Every table of the assistant has a decision about who writes it, by
operation (AssistantTableWrites); a table without one is not built. The
settings are edited by an administrator on a record screen. A conversation
is renamed, archived and deleted by its owner, and started only by the
assistant. Everything else (what was said and done, and what was spent) is
written by the assistant only.
The assistant enforces the decision itself, on any QQQ version: a table has
no capability for what a person may not do to it, so a screen is offered no
button and QQQ defines no bulk action; and the permission checker of every
table refuses the insert, update or delete a table route is about to make,
for every person, with 403 and a message that says why. The assistant's own
writes go through AssistantStore, in code, and are unchanged.
A person's save of a settings table sets no field that is not editable (the
id, the create date, who granted an override, health and test results, index
counters, a scheduled task's approval and run fields): StoredRecords.ofSave,
used by every settings customizer, takes them out of a USER save before a
rule sees them. A client that sends a whole record back keeps working.
Tests go through the real table routes against the fixture application.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
ArchitectureTest refuses a call of oldRecordListToMap anywhere in the
assistant: the map is keyed by the stored key, and the key of a record saved
from a screen is text. StoredRecordsTest pairs a record with its stored row
for an integer, a long and a text key, with the key as text and as a number,
and covers the merged record, refusals and comparisons.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A record screen saves through the UI API's table update route: only the
fields that were changed, with the version that was read. The record's key
comes from the request's path, so it reaches a table customizer as text.
AssistantBudgetCustomizer paired the record with its stored row by comparing
that text with the stored integer key, did not find the row, and judged the
one changed field as if it were a whole new budget: 400, nothing stored. The
same pairing was in every settings customizer: budget overrides, data access
rows, providers, knowledge sources and scheduled tasks were refused the same
way; a profile and a skill were stored with the defaults of a new row written
over what the save had not sent.
StoredRecords now does the pairing for all of them: the key of the record
and of each stored row are read as the type of the table's key field before
they are compared, the values being saved are laid over the stored row each
as the type of its field, and a change is found by comparing values of the
same type. A value that cannot be read as the type of its field is refused.
The rules (the host's ceiling on a cost limit, required fields, allowed
periods and scopes, an override's rules) are judged on the record as it will
be stored, for a save of one field exactly as for a save of all of them.
Tests go through the real route against the fixture application; fourteen of
the sixteen fail without the fix.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
QQQ's createDate and modifyDate behaviors each read the clock, so the two
values of one insert differ by a few microseconds on a clock finer than the
machine this was first run on. The decision claim and the budget event are
written once and never change; an insert customizer on those two tables makes
modifyDate the same instant as createDate. The test of the other tables no
longer asserts equality they cannot give (they are rewritten when a
transcript is redacted).
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Nine tables lacked a standard date: eight had createDate but no modifyDate
(turn events, tool calls, feedback, model calls, data audit, budget events,
client errors, pending action claims) and the budget accounting table had
neither. All nine now declare both, date-time, not editable, set by QQQ.
Every write the QBit makes goes through QQQ's own insert and update actions,
so an insert sets both and an update (an incremented daily usage counter, a
decided pending action, a cleared tool result, a scheduled run that ends, the
accounting row that follows a usage row) moves modifyDate. A table that is
only inserted has modifyDate equal to createDate.
The reference DDL gets an additive block (nullable columns defaulting to the
time of the write, so the previous release keeps working; existing rows get
the table's own creation time). Tests iterate the declared tables, exercise
the store's insert and update paths, and check the DDL creates the columns.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The watchdog ran the deadline logic after two database reads and heartbeat() was synchronized on the monitor that finish() needs, so one blocked database call stopped the deadline, the force-finalize and the final message.end. The deadline now runs first with no database access, housekeeping runs on a bounded abandonable worker, finish() has a guard that ends the stream with an error event, and every phase logs at INFO.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The panel reloads the conversation the moment the stream ends. The
redacted transcript (message parts, pending actions, stored events) was
flushed only after message.end was sent, so that reload read the "[]"
placeholders and showed the newest question and answer as empty bubbles
until the panel was closed and reopened. Persist and flush first, then
emit message.end.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Provider/model registry and OpenAI-compatible adapter (#22), Bedrock and
Vertex (#23), data classification, PHI routing and HIPAA mode (#24).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tools via /qqq/v1 with a loopback stand-in (#14); profiles, skills,
explain access, denials, ArchUnit and the denial matrix (#15); #16; the
demo pending-action card (#18).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Tools are no longer all sent as strict, so Anthropic no longer rejects
every turn with "compiled grammar is too large" (#13).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Assistant Admin record screens (#6), deterministic RobustnessTest (#9)
and the GA-only tag filter (#8), on top of the snake_case rename (#3).
0.1.0-RC.2 on Central is an accidental republish of RC.1, so this
candidate is RC.3.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
0.1.0-RC.2 is already on Maven Central (an RC tag build republished the previous code), so the next release-branch build must publish the following RC. RC tags no longer trigger a publish (/^v[0-9]+\.[0-9]+\.[0-9]+$/ only).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The orb published 0.1.0-RC.1 but could not push this commit (the
CircleCI deploy key is read only).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.