1.19.0

Added

• New jina_read web tool: reads pages through the Jina Reader API with return_format (text/markdown/html/screenshot/pageshot), target_selector, link summaries and a configurable timeout. Works without an API key on the free 20 RPM tier; with JINA_API_KEY the limit rises to 500 RPM. Handles JS-heavy and anti-bot pages that plain extraction cannot.
• New jina_search web tool (key-gated): searches the web via the Jina Search API and returns top results with titles, URLs and content truncated to ~2000 characters per result (title and URL always preserved). Registered only when JINA_API_KEY is configured — the Search API hard-requires a key and bills ≥10k tokens per request.
• Hybrid read_web_page with quality gating and a circuit breaker: the tool is now a smart orchestrator. With JINA_API_KEY the Jina Reader path is tried first; without it, trafilatura goes first and the free r.jina.ai tier acts as the fallback. Every result passes a quality gate (minimum length, raw-HTML fallback detection, JavaScript-garbage heuristics); each path is guarded by an in-memory circuit breaker (3 consecutive failures → 5-minute cooldown → single half-open trial). When both paths fail the gate, the better raw result is returned with a warning instead of nothing. The return contract is unchanged.
• JINA_API_KEY setting documented in the generated configuration reference (Section 12, Search Capabilities).
• New read_telegram_file tool: reads the content of a text-based file (e.g. .md, .txt, .py, .json) previously sent as a document to the Telegram chat. The file_id comes from upload captions or the get_file_info tool; content is served by the new TelegramFileStorage.get_text backed by a shared text-file detector (MIME allowlist first, file-extension fallback — Telegram often reports application/octet-stream for plain-text files it cannot classify). Non-text files (images, binaries) and oversized files are rejected; very long files may be truncated with an explicit marker at the end of the content; tool-level errors (ValueError, FileNotFoundError, StorageError) propagate per the tool error conventions. Only works when file storage is set to Telegram; covered by 10 new tool tests.
• Document upload captions now include file readability hints: the caption metadata exposed to the model for every uploaded document gains file_name, mime_type and an is_text_file flag (computed by the shared detector), so the agent knows upfront whether a file can be read with read_telegram_file.
• Mandatory imperative context-size warning in the system prompt: the advisory warning was replaced with a mandatory instruction to summarize or drop stale tool results before answering.
• Volume-based tool-results warning: when background tool results exceed 25% of the context size, the system prompt lists the top-3 largest offenders (tool name + estimated tokens) and suggests clear_tool_call_history.

Fixed

• Legacy retention_cleanup-<timestamp> scheduler job duplicates are now removed automatically at startup: the old registration code created a uniquely-named retention-cleanup job on every restart (retention_cleanup-YYYYMMDDHHMMSS), leaving dozens of redundant copies in the persistent store that would each fire and run perform_retention_cleanup again. register_retention_cleanup_job now performs a one-shot O(jobs) self-heal pass that deletes any job id matching retention_cleanup-\d{14} from the persistent store (Redis or SQLite) before registering the fixed-id job — system:* and agent:* jobs are never touched — so affected deployments clean themselves up on the next restart without manual intervention.
• LLM thoughts in scheduler-triggered turns now render as native Telegram thinking drafts instead of plain "💡💭" text messages: SchedulerInterface no longer inherits the base implementation that spammed the chat with plain text; it mirrors TelegramInterface by sending an ephemeral <tg-thinking> sendRichMessageDraft through the shared scheduler Bot and clearing it before the final message. Draft failures are logged and swallowed — thoughts are deliberately never delivered as plain chat content.
• Scheduled/self-wake turns can now deliver media and chat actions to Telegram: SchedulerInterface.send_audio, send_video, send_images and send_document no longer raise NotImplementedError (which crashed the job into a generic "Scheduled job failed" note) — they mirror TelegramInterface delivery through the shared scheduler Bot with the payload-fixed chat_id/message_thread_id. Typing / uploading-photo / recording-voice chat actions are sent as best-effort cosmetic hints whose failures are logged and swallowed instead of failing the job; image URL lists fall back to a plain text message with the URLs when the media group upload fails.
• Web pages are no longer returned as mojibake when a server negotiates brotli compression: the shared HTTP helper manually advertised Accept-Encoding: gzip, deflate, br, but httpx has no brotli decoder installed, so servers honoring the header returned brotli-compressed bytes that were passed through undecoded. The header now advertises only the encodings httpx can actually decode (gzip, deflate).
• Selecting an image model no longer overwrites the thread's LLM model: tapping "← Back to providers" in the /image_models menu (shown when a provider has more than 12 image models) previously switched the inline-keyboard state to the LLM flow, so the next tap stored the image model as the thread's chat model until it was re-selected manually.
• Semantic history saving and search no longer fail with a ChromaDB embedding-function conflict error: the custom FastEmbed embedding function did not implement the embedding-function protocol expected by chromadb 1.5.9, so every write to semantic memory failed — background archival tasks died silently and the affected threads' conversation history was lost from semantic memory. A legacy silent fallback to ChromaDB's default (English-only ONNX) embedding model could additionally poison collections.
• The FastEmbed embedding function now fully implements the ChromaDB embedding-function protocol (name / get_config / build_from_config, registered in the Chroma embedding-function registry): collections now persist the embedding function as "known", and its configuration round-trips through save/load.
• Embedding-function-conflict ValueErrors are now wrapped and logged instead of silently killing background archival tasks.
• The fastembed mean-pooling deprecation notice no longer leaks into logs: the store is homogeneous — fastembed 0.8.0 has been pinned since the memory subsystem landed, so the notice is factually inapplicable and was suppressed narrowly (by message and category).
• Emergency summarization now uses the thread's active model instead of the provider default model: previously the request could hit a different, separately-billed model and fail with rate-limit errors that were swallowed by the reactive recovery path, leaving the conversation permanently over the context limit.
• Emergency summarization requests now fit the context window: the summarizer input is truncated under a token budget (background tool-response blobs dropped largest-first, then the oldest turns; the latest turn is never dropped, and in-loop tool-call pairs are excluded as before), and the completion cap is raised to 130% of the configured max_tokens (threaded as an explicit per-call parameter through all provider completion paths).
• The token usage cache is invalidated after a successful emergency summarization (previously stale usage data could re-trigger summarization pointlessly).

Changed

• The silent fallback to ChromaDB's default embedding model is removed: when fastembed is unavailable, semantic memory is now disabled loudly instead of silently switching to an incompatible model (a different vector space that would corrupt search results).
• The embedding model now loads lazily (once, on the first embed) instead of on every get_or_create_collection call (~0.3 s saved per archived message).
• The default local embedding model is unified as sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 — an intentional multilingual default; ChromaDB's built-in default model is English-only.