Status cards and derived-score sheets read one way: they state a verdict and
end there. The metric pages already carried the hand-off; the score sheets
passed neither an opener nor a scope, so they painted no action at all.
- Score sheets (sleep / readiness / recovery / stress / strain) gain the same
auto-sent opener the metric pages use, each scoped to the inputs that drive
the composite rather than to a block it has no equivalent of.
- Workout detail's coach launch carries the workout itself. The chat request
gains an optional workoutId; on the FIRST turn the route narrows by
{ id, userId } and pins ONE bounded evidence section onto the snapshot.
Later turns reuse the pinned snapshot, so per-turn work stays constant in
conversation length.
- /coach?workout=<id> covers the full-page and maximize paths.
The workout block deliberately does not take the fenced transport the stored-
document path takes: a workout row is the user's own sensor record, and the
projection admits no free text. That is enforced, not assumed — sport_type is
a TEXT column, so it is re-asserted against the closed sport union, metadata
is excluded apart from its numeric zone durations read through a Zod slice,
and a recursive guard throws on any non-numeric leaf.
Own-history sport averages move to src/lib/workouts/sport-context.ts so the
narrative and the detail page cannot quote different figures.
The batch route now also resolves the user's timezone and module map through
the arrival seam, so a raw call count on user.findUnique no longer isolates
the read this guard is about. Filtering by the SELECT shape keeps the
once-per-batch contract pinned while letting the unrelated cached lookups
through.
Backfill immunity is proved by replay, not by a hand-picked case: a ten-year
Apple Health export (12772 writes) and a 30-day provider re-sync (58 writes)
must both emit zero events, while a fresh sleep night, workout and first
weight of the day emit exactly one each.
The provider-isolation guard walks the worker's static import graph and fails
if any provider client or resolver becomes reachable. It carries two positive
controls, so a broken entry path cannot make it pass against an empty graph.
The integration test covers what unit mocks cannot: the unique index, the
ON CONFLICT DO NOTHING claim, the losing side of ten concurrent claims, the
forward-only marker, ciphertext at rest, the delete cascade, and the
retention window.
Every ingest path in the product can now emit a typed, salience-classified
event into pg-boss, consumed by one worker that claims a durable per-day
reaction marker and fans out to the reaction surfaces.
The recency classifier runs at the seam, before anything is enqueued, so a
mass backfill — a multi-year Apple Health export, a provider re-sync, a
30-day catch-up — emits nothing at all. Non-salient writes are annotated and
dropped.
The worker makes no provider call by construction: it imports no provider
module, writes only markers, and enqueues nothing that is not budget-gated
downstream. Business refusals return a status rather than throwing, so a
refusal is never retried against a ceiling that will not move until the local
day rolls over.
Seams: sleep (one hook, inherited by all seven transports), workouts (the
batch route, the four provider syncs, the export importer), weight and blood
pressure (both measurement routes), labs (manual, OCR commit, document fact
commit).
The provider workout syncs could not previously tell a created row from an
updated one, so a re-sync of a recent session would have re-fired on every
poll; they now recover the distinction from the row the upsert already
returns.
Reaction lines are AES-256-GCM at rest, following the InsightNarrative
precedent, and the column is registered for key rotation. Markers are pruned
after 14 days by a daily maintenance job.
Also closes the device list's exposure of the same value, refuses a
cross-device refresh rotation, replaces a vacuous expiry test with one
that fails when the check is removed, and asserts the Nightscout host
guard at save time rather than relying on the probe's placement.
The settings mount fires an idempotent consent heal so an account that
predates the gate does not fall into the no-key state. It looked only for
an active receipt, so a deliberate withdrawal — which sets revokedAt —
read the same as never having consented, and the next visit to the AI
settings granted again, writing an audit row for an act the user never
performed. GDPR Art. 7 (3) puts withdrawal on the same footing as the
grant; one undone by a page load is not a withdrawal.
