Files
HealthLog/src/lib/insights/status-shared.ts
T
Marc-André Bombeck b36c7b6697 feat(insights): serve status notes from the new table, only while available
Every status note reader and writer moves to InsightStatusCache: the
generators, the unchanged-data gates, the stale-while-revalidate read,
the ingest debounce, the previous-note prompt block and the
medication-compliance card with its per-medication lines. A stall now
opens a negative-cache window next to the note instead of a stub row
that shadowed it.

Stored model text is never served while its capability is unavailable,
whatever the reason: the status readers and gates check statusText, the
period narrative hides a model-written row while periodNarrative is
unavailable (the deterministic narrative is data and always shown), and
the daily digest drops reaction lines while reactionLines is
unavailable. A read-only miss now reports why (`unavailable` with the
capability's reason) and keeps hasProvider meaning provider presence, so
a withdrawn consent no longer reads as "no provider".

runStatusCompletion takes the capability as a required argument and
re-checks it before the provider chain is resolved, so a job queued
before a switch flipped still stops at the wire. The comprehensive
briefing re-checks briefing the same way. hasUsableStatusProvider is the
one presence probe.

Refs #975
2026-09-24 09:06:22 +02:00

408 lines
15 KiB
TypeScript

/**
* Shared helpers for the seven per-card status generators.
*
* Every `src/lib/insights/*-status.ts` generator carried byte-identical
* copies of `round`, `normalizeSummaryText`, `normalizeLocale`, and (in
* six of seven) `summarizeSeries`, plus an identical "parse the model's
* `{summary}` envelope" block and an identical persist block. This module is the single source of truth so a change
* to the rounding precision, the chart-token scrub, or the cache-row
* shape lands once rather than seven times.
*
* `SupportedLocale` is the full six-locale UI union. The status prompts
* compose one of two reviewed instruction bodies (de / en) but name the
* reader's own language in an explicit output directive, so the reader's
* locale must survive the whole pipeline rather than being collapsed to a
* binary on the way in. The
* medication-compliance generator stores per-medication lines next to its
* summary, so it shares `round`, `normalizeSummaryText`, `normalizeLocale`,
* and `parseSummaryFromContent` but writes its note through `writeStatusNote`
* directly.
*/
import { writeStatusNote } from "@/lib/insights/status-cache";
import { locales, type Locale } from "@/lib/i18n/config";
import { extractJsonObject } from "@/lib/ai/json-extract";
import { stripChartTokens } from "@/lib/insights/chart-tokens";
import {
screenModelOutput,
INSIGHTS_CONTRACTS,
type OutboundReason,
} from "@/lib/ai/safety/outbound-screen";
/**
* The locales the assessment pipeline carries end-to-end — all six the UI
* ships. Aliased to the i18n `Locale` so adding a UI locale widens the
* pipeline in one place rather than needing a second edit here.
*/
export type SupportedLocale = Locale;
const SUPPORTED_LOCALES: ReadonlySet<string> = new Set(locales);
/**
* The presentation contract shared by assessment producers and consumers.
*
* `kind` is the only branch discriminator. In particular, provider
* availability is metadata: it must never hide safe deterministic text.
* Provider/error details deliberately do not belong in this envelope.
*/
export type AssessmentStatus =
| {
kind: "generated";
text: string;
hasProvider: true;
updatedAt: string | null;
retryable: false;
}
| {
kind: "screened-fallback";
text: string;
hasProvider: boolean;
updatedAt: null;
retryable: true;
reason: OutboundReason | "unparseable_output" | "ungrounded_output";
}
| {
kind: "no-provider";
text: null;
hasProvider: false;
updatedAt: null;
retryable: false;
}
| {
kind: "preparing";
text: null;
hasProvider: true;
updatedAt: null;
retryable: true;
}
| {
kind: "revalidating";
text: string;
hasProvider: true;
updatedAt: string | null;
retryable: true;
}
| {
kind: "timeout" | "error";
text: null;
hasProvider: true;
updatedAt: null;
retryable: true;
}
| {
kind: "exhausted";
text: null;
hasProvider: true;
updatedAt: null;
retryable: false;
};
/** Round to `digits` decimal places. */
export function round(value: number, digits = 1): number {
const factor = 10 ** digits;
return Math.round(value * factor) / factor;
}
/** Scrub chart tokens and collapse whitespace out of a model summary. */
export function normalizeSummaryText(value: string): string {
return stripChartTokens(value).replace(/\s+/g, " ").trim();
}
/**
* Validate an arbitrary locale string against the six the UI ships.
*
* A recognised locale passes through UNCHANGED — that is the whole point:
* the former `value === "de" ? "de" : "en"` binary erased a French reader's
* locale before any prompt saw it, so the output-language directive could
* never fire. An unrecognised or missing value defaults to ENGLISH, never
* German; a German default is the bug class this replaces.
*/
export function normalizeLocale(
value: string | null | undefined,
): SupportedLocale {
return typeof value === "string" && SUPPORTED_LOCALES.has(value)
? (value as SupportedLocale)
: "en";
}
export interface SeriesSummary {
