Changelog

Updated: 09/09/2026Word count: 0 wordsReading time: 1 min

For dev and detailed changelogs, see GitHub Releases.

2026-09-01

• [1.2.4] Plugin tools can request ending the Planner after execution; models can stop sending the temperature parameter; WebUI [1.7.3] merges MaiMai Observation into the chat workspace and streamlines memory management
• Maisaka: when a plugin tool returns stop_after_execution: true successfully, the Planner ends after the current tool batch finishes and waits for new messages
• Models: model advanced settings add a "Send temperature parameter" switch (send_temperature, on by default); when off, MaiBot no longer sends any MaiBot-managed temperature — model-level, task-level, or from extra parameters — for compatibility with models that reject this parameter
• WebUI [1.7.3]: the standalone "MaiMai Observation" page merges into the chat workspace; the sidebar now shows chattable sessions and read-only chat streams (marked with an "Observing" badge) in one place, and clicking a chat stream opens its reasoning timeline inline
• WebUI [1.7.3]: long-term memory run status moves into the "More actions" menu, simplifying the top of the memory management page; tuning parameters and import common/advanced parameters collapse into dialogs; the audit timeline layout is adjusted — the audit scope now spans the full row and change summaries merge into the event list
• WebUI [1.7.3]: emoji keyword-recognition prompts integrate into the "Prompt Management" page, with per-language editing, custom versions, and hot reload
• Long-term memory: multiple fixes for person profile refresh, authoritative fact and alias maintenance, and memory import/audit flows; no config changes

2026-08-23

• [1.2.3] Expression config compatibility fix and Maisaka orphan tool message fix
• Expressions: fixed startup failures when an old config still used the vector selection mode; after upgrading, the first launch auto-migrates it to vector_intent (Super Fine) and writes the config back
• Maisaka: fixed orphan tool messages

2026-08-23

• [1.2.2] Expression selection performance and accuracy improved; WebUI resource management and model config experience comprehensively upgraded
• Expressions: MMR diversity reranking rewritten with vectorized implementation, significantly faster on large candidate pools; adjusted vector similarity weight mix, removed lexical overlap scoring to avoid false interference on short texts and CJK content; removed vector selection mode, existing configs auto-upgrade to vector_intent
• WebUI [1.7.2]: model provider form adds custom request headers and a collapsible advanced config panel; task config adds hard timeouts for text and vision tasks, auto-falling back to the next model on timeout; adding a model now defaults to the currently filtered provider; model test image replaced with a standard PNG for broader vision model compatibility
• WebUI resource management: curated expression management refactored — "review/approve" unified as "curate", top tabs replaced by dropdown menu, detail and edit dialogs merged, pagination now supports per-page count selection; jargon list pagination and layout restructured, status filter changed from tabs to dropdown, added prompt template entry; emoji filter and sort unified into the filter card, status switch changed from tabs to dropdown, sort split into field selector + asc/desc toggle, default sorted by registration time (newest first)
• WebUI misc: login page now supports redirect query parameter — embedded pages return to their original location after login; prompt version management adds delete endpoint, auto-restoring the default prompt when the active version is deleted; expression and jargon lists use unified transparent background styling in the retro theme
• Models: temporarily compatible with V4V-class vision model image formats — animated GIF/WebP frames are converted to PNG (first frame only) before submission
• Plugins: "allow/deny" wording in plugin config unified as "read/do not read", more semantically accurate to the actual functionality
• Maisaka: adjusted person reference wording in planning prompts
• Engineering: PyPI index config now supports multi-source fallback (Tsinghua → Aliyun → official), improving install reliability in mainland China

2026-08-19

• [1.2.1] Fixed MCP long-running calls being wrongly timed out, Maisaka final-message compatibility and own-message identification; WebUI now applies saved model config immediately and displays offline adapters correctly
• MCP: fixed Streamable HTTP long-running tool calls wrongly using the HTTP request timeout; reading responses now follows the session read timeout
• Maisaka: fixed compatibility of the final assistant message, now ending with a user message; own messages are now always marked to reduce the model mistaking message sources (removed the self_message_special_mark config option)
• WebUI [1.7.1]: model config is now synced to the runtime immediately after saving, avoiding configs not applying in Docker; fixed chat page monitoring state (errors and thinking order); offline adapters are now displayed correctly
• Adapters: deleting a group chat now also cleans up the explicit allow/deny rules in the adapter policy
• Config: fixed WebUI saving configuration potentially producing invalid TOML