The mint now distinguishes the two callers. The mount heal grants only to
an account with no consent history at all and leaves a revocation
standing; switching document auto-read on is the consent act itself and
may supersede an earlier revocation. The revocation is re-checked inside
the transaction so a withdrawal landing mid-mint is not overwritten.
The revoke endpoint has existed since v1.4.40 but only the native client
called it: on the web, consent could be given and never taken back. The
AI settings now carry the standing decision and the control that changes
it, in all six languages.
The cookie was the Session row's cuid. cuid is built for collision-
resistant identity across distributed writers, not for unguessability —
it encodes a millisecond timestamp and a process-local counter and never
touches a CSPRNG. A session cookie is a bearer credential and has to be
unguessable, so it now carries 32 random bytes and only the HMAC of that
value is stored, matching how api_tokens and refresh_tokens already hold
their secrets.
Compatibility is passive, and deliberately so. A backfill cannot help:
the cookie already sitting in a browser is the cuid, and nothing we
generate server-side can be pushed into it — backfilling would sign
everyone out at deploy. Upgrading a row in place on first use loses to
the parallel requests a single page load fires: each mints its own
secret, one write wins, and the losers cannot re-emit the winner's
secret because only a hash is stored, so the browser can end up holding
a value the row does not honour. Instead, rows minted before the column
keep a NULL token_hash and stay resolvable by id, and are denied the
sliding-expiry extension so they drain within the 30-day session
lifetime. Every login from this release forward is on a secret at once,
and an id stops authenticating the moment its row holds one.
Note for operators: API_TOKEN_HMAC_KEY now also keys session cookies, so
rotating it signs browser sessions out as well as invalidating tokens.
It is already Core-required, so no deployment gains a new prerequisite.
Egress was already correct — the client fetches through safeFetch with
requirePublicHost pinned off the user's opt-in, manual redirects and a
timeout — but the check on the value we STORE was only a side effect of
the live probe the connect route happens to run first. That makes the
guard contingent on the probe's placement rather than on the input.
Run isPublicUrl at input time, gated on the same self-hoster opt-in, so
the URL that gets encrypted onto the row is the URL that was validated.
Cover the alt-notation classes the helper knows about; each case asserts
the probe was never reached, which is what distinguishes the input-time
refusal from the fetch-time one.
Rotation refused a replayed token under a spoofed device id (M-4) but
still rotated a *live* one presented from a different device, and then
inherited the caller's id onto the new row — putting the resulting family
outside the victim device's containment scope. Refuse the mismatch
outright when both ids are present; an absent id on either side stays
unattributable and rotates as before, so a client that sends no
X-Device-Id is not locked out.
The expiry refusal had no coverage at all: the standing test presented a
token that was never issued, so it passed on not_found and stayed green
with the expiry branch deleted. Replace it with the real secret plus a
just-inside-the-window companion, and cover the device binding.
Field-by-field comparison of the handlers against the registered Zod
schemas across measurements, workouts, medications, daily and auth. The
spec was corrected to match the handlers; no runtime behaviour changed.
Measurements:
- GET /api/measurements returns `{ measurements, meta: { total, limit,
offset, … } }` on all seven of its modes. The spec claimed a top-level
`total` that no arm has ever emitted.
- The list rows are stored rows minus the notes ciphertext, or synthesised
day-sum / sleep-segment rows, or pre-folded aggregate buckets. The
resource declared seven fields; the rest are now on the wire contract.
- POST accepts a bare ARRAY body (the combined blood-pressure write) and
answers with an array, and returns 409 on the dedup index.
- GET/PUT/DELETE /api/measurements/{id} carried no path entry at all.
Workouts: both GET operations run the module gate and answer 403
`module.disabled`; the reusable block existed but was never spread. The
batch ingest stays deliberately ungated and undocumented for it.
