The Coach may point at a published /learn guide, but only a slug in the
catalog. That was a prompt instruction with no enforcement — a model could
hallucinate /learn/lower-your-cortisol and ship a dead link. Add a deterministic
post-filter on the assembled reply that keeps a published slug verbatim and
strips any reference whose slug is not in the catalog, tidying the prose around
the removed link. Makes the catalog's impossible-by-construction claim an
enforced guarantee.
The illness-recovery tool and the illness context block carried only labels,
lifecycle, and dates plus the recovery composite — never the recovery-gap, the
metric that drove it, the nadir, the pre-onset deviations, or the red flags the
illness card renders. So a recovery question got the composite, not the gap the
user sees, and a sustained-fever or low-SpO2 escalation was invisible
in-conversation.
Run the existing episode-correlation read-layer (the same engine the card and
the red-flag notifier use) for the most relevant episode — active first, else
the most-recently-resolved — and attach the computed scores to get_illness_recovery,
coverage-gated so a thin signal yields nothing rather than a fabricated number.
Document the illnessScores shape in the Coach prompt so it restates those
numbers verbatim and escalates a red flag rather than reassuring.
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 tool-mode path dropped the launch scope's sources: the inventory probes the
full source set, so a Coach opened from a metric page saw the whole inventory
and could roam, while only the no-tools path honoured the narrowing. Inject a
one-line FOCUS hint naming the launched domain(s) into the tool-mode user turn
so the model prioritises that metric and fetches its figures first, with every
other domain still reachable if the question leads there. Empty on a generic
open, so the prompt prefix is unchanged on that path.
The acute red-flag clause hardcoded its blood-pressure numbers as prose
literals and named no glucose or fever floor at all, so the Coach could
silently drift from the dashboard hero and the notification engine. Compose
the systolic/diastolic, glucose, and sustained-fever thresholds straight from
clinical-floors.ts and echo the previously-missing glucose and fever lines, so
every acute number the Coach states is bound to the one source of truth.
Add an outlook contract (gentle forecast, what-to-expect, anticipatory if-then
— all conditional, ranged, association-framed) composed beside the tone
contract on the Coach and the comprehensive briefing, so a reply can sharpen
expectations without a false promise. Tighten the tone contract against
over-validation (affirm the effort, not an unsafe choice; stay neutral when
there is nothing to praise) and route why/conflict questions through
get_correlations and its coincident-deviation flag before answering.
Coverage asserts the clause numbers equal the constants and the new contract
reaches both surfaces.
The FDR cross-metric driver pairs were reachable only through the
get_correlations tool, so a local/no-tools provider got the coincident
flag but zero discovered drivers and the cross-metric layer silently
degraded. The system prompt's fallback to a snapshot driver field
pointed at a field that did not exist on that path.
Attach a bounded top-3 slice of the same gated, ranked driver output the
tool serves to the no-tools snapshot floor, gated on the derived layer
and shed first under prompt-budget pressure. Local users now reach
parity on the flagship cross-metric capability.
The coincident-deviation read grouped the latest day with a UTC date
slice while the sibling readiness read already used the tz-aware day key.
For a user east or west of UTC, a late-evening or early-morning reading
landed on the wrong calendar day, so a fired flag could compare a vital
from the wrong day against its band and narrate '2 or more vitals out of
band today' on the wrong day.
Thread the user's timezone into the latest-day read and mint the day key
with userDayKey, matching readiness. The Coach derived-snapshot and the
correlations reader pass the real account timezone through. Adds a
regression covering a UTC+2 user whose late-night reading rolls to the
next local day.
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 multi-vital coincident-deviation flag (two or more vitals outside
their usual band on the same day, with the illness-explained reframe)
already reached the period narrative but never the Coach. Attach it,
fired-only, to the derived block so both Coach paths can narrate it: the
tool path via the recovery/correlations tools and the no-tools snapshot
floor. A quiet day adds no entry, keeping the snapshot noise-free.
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.
