Kiki ทุกเวอร์ชันและเวอร์ชันเก่า

- Opening conversations. Recent main and subagent messages can appear while session details and older history are still loading. Active subagents stay marked as working during that recovery. Opening a subagent avoids repeated transcript reloads, and sidebar capability lists load when opened.
- Conversation history. Scroll back through large conversations without losing older text or your reading position. Older-history pages default to 40 turns with a 512 KiB budget; expanding partially loaded messages or tool results automatically continues reading. History tools can return recent turns before preparing the entire conversation. Editing and resending earlier turns no longer leaves duplicate messages, and streaming message batches keep each message's text separate. Session forks skip temporary journal locks, and existing Linux conversations with backslashes in their identifiers remain accessible.
- Saved-conversation context. Rebuild context in an inactive saved conversation without creating a new session. The same session ID, bindings, permissions, and history are retained, current instruction sources are reloaded, and rebuilding sends no model request.
- Desktop exit and reconnect. Tray Quit stops the desktop-managed local service and exits; closing the window leaves the service running. Reconnecting to the verified same-home service preserves your draft and reading position. Externally managed and remote services keep their existing ownership.
- Annotations and composing. Saving selection annotations does less repeated work in large conversations. Add multiple ordered selections with optional comments, including quote-only selections. Applying composer edits keeps the text and caret together instead of moving the cursor unexpectedly.
- Saved memory. Startup automatically imports and validates existing memory in SQLite, preserving original files and the memory on/off setting. Corrected or forgotten memory is removed from later model requests. Complete search results and explicit links in task notes can be reused without repeating the same entry body. Export an entry as Markdown and preview an edited file before importing it; ordinary edits remain undoable, while explicit Forget permanently clears historical payloads.
- Memory Advisor. An optional model can recommend relevant memory within the existing context budget. Choose its model and confirm the destination for each workspace in Settings → Memory before it sends memory out.
- Questions and task completion. Locally answered questions stay at their original position through live updates and conversation reopen. Background question history shows the current answer status. Task completion can wake the main agent after runtime goal pauses and turn exits.
- Storage recovery. After disk-full or temporary storage failures, agents keep retrying pending saves at low frequency. Once storage recovers, the original task and notification finish without repeating model or tool work. Partially written records are checked before only their missing bytes are appended; changed journal bytes remain blocked for inspection.
- Models and context. Main and subagent context overviews update when model limits change. Live history preserves execution settings attached to individual messages. Retained media stays attached to its original message after context rebuilding and compaction. Tool selection no longer reports tools that are already available as policy-denied.
- External engines and recovery. Native vendor tools use the external engine's own permission controls without a duplicate Kiki approval gate; explicit caller restrictions and authorization for Kiki tools remain in force. Send-now controls follow the engine's reported support for in-turn messages, and delivery receipts remain accurate when a turn finishes. ACP permission-mode confirmation, interrupted model streams, external executor startup recovery, and retries for empty compaction summaries are corrected.
- Resuming subagents. Agent