Medications: `nextDueAt` / `nextDueOverdue` are computed by the read paths
and absent from the create/update results, so they cannot be required on
the shared shape (the list row keeps them required). `stockUnitsRemaining`
sums a Decimal(12,4) and is fractional, not an integer. The cadence chips
object documented four fields that do not exist and missed the five that
do. `IntakeSource` admits APPLE_HEALTH.
Auth: the token bundle had its optionality inverted — `user` is always
returned, the token pair only for native callers. Adds the undocumented
403 `oidc_only`, the passkey 403/404, the TOTP enrolment 409s, the
attestation 400, the no-key-registered 409, the security-key 404s, the
refresh `revoke` response shape, and the path/query parameters three
operations read but never declared.
Daily was checked and is accurate; so are the batch ingest, series,
series-batch, bulk-delete, restore and by-external-ids routes.
persistRollupRows scoped its delete to the [min, max] bucket range of the rows
the aggregate RETURNED. That made the repair self-limiting in exactly the case
that needs it: a recompute could only overwrite a bucket the live aggregate
still produces, so a bucket whose source rows had all gone was never in the
delete range and kept its stale aggregate.
Three shapes of the same hole. An empty aggregate skipped the write entirely
(the caller short-circuited on rows.length === 0). A returned range narrower
than the requested window left the edge buckets untouched. A type that lost
every reading in the window dropped out of the returned rows, so the per-type
delete filter could never reach it.
The DAY write hook masked it — it deletes its own bucket by hand when the day
empties — but nothing covered the WEEK / MONTH / YEAR worker, which recomputes
one bucket span at a time: deleting the last reading of a type in a week
dropped the DAY bucket and left the coarser ones stale forever. The coverage
probe still reported hasBuckets = true, so the read-swap served them.
A recompute now prunes the window it was asked for and writes its rows in the
same transaction. The single-bucket write hook keeps its own scope, which is
correct there because it already deletes its bucket when the aggregate empties.
Still one tier: the rollup replaces the legacy read and falls back to live SQL
only on a coverage miss.
The per-document summary job is enqueued at upload time and no-ops while the
documentsAutoAiRead opt-in is OFF. Nothing ever re-enqueued it, so the switch
only applied to documents uploaded AFTER the flip — a user who filled the vault
first and opted in later had documents that would never be read, and the
setting looked inert.
A genuine OFF->ON transition now schedules one bounded catch-up pass over the
user's un-summarised documents. It grants nothing of its own: it only enqueues
the ordinary summary job, which still re-reads the opt-in, re-asserts the
per-document egress consent and reserves the daily token budget. The pass
itself re-reads the opt-in first, so a toggle flipped back OFF before it runs
egresses nothing.
Bounded by an id-cursor walk capped at 200 documents per pass (a summary is a
provider call, unlike the local-compute thumbnail backfill). Idempotent at four
levels: the trigger fires only on a real transition, the per-user singleton
collapses a double toggle, the candidate set excludes summarised documents, and
the summary job skips one that already has a summary.
Every staged fact carried a sourceText the model echoed back, stored verbatim
as 'the exact verbatim span of the document the fact came from'. Nothing
checked that the span existed. A paraphrase, a normalised rendering, or an
outright invention was persisted and shown to the reviewer as a transcription
of their own clinical record — an anchor pointing at nothing, or at a span the
document never contained.
Treat the echo as a lookup key instead. anchorSourceText resolves it against
the extracted text (exact, then whitespace-collapsed and case-folded to absorb
OCR reflow) and stores the span the document actually carries, with its offset.
A quote that matches nothing is dropped and the fact is flagged unanchored;
vision mode has no extracted text at all, so its facts are unanchored by
construction. Rows written before this resolve to unanchored on read — their
quote is preserved but never presented as checked.
The retry suffix still described the strict insight wrapper that was removed
long ago: it demanded citations[] and warnings[] — fields insightResultSchema
does not have — required metricSource, rationale, id and severity on every
recommendation when all four are optional, and named none of the five fields
the live schema actually requires (classification, findings, correlations,
dataQuality, disclaimer).
