Fix piped output truncation, harden scraping, add queue/search/harvest

[eppser]

Ran this against a live event (5 evaluations, ~500 submissions) and hit three bugs. Fixes plus the tooling I needed to review at that volume.

Fixes

1. Piped output truncates

bin/sessionize.js calls process.exit() right after console.log(JSON.stringify(...)). Writes to a pipe are async, so the process exits before the buffer flushes:

$ sessionize session show | wc -c
512
$ sessionize session show > out.json; wc -c < out.json
7139

Every jq example in the README was silently getting truncated JSON. Fixed by exiting from the write callback.

2. Extraction assumed one event's conventions

Titles were found by matching a [Track N] prefix, description and outline by slicing innerText between guessed markers, speaker by a Firstname Lastname regex, and rating criteria were hardcoded to Practical/Originality/Relevance/Clarity. On an event without those conventions you get title: null, description: "", and a speaker name lifted from a form label (mine returned "Media Release").

Extraction now reads the server-rendered hooks that are stable across events: #Session_Id, .ibox-title h3, p.es-description, the .es-categories field table, .es-speakers, and the dt / input.rating-value pairs. Per-event criteria and custom fields come through as data rather than assumptions. Speaker gains tagline, bio, photo, and speaker-level custom fields.

3. Ratings never reached the posted fields

setRatings drove only the jQuery rateYo widget and left the hidden input.rating-value fields (Criteria[i].Rating) untouched. It now writes the input first, then syncs the widget so the UI agrees.

New

• session queue — the full queue via Sessionize's own tabboxes endpoint, paged, with rating state per session. session list only sees the ~20 sessions linked on the rendered page while reporting its count as if complete; it now says so in its output rather than changing behaviour.
• session search <query> — session search scoped to the evaluation.
• session harvest — walks the queue and caches every submission to local JSON (~/.sessionize-cli-harvest/<eventId>-<evaluationId>.json by default). Resumable, and reconnects if a tab or the CDP connection drops mid-walk. Reviewing a few hundred submissions against a file beats a browser round trip each time.
• session rate --dry-run — fills the form and reports what would post without saving. rate also refuses to run without --comment when the evaluation marks it required, instead of submitting a form the server will reject.

Testing

All commands exercised against a live evaluation: queue returned 125/166/92/64/54 across five tracks, exactly matching the counts on the event page; harvest pulled 501 submissions with zero failures; search returned sensible hits; rate --dry-run confirmed the hidden inputs carry the scores and the save button enables, then reloaded without saving. No evaluation was submitted.

One caveat: I drove this through a local Chrome over CDP rather than Browserbase, so the Browserbase connect path is unexercised here. lib/browser.js and lib/auth.js are untouched by this PR.

Not included

extractSessionDetails now returns comment (the editable field's current value) plus commentRequired, where it previously returned a comments string scraped from the page body. Happy to keep the old key as an alias if you'd rather not break that shape.

https://claude.ai/code/session_0183YGef9Gm7vbVSB6HSNRnq