A retried card submit was only repairable while the first attempt's watch was
still active. Once it had fired, expired, or answered in one shot, the retry
re-evaluated the condition and created a second operation.
A watch_submissions ledger, keyed (chat_id, client_request_id), is now written
before the condition is read and carries the outcome once there is one. A retry
looks it up first: a recorded outcome is replayed, a different draft under the
same id conflicts, and only a pending row proceeds - converging on the watch id
reserved up front rather than creating another.
The message-size checks ran after the body had been read, so a request without a
content-length was buffered and parsed in full before being refused. An ingress
cap on the agent's paths now counts the bytes as they arrive, and the chat proxy
reads its body with a ceiling instead of reading it whole first.
The page load reported unread wakes only, so a fresh browser whose watch was
created elsewhere and hasn't fired yet never started polling: the wake landed
without a toast or a dot until a reload. The loader now returns the active-watch
presence too, in one read per page load.
The batch took the 500 soonest-expiring watches of a group every tick, so a group
larger than the cap could leave the rest unchecked until the first 500 expired.
The group is now ordered least-recently-checked first, with a generated
cadence_minutes column and an index so the due predicate no longer re-parses the
spec JSON per tick.
The chip factory and slot vocabulary, the evergreen docs chips, the per-page slot
table and the signal chips each get their own module; registry.ts is now the
public face that re-exports them. A pure move.
dashboardAgentWatchChecks.ts keeps the dispatcher, the failure envelope and the
previous-facts reader, and re-exports the run, queue, error and health families
plus the shared reader contract. A pure move; every import path is unchanged.
The gallery pages, their fixtures and the screenshot script only exist to
look at the agent UI, so they ship separately. The fixtures two unit tests
read — the report view models and the page contexts — stay here.
Head start passed `system` as a bare string, so Anthropic neither wrote nor read
the cache: the ~18.4k-token prefix was billed in full and the agent's step 2 then
paid for a fresh write. Both breakpoints now carry a 1-hour TTL.
The two prefixes were also not identical — `get_run` sat in a different position in
the agent's tool set than in the canonical schema-only one, so they could never have
shared a cache. Reordered, and a fingerprint over system text + tool definitions is
logged per model call alongside the provider's cache-write, cache-read and uncached
input token counts, so a future drift is visible instead of silent.
Landmarks derived from the layout spec must appear in the same order in the server-rendered card
and in the markdown, over healthy, degraded, untrustworthy and empty reports.
The card no longer formats values, resolves codes or classifies footer codes of its own, and a
genuinely-unknown state now gets a neutral icon instead of a green tick.
The text surfaces gain the card's structure (headline, hero evidence, why: block, aggregated
read:, trust flag) and keep their alignment: fixed sparkline column, aligned notes.
Katia Bulatova