mirror of
https://github.com/MBombeck/HealthLog.git
synced 2026-11-18 16:12:11 +00:00
feat(coach): typed grounding ledger + four-rung action ladder (Guard II)
Promote the per-turn magnitude bag to a per-conversation typed Grounding Ledger built from every source the model was actually shown: this turn's tool payloads, prior-turn tool figures, coach-memory goals, the medication schedule, the reference-grounding block, and guided blocks. Assistant prose is never a registration source, so a figure that slipped a rung cannot launder into an authoritative value on the next turn. Complete the action ladder: a reconciled number passes, a reference or education sentence is exempt across all six locales, a positively-typed benign number (date, time, year, ordinal, list marker) passes while an untyped unreconciled one stays on the strip path, and an unreconciled health figure is soft-stripped. Gate the dose continuation exemption on the user's actual schedule: a "keep taking your N mg" is trusted only when N is a dose the user is on; an off-schedule maintenance dose is caught. Extend the risk and percent classifiers to fr/es/it/pl to match the EN/DE depth. No schema migration: the ledger is rebuilt per turn from persisted tool traces and re-rendered prompt blocks.
This commit is contained in:
1 parent
25393d93fe
commit
9c95e713d2
13 files changed
+1506
-94
No files matched your search
@@ -0,0 +1,336 @@
|
||||
import { describe, it, expect, vi, beforeEach } from "vitest";
|
||||
|
||||
/**
|
||||
* v1.32.9 (Coach Guard II / G2) — the Grounding Ledger is CROSS-TURN, so its
|
||||
* two load-bearing properties can only be seen at the route level, over a
|
||||
* conversation with prior turns (the #591 lesson — the pure module cannot see
|
||||
* the route's activation + cross-turn assembly):
|
||||
*
|
||||
* - a prior-turn TOOL figure (persisted as `groundedFigures`) recalled this
|
||||
* turn reconciles and is NOT stripped;
|
||||
* - a number that appears only in a prior ASSISTANT REPLY (never a tool figure)
|
||||
* does NOT ground the next turn — assistant prose is never a ledger source
|
||||
* (D3 / red-team H3). It is still stripped.
|
||||
*
|
||||
* Guards are REAL here (coach-prose-grounding + grounding-ledger unmocked); the
|
||||
* provider loop + persistence are stubbed, exactly like route-tool-mode.
|
||||
*/
|
||||
|
||||
const SNAPSHOT_JSON = '{"bp":{"aggregate":{"mean":128}}}';
|
||||
|
||||
vi.mock("@/lib/api-handler", () => ({
|
||||
apiHandler: <T extends (...args: unknown[]) => unknown>(fn: T) => fn,
|
||||
requireAuth: vi.fn(async () => ({ user: { id: "u1", locale: "en" } })),
|
||||
HttpError: class HttpError extends Error {
|
||||
status: number;
|
||||
constructor(status: number, message: string) {
|
||||
super(message);
|
||||
this.status = status;
|
||||
}
|
||||
},
|
||||
}));
|
||||
vi.mock("@/lib/modules/gate", () => ({
|
||||
requireModuleEnabled: vi.fn(async () => ({ enabled: true })),
|
||||
isModuleEnabled: vi.fn(async () => true),
|
||||
}));
|
||||
vi.mock("@/lib/feature-flags", () => ({
|
||||
requireAssistantSurface: vi.fn(async () => undefined),
|
||||
}));
|
||||
vi.mock("@/lib/api-response", () => ({
|
||||
apiError: (error: string, status: number) => ({ data: null, error, status }),
|
||||
apiSuccess: (data: unknown) => ({ data, error: null, status: 200 }),
|
||||
}));
|
||||
vi.mock("@/lib/logging/context", () => ({
|
||||
annotate: vi.fn(),
|
||||
getEvent: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/logging/redact", () => ({
|
||||
redactSecrets: (s: string) => s,