The retrieval-tool DATA INVENTORY was built against the user's default
narration cluster (cardio/body/mood/medication), so sleep, glucose,
activity, body composition, vascular, SpO2, gait and environment series
reported absent even when the data existed — and the grounding rule then
told the model not to fetch them. Probe presence against the full source
set instead, and generate the metric-series inventory rows from the
complete source-to-section map so every series the user has rows for is
advertised. Per-tool reads still re-scope to their own domain, so the
wider probe never widens a figure read.
Add three retrieval tools whose snapshot blocks already existed but were
unreachable in tool mode:
- get_workouts: recent sessions + per-sport rollup
- get_cycle: phase / prediction / correlation (gated by cycle access)
- get_correlations: the FDR-controlled lagged cross-metric drivers plus
the coincident-deviation flag, surfaced descriptively
Also honour the caller's window in get_labs and get_illness_recovery so a
cross-metric answer no longer silently mixes horizons.
The retrospective narrative was the one model surface that omitted the
shared tone contract, so it read colder than the daily briefing beside it.
Compose the tone fragment into both prompts and warm the hand-written tone
line to name a genuine win when the period earns it, while keeping the
descriptive-never-causal and no-alarm guards intact. The cross-surface
coverage test now asserts the contract reaches this surface too.
Sharpen the Coach's behaviour-change reflexes from the coaching literature,
prompt-only:
- Connect the signals on a why/pattern question — consult correlations and
link them descriptively ("after short-sleep nights your next-morning HRV
tends to read lower") instead of reading metrics in isolation, always an
association worth a small experiment, never a cause.
- Confidence ruler on action turns — after naming one small step, ask how
doable it feels 0-10 and shrink it when that lands low, keeping the choice
the user's.
- Three-beat shape on data-review turns — finding, likely driver, one small
step; name the dominant contributor when citing a derived band.
- Extend the MI micro-moves with developing discrepancy and rolling with
resistance, and add an anti-persuasion check: any suggested change must
serve the user's own stated goal.
One bilingual tone-calibration example added per locale.
The shared contracts cover chronic deferral (dose, diagnosis, drug-level
to a clinician) but had no acute branch. Add a closed-list crisis clause —
chest pain, syncope, sudden severe symptoms, hypertensive-crisis readings,
suicidal ideation — that points to prompt or emergency care without
diagnosing, and keep every other turn calm and non-alarmist. It rides the
single shared-contract source so it lands on the Coach, the per-metric
status cards, the comprehensive briefing, and the period narrative; the
coverage test asserts it on all four surfaces.
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.
The four AI clients threw bare Errors on an empty reply (and, for Codex, on a
stream-level failure) with no httpStatus, so the provider-chain classifier saw
the same null-status shape as a genuine transport failure (ECONNRESET, DNS,
timeout) and could not tell them apart in observability. Attach a sentinel
`httpStatus: 0` plus a `kind` discriminator (`empty_response` /
`stream_failed`) on those throws. The cascade is unchanged:
`isHardProviderFailure` already treats a status <= 0 as a hard failure and
walks to the next provider — this only sharpens the logs.
The Coach SSE turn ran the provider call to completion before the stream
existed and never wired it to the request's abort signal, so a mid-generation
client disconnect still paid the full upstream token cost. Thread an optional
`signal` through CompletionParams and singleUserTurn into each client's
safeFetch (composed with the per-call timeout via AbortSignal.any), and pass
`request.signal` from the chat route's no-tools path. An early disconnect now
tears the upstream request down instead of generating into a closed
connection. Omitting the signal leaves behaviour unchanged (timeout-only).
Drop the unreachable text-only-model reason from the OCR capability
enum and its DTO type. Replace the literal NUL delimiter in the Fitbit
dedup map keys with a named constant so text tooling stops treating the
file as binary (behaviour unchanged). Correct the Coach tool-executor
docblock, which overstated cross-tool snapshot-cache sharing.
Text-mode OCR structures browser-extracted text rather than running a
vision call, so it reserved and reconciled the full vision budget even
though the spend is far lower. Add an ocrExtractText ceiling and reserve
that in text mode, and refund the reservation in full on a clean extract
failure instead of charging the reserved estimate.
