1.2.8
Prompt and context safety
- Hardened prompt context encoding with safe XML attributes, CDATA content, invalid-character normalization, and backward-compatible context parsing.
- Unified editor, browser, canvas, linked-note, chat-selection, and Codex prompt context formatting.
Models and thinking
- Added Pi
maxthinking effort support with model-aware fallback handling. - Added opt-in Codex
ultrareasoning effort support; existing defaults remain unchanged. - Added title-generation language selection, including interface-language follow mode, across supported Providers.
Tools, agents, and history
- Normalized task-management tools into stable Todo snapshots across live Claude streams and history replay.
- Improved ACP diff preservation, Pi web-tool rendering, Codex subagent lifecycle handling, Grok tool-result normalization, and provider-native diff replay.
- Restored Pi Skill command replay from persisted history.
Chat rendering
- Coordinated streaming Markdown renders with snapshot coalescing, visibility-aware throttling, final flushes, and improved thinking-block rendering performance.
Dependencies and security
- Updated Claude Agent SDK, MCP SDK, development tooling, and release-attestation dependencies, including security maintenance updates.
Verification
- Commit: bb49b5f
- Runtime evidence SHA-256: 9794a66df3204b7b886c6703a496a5d210aaa3219f8f78f7e588013ecddc3175
c1a1f785a372cf682796a8610ab7ce9206fef0bb2d3642123285c5ef47972f32 main.js
ce19867a890f8ee65e868db43dd9ce742c9da69bfd615000eacbf01a700053fe manifest.json
364d5511c52eaa24e05bf493bb99a29e83820c1fffcb493fbcc56a1346c2d91f styles.css