fix DSML token leak and use no-thinking model for all non-tool-bound paths

lint

lint

fix 3 layer memory ststem

feat: add youtube_frame_at tool, improve vision URL validation, silence observability logs, and update caveman mode configuration.

feat: implement GAIA benchmark evaluation runner with strict output formatting and vision tool integration.

feat: implement tool error cooldown, caveman system prompt mode, and iteration fail-safe logic

feat: implement web search and file processing tools with robust retrieval fallbacks