After a second independent review of reading files and of the link rule:
- Reading files is off by default. A host turns it on per table
(withFileSource) or for every table (withFileReading); until then the
tool is not offered and a call reads nothing.
- A file is read only when the record it is attached to may be read: where
the host declared it (withAttachedTo, withAttachedToRecordOf), the
assistant must be allowed that record's table and the person that record.
An attachment-style table with no declaration is refused, and the refusal
says what to declare.
- A file kept outside the application is fetched only from an allowed
origin (scheme, host and port; https unless an http origin is listed),
with one deadline for the whole fetch, body included. The client is shut
down after each fetch. A refusal names neither the address nor its host.
- Taking the text out of a document is bounded: what a workbook inflates
to, the rows and cells walked, and time.
- The link rule judges every way a path of the application can be written
(angle brackets, an HTML anchor, relative forms), and a path that is only
written in the content of a file or a record no longer counts as given.
The design note says plainly that the panel is the control for addresses
of other sites and for images.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A review was passed on its count and its lines alone. A review with an
option whose value the process had changed after the card was made, or with
a value to read, went on to store with what the card never showed (proved
by a second independent review).
A review is a screen too: before it is passed, and before it is answered
with "validate", what it has of its own is held to the record of what the
card displayed (screenDiffers), like any other screen. Such a run stops
with nothing stored, and the card to go on with shows the review's inputs
with the values the process holds now and its values to read. The values
the card gave for a review's own inputs are sent with its answer.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A person asked what was in a file attached to a record. The assistant found
the row of the file, said it cannot read files, and wrote an address it had
composed as plain text.
read_file returns the content of a file a record holds, for a person who
may open it. Which fields are files is read from the metadata (a BLOB, a
field with a file-download adornment); no table of another QBit is named.
The record is read as the person, a BLOB is downloaded with the
application's own operation as the person (UiApiGateway.download), and a
file kept outside the application is fetched only from a host the code
allowed (withFileUrlHosts; no credentials, no redirect). Data access and
field exposure narrow it as for every read: a masked field is not read.
Text comes back as text, a PDF or a workbook as its text, bounded, and a
part says it is a part; an image and any other binary is described, never
dumped. The content is untrusted data inside the envelope, and the
tool-call log does not keep it.
A file is shown as a chip that opens the record it is attached to (when the
host said which, and the application lists that table for the person), or
its own record. An address of the application in an answer is kept only
when a tool result, the conversation or the page gave it (AnswerLinks): a
composed path is removed from the streamed and the stored text.
Design: docs/file-contents-and-links.md.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Three cases an independent review proved against the first version, each
with a test that fails without its fix:
- a review with a warning and an error was answered with 'validate first'
before its lines were read, and a process that is not core's streamed one
went on to store. The lines are now judged first, for every process, and
'validate first' is only answered for core's own streamed process; a
review with no lines is not passed.
- a screen outside the process's list of steps (an optional step), with a
defaulted option, was passed although the card never displayed it. The
stored proposal now keeps a record of what the card displayed, and a screen
is passed only against it; a screen the process changed while it ran and
every screen of a state-machine process stop.
- a job the person may not read was reported as running. A job that was not
read is reported as exactly that, and the action fails.
Also: a status check cancels nothing; a run still working is 'started and
has not finished'; the names on a card are in the application's order;
every input is on the card, also one left empty.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The first card showed the first 300 characters of a screen's text and then
passed the screen. It now shows the whole text (up to 2,000 characters), and
a screen whose text is longer is one the card did not show in full: the run
stops there.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A reply that stopped at the output ceiling is read like any other: the tool
calls it had completely formed run, the one it was still writing (its last
part) and any whose input a provider could not read never do and get an
error result that says so. The answer goes on by itself at most
maxOutputContinuations times, each a model call under every limit of any
other; after that it ends INCOMPLETE, never COMPLETE, says what was done and
what was not, and stays in the conversation so that 'continue' goes on.
A provider row lowers the output ceiling of one reply only down to
minOutputTokensPerRequest; it never raises the ceiling the code configured.
An answer shows a record once and at most maxRecordCardsPerTurn record cards
and maxRecordChipsPerTurn record links; a confirmation card is never left
out. The system prompt asks for one proposal for a set and compact answers.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
docs/one-card-for-a-set.md is the design and the safety argument for parts A
to D; README, UPGRADE, CHANGELOG and the agent notes follow it. UI API
extension assistant 1.11.0 (additive): AssistantActionV1.recordCount,
records and recordsBefore for a run over a set.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A person the application refuses is refused before a card where the
application says so beforehand; the confirm-time refusal (a permission that
goes between the card and the confirmation) keeps its tests, with the UI's
own status and message. The tool-set file changes in one description
(run_process), replaced as text.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A confirmed run_process goes through the process's own init and step
operations as the person and passes a screen for them only when it shows
nothing their card did not: its text and inputs were on the card with the
values the process has now, and a review counts exactly the records of the
card and, with every record validated first, has no warning and no error.
Anything else comes back to the person with a new card and nothing stored.
The set is frozen by primary key and by one digest over its records as the
person read them; a set that changed is not run. The card keeps what the
records were (how many, a sample of 20) and lists the whole set by name.
The assistant no longer creates a job: when the application itself hands the
work to a background job, the job is the person's, it is waited for, and its
outcome (rows done, failed and why) is the card's. A job the application
created waiting is cancelled and the run reported as not done.
A change to records of a table is proposed only to a person the application
lets make it (what it tells their own screens); otherwise they are told so
and there is no card.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>