The opener hint, the first-name clause, resolved follow-ups, the table rules
and the final-round line used to be written into the system prompt, so the
cached prefix (tools plus system) was rewritten on every turn and again on
the forced final round. They travel on the user side now: the per-turn lines
in the user message, the table rules as a note after the round that produced
the first table, the final-round reason as the last user line. The system
prompt depends on the locale, the Coach preferences and the about-me text
only.
Refs #1217
A catch that drops what it caught is where the silent nightly failures
started. Every `catch` and `.catch()` under src/lib/jobs now rethrows,
fails the job, writes to a ledger an operator reads, counts into the run's
failure facts, hands the error back to its caller, or logs at warn under a
stable action. The enqueue helpers that swallowed a failed send (briefing,
status, narrative, morning digest, document index / summary / thumbnail,
workout insight, boot-time sweeps), the Coach nudge's provider fallback,
the workout and reaction-line cleanups, the advisory-lock release, the
backup attempt bookkeeping and the cycle reminder's dedupe reads now leave
a warn line instead of nothing or an info annotation.
`error-level-discipline-guard` holds the rule. The catches that stay
silent are listed with their reason (the pgboss readers' documented null on
a web-only deployment, the key-ring trial, the restore drill's per-account
result, an already-deleted Telegram message); a new silent catch fails, an
entry that no longer matches fails, and each entry is shown to be
load-bearing by removing it.
A Coach turn cut short by its token, time or round budget writes
`coach.turn.budget_stop` at warn.
Own-plan turns get 165k tokens, 200 s and 16 rounds, operator turns 55k,
120 s and 8; the final-answer reserve, the concurrent-turn slot and the
step cap move with them.
Codex requests carry a per-conversation prompt_cache_key and a stable
session id, api.openai.com gets the same key, Anthropic usage counts
cache reads and writes in the gross, and local servers report cached
prompt tokens, so the turn budget discounts cached input on every
provider.
The Coach loop's final round sent toolChoice "none" without the tool
definitions. With tool_use/tool_result blocks in the history Anthropic
refuses such a request (400), and OpenAI Chat Completions refuses
tool_choice without tools, so every budget, time, cap and no-progress stop
ended in an error on those wires. The final round now carries the tools
with toolChoice "none", which every wire accepts, thinking included.
Codex kept reasoning items as provider state even when the request asked
for no encrypted reasoning; replaying such an item under store:false is a
400. Items are kept only when the request included
reasoning.encrypted_content and the item carries it.
The reasoning dialect cache learned from any 4xx that mentioned
reasoning or thinking (auth, rate limits, model names such as
deepseek-reasoner) and from replay errors, and kept the downgrade for the
life of the process. Learning now requires an HTTP 400 that refuses a
sent parameter, ignores replay errors and the model name, and a learned
downgrade expires after an hour.
Build the memory block for the providers the turn sends to, send it on
both paths, keep the facts it recalled in the owner's trail, and offer a
health fact the background found as the answer's one note. Every turn
enqueues the memory refresh; the job's debounce and singleton key decide
when it runs. A remembered fact points back to the stored user message.
Unit tests for the turn budget and the no-progress brake, the budgeted
loop (every stop reason, the round-by-round ledger, repeated calls,
questions, comparisons, reasoning and provider state), the trail recorder
and its screen, clarifications through the tool and the brake, the
comparison table and chart, the trail column and its owner-only read, the
reply frames, and the turn's frames end to end through the route. An
integration test proves the trail against a real database. Eval
scenarios cover a multi-round why question, a period comparison with its
chart, a question asked through the tool, and keep the charts the
existing table scenarios already showed.
Take the effective reasoning level from the route rather than resolving
it in the turn; the turn only caps it when the operator pays.
Replace the fixed three tool rounds with a budget fixed by who pays:
120k tokens, 150 s and 12 rounds on the person's own plan, key or local
model, 40k, 90 s and 6 rounds when the operator pays. Cached input counts
at a tenth. The final answer is reserved from the start, so a turn whose
budget, clock or round cap runs out answers from what it read. A
no-progress brake stops a turn that repeats calls or reads nothing new,
and a repeated call is never run again.
The day's ledger is booked round by round: round one and the final
reserve when the turn starts, each further round right before it runs,
every round settled as it returns, the unused reserve given back at the
end. A round the daily ceiling refuses ends the turn with an answer.
Every phase goes out live as an activity frame (thinking, fetching,
summarising, asking, remembering, planning, the stop, the answer), with
screened reasoning titles and summaries; the structure rides the
provenance and the text is sealed in trail_encrypted, served by
GET /api/insights/chat/{id}/messages/{messageId}/trail. Tables go out as
interim results as soon as they are read.
Add ask_clarification with a server brake, remember_fact and
propose_plan over the memory contract, model-free decision turns,
compare_series with the compare chart, eight tables per message,
assumptions with change-assumption chips, a reasoning level per turn and
at most two concurrent turns per person.
