0.129.0: Analysis reads long runs past 100,000 items; traces past 32 MiB get their own advice; shared-world skips counted apart

humanish 0.129.0 fixes three things long studies reach. Analysis no longer refuses a run by how
many events, trace items or screenshot frames it holds; it still reads each file only up to 16
MiB. A trace or live file past 32 MiB, the most the trace loader reads, is refused with advice
that fits, a running participant whose live file passes 32 MiB keeps the items the Observer page
already shows, and the Observer server reads run.json under that limit. Shared-world runs count a
participant skipped before its session apart from those that used the app, and no longer count a
skip as overlap. This is a minor release because a shared-world skippedReason is now a sentence
where it was handoff-timeout or host-ended-before-handoff.

Fixes

• Analysis no longer refuses a run by how many run events, trace items or screenshot frames it holds. It refused a run.json with more than 100,000 events or a participant trace with more than 100,000 items (ANALYSIS_SOURCE_INVALID), and an evidence entry whose screenshot frame passed 100,000 (ANALYSIS_INPUT_INVALID). Analysis still reads run.json and each trace file only up to 16 MiB, and refuses the whole run when one of them is larger. A trace reaches 16 MiB after about 3,570 model turns at the 4,700 bytes a turn the bundle estimate counts, 9.9 hours at its 10 seconds a turn, and after about 8,540 turns at the live fit of 1,965 bytes a turn. verify and the Observer read each file up to 32 MiB, so they still accept a run between 16 and 32 MiB that analysis refuses (#1840). (#1824, #1843)
• A trace or live file that run.json records past 32 MiB, the most the trace loader reads of one file, is refused with a message that says so and that restoring a copy does not help; it said to restore the run folder from a copy. In the served Observer, a running participant whose live file passes 32 MiB keeps the items the page already holds, with a live-items.not-read warning that later items are not read; the page dropped every item and showed the trace as unavailable. A page opened after the file passed the limit, and every page once the run finishes, shows that trace as unavailable with the file's size. The Observer server reads run.json and a saved observer-data.json only up to 32 MiB, for its feed and page and for run.json asked for by name, and answers 404 with a reason naming the limit past it; it read both whole. Other run files asked for by path, a trace or live file among them, are still sent whole. (#1825, #1844)
• A shared-world participant the run skips before its session starts, such as a follower stopped at the handoff because the host never reached a lobby, is now counted apart in the review summary in run.json and review.md. A run of 3 whose 2 followers timed out at the handoff reads "1 participant used one public app. 1 of 1 reached the goal. 2 participants were skipped before their sessions started; the gaps say why." It used to count all 3 as using the app. The skipped participant's humanish run line reads blocked and ends on its skip reason, and its roles[].status in --json is blocked with a skippedReason. Both said failed. The skip reason is a sentence, such as "skipped: the host did not reach a lobby within the handoff deadline; no desktop was created", in the form the computer-use skips use. The gap line in run.json carries the same sentence where it read player-2: handoff-timeout. The skipped participant's own summary in run.json simulations[] ends on that sentence too, such as "Persona 02 was skipped: study budget reached before this participant started: ...; no desktop was created." A budget skip read "Persona 02 would use the shared app; no session ran.", as on a dry run. (#1831, #1845)
• A skipped shared-world participant no longer counts as using the app at the same time as another. Its window sat inside a session that was still open, so a run whose 2 followers timed out at the handoff printed "Participants used the app at the same time" while only the host used the app, and overlapProven in --json was true. The run line, overlapProven, maxSimultaneousLanes, the review summary and the concurrency event in run.json now count only the participants that used the app. run.json's laneWindows still hold a window for every declared participant. The in-progress summary, written before any participant starts, reads "3 participants are planned, all in one public app." where it read "3 participants are using one public app at the same time.", as the computer-use one does since 0.127.1. (#1831, #1845)

Checked before release

• CI on this PR (sharded gate on both Node versions, site, Observer proofs, TUI, guest desktop, secret scan) is the gate of record. pnpm gate on the release commit passed.
• The post-publish check rehearsed on the release tarball (pnpm packed:check): all 125 rows passed. The registry and unpkg rows run after publish.
• pnpm release:dogfood: a participant installed the tarball, ran the offline first-run study (share_ready, 16 of 16 checks, $0) and passed.
• Live on the release tarball (hosted E2B, Chrome, keys from the user store):
  • try-live with analysis: passed, analysis wrote its findings, verify passed.
  • humanish observe --all --serve on that run: run.json and observer-feed.json answered 200. With run.json padded to 33,554,433 bytes, both answered 404 with "run.json is larger than 33554432 bytes, the most humanish reads of it." Verify passed again after the file was restored.
• Before merge, each change had a Claude review (Kent Beck and John Ousterhout lenses) and a fix round. The trace and serving change (#1844) had two Codex reviews; round 2 found no runtime defect. Analysis was measured on a synthetic 25-participant, 8-hour run: about 118k input tokens per cohort request at 10 s and 6 s turns.

Not verified

• Analysis past 100,000 events or items in a live run; the tests and the synthetic run cover it.
• A live file passing 32 MiB in a running study, and the page's live-items.not-read warning in a browser.
• A live shared-world run: the public route needs a multiplayer app the operator owns. The skip, overlap and summary changes are covered by route tests.
• Consumers outside this repository of the old shared-world skippedReason values.
• macOS and Windows.

Package

humanish@0.129.0: 1004 files, shasum ee0d4cdfafa0f145b5555c5e7b8dba8681087472.

🤖 Generated with Claude Code