Repository navigation
2.0.10
This release includes all merged changes since v2.0.7, covering PRs #41, #42, and #43 and the application changes in versions 2.0.8 through 2.0.10.
Audio transcription
- Telegram voice and audio messages become prompts in the current thread. Captions are preserved, including audio albums. Audio sent as a document remains an attachment for on-demand transcription.
- Add
transcribe_audiofor chat file IDs and absolute workspace paths, using OpenRouter withqwen/qwen3-asr-1.7bby default. Support optional language hints and WAV, MP3, FLAC, M4A, OGG/Opus, WebM, and AAC files up to 20 MB. - Validate audio byte signatures before uploading, including when a format override is supplied. Detect audio with missing metadata and enforce the actual downloaded size.
- Save long transcripts in the database and expose bounded previews and continuation pages without another transcription request. Continuations respect user, thread, message, and fork visibility.
- Skip downloading and transcribing redelivered Telegram updates that already have an accepted turn. Failed updates remain retryable, and new messages may reuse the same audio file.
- Retry HTTP 429, 502, 503, and 504 responses up to twice on the selected model, honoring
Retry-Afterwithin the shared timeout. Retries do not switch models./stopcancels downloads, transcription, and retry waits. - Report download failures, oversized files, rate limits, and empty speech accurately. Failed or cancelled transcription does not register new audio files or start an assistant turn, and existing shared files remain intact.
- Show a localized microphone transcription status without a filename. Voice/audio messages do not show a download status or an "Audio transcribed." success message. Ordinary file handling remains available.
Telegram activity and skill status
- Show an eyes reaction on accepted source messages until the response finishes, fails, or is cancelled. Prefix topic titles with an hourglass while work is pending.
- Run indicator updates in the background so slow Telegram API calls do not delay generation. Persist synchronization state and recover cleanup after failures and restarts, including overlapping bot processes and expired leases.
- Preserve manual topic renames, reconcile changes during in-flight requests, and ignore echoed activity-title events. Skip reactions that Telegram cannot apply to deleted messages or restricted chats.
- Display the loaded skill name in tool status and summaries, with escaped names and clearer ordinary file-read labels.
Image-generation intent
- Require clear user intent before synthesizing an image or making a generative edit. Factual-image requests and work with existing images use retrieval and ordinary editing tools.
- Update presentation guidance to use supplied or retrieved images and editable diagrams. A presentation topic or need for decoration alone does not authorize generated artwork.
- Add
live:pi-image-intent-checkto check the model's first action for retrieval and synthesis requests without executing the selected action.
Upgrade and verification
- Deploy application
v2.0.10with the published E2B imageai-tg-bot-tools:v2.0.10, using 2 vCPU and 2 GiB RAM. LeaveE2B_TEMPLATEunset for the version-derived default, or update an explicit override to this exact image. KeepE2B_DEPLOYMENT_IDunchanged. - Database initialization adds durable activity synchronization, topic-title state, and saved audio transcripts. No new required environment variables are introduced. Set
OPENROUTER_TRANSCRIPTION_MODELto override the Qwen default andTRANSCRIPTION_TIMEOUT_MSto override the 120-second timeout. - Typecheck, the full local suite of 597 passing tests with 2 skipped, and build passed on release commit
483e75a. GitHub CI also passed on that merged commit, including PostgreSQL integration coverage. A live Qwen request confirmed transcription through the JSON API. npm run e2b:releasevalidated the existing immutableai-tg-bot-tools:v2.0.10image and passed the live E2B smoke test before this GitHub release was published.