The saved Coach window only seeded the snapshot and each tool's default.
A tool could still read back to all time, the availability probe reported
counts and means for history outside the window, and several blocks read
fixed windows of their own (labs 12 months, correlations 180 days, the
coarse tail, the all-time extremes) whatever the setting said.
coachPrefs.defaultWindow is now enforced as a limit through one module,
history-reach.ts: the turn context clamps the client window, the snapshot
clamps its window and gives every fixed-window block either the shorter
window or, when the block's own window cannot fit, leaves it out. Every
retrieval tool clamps its window, an earlier period or a stored table that
reaches past the limit answers outside_reach, and the probe reads only
inside the limit and names older history as outside_reach without any
figure. Follow-up chips that would read past the limit are not offered.
The default (allTime) is no limit and reads as before. MCP clients pass no
limit and keep their own window arguments.
Adds the twelve-month option, rewrites the picker text to state what is
read with the numbers the readers use, shows the same options in the
managed-profile settings, counts cluster members with a singular, and
drops the opacity on switched-off rows that put their text below WCAG AA.
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).
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 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.
The owner split left an operator-funded chain with no ceiling at all. Its arm
compared operator_tokens, and that counter returns to roughly zero on every
reconcile the user's own plan settled: on the 2026-09-11 evidence the shared
key answered 500 and codex served, so the counter never grew, every job
reservation was admitted forever, and a runaway client loop could write an
unbounded row. A pre-existing concurrency contract caught it from the other
side, two reservations that must not both pass one cap both passing.
reserveBudget now refuses when the owner's counter is over its cap OR the day's
mixed total is over the abuse ceiling, and a background surface gets its share
of that ceiling too, so background spend is bounded on the total and not only
under a counter the operator's key never fills. The surface is a required
argument, so the type checker enumerated every reservation site.
The hop guard learns the same lesson: an operator-funded fallback hop is now
rationed at the calling surface's ceiling rather than the full operator cap. A
job admitted under the job share could previously spend the whole 200 000 once
its chain fell back onto the shared key, which is the original defect reached
by the fallback path.
The job share applies only where the operator pays for the chain. Multiplying
whichever ceiling the chain resolved halved background generation for a
self-hoster on a local model, who costs the operator nothing.
Refusals now name the ceiling they hit and whose it was. A refusal used to emit
the day's mixed total and nothing else, which on an operator refusal is not the
counter that tripped.
A Coach retrieval tool that read the window and found nothing returned
`reason: "no_data"` — the same code for two situations that have nothing in
common: the record holds nothing for that metric, and the record holds a
substantial history for it, all of it older than the window that was searched.
Nothing downstream could tell them apart, so the Coach said the honest-sounding
sentence for the wrong case, and then, having been told there was nothing there,
changed the subject. Someone with a year of imported glucose readings was told
the readings did not exist, on the same account whose glucose page was showing
them.
The window cap stays. It exists because a multi-year history does not fit a
prompt budget, and widening it would only move the same failure. What was
missing is the ability to say a different true sentence.
- `tools/availability.ts` — one grouped aggregate per miss, no rows
materialised, resolving four states: `no_data` (probe confirmed empty),
`outside_window` (rows exist, older than the searched window; carries the
count, the date range in the user's own timezone, and a bounded per-series
mean/min/max), `unavailable_in_scope` (rows exist inside the window and the
domain still produced none — a module is off), and `no_data_unconfirmed` (the
probe itself failed; absence was never established, so it is not asserted).
- Every empty read in the executor routes through it. A bare `no_data` can no
longer be returned without the record having been checked.
- The DATA INVENTORY carries the same distinction, so a domain the window
cannot reach is advertised with what it holds instead of as absent — the model
no longer skips the fetch, and a vague follow-up keeps the subject.
- Both instructions that told the model to report absence and move on now
branch on the reason code, and forbid answering about one metric by reporting
another.
- The out-of-window aggregate joins the authoritative set for the prose
verifier, so the sentence the fix exists to make sayable is not elided.
- A cross-cutting narration block that throws is now counted rather than
vanishing into a `null` that looks like honest absence.
- The medication-compliance page tells a failed read apart from an account with
no medications, instead of inviting someone to re-enter what it already holds.
Refs #648
Add three per-user AI settings, all configured in the app (no new env vars),
behind one migration (0198):
- Response timeout (seconds). A per-user upstream timeout threaded onto the
provider call. A local/self-hosted backend can take more than the legacy
60 s default to load a model on the first request, which timed the Coach
out before the first token. Defaults to a generous ~180 s when unset;
validated to 10–600 s.
