0.151.0: Claude participants compact long sessions

humanish 0.151.0 lets a Claude participant held to a scheduled end, such as an hour or a working
day, keep going. Its sessions now compact on the server at 60,000 input tokens: the participant
writes first-person notes and carries on from them. Before, a session stopped near the Messages
API's 32 MB request limit. In a live 65-minute Claude Haiku 5.5 session, requests stayed under
3.6 MB across 5 compactions. A Claude participant's default output limit rises from 16,000 to
32,000 tokens to leave room for the notes. A participant whose study names no mission now reads
that it is trying a web app for the first time, where it used to read that it was testing one,
so the participant prompt digest changes for every study.

Changed

• A computer-use participant whose study names no mission now reads "You are trying a web app for
  the first time ... Use it the way you actually would, until you have decided whether you would
  keep using it, then stop." It used to read "You are testing a web application", and with that
  mission 4 of 8 live reports described the data they entered as test or placeholder data; with
  the new one, 1 of 8 did. A shared-world participant with no mission is now told it is one of many
  people using a shared app. Other text a participant reads drops harness and tester words: the
  per-participant focus line reads Your focus: (was Lane focus:), the inbox lines no longer say
  "test identity" or "harness limitations", the dwell note says how long the participant stayed on
  the page, and a Claude Code participant's notes begin Note:. The participant prompt digest
  changes for every study. Evidence:
  docs/evidence/computer-use/persona-adherence-2026-10-26.md.

Fixes

• A Claude participant held to a scheduled end, such as an hour or a working day, no longer stops
  at the Messages API's 32 MB request limit. Once a request carries 60,000 input tokens, the
  server compacts the session: the participant writes first-person notes on what it set out to do,
  what it did and saw, and what it still means to do, and carries on from them. Later requests
  start from those notes and no longer send the screenshots before them. The trace's conversation
  record gives each request's size in requestBytes, marks the compacted request
  compacted: true and counts the replies the notes replaced in summarizedTurns. The compaction
  is billed as an inference of its own, with its own row in tokenUsage.turns, and the cost
  estimate and spend cap include it. The notes count toward the request's output limit, so a
  Claude participant's default maxOutputTokens rises from 16,000 to 32,000: on a 54,000-token
  session the notes took 23,660 to 26,625 tokens. Notes that reach the limit stop short, so they
  are left out with the reasoning built on them, and the next request compacts again; a closing
  request does not compact. The server's tool-result clearing is no longer sent: it cleared one
  screenshot per request and left the billed input nearly unchanged. A session whose screenshots
  pass 30 MB before the server compacts still stops before the limit, naming the size.