So the one corrective pass steered the model away from the contract it was
being corrected towards. A reply that obeyed the correction could not
validate, fell through parseComprehensiveResult's raw-object escape hatch,
and the downstream surfaces read an unvalidated payload.
Describe insightResultSchema and nothing else, and carry the contract as data
in RETRY_CONTRACT_EXAMPLE so a test can assert prompt and validator agree.
A READY row whose ciphertext will not open reports UNAVAILABLE, so the view
does not render a heading over nothing. Nothing asserted that, and removing the
degrade left the suite green.
Opening a document said "the summary is being prepared" and kept saying it.
The view had one fallback for a null summary column and used it for every
outcome: a document nobody ever enqueued, a missing vision provider, a spent
budget, a withdrawn consent, an unreadable file, a provider error, and a
summary the outbound screen withheld all rendered as work in progress.
The common case is not the screen. The background job is enqueued at exactly
one place — a fresh upload with the auto-read opt-in already on — and no
backfill reaches anything else. Every document uploaded before the toggle went
on has a null column no job will ever fill, and the view promised all of them a
summary that was never coming.
Record the outcome as `summary_state` (migration 0254, additive, defaults to
NONE which is truthful for every existing row) and branch on it. Only PENDING,
set when a job genuinely exists, may claim generation is running. NONE,
WITHHELD and UNAVAILABLE each say what happened and offer to generate one.
The withhold keeps its rule and loses its dead end. Blocked prose is still
never shown and never stored — only the fact of the refusal is kept, and it is
no longer terminal, so asking again re-runs the gauntlet. That was the original
intent; leaving the column null defeated it, because null read as "generating"
and the guard meant nothing would ever overwrite it.
The on-demand summary now persists what it produces. It is the only repair path
for the documents the job never ran for, and re-deriving the same paragraph on
every open costs a provider call for nothing. Generation is still only ever a
click — nothing runs because a view rendered. Transcription stays session-only.
Also drop the "read by AI" line from the detail footer.
The Today hero stayed on its all-clear line while a dose was actually due:
the dose-window item only fired once the slot was already overdue, so a
dose the user could take right now produced no rail item at all. Emit the
item from the moment the dose becomes takeable — the on-time window's lead
plus the early grace, the same instant logging it starts crediting the
slot — and give the two faces distinct wording and status: a due dose reads
calm and informational, an overdue one keeps the warning. A settled day
still emits nothing, and a cached block whose anchor has since passed keeps
rendering neither face. Push reminders are untouched; they still fire only
on overdue.
The briefing grounding gate discarded the whole block over one number it
could not verify, and it could not verify plenty of legitimate ones: the
allow-set was a hand-maintained list of feature blocks that had already
drifted from the prompt three times, and still omitted medication
compliance, the context counters, the correlation coefficients, every
coverage object and the comparison snapshot — all of which the prompt hands
the model verbatim. Derive the allow-set by walking the prompt payload
instead, so grounded means exactly what it says: a figure the server gave
the model. A fabricated number still matches nothing.
Fix the disposal to match. A content-failed retry left a hole while a
transport-failed one fell back to the previous cached briefing, though in
both cases the rejected text is discarded and the previous briefing passed
this same gate when it was written. Fall back on both paths, and only
report the briefing as withheld when there is genuinely nothing to show.
The on-demand route had no fallback at all; it now mirrors the background
generator that writes the same cached payload.
Importing the plural helper directly from the dashboard trend tile cost
~78 KB gz across the route chunks: the module is tiny, but a new edge
from a widely-instantiated component gets duplicated into every chunk
that contains it rather than hoisted. The provider is already in every
chunk, so expose a plural-aware tCount() there and let components call
that instead. Total client JS goes back under budget.
Also fix the two tests the honest-loading-state change invalidated: the
mental-wellbeing landing tests seeded a never-resolving read and so now
render skeletons rather than instrument cards, and the achievements page
needs a router in scope for its module-gate redirect.