The o1/o3/o4 prefix match also caught o1-mini, o1-preview and o3-mini,
which have no image input. A user pinned to one was told OCR works, then
the provider rejected the image block. Treat an o1/o3/o4 slug as
vision-capable only when it is not a -mini or -preview variant, applied
to both the OpenAI and Codex gates.
The F1 coach tools each rebuild a snapshot with a distinct single-source
scope, so the 60s snapshot LRU (keyed on the source list) never shares their
reads. On a multi-metric turn the model can fan out up to six tools that each
fire `extractFeatures` (a user lookup plus a windowed measurement read plus the
all-time extremes) and a prefs `findUnique` concurrently — up to twelve heavy
reads against the shared Prisma pool at once.
Memoise both reads on the request-scoped event cache, keyed on userId (and
windowDays for the features read), so the fan-out collapses to one of each
within a request while staying correct across requests. Project the feature
extractor's measurement read to the six columns its consumers actually use
(type, value, measuredAt, sleepStage, source, deviceType) so the 200k-row-class
PULSE / glucose windows stop pulling every column. Outputs are unchanged.
The Anthropic client derived `finishReason` solely from `stop_reason`. The
Coach tool loop continues only when `finishReason === "tool_calls"`, so if
Anthropic ever returned tool_use blocks under a different (or absent)
stop_reason, the reason resolved to undefined, the loop dropped the tool
request, and the empty tool_use-only reply surfaced as the final answer —
the requested data was never fetched.
Set `finishReason` to "tool_calls" whenever tool_use blocks are present,
clamped ahead of the stop_reason map, so the presence of a tool request is
authoritative regardless of the reported stop reason.
The OpenAI client pinned `response_format: { type: "json_object" }` on every
call. The Coach tool loop's forced-final round runs with `tool_choice: "none"`
and expects plain prose, but JSON mode coerced the reply into a JSON object —
the user saw a JSON blob or an empty reply instead of the coach's answer.
Send `response_format` only when the caller opts in with `responseFormat: "json"`
and no tools are present, mirroring the Anthropic client's prefill gate. The
JSON-consuming callers that relied on the implicit mode (insight generate +
briefing retry, medication extraction, the connection-test probe) now pass the
flag explicitly; the prose Coach path and every tool round stay out of JSON
mode. Codex already omitted the flag and is unaffected.
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.
The Coach snapshot builder returns the structured, post-degrade record it
serialises to JSON alongside the JSON string, so a single domain block can be
read out without re-parsing. Add an optional tool-call trace to the provenance
envelope (which retrieval tools ran + whether each found data) and mirror it in
the OpenAPI provenance schema.
Groundwork for on-demand retrieval: the retrieval tools slice one section out
of this record, reusing the builder's gates and I/O verbatim.
Replace the CompletionParams { systemPrompt, userPrompt } string pair with a
structured { system, messages[] } request across all five AI clients. One
refactor unlocks three things: provider prompt-caching (Anthropic cache_control
on the stable system prefix, OpenAI/Codex automatic prefix cache, with
cachedInputTokens surfaced for observability), the tool-calling wire surface,
and clean multi-turn.
Each client maps the new shape to its own wire and degrades unsupported
features silently: the local OpenAI-compatible server never receives a tools
field (most reject unknown keys), and prompt-caching is best-effort everywhere.
CompletionResult gains toolCalls, finishReason and cachedInputTokens; the tool
types and per-provider mapping ship now, the tool-use loop does not.
Vision inputs fold into the user turn's content parts via singleUserTurn, with
each provider's pre-existing block ordering preserved. The corrective-retry
paths append to the last user message rather than a flat string. The Coach
keeps assembling one user turn so its snapshot-once and grounding behaviour is
unchanged. No wire schema visible to API consumers changed.
Behaviour is unchanged across every surface; the full suite stays green.
The 0–14d raw band folded same-day readings on a UTC calendar key, so a
late-evening reading slipped onto the wrong day for non-UTC users and
across DST transitions. Key it on userDayKey(tz) instead and make the
timezone a required option so a caller cannot silently fall back to UTC.