points: number;
start: number;
end: number;
delta: number;
mean: number;
min: number;
max: number;
}
/**
* Fold a value series into a single start/end/delta/mean/min/max
* summary. Returns null for an empty series.
*
* v1.4.33 — fold sum/min/max into a single walk. The previous
* `Math.min(...series.map(...))` / `Math.max(...series.map(...))`
* spread tripped V8's ~125 000-arg ceiling on the bound /api/analytics
* path; see `.planning/round-v1433-analytics-500-report.md` §"Carry-
* over". These helpers are fed bounded windows today so the crash
* never reached them, but the spread allocates a transient args array
* on every call — the fold is both stack-safe and cheaper.
*/
export function summarizeSeries(
series: Array<{ value: number }>,
): SeriesSummary | null {
if (series.length === 0) return null;
const first = series[0].value;
const last = series[series.length - 1].value;
let sum = 0;
let minVal = series[0].value;
let maxVal = series[0].value;
for (const entry of series) {
sum += entry.value;
if (entry.value < minVal) minVal = entry.value;
if (entry.value > maxVal) maxVal = entry.value;
}
return {
points: series.length,
start: round(first, 2),
end: round(last, 2),
delta: round(last - first, 2),
mean: round(sum / series.length, 2),
min: round(minVal, 2),
max: round(maxVal, 2),
};
}
/**
* Strip a Markdown code fence and isolate the JSON object body.
*
* The Anthropic and local providers have no native JSON mode (only the
* OpenAI-family clients send `response_format: json_object`), so a
* compliant model still routinely wraps its `{ "summary": ... }` reply in a
* ```json ... ``` fence or prefixes it with a sentence. `JSON.parse` then
* throws and the caller would surface the raw fenced string as the
* user-facing assessment. The extraction itself lives in
* `src/lib/ai/json-extract.ts`, where the Anthropic client applies it on the
* way out; this name stays for the insights callers.
*/
export function stripJsonFences(content: string): string {
return extractJsonObject(content);
}
/**
* Outcome of reading a status completion for its user-facing summary.
*
* - `summary` — a `{ "summary": "…" }` envelope parsed to a non-empty
* string. Render it.
* - `prose` — the model ignored the JSON contract and returned a plain
* sentence. Render it verbatim (back-compat with non-JSON-mode providers).
* - `unparseable` — the model ATTEMPTED the structured envelope but it came
* back malformed, truncated, or summary-less. There is NO safe text to
* render: the caller must fall back to its deterministic stub. This is the
* state that stops a partial `{ "summary": "…` (braces, property name,
* quotes, literal `\n`) from ever reaching the card — the class of bug
* reported against the resting-pulse Assessment.
*/
export type SummaryExtraction =
| { kind: "summary"; text: string }
| { kind: "prose"; text: string }
| { kind: "unparseable" };
/**
* Parse a candidate string as a `{ summary }` object. Reports whether the
* string was valid JSON *object* at all (`parsed`) separately from whether it
* carried a usable non-empty `summary` — so the caller can tell a malformed /
* truncated envelope apart from a well-formed object that simply lacks the
* field.
*/
function parseSummaryObject(candidate: string): {
parsed: boolean;
summary: string | null;
} {
try {
const obj = JSON.parse(candidate) as unknown;
if (obj !== null && typeof obj === "object" && !Array.isArray(obj)) {
const value = (obj as { summary?: unknown }).summary;
if (typeof value === "string" && value.trim().length > 0) {
return { parsed: true, summary: value };
}
// Well-formed object, but no usable summary (e.g. `{}` or `{summary:""}`).
return { parsed: true, summary: null };
}
// Valid JSON, but not a summary object (array / number / bare string).
return { parsed: false, summary: null };
} catch {
return { parsed: false, summary: null };
}
}
/**
* Did the completion OPEN like a JSON object? A genuine prose sentence never
* starts with `{` (a stray brace mid-sentence does not count), so a `{`-led
* payload that failed both parse attempts above is a malformed / truncated
* envelope — never surfaced raw. Strips a leading code fence first so a
* fenced-but-truncated object is caught too.
*/
function looksLikeJsonObjectAttempt(content: string): boolean {
const afterFence = content.replace(/^```(?:json)?\s*/i, "").trimStart();
return afterFence.startsWith("{");
}
/**
* Read a model completion for its user-facing summary and classify the result.
*
* Tries the content as-is first (the common, fence-free case), then a
* fence-stripped retry for providers without a native JSON mode. A completion
* that attempted the `{ summary }` envelope but came back malformed / truncated
* / summary-less resolves to `unparseable` rather than falling through as raw
* text — the caller then serves its deterministic fallback. Only a completion
* that never attempted the envelope (genuine prose) passes through verbatim.
*/
export function extractAssessmentSummary(content: string): SummaryExtraction {
const trimmed = content.trim();
// Attempt 1 — the content is itself a JSON object.
const direct = parseSummaryObject(trimmed);
if (direct.parsed) {
return direct.summary
? { kind: "summary", text: direct.summary }
: { kind: "unparseable" };
}
