v2026.8.21

Breaking Changes

Fixed

• Interactive submissions now render a local pending user echo immediately, reconcile it with the canonical message_start, and remove it for rejected or extension-handled input without writing render-only state to session history.
• The claude-sdk-oauth lane now surfaces Claude policy refusals (for example cybersecurity refusals) as an immediate, user-visible error naming the refusal category and explanation, instead of hanging until the ~90s stream watchdog timeout. Refusals are classified as non-retryable, so they no longer enter the timeout-retry ladder or account failover (#1052).
• Contended settings-lock retries now sleep via Atomics.wait instead of busy-waiting, retry-fallback canonicalization is memoized per error burst, and cursor-cli-oauth/claude-sdk-oauth settings loads are cached by mtime+size. Together these eliminate the settings-lock CPU-spin that froze the TUI at ~100% CPU under provider-error storms (#1056).

Added

Changed

• Settings reads no longer acquire the settings lock: writes publish atomically via a same-directory temp file plus rename, so read-only settings loads skip lock acquisition entirely and can never observe a torn write. Concurrent writers still serialize on the lock and re-merge against the winner's content.

• Refreshed dependency pins, including @anthropic-ai/claude-agent-sdk 0.3.238, jsdom 30, undici 8.10.0, marked 18.0.10, highlight.js 11.12.0, grok-mermaid 0.2.3, minimatch 10.2.6, ws 8.21.3, and typebox 1.3.16, and removed the unused @mistralai/mistralai and @types/ms entries.

• The monitor tool's description, schema text, prompt guidance, and terminal docs now state the verified contract (PTY output with stderr merged, event-only filtering, dedup and pause semantics) and include worked recipes plus an anti-pattern reference.

Fixed

Removed