v0.1.0-alpha.2

New features

• Interactive reply auto-continuation. In interactive mode, when the assistant ends a turn with a reply, Drea asks the model to classify the reply as either a status update or a request for user input. Status updates (CONTINUE_AUTONOMOUSLY) resume work automatically; questions, choices, blockers, or ambiguous replies (PAUSE_FOR_USER) wait for you as before. The classifier is implemented as an internal classify_reply tool so it works on endpoints that force JSON output. Bounded by --reply-chain (default 15).
• record_thoughts tool. The model can persist reasoning, plans, and intermediate conclusions into the conversation context. Because it is read-only, it never prompts for approval. If the model streams reasoning but does not emit record_thoughts, Drea follows up and asks it to do so.

Improvements

• JSON-mode reliability. To avoid the awkward pause where the user sees reasoning stream live but then the answer appears all at once, the client also streams two things live: plain prose that does not look like JSON, and the message field of JSON-mode reply pseudo-tools, decoded incrementally by a small replyExtractor. Real tool-call JSON is still withheld until the stream ends so partial JSON does not scroll by.
• Auto-approve split. The old --auto master switch is joined by per-category flags --auto-write, --auto-git, and --auto-run, so you can allow file edits without unattended shell commands. The /auto slash command supports toggling individual categories.
• Prompt caching. When a --prompt-cache-key is set, Drea forwards it to compatible gateways and reports cache hit/miss status. Context compaction was re-enabled to work correctly with cached prefixes.
• Live reasoning view. The model's chain-of-thought is rendered as a rolling live view so you can follow its thinking as it streams.
• Better transient-error messages. Rate-limit and retry error messages are phrased so the model understands they are API-side issues, not problems with the workspace or its own actions.

Documentation

• README and TECHNICAL are updated to cover prompt caching, the live reasoning view, reply continuation, record_thoughts, the auto-approve split, and JSON-mode live streaming.