"As a table" used to drop the table's chart, so the reply offered no way
back to it. The copied table now keeps its chart and carries view: table,
which the panel opens on the table (the chart one tap away on the toggle)
and the chips read to offer "as a chart". show_result with view table
does the same. The field is optional on the result metadata; a client
that ignores it shows the chart first as before.
The Coach's recent nights carried the wake instant in UTC as their date,
while the sleep table under the answer keys a night by the local day it
ends on. East of UTC the two named different days for the same night.
The snapshot now gives the night's own key.
The recheck and follow-up rules now ride the tool-mode prompt only once
the conversation holds a table, or from the round after the turn produced
its first one. The clarifying-question rules are shorter and English for
every locale, like the other instructions to the model; the German copy
of them and of the recheck rule is gone. A turn without tables sends
about 1 300 characters less on every round.
The golden transcripts change only in the system prompt's hash and
length (4 509 to 3 187 characters; 4 728 to 3 280 for the German turn).
Table refs, step ids, chip ids, a chip's anchor ref, clarification choice
ids and a reused table's source are now validated against their exact
server-minted patterns when a stored reply is read back. Some of them are
written into the model's context on a later turn (the earlier-tables list,
the "keep looking" fetched list), so a stored row edited by hand can no
longer carry text into a prompt through them.
show_result now passes the same scope gate a fresh read does before it
decrypts a stored table: a metric the person has since excluded from the
Coach, or one outside the conversation's scope, answers
unavailable_in_scope and nothing is sent. The list of earlier tables in
the context drops those metrics too, so the model is not told they exist.
The exclusion (the person's own list plus switched-off modules, and no
adherence when medications are excluded) is the snapshot builder's, now
one exported helper both paths use.
"Keep looking" no longer copies the unfinished question into the
server-written context lines. The question stays in the transcript as the
person's own turn; the line only points at it. Lifted into the system
role, whatever the question said would have carried the system's
authority.
A turn now reads its conversation's newest messages once and shares the
rows between the chip, clarification, "keep looking" and repeat-question
checks, instead of three identical reads.
Earlier tables keep their m<k> names once a conversation outgrows the
window a turn loads: the index counts the assistant messages before the
window too, so a name the model read on an earlier turn still points at
the same table.
The backup validation now reads the per-message table ceiling from
results/refs, which imports nothing but types. Reading it from
results/project pulled the tool executor and the message store into the
snapshot builder's import graph, and the free-text sanitiser guard then
walked files that never build a prompt.
The grounding ledger derives differences only between figures of one
payload, so a correct "up 4 mmHg on last week" drawn from two separate
table reads was withheld as unverified. A table of the period before or of
a year ago now reads the current window too, and its summary carries the
current figures and the change (delta and percent change) beside the
earlier ones. A wrong delta is still flagged.
The live evaluation now captures the tool calls a generation makes, with
their arguments, and hands them to the judge. runScenarioLive puts a
scenario to the model in the chat turn's own context (system prompt,
tool-mode and dialog addenda, DATA INVENTORY, EARLIER TABLES, transcript)
over the real tool catalogue, answers each call with a stand-in result,
and reads the reply through the turn's clarification and chip parsers.
runScenarioJudge grades each scenario a few times with the same
evaluateScenario the deterministic suite uses and reports a pass rate per
scenario; the server-only scenarios and checks are left out.
Like the existing judge it runs only with COACH_EVAL_API_KEY, from the
nightly workflow's non-blocking step, and never throws.
A withheld table now leaves a line saying why (module off, or the stored
tables could not be read) and a failed read shows an error row with a
retry, so the Data used count matches what is listed. A table copied from
an earlier answer names that answer's date under its title. The preview
shows the latest twelve periods of a time series, and the truncation note
says the latest rows are kept, which is what the server keeps. The title,
copy and view controls sit above the scrolling table with the title kept
as a hidden caption, and a reply with tables takes the full column width.
The step headers and the related-metric chip read as label: value in all
seven locales, so a metric name no longer sits capitalised mid-sentence.