The coarser DAY/WEEK/MONTH/YEAR bands stay UTC-anchored on purpose: the
rollup writer and the live aggregator both bucket on the UTC calendar
day and are parity-locked for the read-swap fallback. Document the split
and thread the resolved user timezone from the coach snapshot caller.
The codex vision gate accepted any gpt-5 prefix, which included the gpt-5-codex
specialist slugs that the surrounding contract documents as text-only. Each lab
OCR upload on such a slug burned a doomed image attempt before falling back.
Exclude the -codex slugs from the gate so multimodal gpt-5.x slugs still report
vision while the codex specialists stay on the text path.
Add a labs block to the coach snapshot so the assistant can answer
questions about the user's lab results. Each entry is the most recent
reading per biomarker over the last 12 months (capped), carrying the
resolved name, value, unit, reference range, in/out-of-range verdict,
and measured date. Qualitative readings (e.g. negativ) are included
with a neutral range status. Values resolve server-side via the same
helpers the labs API uses, and the system prompt instructs the model to
quote them verbatim and never invent a lab number, so the block stays
grounded. The free-text qualitative result is sanitised before it can
reach the prompt; the encrypted note is never read.
Image input rides the ChatGPT-OAuth codex/responses endpoint via the
input_image content block (protocol spec 2b) — no vision API key. The
codex client folds params.images into input_image blocks alongside the
input_text; text-only calls keep the prior byte-identical wire. The
vision gate now resolves codex capability against the working slug
(GPT-5.x / gpt-4o-class multimodal) instead of a blanket false, with the
OCR capability resolver feeding the codex slug rather than the user's
aiModel. A non-multimodal slug or an image-block rejection degrades to a
clean extract failure / local-OCR text mode.
The safety-contract matrix parity test checks rule bodies are non-empty;
this asserts the resolved getCoachSystemPrompt(locale) string carries the
grounding marker and the GLP-1 dose-safety contract for all six UI
locales, closing the gap where a non-EN/DE locale could silently fall
back to a weaker EN-plus-footer body.
BP, pulse, weight, BMI, mood and medication-compliance still asked the
model to derive the recent-vs-baseline comparison from raw buckets — the
exact arithmetic small models do unreliably, and the pattern the
base-system prompt already advertises as precomputed.
Build the SIGNAL block (current / baseline / signed delta / normal-swing
verdict) with buildMetricSignal from each card's graded series and inject
it into the snapshot; BP carries two signals, compliance one per
medication. Each card's prompt section now leads from the signal block
and is told to state, not recompute. The server-computed compliance %
is unchanged.
The single multi-type window read had no take bound, so a lastYear /
allTime scope over a high-frequency type (PULSE / glucose are
200k-row-class) loaded a year of rows just to fold them into the ~21
daily + ~10 weekly buckets the prompt shows. Read newest-first with a
6000-row cap and restore ascending order; the coarse MONTH/YEAR tail
still comes from the rollup tier, so deep history survives.
detectRefusal guards only the inbound user message; the model's reply
was streamed verbatim. A dose-prescription or a fabricated clinical risk
score that slipped past the system-prompt contracts reached the user
unfiltered.
Add a deterministic outbound screen on the assembled reply (which the
route already buffers before streaming): tight dose-change patterns plus
risk-score patterns, EN/DE. On a trip the turn is replaced with a calm,
grounded fallback and any reminder suggestion / key-value provenance from
the discarded reply is dropped.
The daily-briefing block (paragraph / signalsOfDay / keyFindings) was
schema-shaped but never number-checked the way recommendations[] are: a
restated figure could drift from the snapshot with no server rejection.
Add a grounding cross-check that asserts every number the briefing prose
states traces to a features.signalsOfDay figure (latest / avg / delta)
within a rounding tolerance. Wire it into both generation sites with one
corrective retry, then strip the briefing rather than persist a
fabricated number. Also gate the higher-temperature paragraph reroll on
the same check so a phrasing reroll cannot drift a figure.