|
||||
redactOptional: (s: unknown) => s,
|
||||
}));
|
||||
vi.mock("@/lib/auth/audit", () => ({ auditLog: vi.fn() }));
|
||||
vi.mock("@/lib/db", () => ({
|
||||
prisma: {
|
||||
user: { findUnique: vi.fn(async () => ({ coachPrefsJson: null })) },
|
||||
coachConversation: { findFirst: vi.fn(async () => ({ id: "c1" })) },
|
||||
},
|
||||
}));
|
||||
|
||||
const { checkRateLimit } = vi.hoisted(() => ({
|
||||
checkRateLimit: vi.fn(async () => ({ allowed: true })),
|
||||
}));
|
||||
vi.mock("@/lib/rate-limit", () => ({ checkRateLimit }));
|
||||
vi.mock("@/lib/i18n/server-locale", () => ({
|
||||
resolveServerLocale: vi.fn(async () => "en"),
|
||||
}));
|
||||
|
||||
const { runStreamingRawCompletionWithFallback } = vi.hoisted(() => ({
|
||||
runStreamingRawCompletionWithFallback: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/provider-runner", () => ({
|
||||
AllProvidersFailedError: class extends Error {},
|
||||
runStreamingRawCompletionWithFallback,
|
||||
}));
|
||||
|
||||
const { resolveProviderChain } = vi.hoisted(() => ({
|
||||
resolveProviderChain: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/provider", () => ({
|
||||
resolveProviderChain,
|
||||
resolveProvider: vi.fn(),
|
||||
}));
|
||||
|
||||
const { assertConsentForChain } = vi.hoisted(() => ({
|
||||
assertConsentForChain: vi.fn(async () => undefined),
|
||||
}));
|
||||
vi.mock("@/lib/ai/consent-guard", () => ({ assertConsentForChain }));
|
||||
vi.mock("@/lib/ai/prompts/insight-generator", () => ({ PROMPT_VERSION: "x" }));
|
||||
vi.mock("@/lib/ai/ai-budgets", () => ({
|
||||
AI_BUDGETS: { coach: { maxTokens: 1500, temperature: 0.4 } },
|
||||
}));
|
||||
|
||||
vi.mock("@/lib/ai/coach/types", async () => {
|
||||
const actual = await vi.importActual<typeof import("@/lib/ai/coach/types")>(
|
||||
"@/lib/ai/coach/types",
|
||||
);
|
||||
return actual;
|
||||
});
|
||||
|
||||
const { fetchConversationWithMessages } = vi.hoisted(() => ({
|
||||
fetchConversationWithMessages: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/persistence", () => ({
|
||||
appendMessage: vi.fn(async () => ({ id: "m1" })),
|
||||
createConversation: vi.fn(async () => ({ id: "c1" })),
|
||||
fetchConversationWithMessages,
|
||||
listConversations: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/coach-memory-shared", () => ({
|
||||
enqueueCoachMemoryRefresh: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/facts", () => ({
|
||||
storeDeterministicFacts: vi.fn(async () => undefined),
|
||||
}));
|
||||
|
||||
// Guard II — the schedule read. Empty here; the schedule-gated dose rule is
|
||||
// unit-tested in the outbound-screen suite.
|
||||
vi.mock("@/lib/medications/scheduled-doses", () => ({
|
||||
getScheduledDoseValues: vi.fn(async () => []),
|
||||
}));
|
||||
|
||||
const { reserveBudget, reconcileSpend } = vi.hoisted(() => ({
|
||||
reserveBudget: vi.fn(async () => ({ allowed: true, reserved: 3000 })),
|
||||
reconcileSpend: vi.fn(async () => undefined),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/budget", () => ({
|
||||
buildDateKey: vi.fn(() => "2026-07-23"),
|
||||
reserveBudget,
|
||||
reconcileSpend,
|
||||
resolveDailyCap: vi.fn(() => 2_000_000),
|
||||
}));
|
||||
|
||||
const { detectRefusal } = vi.hoisted(() => ({
|
||||
detectRefusal: vi.fn(() => ({ refuse: false })),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/refusal", () => ({ detectRefusal }));
|
||||
vi.mock("@/lib/ai/coach/outbound-guard", () => ({
|
||||
screenCoachReply: vi.fn(() => ({ block: false })),
|
||||
