Asked how many capacity commitments were on the book, Piggy answered "3".
Production holds 5. It had called the idle-capacity tool, which filters to
blocks above an idle threshold, and read the length of that list as the size
of the book.
The system prompt already forbade this in terms — "never report a filtered
count as a total; pig_get_idle_capacity returns the blocks with idle hours,
not the book" — and the model did it anyway. That is the second time this
argument has been lost in the prompt, so it is settled in the payload
instead: a result that cannot describe its own scope will be misread
eventually, however firmly the prompt objects.
Every tool that returns a count or a collection now carries one shape:
what it covers, how many matched, out of how many, under which filters, and
whether the list was truncated. The denominators are read from the database
rather than inferred. The pre-formatted headline states the scope too, since
that is the sentence a small model quotes most readily — the idle tool now
opens "3 of 5 live capacity commitments on the book", which is the sentence
that makes the original mistake impossible to phrase.
Two details worth keeping. Record reads enumerate rather than filter, so
their scope states a boundary instead of a ratio: these are that record's own
figures, never book-wide totals. And the workspace summary's idle threshold
is deliberately recorded as 0, distinct from the idle tool's 0.25 — that
mismatch is why three different idle figures appeared across the UI, and
naming it in the data is how it stops being invisible.
Verified against the live model: the failing question now answers 5, demand
deals 13 and contracts 20 — each drawn from a payload whose filtered figure
was smaller — while "which blocks are sitting idle" still names exactly the
blocks that are.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Piggy answered in raw markdown, threw away every tool result it streamed,
and fought the reader's scroll on every token. The three surfaces that
made it worth having — what it read, how it reasoned, what it cost — were
all on the wire and none of them reached the screen.
The transcript is now composed of five parts under components/piggy:
answers render through streamdown, the container sticks to the bottom
without pinning the reader there, tool steps say what they read and link
to the record, and each turn carries its model and token count. Three
lifecycle bugs went with them: Stop left a permanent spinner, a truncated
stream was indistinguishable from thinking, and a failed send destroyed
the message it failed to send.
Underneath, the inference path grew timeouts, jittered retries on 429 and
5xx, tolerance of the malformed frames a 30B model emits, and an
agent_runs row per turn so chat spend is observable. The system prompt now
states that a field ending in Cents is cents — without it nemotron renders
costPerGpuHourCents: 189 as "$189 per GPU-hour", which is a 100x error on
the most scrutinised number in the room.
The demo book was arithmetically incoherent: every deal's value
contradicted its own allocation revenue by up to 3.6x, nothing had ever
closed, no customer had any paper, and the marketplace was empty. Deal
value is now derived from the allocation, the book clears 5.3% across five
blocks with one deliberately underwater, and the renewal, compliance and
agent-provenance machinery finally has rows to act on. A --clear that
deleted every obligation, SLA term and capacity request in the database
regardless of origin is scoped to the demo's own ids.
Around that: accounts have a detail page, ⌘K searches the book, Settings
can mint the API keys it always claimed to, and deploy.sh actually ships
the agent instead of silently skipping its compose profile.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
claude