The reminder, coach and plural fixes were pinned; the three formatter
fixes were not. Add the duration abbreviations across all six locales
and a structural check that the cycle calendar, device-score tile and
sleep-stage axis pass a resolved app locale to Intl rather than
undefined, which silently resolves to the browser locale.
Four locales received English on surfaces whose translations were
already complete, because the resolving code was a de/en binary.
- Medication reminder push and the Telegram inline-keyboard labels
admitted only de/en, so es/fr/it/pl got the English template on every
channel. Route the resolver through a shared coerceLocale().
- The Coach's trailing "reply now in X" clause said English for any
non-German locale, overriding the correct system-prompt directive.
Read the language name from a shared map instead.
- Coach refusal copy was a de/en constant pair, answering four locales
in English at the one turn the Coach is declining to help. Move the
copy into the bundles; the EN/DE constants now derive from them.
- The sleep-stage chart formatted axis ticks as en-US for four locales;
the cycle calendar and the device-score tile followed the *browser*
locale rather than the app locale.
- Three components carried a byte-identical duration formatter that
branched on German. Lift it to one key-driven helper.
Polish needs a third plural tier: 2-4 takes a "few" form a two-form
scheme cannot express, so "2 godzin temu" read as "2 of-hours ago".
Add a pluralKey() helper driven by Intl.PluralRules and a Few tier to
the relative-time and stale-hint keys. The five locales whose rules
have no "few" category resolve exactly as before by construction, and
their Few values mirror Other. Teach the reverse-coverage guard to
expand a pluralKey() base into its three tiers.
generateInsight() parsed against the strict aiInsightResponseSchema,
cross-checked that every recommendation's metricSource appeared in
citations[], and threw InsightSchemaError on a second failure. Its only
caller was runWithFallback(), which had no production caller — every
live surface runs runRawCompletionWithFallback. So none of it executed.
Wiring it into the live path was the alternative and it is not viable:
the live insightResultSchema has no citations[] field at all and treats
metricSource as optional, so the cross-check would have rejected
essentially every generation rather than tightening anything.
Removed: runWithFallback and its param/result types, generateInsight
and its private helpers, findUncitedRecommendations,
findRecommendationsMissingRationale, the never-thrown InsightSchemaError
and the isHardProviderFailure branch that excluded it, and
computeConfidence — a deterministic override that only ran inside the
dead wrapper.
Kept: buildRetryCorrectionMessage, which comprehensive-generate sends on
its own retry leg; the file it lived in is renamed to match what it now
holds. aiInsightResponseSchema stays as the shape that message
describes. Note the retry prompt asks for a stricter payload than the
live validator requires — that mismatch predates this and is left alone.
computeCitationCoverage is untouched and still live: it is a metric, not
a gate, and v1.30.24 wired it into the comprehensive post-parse step.
Chain-runner coverage the deleted tests carried (last-working cache,
empty-input guard) is ported onto runRawCompletionWithFallback, which
shares the machinery.
Three things the write surface claimed but did not hold.
Scope: the run bodies never re-checked ctx.canWrite. Enforcement was
real but lived entirely in the registration decision, and the test
asserted registration only — so a refactor that registered the write
tools unconditionally and meant to refuse at call time would open the
surface with the suite still green. Wrap every run body in a scope
guard at the registry level, so a new write tool inherits it, and drive
the run bodies directly from a read-only context in the test.
Free text: medication names, doses, treatment classes and lab analyte
names were interpolated into tool prose with no delimiter. A lab
analyte name is transcribed out of an uploaded document, so a hostile
document chooses it, and the write tools ride the same wire. Fence the
assembled prose with the shared data-fence markers and scrub short
result titles, which would leak markers into a host's result list.