coachOutboundFallback: vi.fn(() => "fallback"),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/system-prompt", () => ({
|
||||
getCoachSystemPrompt: vi.fn(() => "SYSTEM"),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/about-me", () => ({
|
||||
getSelfContextTextForUser: vi.fn(async () => null),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/snapshot", () => ({
|
||||
buildCoachSnapshot: vi.fn(async () => ({
|
||||
snapshotJson: SNAPSHOT_JSON,
|
||||
sections: { bloodPressure: { aggregate: { mean: 128 } } },
|
||||
provenance: { windows: ["last30days"], metrics: ["bp"] },
|
||||
referenceGrounding: "REFERENCE RANGES",
|
||||
})),
|
||||
}));
|
||||
vi.mock("@/lib/workouts/hr-series", () => ({
|
||||
buildWorkoutHrSeries: vi.fn(async () => null),
|
||||
}));
|
||||
vi.mock("@/lib/workouts/zones", () => ({
|
||||
computeZones: vi.fn(() => null),
|
||||
hrMaxFromAge: vi.fn(() => 185),
|
||||
parseWhoopZoneDurations: vi.fn(() => null),
|
||||
}));
|
||||
vi.mock("@/lib/workouts/sport-context", () => ({
|
||||
buildSportContext: vi.fn(async () => null),
|
||||
}));
|
||||
|
||||
const {
|
||||
buildCoachDataInventory,
|
||||
renderDataInventory,
|
||||
renderFocusHint,
|
||||
runCoachToolLoop,
|
||||
} = vi.hoisted(() => ({
|
||||
buildCoachDataInventory: vi.fn(async () => ({
|
||||
entries: [],
|
||||
restMode: false,
|
||||
cycleEnabled: false,
|
||||
window: "last30days",
|
||||
probeScope: { sources: ["bp"], window: "last30days" },
|
||||
})),
|
||||
renderDataInventory: vi.fn(() => "DATA INVENTORY\n- blood pressure: present"),
|
||||
renderFocusHint: vi.fn(() => ""),
|
||||
runCoachToolLoop: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/tools", () => ({
|
||||
COACH_TOOL_DEFS: [{ name: "get_metric_series" }],
|
||||
MAX_ROUNDS: 3,
|
||||
buildCoachDataInventory,
|
||||
renderDataInventory,
|
||||
renderFocusHint,
|
||||
buildToolModeAddendum: vi.fn(() => "TOOL ADDENDUM"),
|
||||
runCoachToolLoop,
|
||||
}));
|
||||
|
||||
const { parseKeyValuesSentinel } = vi.hoisted(() => ({
|
||||
parseKeyValuesSentinel: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/keyvalues", () => ({ parseKeyValuesSentinel }));
|
||||
const { parseSuggestReminder } = vi.hoisted(() => ({
|
||||
parseSuggestReminder: vi.fn(),
|
||||
}));
|
||||
vi.mock("@/lib/ai/coach/suggest-reminder", () => ({ parseSuggestReminder }));
|
||||
vi.mock("@/lib/ai/coach/suggest-gate", () => ({ gateSuggestion: vi.fn() }));
|
||||
vi.mock("@/lib/validations/coach-prefs", () => ({
|
||||
parseCoachPrefs: vi.fn(() => ({ defaultWindow: undefined })),
|
||||
DEFAULT_REMINDER_SUGGESTION_PREFS: {},
|
||||
}));
|
||||
|
||||
const { appendMessage } = await import("@/lib/ai/coach/persistence");
|
||||
|
||||
const sse = vi.hoisted(() => ({ done: Promise.resolve() as Promise<unknown> }));
|
||||
vi.mock("@/lib/sse/create-stream", () => ({
|
||||
createSseStream: (
|
||||
producer: (c: {
|
||||
signal: { aborted: boolean };
|
||||
enqueue: () => void;
|
||||
}) => void | Promise<void>,
|
||||
) => {
|
||||
sse.done = Promise.resolve(
|
||||
producer({ signal: { aborted: false }, enqueue: () => {} }),
|
||||
);
|
||||
return new ReadableStream();
|
||||
},
|
||||
}));
|
||||
|
||||
import { POST } from "../route";
|
||||
|
||||
const post = POST as unknown as (req: Request) => Promise<Response>;
|
||||
|
||||
function chatReq(body: Record<string, unknown>): Request {
|
||||
return new Request("http://localhost/api/insights/chat", {
|
||||
method: "POST",
|
||||
headers: { "Content-Type": "application/json" },
|
||||
body: JSON.stringify(body),
