0.145.0: A remote runner's keys stay in E2B's proxy

humanish 0.145.0 keeps the OpenAI, Gemini and E2B keys out of a remote E2B runner's environment.
E2B's egress proxy adds each key to the runner's requests on its fixed host, and the runner holds
placeholders; code in the runner can still read the keys back through the E2B API, and a Modal
runner still receives them as environment variables. Study analysis with gpt-6-luna sends low
text verbosity, which made it about four times faster on the benchmark recordings, and an
observation of what a participant said may now also cite the capture they spoke about.

Security

• A remote run on E2B no longer puts the OpenAI, Gemini or E2B key in the runner's environment.
  Each one stays in an E2B header transform on its fixed host (api.openai.com,
  generativelanguage.googleapis.com, api.<E2B_DOMAIN>), the run gets a placeholder, and Node in
  the runner trusts E2B's proxy CA through NODE_EXTRA_CA_CERTS. The line before the run names
  the keys E2B's proxy holds apart from those that enter the runner as environment variables.
  Code in the runner can still read the keys back from E2B's sandbox info endpoint, which returns
  the transforms as they were set. A Modal runner, which has no header transform, still receives
  its keys as environment variables.

Changed

• An analysis observation that reports what a participant said may now cite the captures or
  actions they were speaking about, as long as it cites at least one of their own messages.
  Before, any such extra citation failed the whole analysis, which caused nearly all
  gpt-6-luna rejections on the benchmark recordings
  (participant statements that cite captures).
  The analysis prompt is unchanged.
• Study analysis with gpt-6-luna sends text.verbosity: "low". On the Taskly benchmark
  recordings at high effort, its median analysis took 110 s in place of 406 s, 22 of 24 analyses
  passed the evidence checks in place of 18 of 24, and the cost stayed under $0.01
  (gpt-6-luna with low text verbosity).
  At medium effort the median took 80 s in place of 41 s. Other models send the API's default
  verbosity.
• The Taskly benchmark rubric is version 2. It reads more wordings of D1 (Clear completed does
  nothing) and D5 (Save leaves the row in edit mode), so analysis recall on the 2026-10-21
  recordings rises for every analyst except gemini-4-argon, for example gpt-6.1-sol at medium effort from 28/60 to
  36/60. Invented claims are unchanged.