The "compare with the period before" chip now needs a first reading
before the window's start, the same history check the other history chips
use, and is never offered on an all-time table, which has no period
before it: the table tool reads all time as the current period whatever
it is asked.
Tables are now projected outside the provider's failure path: a
projection defect costs the turn its tables and is annotated, instead of
refunding tokens already billed and dropping the reply.
The at-rest ceiling for a message's tables is applied once, when the
dialog is assembled, before anything streams. A table trimmed only at
write time used to show live and read back as withheld after a reload.
Ten scenarios, each in German and English, describe what a turn must do
with a question: which tool it calls and with which arguments, whether it
may read anything new, whether it asks back, declines a dose, offers
chips only for what the record holds, answers a chip without a provider,
offers to keep looking after a forced answer, and keeps an earlier
table's title out of the model's context. evaluateScenario grades a run
against them.
The deterministic suite drives POST /api/insights/chat with a scripted
provider; the tool loop, executor, table tool, chips, clarification
filter, method line, reply guards and stream run for real. Every check
also runs against a misbehaving model or a broken record and must fail
there.
route-followups.test.ts pins the front door for chips and answered
questions: validation, the rate limit and capability gate on a reuse
turn, the stored chip winning over anything the request adds, stale
chips and answers degrading to a plain message, and the continue line.
A total's weeks and months are now the sum of its days, labelled as a
total, and a level's stay a mean. All time is read as UTC calendar months
from the rollup tier's MONTH fold, or the same fold of the live table when
the tier has no rows, so a month no longer reads as a mean of daily totals
for a short history and a monthly total for a long one. Months are keyed
by their UTC start instead of being re-read in the user's zone, which
named every month one early west of UTC.
The chart's long ranges had the same shift: tier rows are now stamped at
noon UTC of the bucket's first day, which falls inside the bucket's own
week and month in every zone.
Step length and walking speed keep two decimals, loud-sound events count
as a total in whole events, and HRV counts a night measured by both
estimators once and charts the estimator the person mostly has. The model
summary names what a period's value is and gives a total's range total.
The outbound screen matched "Erhöhe auf 2,4 mg" but not "Erhöhe deine
Dosis auf 2,4 mg": the German change verbs had no slot for an object
between the verb and the target, unlike the English "increase your dose
to". The raise and lower patterns now accept one.
Asking for an earlier table "as a chart" went through show_result, which
builds the requested view, but collecting the turn's tables then reset
every chart to null and only derived a new one for fresh tables. The table
came back without its chart. A table shown again now keeps the view it was
given.
When the tool loop forces an answer at its round cap, the reply leads
its chips with "Keep looking". Tapping it runs a normal turn, under the
budget gate as ever, with two server-written context lines: the question
that was left unfinished (at most 1000 characters) and what the forced
turn already read, each table by the name show_result takes, so the
continued turn shows those again instead of fetching them twice.
An answer is continued at most once. The continued reply records the
message it continues in metricSource.continuationOf and offers no
further "Keep looking", even when it is forced again; a chip on a reply
that is itself a continuation is refused.
Up to three chips, derived on the server from what the turn read: the
other view of a table with a chart, the period before, a year earlier,
a wider window, and a metric a discovered pattern links to. A chip is
offered only when its read was done this turn and the record holds the
metric; the year-ago and wider-window chips also need the record's
first reading to reach back that far. The model may name chips in a
FOLLOWUPS block, but only to reorder ones the rules already allow; every
label is rendered from the catalog.
A tapped chip resolves against the chips stored on the conversation's
latest reply. "As a chart" and "as a table" are answered from the
stored table without a model call or a budget reservation: the rows are
copied unchanged and the message is tagged providerType "reuse". Other
chips run a normal turn with a server-written context line. A chip that
is no longer current falls back to a plain message.
A setting turns the chips off (absent means on).
A table over time is drawn as a line for a level metric and as bars
for a total, with at most two series in one unit. Category counts get
bars, horizontal above six categories, the top eight and one other.
Fewer than two or more than 400 rows, or no numeric column, stay table
only. show_result's view now counts: table drops the chart, chart turns
an earlier day table into counts per range with a histogram (fixed
widths for the common metrics, Freedman-Diaconis otherwise, none below
ten values).
