Skip to content

v3.5.0

Choose a tag to compare

@cursor cursor released this 11 Nov 09:12
· 8 commits to main since this release
b911465

Changed

  • Claude CI review is opt-in through the claude-review label and cancels stale runs. Ordinary validation does not call an LLM.
  • Added an explicit dashboard profile for local agent/tool activity monitoring; it is also included in full.
  • Twenty core skill entrypoints now use bounded workflows with available agents. Entrypoint-byte comparisons and their limits are recorded in the verification guide.
  • Default-profile documentation now distinguishes catalog contents, installed components and hook registrations. Removed unsupported token-saving multipliers, completion rates and automatic swarm guarantees from current product copy.
  • Refreshed localized introductions and replaced legacy workflow GIF embeds with a locally rendered, editable HyperFrames product GIF and source-controlled runtime diagrams. Historical media are explicitly labeled.
  • Synchronized package and plugin metadata at 3.5.0. Optional dependencies, data destinations and security boundaries are documented.

Fixed

  • Linux CI exposed a bundled transcript-parser CLI side effect in the PreCompact hook. Parsing is now a pure shared module with a separate CLI wrapper; canonical-path package tests reproduce the previous failure on macOS too. Continuity ledger/handoff paths reject symlink escapes and save messages reflect actual writes.

  • Independent follow-up review found two additional regressions before publication: executable skill permissions are preserved (including legacy repair and user-mode conflicts), and agent-memory paths reject symlink escapes before reads/writes. A bounded cooperative lock protects append/retention from lost parallel updates.

  • Remote installer update failures stop without removing the checkout. Normal Claude and Codex installs reconcile unchanged owned files, preserve modified/unmanaged files and back up replacements.

  • The installed CLI and npm vibeco alias are available through explicit launchers; profile changes use the recorded source repository.

  • Hook input readers accept canonical stdin events and tool_response. Agent memory honors user/project/local scopes, and recall requires relevant content.

  • Hook context accounting separates sessions and coordinates writes. Explicit deduplication avoids repeating unchanged optional reminders; truncation preserves valid text.

  • Optional native SQLite loading works from emitted ESM bundles and reports missing/incompatible dependencies. Braintrust extraction and GitHub-mutating hooks require explicit opt-in flags.

  • The optional dashboard binds to loopback and validates browser origins for HTTP and WebSocket requests. Credential-path matching covers additional supported shell forms while remaining a best-effort guard.

Verification

  • See the verification record for reproducible checks, measurements and their limitations. Historical changelog claims below describe their original releases and are not current benchmark evidence.