0.142.0: Remote runs, Gemini and GPT-6 analysis models, and a view-only live stream

humanish 0.142.0 can run a study remotely: humanish run <study> --remote puts the study's own
process in a sandbox, so a long study keeps going after this machine sleeps, and observe --run <id> reconnects to it. Study analysis can run on gpt-6.1-sol and gpt-6-luna as well as the
default gpt-6-astra, and on Google's Gemini API with humanish keys set gemini. A hosted
desktop's live stream is now view-only on the VNC server itself: before, anyone holding a
participant's stream URL could type into that desktop.

Added

• Analysis runs on gpt-6.1-sol ($2 / $10 per 1M tokens) and gpt-6-luna ($0.10 / $0.50) as well
  as the default gpt-6-astra ($10 / $50). Pass humanish analyze --model gpt-6-luna or set
  review.analysis.model; admission, the cost cap and the recorded cost use each model's own rates
  and image-token sizing. analyze --help lists the accepted models. The default is unchanged.
• Study analysis can run on Google's Gemini API. humanish keys set gemini stores GEMINI_API_KEY
  (GOOGLE_API_KEY is read when it is unset), and review.analysis.provider: gemini or humanish analyze --provider gemini sends the evidence packet and captures to Google on gemini-3.8-flash.
  Admission, the cost cap and priced usage work as on OpenAI, and a missing Gemini key never falls
  back to OpenAI.
• humanish run <study> --remote (and watch --remote) runs the study's own process in a
  sandbox on its participants' provider, so a long study keeps going when this machine sleeps or
  its terminal closes. The command plans the study here, installs this humanish in the runner,
  serves the run's live Observer on 127.0.0.1 through the provider's port authentication, and
  copies the finished run back, each file checked against its digest. Ctrl-C stops the run and
  reclaims its sandboxes; after a closed terminal, humanish reclaim --run <id> stops the runner
  and every participant, found by the run's tags. The runner receives only the keys the study's
  plan reads. It takes computer-use studies on e2b-desktop or modal-desktop with a model API
  key and a public app-url subject, and refuses others before creating anything.
• humanish observe --run <id> on a remote run that has not been copied back reconnects to its
  runner: it serves the live Observer again, follows the run to its end, copies it back and stops
  the runner, as the launching command would have.

Changed

• Key discovery now treats a key's second spelling as the same key: a project's
  .humanish/local/provider.env no longer fills GITHUB_TOKEN beside an exported GH_TOKEN.
• An analysis that ends in an HTTP error now names the status code in its warnings.

Security

• A hosted desktop's live stream is view-only on the VNC server itself. E2B and Modal started
  x11vnc without -viewonly, and the stream URL's view_only setting applied only to the noVNC
  page, so anyone holding a participant's stream URL could type into or click on that desktop.
  humanish now switches the server to view-only after the stream starts and confirms it; if the
  server does not confirm, the run continues with screenshots and no live stream.

Not changed

The participant's model and what the default analysis reads are unchanged, so this release does not rerun the efficacy benchmark; the 0.133.0 result stands. A comparison of analysis models on fixed recordings is in progress and will set the default analysis model in a later release.