v0.12.0

Minor Changes

• 8aae1b2: Remove references to eu api endpoint

• 2b16772: subtext-sightmap: upload the .sightmap/ corpus with the sightmap CLI
  (sightmap export --url <sightmap_upload_url>) instead of the bundled Python
  collector, and remove collect_and_upload_sightmap.py.

  sightmap export routes the upload through the Go loader — the single source of
  truth, shared with the server-side reader — and POSTs the whole canonical wire
  (components incl. view-scoped, views/routes, requests, messages, memory, tags), so
  review snapshots and signals now also carry view and network annotations, not just
  components. Drops the Python 3 / PyYAML dependency. The inline review-open sightmap: array stays as the small-set / no-binary fallback. Also documents that
  the upload token is single-use and time-limited (upload immediately after
  review-open).

Patch Changes

• fc74464: Document what the review-search index actually covers, and drop an example that could never match. The index holds page navigations, custom events, and network requests that failed (status >= 400). Successful requests, clicks, and console messages are not indexed and can never match — so a predicate over a 2xx status matches nothing, and a not_has over one is vacuously true for every session.

  The skill's worked example did exactly that: it used not_has over a 2xx checkout/pay response to mean "never got a successful payment", which is satisfied by every session in the window regardless of behavior. Replaced with a verified example that builds its absence check over navigations instead, and that also demonstrates junction nesting (operands accept any node, not just has) and a custom event predicate.

  Also documents the predicate readback. Every response echoes how the server parsed the tree, which is the only way to tell a mis-built-but-parseable query from a genuinely empty result — zero matches on its own could mean a wrong predicate, too narrow a window, or an unindexed signal.

  This coverage rule appears in neither the tool schema nor the tool description; it surfaces only in the zero-result help text, so a query that returns rows never reveals it.

• fc74464: Correct the documented result ordering for review-search. The skill said results were not ordered by start time and told callers to sort them, which read as "unordered". They are ordered — by last activity, most recent first. Start time is not the sort key, and because the response displays only started, the ordering looks arbitrary in the output when it isn't. Also notes that the top of the list turns over quickly on a busy org, so an identical query re-run seconds later can return a different set.

• fc74464: Trim subtext-search to what the review-search schema and its error messages don't already carry. Removed the parameter and operator inventory (match kinds and their fields, string/int operators, between inclusivity, method case sensitivity, limit bounds, empty-match semantics) — all of it is in the self-describing tool schema, and the match-kind and since-format rules additionally reject with messages that name the rule and the fix.

  What stays is what inspecting the tool can't tell you: when to reach for search over review-open or review-list-sessions, the handoff into review, result ordering, per-call cost, the mutual exclusivity of since/time_range (the schema's per-field "one of" wording doesn't convey that passing both is rejected), and the three where shapes whose rejections surface as raw unmarshal errors naming an internal type (and/or taking an operands object, count taking an object, not_has being a leaf rather than a junction).

  Also disambiguates "URL", which previously read as a reason to skip search whenever one was on hand. A Fullstory session URL names one recorded session and should be opened directly; an app URL like /checkout is something sessions visited or requested, which is a search predicate and a reason to search rather than skip it.