v0.9.4

New

• agenthud summary --model <name> forwards --model to claude -p. Summarization is a low-reasoning task, so cheaper models almost always suffice — sonnet (~40% cheaper than Opus, 1M context) or haiku (~80% cheaper, 200K context). Accepts a short alias or a full model id. Default unchanged (claude's own default model).
• Oversize report warning. The input stats line now includes an estimated token count. Reports above ~300K tokens print a loud warning and, in interactive mode, prompt one more time before sending. Range mode with -y only prints the warning and proceeds.

Fixed

• report --include validates unknown types. A typo like --include response,bas used to be silently accepted (no match, no error). Now unknown tokens trigger an error listing the offending values and the full list of valid types. Missing value after --include also errors.