2026-08-18

• [1.2.0] Maisaka: Replyer now uses different reply modes per scenario for more diverse replies; improved Replyer organization
• Reply effect evaluation upgraded (currently v6): responsiveness no longer considers reply speed, removed the raw total score without clear semantics, records without related info are marked "completed / no info", records that didn't finish the observation window are marked "incomplete" and excluded from scoring; score distribution is now a per-sample scatter plot; supports deleting / clearing score records
• Models: official support for the Response endpoint; model context and output upgraded to a flat Item-first structure, keeping body, reasoning, function calls, tool results and provider-native activities separate
• Expression: fixed repeated abnormal restarts after expression vector index corruption; optimized online index maintenance (incremental assignment, k-means++, lock-free atomic writes)
• Adapters: bot platform accounts now persist the identity reported by the adapter, so multiple accounts on one instance are reliably recognized; adapters can auto-discover their ID; access policy adds independent group and private default actions (allow by default, switchable to deny)
• WebUI [1.7.0]: new dedicated adapter management and unified command management pages; improved model configuration layout; fixed local model testing bug; settings page shows discovered adapter accounts with online status and soft disable / restore; group frequency can be set per mode (with wildcard and default config); reasoning logs now show input, output and total tokens
• Messages: nested forwarded messages can now be viewed
• Plugin SDK: send.text, send.emoji, send.image, send.forward, send.hybrid, send.command and send.custom support return_details=True to get the platform-confirmed final message ID
• Improved startup onboarding

2026-08-04

• [1.1.4] Models: added support for the OpenAI Responses API (text, images, structured output, function tools, native tools, streaming events and usage stats); added native web search for DeepSeek v4 flash with related parameters
• Maisaka: Responses native web search summaries (query, action, status and source count for the round) shown in the monitor and regular logs
• WebUI [1.6.3]: new native detailed statistics page (keeps the old HTML report, interactive filtering by model/module/request type/chat flow, trends and performance metrics)
• Plugin management: shows conflicting directories of duplicate plugins with explicit load failure reasons; refreshes runtime state immediately after enabling a plugin; cleans up empty plugin root directories on startup
• Fixed being unable to add a new model provider when both model and provider lists are empty; improved plugin market card layout

2026-07-28

• [1.1.3] WebUI: optimized sidebar hover behavior, page colors and layout, new storage management page
• Maisaka: fixed Planner native reasoning incorrectly passed as body to Replyer; added typo correction message references
• [1.1.2] WebUI: optimized homepage cards
• [1.1.1] Main program: statistics charts split into customizable cards, fixed memory growth from full model call detail loading
• WebUI: LLM request error classification in reasoning view, global AI search upgraded to draggable multi-turn Agent overlay
• Chat: fixed session teardown on page switch, default nickname "Human", user avatar and emoji support
• Plugin list now layered by load status; homepage version and card layout streamlined
• Maisaka: added reply.before_post_process Hook for per-reply text post-processing control
• MCP: process-level shared server connections with hot reload, improved WebUI MCP configuration

2026-07-22

• [1.1.0] Main program: optional interactive terminal input with /clear, /pm, /offline, /online commands for chat and adapter management
• A_Memorix: long-term memory lifecycle (decay/freeze/restore/protect/recycle bin), improved retrieval quality and character profiles
• Legacy memory migration fixes: orphaned associations, timeline selection, entity renaming issues
• Maisaka: separated behavior style from persona, fixed cross-day time reminder interrupting tool chains
• WebUI: fixed frequency display precision, QQ number config, model rename, and homepage animation issues
• Plugins: automatic compatibility check after host update, tightened Host version range

2026-07-09

• [1.0.12] Improved Planner-to-Replyer information transfer and reduced duplicate replies
• WebUI: more reliable offline observation records, custom API model lists, multiple model configurations, data import/export, and upgrade announcements
• Initial setup now guides users to replace the temporary startup Token with a persistent Token
• Messaging: the host can control adapter admission; fixed handling of oversized emoji images

2026-06-12

• [1.0.0] Systematic upgrade! Maisaka inference engine refactored with Planner-Replyer deep integration
• Thinking effort mechanism: dynamically controls reply time and length
• A-Memorix Memory Engine v1.0: knowledge graphs, character profiles, chat summaries
• Feedback correction system: automatically corrects outdated memories
• MCP built-in plugin; global memory configuration added
• WebUI: Model preset marketplace, comprehensive security hardening, frontend auth refactoring
• For a more complete illustrated explanation, see the MaiBot 1.0.0 Update Feature

2026-01-11

• [0.12.2] Optimized private chat wait logic, force quote reply on timeout
• Fixed disconnection issues with some adapters, optimized memory retrieval logic

2025-12-31

• [0.12.1] Year-end summary feature (WebUI), optional LLM judgment for quote replies
• Expression optimization: automatic and manual evaluation support
• Reply and planning records viewable in WebUI
• Global memory blacklist: exclude specific group chats from global memory

2025-12-21

• [0.12.0] Thinking effort mechanism: dynamic reply time and length control
• Planner and Replyer integration, new private chat system
• MaiMai dreaming feature, MCP plugin as built-in
• Global memory configuration added

Earlier Versions

For changelog of earlier versions, see GitHub Releases.