docs: ZeroGPU rate-limit expectation in UI header

harness: tidy_list — drop orphan numbers, fix 1.+bullets, dedupe N\nN. pattern

byte-cap: truncate reply at sentence/word boundary ≤150 B

harness: recover query/text from malformed args + run second turn anyway

fallback: surface tool_result as Reply when second-turn is empty/rate-limited

two-turn: clean assistant message + higher second-turn temp + retry

orchestrator: simulate tool result + second-turn final reply

harness: name-resolver + always surface malformed intent

app: use shared tool_extract module

share: tool_extract — robust 3-format harness

harness v3: defensive guards + _clean_visible + safe_generate

harness v2: strong JSON repair + temp 0.1 + no repetition_penalty

harness: robust tool_call extraction (3 formats + JSON repair)

fix: force CPU for embed model (ZeroGPU forbids CUDA init outside @spaces.GPU)

gradio>=5.0 (fix huggingface_hub HfFolder import)

bump gradio sdk_version to 5.49.1

deps: transformers + spaces + sentence-transformers

app: ZeroGPU + transformers (3 modes: LoRa, free chat, RAG)

README + ZeroGPU config

requirements

app: 3-mode Gradio (LoRa radio + free chat + RAG)

initial commit