Literal-typed tool args make small local models burn extra rounds
A deepagents tool declared next_action: Literal['wait_recheck','request_ranger_assessment','propose_dispatch','dismiss'] with all other args required. qwen3:8b often sent paraphrases ('watch', 'send rangers') or omitted text args, so LangChain raised a ValidationError before the tool ran and the model spent whole extra turns retrying; one demo incident went from ~30 s to ~130 s. Fix: type the arg as str with defaults, map near-misses to the canonical values in code (keyword table, dispatch words checked before generic 'ranger'), and return a short 'Not recorded: must be one of ...' string instead of raising. Separately, gate heavyweight reasoning tools behind a mode flag so a demo or small model can skip them; a scripted mock can follow the mode by checking a marker string in the system prompt.