- Fix multiline system prompts, memory notes, and saved sessions collapsing or overflowing when edited in phone settings.
- Move Memory Notes next to the Enable Memory setting.
- Update Memory and Choice guidance to save useful durable facts and avoid unnecessary choice prompts.
- Move new-session vibration feedback to when the confirmation opens.
- Replace prompt-generated JSON tool requests with OpenAI-compatible function tools, streamed tool calls, and matching tool-result messages. (should make models better at toolcalls)
- Return malformed tool arguments to the model as errors, report unsupported tool models clearly, and keep streamed prose instead of executing ambiguous mixed responses.
- Normalize unsupported punctuation and emoji into readable watch-safe text.
- Improve location results with fresher high-accuracy coordinates and street, postal code, and locality details.
- Add a model-aware provider picker to the phone settings.
- Add OpenRouter routing choices for automatic, cheapest, fastest generation, and fastest response-start provider selection.
- Fix current-location weather requests and back-to-back JSON tool calls.
- Add light Markdown rendering for assistant replies and saved conversations.
- Fix each tool appearing only at the end by showing it when execution starts and retaining it above the final answer.
- Fix rapid live updates outrunning the watch renderer and leaving blank space before the final response.
- Fix tool-call history using a different protocol from the JSON format requested from models.
- Fix streamed replies pulling the chat back to the bottom after the user scrolls up.
- Add a one-time first-run notice with a clear Early access heading and directions to support and control documentation. And add a phone-setting control that shows the first-run notice again on the next watch-app launch.
- Update tool activity live as each tool starts and show location, weather, search, scrape, and calculation details.
- Fix watch message pacing and send tool history separately from reply text to reduce dropped response starts.
- Fix split emoji in streamed replies, preserve Markdown dividers in saved sessions, and time out unanswered choice prompts.
- Count completed messages and successful searches, reduce credits requests, and improve free-model capability lookup.
- Replace partial streamed text with the error when a response fails.
- Preserve failed user messages and record why the assistant did not respond as context for the next request.
- Preserve emoji supported by Pebble Gothic fonts while replacing unsupported emoji.
I normally wait a little between releases, building up some actual changes, making it worth it, And I had 2 or 1 and then discovered some bugs (tool-call wise) and so you get a smaller 0.3.2 A little early.
- Require confirmation before clearing a conversation, with a safe return to the existing chat.
- Start dictation automatically after confirming a new session.
- Show each tool call on its own line in the watch conversation before the final response.
- Preserve tool calls and results as context for later conversation turns. (The big one, as it helps the agents a TON when asking follow up questions)
Finaly tested it on my old Pebble Time, and there was issues.
- Return tool results to the model with standard tool roles and matching call IDs instead of user messages. (Makes it easier when looking at OpenRouter logs)
- Remove the touchscreen hold gesture for clearing conversations to prevent accidental activation. (Was the only activation in my experience)
- Fix tool and choice navigation keeping selected rows or their context headers off-screen.
- Show submitted choice questions and answers in the watch conversation.
- Show remaining credits, messages, and searches on Pebble Time and Pebble 2 while keeping full Home Stats on Pebble Time 2. (Wow, I finally tested it on my old Pebble Time and found issues, who could have guessed)
- Clarify Home Stats platform visibility and automatic monthly resets in phone settings.
- Reduce the gap between the home title and statistics on all watch sizes. (Was annoying on the small watch size)
- Align in-chat status messages with the final assistant text position. (Had been bugging me for a while but was not a big problem)
- Add a choice tool with watch-based navigation, cancellation, and dictation for custom answers.
- Add an opt-in Firecrawl tool for scraping readable content from web pages.
- Support weather lookups using locations instead of only city names.
- Convert location lookup into an on-demand tool and return failures to the model without halting. (Lets you keep it on if you, like me, turn off location when you aren't using it)
- Include tool instructions and JSON fields in the system prompt only when each tool is enabled.
- Support repeated, ordered tool-call rounds with parallel execution and streaming fallback handling.
- Improve cancellation and message delivery with request IDs, bounded retries, preserved queues, and stale-choice protection.
- Add a Timeline toggle to phone and watch settings.
- Update one saved session record per conversation instead of creating one for every turn.
- Fetch current currency rates from Frankfurter and reject incompatible unit conversions.
- Protect privacy by removing conversation content, queries, and URLs from persistent logs and updating related documentation.
- Fix tool-setting navigation that could leave selected rows off-screen or behind the scroll indicator.
- Add an opt-in Health tool for activity, sleep, calorie, and heart-rate data over inclusive date ranges.
- Return final assistant answers as plain text while reserving JSON for tool requests.
- Add model-aware reasoning controls using capabilities reported by OpenRouter.
- Support replacing existing notes via structured memory tool
- Streaming responses + tool calls
- Weather integration
- Vibrate after first streamed token
- Search fix
- UI update
- Tool fixes
- Cleanup of the codebase
Quick little fix of a dum system prompt
Streaming fallback bug fix
Cancel queue clearing
AppMessage size/safety improvements
More robust JSON parsing
Local time in system prompt
30-second streaming watchdog
Reasoning trace awareness
Internal history → conversationHistory rename