|
||||
});
|
||||
}
|
||||
|
||||
/** Force the model's assembled prose through the sentinel/suggest parsers. */
|
||||
function stubReply(prose: string, toolResults: unknown[]): void {
|
||||
parseKeyValuesSentinel.mockReturnValue({
|
||||
prose,
|
||||
keyValues: [],
|
||||
malformed: false,
|
||||
malformedEntries: [],
|
||||
});
|
||||
parseSuggestReminder.mockReturnValue({ prose });
|
||||
runCoachToolLoop.mockImplementation(async () => ({
|
||||
result: { content: prose, tokensUsed: 80, model: "m" },
|
||||
workingProviderType: "anthropic",
|
||||
totalTokens: 80,
|
||||
rounds: 1,
|
||||
toolTrace: [{ name: "get_metric_series", present: true }],
|
||||
toolResults,
|
||||
}));
|
||||
}
|
||||
|
||||
function assistantContent(): string {
|
||||
const calls = (appendMessage as ReturnType<typeof vi.fn>).mock.calls;
|
||||
const assistant = calls.find(
|
||||
(c) => (c[0] as { role: string }).role === "assistant",
|
||||
);
|
||||
return (assistant?.[0] as { content: string }).content;
|
||||
}
|
||||
|
||||
describe("coach chat — cross-turn Grounding Ledger (G2)", () => {
|
||||
beforeEach(() => {
|
||||
vi.clearAllMocks();
|
||||
reserveBudget.mockResolvedValue({ allowed: true, reserved: 3000 });
|
||||
detectRefusal.mockReturnValue({ refuse: false });
|
||||
checkRateLimit.mockResolvedValue({ allowed: true });
|
||||
assertConsentForChain.mockResolvedValue(undefined);
|
||||
resolveProviderChain.mockResolvedValue([
|
||||
{ providerType: "anthropic", instance: {} },
|
||||
]);
|
||||
});
|
||||
|
||||
it("recalls a PRIOR-turn tool figure (persisted groundedFigures) without stripping it", async () => {
|
||||
// Turn 1 fetched systolic 128 and persisted it as groundedFigures.
|
||||
fetchConversationWithMessages.mockResolvedValue({
|
||||
id: "c1",
|
||||
attachmentCount: 0,
|
||||
summary: null,
|
||||
messages: [
|
||||
{ role: "user", content: "How is my BP?" },
|
||||
{
|
||||
role: "assistant",
|
||||
content: "Your systolic averaged 128.",
|
||||
metricSource: { groundedFigures: [128] },
|
||||
},
|
||||
],
|
||||
});
|
||||
// This turn: a FRESH sleep tool activates the verifier; the model recalls
|
||||
// the prior 128 AND cites the fresh 440.
|
||||
stubReply(
|
||||
"Your sleep averaged 440 minutes, and your systolic 128 average still holds.",
|
||||
[{ present: true, data: { metric: "sleep", aggregate: { mean: 440 } } }],
|
||||
);
|
||||
|
||||
await post(chatReq({ conversationId: "c1", message: "and my sleep?" }));
|
||||
await sse.done;
|
||||
|
||||
const content = assistantContent();
|
||||
expect(content).toContain("128");
|
||||
expect(content).toContain("440");
|
||||
expect(content).not.toContain("[unverified]");
|
||||
});
|
||||
|
||||
it("does NOT let a number from a prior ASSISTANT REPLY ground the next turn (D3 / H3)", async () => {
|
||||
// The prior assistant reply contains 158, but the persisted tool figure was
|
||||
// 128 — 158 was never a tool figure (a rung leak / fabrication). The ledger
|
||||
// must not read the assistant prose, so 158 stays ungrounded and is stripped.
|
||||
fetchConversationWithMessages.mockResolvedValue({
|
||||
id: "c1",
|
||||
attachmentCount: 0,
|
||||
summary: null,
|
||||
messages: [
|
||||
{ role: "user", content: "How is my BP?" },
|
||||
{
|
||||
role: "assistant",
|
||||
content: "Your systolic spike to 158 mmHg is worth watching.",
|
||||
metricSource: { groundedFigures: [128] },
|
||||
},
|
||||
],
|
||||
});
|
||||
// This turn: a fresh BP tool returns the real 128; the model repeats the
|
||||
// fabricated 158.