The vision allowlist matched only the 3.5/3.7/4 families, so a
self-hoster pinned to claude-3-opus / claude-3-sonnet / claude-3-haiku
was wrongly blocked from lab OCR though those models read images. Add
the bare claude-3-{opus,sonnet,haiku} prefixes.
Extend CompletionParams with optional images[] and documents[] so the
provider clients can carry an uploaded lab-report photo or PDF. Anthropic
folds image and native document blocks into the user turn; OpenAI and the
local OpenAI-compatible client fold image_url blocks. Text-only callers are
untouched. Add supportsVisionForConfig / supportsPdfForProvider and an
ocrExtract budget for the Lab-OCR ingestion path.
Carry an additive usage envelope on the SSE done frame
({ totalTokens, model }) and persist tokensUsed + model onto
coach_messages (migration 0182, nullable, no backfill) so the
per-message token footer survives a conversation reload. Add an
optional reasoning frame the client can render inside the thinking
disclosure when a reasoning-capable provider emits a summary; providers
without reasoning emit none. Extend CoachMessageDTO and regenerate the
OpenAPI contract. Server-authoritative: clients display the count, never
recompute it.
One-shot prettier pass so the new format:check gate starts from a clean
baseline. Formatting only — no behaviour change. Generator-owned and
vendored artifacts stay untouched via .prettierignore.
Remove the unused output-key reverse map the batch path never read, and
have the budget table satisfy the AiBudget interface directly so the type
is enforced rather than orphaned.
Add generateStatusBatchForUser: it prepares the seven specialised status
snapshots once, sends one prompt for a { perMetric } envelope covering only
the metrics with data, and fans each summary back into the same per-metric
cache rows the standalone generators write. One JSON-correction retry guards
the parse; any omitted metric or a failed batch degrades to that card's own
single-call path, so a partial response never goes dark. The nightly warm and
the forced warm route through the batch in production while the injected
per-card seam keeps the cron tests intact. Adds the status-batch prompt
scaffold composed from the shared assessment contract.
Introduce a prepare -> run -> finalize contract for the per-metric status
cards so the snapshot build and the provider call can be driven separately.
The public generator keeps its signature and one-call-per-metric behaviour;
weight-status now resolves through the shared single-card path. Add the
status-batch AI budget entry sized for one combined seven-metric call.
`ALLOW_LOCAL_AI_PRIVATE_HOSTS` was a binary flag that, when `true`,
opened every private/internal host at once — including the cloud
metadata endpoint and any internal panel. Replace the binary check at
all three sites (write-time validation, local client, test resolver)
with a shared `isLocalAiHostAllowed` helper: a comma-separated host list
permits only those exact hostnames, while the legacy `true` still
permits any private host for backward compatibility. The safeFetch /
isPublicUrl floor is untouched; the allowlist only widens what the
operator deliberately permits. Docs and the SSRF-pin inventory test are
updated accordingly.
The daily token cap was read before the provider call and recorded only
after a successful reply, so concurrent requests could all read
`spent < cap` and all hit the provider, and an empty/sentinel reply
recorded zero spend despite burning upstream tokens.
Reserve the budget atomically before the call: a single upsert-increment
returns the new total, so concurrent requests serialise on the row and
cannot both pass the cap; over-cap refunds the reservation and surfaces
`coach.budget.exceeded` as an SSE error frame. Reconcile the reservation
to the actual token count immediately after the provider returns —
before any empty/sentinel short-circuit — so burned tokens are always
recorded. An integration test proves two concurrent requests cannot both
exceed the cap and that an empty reply still records its cost.
The per-worker in-process "last working provider" hint stays volatile
and is deliberately NOT folded into the durable Postgres health ledger.
The ledger already shares the load-bearing skip signal across workers;
the only uncoordinated piece is a soft reorder whose worst case is one
redundant first-hop per worker per hour. Promoting it to a per-request
ledger write would add hot-path write amplification for a negligible
saving. Documented inline so the trade-off is explicit.
Marc-André Bombeck