Confirm: the header called the confirm flag a PROTOCOL-LEVEL CONFIRM
GATE. It is a client-asserted boolean with nothing binding it to a
prior preview. Binding it to a server-issued preview token would prove
only that a preview call happened, which any client can do, so it buys
no human-in-the-loop for a round trip and server state the stateless
transport does not want. Describe it as the convention it is and name
what actually bounds the blast radius.
enforceBudget and recordSpend have no callers left. Both defects they
carried recurred every time a surface copied them: a TOCTOU window
between the read and the bump, and a cap parameter defaulting to the
operator ceiling so an omitted argument rationed a self-hoster paying
for their own key. Deleting them means a future surface cannot reach
for the older, simpler-looking helper and reintroduce either.
getDailyTokenSpend goes with them — it existed to feed the read half.
The ledger-poisoning clamps recordSpend carried (NaN, negative,
fractional token counts) move to reserveBudget, where the property now
lives. Refresh the doc comments that still narrated the old flow.
Pin the four guarantees at the new chokepoint and at both call sites: spend
lands on the ledger, over-cap is refused before any egress, a BYOK chain is
measured against the user-plan ceiling rather than the operator's, and a
failed call refunds in full. Also pin the ordering the consent gate depends
on — a consent-blocked user is never charged.
The fixture described a clean resting series but supplied rows for only
twenty of the thirty days it fed PULSE for. Now that the resolver merges
per day, the ten resting-free days are estimated from their own readings
— and on a day whose only readings are a 150 bpm workout the honest
estimate falls outside the band. Match the spans so the case tests what
its name claims; gap-day behaviour is pinned in resting-pulse.test.ts.
The status tier funnels every generation through runStatusCompletion, which
reserves against the day's ledger before egress and reconciles after. The
comprehensive briefing did not: the shared generator and the on-demand
POST /api/insights/generate route both called the provider chain directly, so
none of that spend reached coach_usage and no daily ceiling applied to it.
Route every briefing provider call — first pass, JSON-shape retry,
grounding-correction retry, and the daily paragraph re-roll — through one
chokepoint that reuses reserveBudget / reconcileSpend and resolveDailyCap. The
reservation lands after the consent gate, so a consent-blocked user is never
charged, and a failed or timed-out call is refunded in full.
Over-cap is a typed skipped outcome for the off-request generator (no upstream
fault to retry against, so no failure marker and no retry enqueue) and a 429
for the on-demand route.
Each case fails on the pre-fix code: the rollup tests on an unaligned
recompute window, the batch tests on the older stats: snapshot winning,
and the resting test on a series that stops at the fold horizon.
resolveRestingPulseSeries returned the resting branch as soon as a
single RESTING_HEART_RATE row existed and never consulted the PULSE
proxy again. Since the retention fold began minting derived resting
rows for the days it folds, a proxy account carries derived rows for
old folded days and nothing but raw PULSE for the recent ones — so the
visible series stopped dead at the fold horizon.
Merge per day instead: a day with a resting row uses it, a day without
falls back to the proxy. which stays a three-value labelling signal and
reports proxy as soon as one point is estimated, so a partly-estimated
series is never labelled resting heart rate.
Widen the intraday baseline read to pull PULSE history when the newest
resting row is stale, not only when the rows are few — the count gate
alone pinned the baseline to the fold horizon.
recomputeUserRollups filtered measurements by a caller-supplied from/to
but grouped them by date_trunc, and persistRollupRows then deletes every
rollup row in the bucket range it is about to write. A from that landed
mid-bucket therefore replaced a complete bucket with an aggregate
computed only from the readings after that instant.
Every caller passes a moving instant — now minus ninety days on the
read-path refresh, an arbitrary instant on the coverage fallback, now
minus five years on the boot backfill and the restore path — so a
different oldest bucket was corrupted on each invocation and nothing
repaired it. Snap both edges out to whole buckets via bucketSpan, whose
own contract note already required exactly this of every recompute.
The red-team corpus probed only user messages and model replies. Neither is
the channel that reaches the prompt unguarded: a lab analyte, panel or unit
name, or a medication label, none of which the account holder authors when
the row was committed from an uploaded document.
