or the custom row said an unlisted ID could be typed.

• Trigger and search placeholders now say "paste a model ID"; the popover carries a persistent footer spelling out the Enter gesture.
• The free-form row reads Use "<id>" with a CUSTOM badge instead of a bare ID indistinguishable from a catalog entry.
• LM Studio / Ollama / OpenAI-compatible get an in-place hint under the field, since only those catalogs are knowingly incomplete. Providers with an authoritative catalog stay uncluttered.

Two bugs fell out of the rewrite:

• Enter committed the raw search text over whatever cmdk had highlighted, so arrowing to a listed model and pressing Enter saved the search string instead. Dropping the interception lets cmdk select normally; it re-selects the first item on every keystroke, so the free-form row still wins Enter while typing.
• Pasted IDs were stored untrimmed, and the provider then rejected the surrounding whitespace.

Verified against the running extension over CDP: free-form Enter (even when a fuzzy near-match is listed), arrow-then-Enter, mouse selection with context-window autofill, and whitespace trimming.

fix(app): run the model picker's fuzzy search on the trimmed ID

Trimming the pasted model ID fixed what got committed but left the Fuse query on the raw text, so the two halves of the picker disagreed about what the user had typed — and the free-form row that used to paper over the gap is now correctly suppressed for IDs the catalog already has.

Pasting gpt-5 scored the padded pattern better against gpt-5.5 than against the exact gpt-5, so Enter saved a different model and logged it as a deliberate pick. Pasting o3 matched nothing at all, which emptied a list whose trimmed form is a direct hit and left Enter with no row to commit and no way out but deleting the whitespace by hand.

Both derivations now run through getModelPickerRows on the trimmed ID, which also makes the picker's "always at least one row" invariant true — an exact catalog ID scores 0 and ranks itself first, and anything else is by definition custom — so dropping the empty state stays correct.

Regression tests drive the real Fuse config over the shipped catalogs. Verified live over CDP: gpt-5 commits gpt-5, o3 commits o3, and a padded unlisted ID still commits trimmed via the free-form row.

fix(app): reset the context window when a custom model is chosen

The autofill effect only wrote contextWindow when the catalog could size the model, so a custom ID silently inherited the previous model's window. Selecting gpt-5.5 and then pasting an 8k local model saved a 1M-token window, which overflows on the first long chat.

Custom models now fall back to the 128k default: still a guess, but a safe one, and the field stays editable. Catalog models are unaffected, and a zero-length catalog entry now takes the default too rather than saving a window of 0.

This predates the branch, but free-form entry was a hidden affordance before and is the promoted path now, so the bug is newly reachable by design rather than by accident.

The hint test named the provider itself instead of asking the template for it, so it asserted a sentence the dialog never renders; it now locks the real one.

What's new:

Auto-updates on Windows 🎉 — Long pending, finally here! BrowserOS now updates itself on Windows, just like on macOS. No more downloading and manually installing every release

Fixed the recurring Windows pop-up — Squashed the bug where a dialog kept popping up every ~15 minutes on Windows, and improved its reliability overall. Please let us know if you still see it!

Third-party agent integrations — Fixed a bunch of bugs and issues with agent integrations like Codex and Claude Code — connecting and staying connected is now much more reliable.

fix(agent-mcp-manager): isInstalled + link() gate use installCheckPaths, not systemPaths only (#1864) (e12a232)

feat(claw-server): self-heal MCP config paths at boot (#1863) (1981603)

fix(agent-mcp-manager): write Antigravity config to .gemini/config/, not .gemini/antigravity/ (#1862) (61ad3eb)

fix(tests): kill the Bun mock.module cross-file pollution class (Levels 1 + 3 of the reliability audit) (#1856) (84680b6)

chore(release): build v0.47.18 [skip ci] (214a1a5)

ci(tests): retry each suite once on failure (#1851) (8f610bf)

fix(claw-server): paginate listTasks by session via HAVING, not row WHERE (#1827) (4df440d)

fix(claw-server): default audit created_at in JS so it records on SQLite < 3.42 (#1846) (af6a34b)

fix(audit): group Agent filter by slug, not per-connection agentId (#1849) (e0b9fa9)

chore(release): build v0.47.17 [skip ci] (33abee4)

fix(bos-build): download all arch server bundles for universal builds (#1843) (7c44489)

docs: updated README (#1842) (868a33d)

chore(release): build v0.47.16 [skip ci] (fd2221a)

feat(claw-server-rust): claw stack parity + agent-mcp-manager crate (#1839) (ff8ec1e)

refactor(claw-server): fold session-naming into BASE_EFFECTS, nudge state on identity (#1838) (2a269f9)

Showing the latest 15 of 19 changes. View the full changelog.