- True token streaming for the local OpenAI-compatible client. The client now
POSTs with stream:true and parses the SSE delta chunks, emitting tokens as
they arrive, with a per-idle-gap timeout instead of a whole-call ceiling and
a graceful fallback to the buffered path when a server cannot stream. The
Coach route runs the provider call inside the response stream behind a
keepalive heartbeat so a reverse proxy no longer idle-drops the connection
during model load or generation. Grounding, refusal, and sentinel-stripping
still run on the complete reply before any token reaches the client.
- Dedicated document-scan (Lab-OCR) provider. Off by default — the OCR path
uses the main provider chain. When enabled, lab-report processing resolves a
separate provider/model/key, with a hint when the chosen model is not
vision-capable. The new key column is AES-256-GCM at rest and registered in
the encrypted-columns registry and the rotation script.
i18n: EN + DE translations for the new Settings → AI strings (other locales
carry interim copies to keep locale parity green).
Each retrieval tool re-scoped the snapshot to a single source, so its cache key
never matched the inventory's full-source build — a four-tool turn paid four
extra full snapshot builds on top of the inventory's. Thread the inventory's
probe scope to every tool so a per-tool read lands the one cache entry the
inventory already primed, collapsing the turn to a single build (a window
override still gets its own correct build). Raise the loop's round cap from two
to three so a sequential cross-metric why-chain isn't starved; the per-round
token budget is unchanged and the absolute ceiling tracks one above so the
final round still forces prose.
The daily AI-token gate was a flat 25,000/day for every provider. That cap
exists to bound the OPERATOR's API bill, but a ChatGPT-OAuth (Codex) or BYOK
chain egresses on the USER's own plan/key and costs the operator nothing, so
gating it on the operator-cost ceiling is wrong. A single gpt-5.x reasoning
turn legitimately reports 20k-40k gross total_tokens (re-sent system prompt +
inventory + tool defs per round + hidden reasoning, summed across rounds), so
one or two turns exhausted the flat cap and locked the user out of a plan they
pay for.
Classify the chain's cost owner by its primary provider via resolveDailyCap:
an admin-openai primary (operator pays) keeps the operator-cost cap; every
user-egress primary (codex/openai/anthropic/local) gets a generous
abuse-only ceiling. Threaded into the Coach chat gate and both OCR-extract
modes.
Raise the operator-cost cap 25,000 -> 200,000 so the operator-key path also
survives a normal day of reasoning turns, keeping gross-token accounting
rather than reworking the reconcile math.
Subtract cached input tokens at reconcile: the Responses-API gross total still
includes prompt-cached input the user did not re-pay for. The codex client
already parsed cached_tokens; sum it through the tool loop and bill
total_tokens minus cached.
The Daily Briefing strips any number absent from its server-computed
figures; the Coach's tool path had no equivalent, so a transcription or
paraphrase drift (a tool returns systolic 128, the reply says ~138) could
ship. Collect the numeric leaves from every present tool result this turn,
extract the numbers the reply asserts, and soft-correct any that match no
fetched figure to a neutral placeholder — annotated, non-blocking, and a
no-op on a qualitative turn or the no-tools path where there is nothing to
grade against. The bounded retrieval loop now returns the present results'
payloads so the check has an authoritative figure set.
Thread the request abort signal through runCoachToolLoop into each
per-round completion so a mid-generation client disconnect tears down
the upstream call instead of paying the full token cost. Completes the
no-tools-path wiring to cover the tool-retrieval path as well.
Regenerate the OpenAPI spec for the trimmed OCR-capability reason enum.
Replace snapshot-stuffing in the Coach turn with a small base context plus
on-demand retrieval. The model receives the full system prompt, a tool-mode
grounding addendum, and a tiny DATA INVENTORY manifest listing which domains the
user has logged and which tool fetches each; the figures themselves are pulled
only when a question needs them.
Six read-only tools wrap the existing server-authoritative snapshot builder:
get_metric_series (blood pressure / weight / pulse + the additive series),
get_glucose_panel, get_sleep, get_medication_compliance, get_labs, and
get_illness_recovery. Each returns a structured { present: false } for an absent
or opted-out domain rather than throwing or returning an ambiguous empty list,
so the model can always tell missing data from real data. Module and cycle gates
are enforced at the builder, so opted-out domains are unfetchable.
A bounded loop runs the call -> tool-result -> answer cycle: at most two rounds
with a hard cap of three that forces a final answer, parallel tool execution,
and only the final round streamed. The day's budget reserves the per-call
ceiling times the round count and reconciles the summed tokens; the atomic
reserve and reconcile primitives are unchanged. All inbound gates (rate limit,
refusal, replay scan, consent, budget) stay before the loop.
The legacy snapshot path stays alive as the no-tools fallback: tool mode is on
only when every provider in the chain supports tool-calling, so a chain that
includes a local model falls back to the prior behaviour verbatim. The tool
trace is persisted onto the assistant message provenance.
Marc-André Bombeck