The method line merged a repeated read by keeping the larger count. With
show_result that could let an earlier answer's count stand over the read
this turn just made. A fresh read of the same source, window and period
now always wins, whichever order the steps ran in, and the two are never
added. An entry that rests only on an earlier table says so ("from an
earlier answer"), including a reused table no step accounts for.
The new catalog string is in all seven locales.
get_metric_table now gets a step like every other read: the domain is
its validated metric, and once the table is in, the window, period and
granularity are the ones the table was built with (after the executor's
defaults). The count is the readings the table folds, taken from the
summary the server computed, so a read past the sixth table still counts.
A step whose call produced a table now names it (resultRef), for the
projected tables of the older tools as well.
show_result gets its own step. It reads nothing new: while it runs it
names no domain, and once the stored table is found it takes the source
of the read that built it and the name the copy carries in this turn. A
table that could not be shown is a failed step with the reason, never an
empty read.
The property test now drives show_result too and checks resultRef; the
coverage test gives each tool the smallest arguments it accepts instead
of passing empty arguments to a tool that requires a metric.
get_metric_table reads one metric as days, weeks or months through the
chart reader, cut in the user's zone and collapsed to one source per day,
with null for a period without a reading. The model reads a bounded
summary; the full table rides the turn, encrypted at rest. show_result
shows a table an earlier reply of the same conversation fetched, resolved
only against that conversation's own messages. Workouts, labs, sleep and
adherence results project into tables, earlier tables are named in the
context by metadata alone, and result:rN marks are stripped from the reply.
The evidence disclosure under a Coach answer now carries one muted
line: which sources and windows the turn read, how many readings, and
how they were aggregated ("Blood pressure, last 90 days: 142 readings,
weekly averages · Sleep, last 30 days: no readings in this window").
The server builds it from the recorded steps and tables in the request
locale; it holds catalog strings and counts only, never a reading, and
a source that turned up nothing says so.
The tool-mode prompt gains a recheck rule: when the person questions a
figure, name the source and window, read it again (or reuse the stored
table), and correct the answer if the two differ.
Each tool call now emits a running step as it starts and its final
status as it settles: found with a server-counted number of readings,
empty with the reason, or failed. The no-tools path shows one step for
the record summary. Steps carry catalog labels rendered in the request
language, closed-enum domains and windows, and integers only; a lab
analyte, a value or a model-written argument never reaches one. At most
twelve per turn. They persist on the provenance, so a reload shows the
same list.
The conversation window reaches the step so a call without a window
names the one it actually read. The no-tools transcripts gain the
summary step; a new transcript pins found, empty and failed steps ahead
of the first token.
A reply may end with a ---CLARIFY--- block (kind metric, window or
context, 512-byte cap). The server keeps only metric choices the record
holds, drops the question when fewer than two remain, takes window
choices from the presets, screens the question text like any reply, and
never offers a second question right after one. Choice labels come from
the catalog, never from the model.
An answer arrives as clarification { messageId, choiceId? }; it resolves
against the stored question on the conversation's latest message and
adds one server-written line to the next turn's context. The card sits
above the composer, announces politely and takes no focus.
A reply that is only the block, with no question before it, is now an
empty reply instead of falling back to the unstripped text.
Adds every string the dialog needs up front: step labels, table titles,
columns and controls, the method line, the follow-up chips and their
setting, the clarification card, and the caption of an answer reused from
a stored table. A key catalog shared by server and client names each one;
a test holds the catalog and the bundles equal in both directions.
GET /api/insights/chat/{id}/messages/{messageId}/results returns the
tables one assistant message read, narrowed to the conversation's owner
like the conversation read, with a table withheld as module_disabled when
its domain's module is off and as unavailable when it cannot be read. A
data read: no AI capability.
The published document now describes every stream frame as the
CoachStreamEvent union, including the suggestedAction frame it never
named, the new step, result, followUps and clarification frames, their
components, the provenance fields, the followUp and clarification request
fields, and the reuse provider type.
