Count, filter and page the Observer grid and bound watch output at 100 participants

[danielgwilson]

Slice 4 of #1737: the Observer grid and humanish watch at 40 and 100 participants. No humanish.observer-data.v1 change.

Measurements

Fixture: observer/tests/scale-fixture.ts builds a finished computer-use run of N participants with the producer (buildObserverData) from the recorded one-participant run in tests/golden/labs/live.json (50 turns each, own screenshot folder each). Personas rotate over 3 ids; every ten participants end as 6 reported complete and 1 each blocked, gave up, interrupted, failed. Headless Chromium, built artifact served over loopback. The measurement script is scratch and not committed.

	40 @1440	40 @390	100 @1440	100 @390
artifact with data, before / after	3,329,516 / 3,331,001 B	same	6,538,638 / 6,540,123 B	same
observer-data.json	2.14 MB	2.14 MB	5.35 MB	5.35 MB
page 1 painted, median of 5 alternating loads, before / after	756 / 731 ms	533 / 493 ms	763 / 726 ms	701 / 691 ms
grid page 1	36 cards	36	36	36
paging	36 + 4	36 + 4	36 + 36 + 28	36 + 36 + 28
status counts above the grid, before / after	none / 24 · 4 · 4 · 4 · 4	same	none / 60 · 10 · 10 · 10 · 10	same
persona, before / after	search text only, 13 / select, 13	same	search only, 33 / select, 33	same
status filter Blocked	4 cards	4	10	10
pin on last page	moves to page 1	same	same	same
comparison of 3	opens	opens	opens	opens
report (analysis citing 360 / 900 entries)	opens	opens	opens	opens
first card after Next page from the bottom, before / after	in view (page 2 is 4 cards)	9877 px above / in view	5389 px above / in view	9877 px above / in view

The timing rows come from one alternating before/after run on a shared, loaded box. An earlier unloaded run measured 316-486 ms before the change. Before and after are within noise. Served polling at 100 participants costs about 300 ms of long tasks per 20 s (four 5 s polls of the 5.35 MB JSON) before and after.

humanish watch <study> --dry-run with 100 participants. The before column was measured with the cap patched locally to 100 and then reverted. The after column was measured on main with the cap at 100 (#1739), unpatched:

	before	after
stdout lines	220	122
of which participant lines	100	2 (count line, not listed:)
of which persona-background warnings	100	100 (src/study, not in this PR)
stderr lines (plan, one per participant)	103	103 plus 2 lines of the Observer artifact build (participant-runs.ts, not in this PR)

A live-shaped result (tests/cli/computer-use-participants-output.test.ts) prints 31 stdout lines at both 40 and 100 participants. Before this PR the same 100-participant result printed one participant line per participant with its whole closing message, plus 100 copies of the screenshots warning.

What changed

• observer/lib/grid-filters.ts is a new module that owns the grid's filter fields. Its interface is GridFilters (with persona optional, so a preference saved before persona existed still reads), NO_FILTERS, isGridFilters, filterParticipants, activeFilterCount and statusCounts. The default, the storage guard and the predicate move out of app.tsx, and the badge count and clear literal move out of grid-options.tsx. Before this PR a new field touched those five places.
• GridStatusSummary (in components/grid-options.tsx) adds one toggle button per status label with its count, shown when participants ended more than one way. Pressing a count filters the grid, and pressing it again clears the filter. It counts statusLabel, the key the status filter matches, so a count equals the cards it shows. StudyGrid renders it through a new statusSummary slot.
• A Persona select in the view options, shown when the run has more than one persona. It reads stream.sim.personaId, which the contract already carries.
• Paging: the pager scrolls the grid section to the top of its scroller, and a filter change returns to page 1. Windowing was not needed: 36 cards per page render in under 0.5 s, and paging already worked.
• formatCuaStudyHuman (src/cli/commands/study-format.ts), the output of watch and run:
  • Up to 16 participants print one line each, as before.
  • Above 16, it prints one count line by status, the first 16 participants that did not pass, and a not listed: line.
  • Every participant line keeps its closing message on one line of at most 160 characters.
  • warningLines prints a repeated warning once, in all four route formatters.
  • --json output is unchanged.

Checks

• Red first, through the rendered App in jsdom: observer/tests/scale.test.tsx, 12 tests at 40 and 100 participants. Before the change, the status row, persona and page-reset tests failed (6 of 11 at that point). Mutations caught: page reset removed, persona clause removed, persona required by the guard.

• tests/cli/computer-use-participants-output.test.ts, 6 tests, red before the change:

  • five on results built from tests/golden/routes/computer-use-fanout-live.json;
  • one through humanish watch <study> --dry-run at 40 participants.

  Mutations caught: the bound raised to 17 or to 100, the dedupe removed.

• New browser case grid-pages-phone (80 participants at 390 px): it fails without the scroll and passes with it.

• pnpm --filter humanish-observer typecheck, then the observer suite after a build: 491 passed.

• pnpm build, then the four observer proofs passed: iframe; browser, 74 cases; chrome, 4; reliability, 2. The observer suite and the proofs last ran on this branch rebased onto 41a72af. The two later rebases (Let a study have 100 participants and run them within the E2B plan #1739, Fix the four TUI gaps from the 0.116.0 release notes #1735) changed nothing under observer/.

• Artifact: 1,185,473 to 1,187,030 bytes before data, against a 1,500,000-byte budget.

• pnpm release:check passed on the pushed head: vitest 7987 passed and 12 skipped, TUI 192 passed.

• vocabulary.lane fell from 91 to 90, so this PR lowers its cap.

Needs attention

• The persona-background warning per participant (src/study/warnings.ts) and the plan line per participant on stderr (src/routes/computer-use/participant-runs.ts) still grow with the roster. Both files belong to the participants-cap session (A study can't simulate a full day: computer-use studies stop at 16 participants #1737 slice 1), which was told.
• The shared-world formatter still lists every participant on its own line.
• A live run over 16 participants records its analysis skip (AUTOMATIC_ANALYSIS_PARTICIPANT_LIMIT, Let a study have 100 participants and run them within the E2B plan #1739) only in the CLI result. withParticipantLimitSkip writes no automatic job record, so the Observer of that run has no Findings view and no sentence saying why. I found this by reading src/run/route-shell.ts; no run was made. The Observer has no detail text for that code either. The cohort analysis of slice 3 removes the skip.

Not verified

• No live run and no real 40- or 100-participant bundle. The fixtures replicate one recorded participant.
• The before watch measurement at 100 used a locally patched cap, because the cap was 16 on main at the time.
• The browser timing comes from one shared box under load from other sessions.
• Served polling of a 5 MB observer-data.json every 5 s was measured, not changed.
• The report was checked with a synthetic analysis citing up to 900 entries. The cohort analysis of slice 3 does not exist yet.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
humanish	Ready	Preview	Oct 9, 2026 7:04am UTC