Chimera v0.11.0 — CSS extraction, resumable crawl, speech-to-text

Scraping polish + the honest audio integration.

Added

• Deterministic CSS-selector extraction before the LLM. extract now takes selectors (field → CSS, e.g. {"price": ".price", "link": "a.more::attr(href)"}): for a known page template those fields are pulled with BeautifulSoup — free, exact, no LLM — and the safe quarantined LLM fills only what a selector missed (crawl4ai's cheapest-tool-first idea).
• Resumable crawl. crawl checkpoints its frontier + visited set to disk after every page (atomic) and appends pages to a .jsonl, so a crawl interrupted at page N resumes from N+1. limit is the total target across resumes; state auto-clears on completion.
• transcribe_audio — speech-to-text. The symmetric partner to the existing image-generation and text-to-speech tools. Local faster-whisper via the new stt extra (offline/private), else the hosted OpenAI Whisper API.

An honest note on scope

This release also studied five ML frameworks (Keras, PyTorch, TensorFlow, OpenCV, scikit-learn) and five audio/vision/generative projects (faster-whisper, pipecat, wav2letter, FLUX, Stable Diffusion). The honest conclusion: Chimera is an agent, not a framework or a model. It orchestrates these — calls them via API, or runs them in its code sandbox — it does not, and cannot sensibly, reimplement PyTorch or "beat" Stable Diffusion (that would require training billion-parameter models). Speech-to-text was the one clear missing primitive, so it's added; everything else is already reachable through the execute_code sandbox (the agent can import sklearn, cv2, torch). Orchestration multiplies the agent; reimplementation would only produce a slower copy.

pip install --upgrade chimera-agent