Each later piece of the dialog gets its own module, answering nothing for
now: the step builder, the table projection and chart choice, the result
ref marks, the follow-up derivation, sentinel, resolution and continue
chip, the reuse turn, the clarification parse and answer, the method line,
and the four prompt addenda. The turn pipeline calls them where they
belong, so the frames go out in the documented order (step, token,
provenance, result, cards, clarification, followUps, done) and a blocked
turn drops its tables, chips and clarification. With every piece empty
the stream and the prompt stay byte for byte as before.
The retrieval loop takes onCallStart and onCallSettled callbacks, fired
for every call with a turn-wide index; a throwing callback is swallowed.
The trace carries each call's schema-validated arguments (never the raw
string, never persisted), the result says whether the answer was forced
at the round cap, and a seedCalls parameter is accepted but not read yet.
A new nullable column, coach_messages.results_encrypted, holds the tables
of values a turn read as AES-GCM ciphertext; only their metadata rides the
plaintext provenance. readMessageResults serves them to the owner, each
table whole or withheld with a reason. The provenance restore learns the
steps, method, table metadata, chips, clarification and the forced-answer
marker, entry by entry against the wire schema.
The column is registered for key rotation, carried through both backup
formats (ciphertext in a disaster-recovery file, readable JSON in a
portable one) and goes with its row on conversation delete and wipe.
Adds the TypeScript types for the four additive stream frames and their
provenance fields, a Zod mirror for the published document held equal to
the types by a type test, and a guard that no newer frame reuses a key the
shipped native client decodes. The chat request accepts followUp and
clarification, inert for now.
The route keeps the front door: auth, the coach capability, the body
read and schema, the rate limit, the locale and the inbound refusal
screen. Everything after it moves into one module per step:
conversation, context, chain, budget, model, reply-guards, cards,
provenance, persist, sse, errors and pipeline. The capability re-check
at the egress site stays owned by the route and is handed to the
pipeline, so the capability inventory still reads both calls there.
The stream producer now writes through a TurnEmitter created inside
createSseStream and passed to the model step, which is where live
progress frames will come from. Nothing writes to it early yet.
No behaviour change: the golden transcripts pass byte-identical, the
existing route tests pass unmodified, and every annotate, audit and
error-code name appears as often as before.
Guards follow the code. The fence call-site check reads turn/context.ts,
the German-default locale ban covers every file in turn/, and the
fenced module-graph test now asserts that turn/sse.ts and turn/errors.ts
reach neither the tool registry nor the snapshot builder. That test
named lib/ai/coach/tools.ts, which stopped existing when the registry
became a directory, so its negative check could not fail; it points at
tools/index.ts now, and a new sanity check fails if the tool route can
no longer reach the path it names.
Making `=true` grant nothing would stop a single-user self-host's local
model on a patch update. It is deprecated for one release cycle instead,
in a narrowed and safe form: it still grants a private host, but only to
the operator's own configurations (the instance-wide admin provider and
AI settings saved on an admin account), and only through the pinned
operator-approved dispatcher, so redirects stay forbidden and metadata,
link-local and unspecified addresses stay refused. A non-admin account's
own base URL needs an exact origin in AI_PRIVATE_ORIGINS.
Provider resolution passes the owning account's role to the client
(`operatorTrusted`); the save route and the connection test apply the
same rule to the account being configured. The boot summary now says the
setting is deprecated and names the replacement, and the admin AI
settings page shows the exact AI_PRIVATE_ORIGINS line for the private
origins saved on the instance (every account's Local and gateway URL and
the admin base URL, scheme, host and port only), so the operator can copy
it and drop the old setting.
The list of provider tags that send data out through a credential the
operator holds (admin-openai, admin-codex) existed four times: in the
consent gate, in the capability resolver, in the shared-provider health
read and in the admin provider-health ordering. A tag added to one and not
the others would be consented for in one place and not the other. The list
and the local-provider test now live in src/lib/ai/provider-egress.ts and
every reader imports them.
The guard meant to keep one copy matched a variable name, so the copy in
the resolver, under another name, never tripped it. It now matches the
list itself in either order and names the one file allowed to hold it.
Marc-André Bombeck