The free-text guard's file list was hardcoded and had rotted — three of its
four entries were no longer reachable from the snapshot, and it covered one
of the dozen files that actually emit free text, which is why the lab fields
went unnoticed. Derive it from the import graph instead; it now covers 75
files, with a floor assertion so a resolver break cannot silently empty it.
Deriving the list immediately surfaced a second leak: features.ts shipped a
raw medication name and dose into the snapshot medications block.
Also fix sanitizeForPrompt itself. It deleted control characters, which
welded neighbouring tokens together and destroyed the word boundary its own
injection patterns anchor on, so a payload preceded by a newline survived
because of the newline. Replace with a space and collapse before matching.
The Coach SNAPSHOT was spliced into the prompt bare. Its JSON leaves are not
all server-computed: lab analyte, panel and unit names are transcribed by a
model out of an uploaded PDF and stored verbatim as the biomarker name, so a
hostile document chooses those strings. The about-me block has been fenced
since v1.16.1; extend the same pattern rather than adding per-field escapes.
- Move the fence markers and the wrapping primitive into a shared data-fence
module. Content is scrubbed against every known marker, not just its own
pair, so text in one block cannot forge another block's boundary.
- Fence the SNAPSHOT with an explicit data/instruction frame.
- Fence the OCR text handed to the document extractor; the previous
'OCR TEXT:' label was a label, not a boundary.
- Sanitise analyte, panel and unit at the labs source, which feeds both the
snapshot prompt and the labs_read tool result, and the medication label in
the adherence storyline.
- Capture a ---REMEMBER--- sentinel as proposed rather than active. It was
the only model-driven write with no confirm step, and the reminder recall
block fed it back into every later turn. Proposed reminders are already
excluded from that block and already carry a confirm control.
The status-provider suite drives the real budget module against an
in-memory stand-in for the day's usage row, so it exercises the atomic
reserve/reconcile path rather than asserting a mock was called: spend
lands on the ledger, an exhausted day is refused before any egress, a
failed generation is refunded, and a chain on the user's own key is
measured against the user-plan ceiling rather than the operator's.
The snapshot suite walks a 60-turn conversation and pins the full
snapshot to exactly two turns — the first and the elision crossing —
including the case where a lost reply shifts the turn parity.
The provider chokepoint every status and reference generator funnels
through ran with no budget accounting at all. Nine generator families
egressed behind it — the specialised status cards, the generic metric
cards, biomarker cards, the batched assessment, the derived scores, the
period narrative, and the off-request Coach memory workers (rolling
summary, fact extraction, plan proposals) — and none of that spend landed
on the daily ledger or met any ceiling. On an operator-key account it was
unmetered operator spend.
Reserve and reconcile around the provider call, using the same atomic
upsert the Coach and document paths use. The ceiling follows the cost
owner: a chain whose primary is the operator's own credential draws the
operator ceiling, everything else draws the user-plan ceiling, so a
self-hoster on their own key is never rationed by the operator's bill.
The status enqueue keyed its de-dupe singleton on (user, metric, locale).
The locale on a status read is client-chosen, so each of the six
supported locales opened its own slot and one legitimate generation pass
could be multiplied into six. Key on (user, metric) and carry the locale
in the payload instead.
The Coach re-shipped the full snapshot on every turn past the history
cap, not just the turn that crosses it, so a long conversation paid the
per-turn snapshot cost the snapshot-once model exists to remove. Re-ground
on the crossing only, tested as a narrow window rather than an equality
because a turn advances the count by one or two depending on whether the
reply persisted.
Give /api/ai/test a daily ceiling and a ledger entry. Its burst bucket
bounded rate but nothing cumulative, and with an empty body it resolves
the same chain generation uses — which can be the operator's credential.
Correct the memory-refresh worker comment, which asserted a budget check
inside the chokepoint that did not exist.
Marc-André Bombeck