// Attempt 2 — strip a code fence / surrounding prose and retry.
const stripped = stripJsonFences(trimmed);
if (stripped !== trimmed) {
const viaFence = parseSummaryObject(stripped);
if (viaFence.parsed) {
return viaFence.summary
? { kind: "summary", text: viaFence.summary }
: { kind: "unparseable" };
}
}
// Neither parse produced a summary object. If the payload OPENED like an
// object, the model attempted the envelope and it came back malformed or
// truncated — resolve to the controlled fallback, never the raw braces /
// property name / quotes / literal `\n`. Genuine prose falls through.
if (looksLikeJsonObjectAttempt(trimmed)) {
return { kind: "unparseable" };
}
return { kind: "prose", text: content };
}
/**
* Pull the `summary` string out of a model completion. The status prompts
* return a `{ "summary": "…" }` envelope; a model that ignores the contract and
* returns bare prose falls back to the raw content. A malformed / truncated /
* summary-less structured envelope resolves to an EMPTY string (never the raw
* JSON) so no caller can surface partial structured output.
*
* Prefer `extractAssessmentSummary` in new code: it distinguishes an
* unparseable envelope from genuinely empty prose, which this thin wrapper
* collapses to `""`.
*/
export function parseSummaryFromContent(content: string): string {
const extracted = extractAssessmentSummary(content);
return extracted.kind === "unparseable" ? "" : extracted.text;
}
/**
* Store one status note (`InsightStatusCache`, encrypted) for today and return
* when it was written. The medication-compliance generator also stores its
* per-medication lines and goes through `writeStatusNote` directly.
*/
export async function persistStatusInsight(args: {
userId: string;
cacheAction: string;
todayKey: string;
text: string;
/**
* Fingerprint of the data snapshot this note was written from (see
* `snapshot-hash.ts`). The regeneration gate compares the next snapshot's
* hash against it and skips the provider call when nothing changed.
*/
snapshotHash?: string;
/**
* Cheap fingerprint of the SALIENT INPUTS (per-type count + newest
* measuredAt) for slow-moving metrics. The input gate compares a freshly
* probed fingerprint against it BEFORE the heavy snapshot build.
*/
inputHash?: string;
}): Promise<string> {
const generatedAt = await writeStatusNote({
userId: args.userId,
cacheAction: args.cacheAction,
todayKey: args.todayKey,
text: args.text,
snapshotHash: args.snapshotHash,
inputHash: args.inputHash,
});
return generatedAt.toISOString();
}
/**
* Outcome of turning one status-card completion into a shippable assessment.
* `blocked` carries the tripped contract so the caller can annotate before it
* withholds.
*/
export type StatusSummaryOutcome =
| { ok: true; kind: "generated"; text: string }
| {
ok: false;
kind: "screened-fallback";
reason: OutboundReason | "unparseable_output";
};
/**
* The single finalize step every per-metric / biomarker / derived-score status
* generator runs on its completion: parse the `{ summary }` envelope, scrub
* chart tokens and whitespace, then SCREEN the prose against the insights
* safety contracts.
*
* Why the screen belongs here. Before this existed, the only transform between
* a provider and a persisted status assessment was `normalizeSummaryText`,
* which is whitespace-only — so a dose-change imperative or a fabricated
* clinical risk score written onto a metric card persisted as that day's cached
* assessment while the identical sentence in the Coach was caught and replaced.
* Every generator shares this chokepoint, so a card cannot be added that
* silently skips the screen.
*
* SURFACE POLICY — WITHHOLD. Unlike the Coach, a status card is generated in
* the background and read from cache; nobody is waiting on the turn. Persisting
* a refusal is wrong (it would pin "I can't advise on doses" to the user's
* Weight card for the day, which is noise on a surface they did not ask a
* question on) and persisting the screened prose is obviously wrong. So the
* caller serves its existing deterministic stub — the same signal-grounded line
* the card shows when no provider is configured — and persists no model text.
* The card degrades to a state the UI already renders honestly, and the short
* negative-cache window lets the next run try again.
*
* The contracts include `causal` here: GROUND RULE 12 declares surface
* `insights`, and these cards are that surface.
*/
export function finalizeStatusSummary(
content: string,
locale: SupportedLocale,
): StatusSummaryOutcome {
const extracted = extractAssessmentSummary(content);
// A malformed / truncated / summary-less structured envelope has no safe
// text to render. WITHHOLD — the caller serves its deterministic stub. Only
// the `unparseable` discriminator is logged, never the assessment content.
if (extracted.kind === "unparseable") {
return {
ok: false,
kind: "screened-fallback",
reason: "unparseable_output",
};
}
const text = normalizeSummaryText(extracted.text);
if (!text) return { ok: true, kind: "generated", text: "" };
const decision = screenModelOutput(text, locale, INSIGHTS_CONTRACTS);
if (decision.block && decision.reason) {
return {
ok: false,
kind: "screened-fallback",
reason: decision.reason,
};
}
return { ok: true, kind: "generated", text };
}