Repository navigation
v2.0.2
Released: 2026-11-05
This release ships a webhook-driven automation system, LLM provider failover registration, a unified logging pipeline that captures every platform SDK, and a full correctness pass over the robot/ core and platform adapters — the result of a complete audit of the message-handling layer. It includes 8 community PRs and more than 20 bug fixes.
Highlights
- Webhook triggers — any external system (CI, monitoring, scripts) can now drive MuseBot with a single HTTP POST; the rendered prompt runs through the same command pipeline as cron jobs and delivers the answer to any configured platform.
- LLM failover — register fallback providers so a provider outage degrades gracefully instead of failing the request; fallbacks can be pinned per
provider:model. - Every SDK now logs through MuseBot's logger — Lark, Slack socket mode, gRPC, DingTalk, Discord, WeChat, Gemini, Milvus, DashScope and LangChain output no longer bypasses the unified log file.
robot/stability audit — 20+ fixes across Telegram, WeChat, DingTalk, Discord, QQ, Slack, cron and the core robot, including several crashes and one bug that broke WeChat passive replies entirely.
New Features
Webhook event triggers (@yincongcyincong)
POST /webhook/{token} accepts an external event, renders its payload into a prompt template, and asynchronously executes it against the webhook's configured target (any supported platform, group or user). Requests are acknowledged with 202 immediately; execution reuses the same synthetic-robot path as scheduled cron jobs, so all commands, LLM routing and reply delivery behave identically.
- Token-based URLs — no shared secret in the query string
- Full CRUD via the admin API
- Bilingual docs:
static/doc/webhook.md/webhook_ZH.md
LLM failover registration (@yincongcyincong)
Requests can now declare a failover chain of LLM providers. When the primary provider errors, the request transparently retries down the chain. Fallback entries support provider:model pinning (e.g. openai:glm-5.3-flash) or per-provider default models.
Outcrawl as a task web search provider (#75, @heydryft)
Tasks can now use Outcrawl as their web search engine alongside the existing providers.
Static stickers read as images (#67, @MaizioO123)
Telegram static (non-animated) stickers are now sent to the LLM as images instead of being skipped, so vision models can see them.
Gemini video model refresh (@yincongcyincong)
Updated Gemini video generation support to the latest models.
Bug Fixes
Telegram
- Update polling client had no timeout (#68, @MaizioO123) — a stalled connection could block update processing indefinitely; the polling client now has a proper timeout.
- Replies to ordinary bot messages were mishandled (#69, @MaizioO123).
- Race on the global bot client (@yincongcyincong) — the shared
TelegramBotinstance is now guarded by a read/write mutex; concurrent sends from cron, webhooks and the update loop no longer race. - Crash when the bot client is not ready (@yincongcyincong) — scheduled and webhook-triggered sends now check for a nil bot client instead of panicking.
- "Thinking…" placeholder stuck forever after an image failure (@yincongcyincong) — failed image generations now edit the placeholder with the error (and fall back to a plain message if editing fails).
- Wrong message for empty MCP output (@yincongcyincong) — the MCP empty-content branch sent the task message instead of the MCP one.
- Passive-mode replies were never delivered (@yincongcyincong) — the finished answer is now always stored for the callback to pick up; previously every passive answer returned HTTP 500 to WeChat.
- Duplicate-callback race (@yincongcyincong) — when a duplicate message arrived while the first was still processing, the handler logged with a
Robotthat had not been assigned yet.
DingTalk
- Inverted expiry check in the access-token cache (@yincongcyincong) — valid cached tokens were discarded; the cache now actually hits.
Discord
- Panic on interactions without a direct user object (@yincongcyincong) — the username lookup now falls back from
interaction.Usertointeraction.Member.User, fixing a nil dereference on member-scoped interactions.
QQ (OneBot)
- Nil robot from an unparseable payload took down the handler (@yincongcyincong) — the HTTP handler now checks the parsed robot before dispatching instead of letting
recover()swallow it into a generic panic log.
Slack
- Mention check in public channels (@yincongcyincong) — the bot now requires an explicit
@botmention in public/group channels (C*/G*) before responding.
Cron & scheduling
- Return an error instead of panicking when the scheduler is not up yet (#76, @djs-91).
- Nil-scheduler startup window (@yincongcyincong) — the scheduler is now created and started before the DB load (with retries), so chat-side cron commands never see a nil
Cron; a failed load leaves an empty-but-working scheduler.
Core robot
- URLs mangled by prefix stripping (@yincongcyincong) —
https://was stripped afterhttp://, turninghttps://…intos://…; both schemes are now removed in the correct order. CreatePhotoerrors reported the wrong media type (@yincongcyincong) — the configured type was shown instead of the unsupported one actually received.- Voice transcripts: DB errors leaked to users (@yincongcyincong) — transcription succeeds, persistence is now best-effort; users get the transcript instead of a database error message.
- Default command handlers were silently dropped (@yincongcyincong) —
ExecCmdnow invokes the registered default handler.
Video generation
- Handle Gemini video operation results safely (#72, @Danche23).
- Close Aliyun video responses between polls (#74, @Danche23) — fixes connection/file-descriptor exhaustion during long video generations.
Logging
A full audit of SDK logging, landed by @yincongcyincong:
- Lark and Slack socket mode clients are now wired into the logger (previously the Lark client wrote to its own private stdout logger, and the Slack option was a no-op).
- gRPC output is captured through a
LoggerV2adapter (info demoted to debug, fatal exits after logging);grpcmoved to a direct dependency. - The context-aware logger methods were rebuilt on the real logger instance — they previously routed to
zerolog.Ctx(ctx), which silently dropped events. Every log line now carriesbot_nameandlog_id, and one log id covers a whole cron fire / webhook run end to end.
Documentation & Build
- Keyless remote MCP example with deed.page (#66, @f-o-x11).
- Docker image builds resilient to restricted networks (@yincongcyincong).
Contributors
Many thanks to everyone who contributed to this release:
- @yincongcyincong
- @djs-91 — #76
- @heydryft — #75
- @Danche23 — #72, #74
- @MaizioO123 — #67, #68, #69
- @f-o-x11 — #66
Upgrade Notes
- Restart required — all fixes take effect on restart.
- WeChat passive mode (
wechat_active=false): replies now actually reach users; no config change needed. - Slack public channels: the bot now only responds when explicitly @mentioned — verify your workflows if they relied on the old behavior.
- Webhooks: create your first webhook via the admin API; see
static/doc/webhook.md(EN) /webhook_ZH.md(中文). - Failover: optionally set
-failover_providers(e.g.openai:glm-5.3-flash) to keep chats working through provider outages.