v0.18.0

Release v0.18.0

New Features

• WeChat agents now accept voice, video, and file messages in addition to images. Users can send an audio note, a short video, or a document and the agent will receive the content directly instead of a "type not supported" fallback.

Improvements

• Expanded the catalog of selectable LLM models — the latest DeepSeek, Claude Opus, Kimi, and MiMo releases are now available when configuring agents.
• When an agent's underlying model returns an empty response (for example, Gemini rejecting a malformed tool call), the conversation thread now recovers automatically instead of becoming stuck. The offending turn is cleaned up, and internal logs capture the surrounding tool-call history so engineers can diagnose the cause.

Other

• Refreshed internal dependencies across the Python backend and Go integrations.

Full Changelog: v0.17.60...v0.18.0

v0.17.60

Release v0.17.60

Bug Fixes

• Fixed a long-standing issue where images sent through WeChat came through as unreadable garbage and the AI refused to process them. Inbound images now decrypt correctly, so users can share photos with their agents and get a real response.

Full Changelog: v0.17.59...v0.17.60

v0.17.59

Release v0.17.59

Bug Fixes

• Fixed an issue where images sent through WeChat could be stored incorrectly and cause the AI to reject them, with the error then blocking any further messages in the same conversation. Inbound images are now validated before being forwarded to the AI, and unrecognized or unsupported formats are dropped cleanly instead of poisoning the conversation.

Full Changelog: v0.17.58...v0.17.59

v0.17.58

Release v0.17.58

Improvements

• Improved reliability for agents with large skill catalogs. Built-in capabilities — current time, long-term memory, posts, activities, and sub-agent calls — now remain available even when the automatic tool-selection layer narrows the active tool set. Previously these core tools could be filtered out, causing agents to repeatedly attempt the same failed call without recovering.
• Raised the threshold at which the automatic tool-selection layer activates and refined the counting logic so built-in provider tools no longer push borderline agents into the selection path prematurely.

Full Changelog: v0.17.57...v0.17.58

v0.17.57

Release v0.17.57

New Features

• Added time-limited share links for chats and posts. Team members can now generate a public URL that lets anyone — including recipients without an account — view a chat transcript or post for three days.
• When posts are delivered to WeChat or Telegram, the push message now contains a share link instead of a login-gated URL, so off-platform recipients can open the content directly.

Improvements

• Each public view increments a view counter on the share link, and creator metadata (user and team) is retained for future reporting.

Full Changelog: v0.17.56...v0.17.57

v0.17.56

Release v0.17.56

Improvements

• Tuned the threshold at which the automatic tool-selection layer activates, so agents with mid-sized skill sets now run without the extra selection step and only agents with larger skill catalogs trigger it.
• Routine dependency updates.

Full Changelog: v0.17.55...v0.17.56

v0.17.55

Release v0.17.55

Improvements

• Added diagnostic logging that records which image URLs are forwarded to the model, making it easier to investigate image-processing errors in production.

Full Changelog: v0.17.54...v0.17.55

v0.17.54

Release v0.17.54

Improvements

• Silenced noisy Gemini-related warning logs (schema compatibility and automatic function calling notices) that were flooding production logs without adding operational value.

Full Changelog: v0.17.53...v0.17.54

v0.17.53

Release v0.17.53

Bug Fixes

• Fixed several issues in the WeChat integration that prevented inbound images from reaching the LLM and caused spurious "push failed" alerts even when messages were successfully delivered.

Full Changelog: v0.17.52...v0.17.53

v0.17.52

Release v0.17.52

New Features

• Alert forwarding for Go integrations: The Telegram and WeChat integration services now automatically forward error-level events to the configured alert channel (Telegram or Slack), matching the behavior already in place on the Python side. Both stacks share the same per-minute alert budget so operators are not flooded.

Full Changelog: v0.17.51...v0.17.52