|
||||
stubReply("Your systolic spike to 158 mmHg persists this week.", [
|
||||
{ present: true, data: { metric: "bp", aggregate: { avgSys30: 128 } } },
|
||||
]);
|
||||
|
||||
await post(chatReq({ conversationId: "c1", message: "still high?" }));
|
||||
await sse.done;
|
||||
|
||||
const content = assistantContent();
|
||||
expect(content).not.toContain("158");
|
||||
expect(content).toContain("[unverified]");
|
||||
});
|
||||
});
|
||||
@@ -103,9 +103,14 @@ import {
|
||||
} from "@/lib/ai/coach/tools";
|
||||
import { parseKeyValuesSentinel } from "@/lib/ai/coach/keyvalues";
|
||||
import {
|
||||
findUnverifiedCoachNumbers,
|
||||
findUnverifiedCoachNumbersInLedger,
|
||||
stripUnverifiedNumbers,
|
||||
} from "@/lib/ai/coach/coach-prose-grounding";
|
||||
import {
|
||||
buildGroundingLedger,
|
||||
figuresForPersistence,
|
||||
} from "@/lib/ai/coach/grounding-ledger";
|
||||
import { getScheduledDoseValues } from "@/lib/medications/scheduled-doses";
|
||||
import { scrubUnknownLearnLinks } from "@/lib/ai/coach/learn-link-guard";
|
||||
import { parseSuggestReminder } from "@/lib/ai/coach/suggest-reminder";
|
||||
import { gateSuggestion } from "@/lib/ai/coach/suggest-gate";
|
||||
@@ -270,6 +275,14 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
// ── Conversation resolution ──────────────────────────────────
|
||||
let workingConversationId: string;
|
||||
let priorTurns: CoachTurn[] = [];
|
||||
// v1.32.9 (Coach Guard II / G2) — cross-turn Grounding Ledger sources. Prior
|
||||
// USER messages (numbers the user themselves stated) and the persisted
|
||||
// per-turn tool figures of prior turns (`groundedFigures` — server-computed
|
||||
// magnitudes, NEVER the assistant's prose, per D3). Both feed reconciliation
|
||||
// so a figure recalled a turn later is not stripped, without a laundering
|
||||
// path through assistant narration.
|
||||
let priorUserMessages: string[] = [];
|
||||
let priorToolFigures: number[] = [];
|
||||
// v1.11.1 — rolling summary of the elided older turns, read stale-while-
|
||||
// revalidate; null for a fresh conversation or when none is on file.
|
||||
let priorSummary: string | null = null;
|
||||
@@ -316,6 +329,16 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
role: m.role,
|
||||
content: m.content,
|
||||
}));
|
||||
// Ledger cross-turn sources (D3-safe): user-authored numbers + the
|
||||
// persisted tool figures of prior turns. Assistant PROSE is never read.
|
||||
priorUserMessages = existing.messages
|
||||
.filter((m) => m.role === "user")
|
||||
.map((m) => m.content);
|
||||
priorToolFigures = existing.messages.flatMap((m) =>
|
||||
m.metricSource?.groundedFigures
|
||||
? [...m.metricSource.groundedFigures]
|
||||
: [],
|
||||
);
|
||||
|
||||
// v1.4.43 W13 M-3 — replay-injection guard. `detectRefusal` runs
|
||||
// only on the inbound `message` per turn, so an injection that
|
||||
@@ -409,6 +432,11 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
// v1.16.0 — composed self-context: structured questionnaire fields
|
||||
// plus age/gender merged in from the User profile.
|
||||
const aboutMe = await getSelfContextTextForUser(userId, locale);
|
||||
// v1.32.9 (Coach Guard II / G3) — the user's active medication doses, for the
|
||||
// Grounding Ledger's `schedule` source AND the schedule-gated dose
|
||||
// continuation exemption. Fail-open: a read failure yields an empty set, which
|
||||
// keeps the Guard I phrase-anchored dose behaviour.
|
||||
const scheduleDoses = await getScheduledDoseValues(userId).catch(() => []);
|
||||
// v1.22 (B2/F6) — the canonical system-prompt module is owned elsewhere, so
|
||||
// the two memory/action clauses are appended at assembly time here: one
|
||||
// teaches the model to emit `---REMEMBER---` (durable "remind me" capture),
|
||||
@@ -890,7 +918,7 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
// the system-prompt GLP-1/grounding contracts. On a trip the turn is
|
||||
// replaced with a calm, grounded fallback and any reminder suggestion /
|
||||
// key-value provenance is dropped — the user never sees the unsafe text.
|
||||
const outbound = screenCoachReply(replyText, locale);
|
||||
const outbound = screenCoachReply(replyText, locale, scheduleDoses);
|
||||
if (outbound.block && outbound.reason) {
|
||||
replyText = coachOutboundFallback(outbound.reason, locale);
|
||||
annotate({
|
||||
@@ -919,32 +947,61 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
// (`snapshot.sections`, which already carries the correlations-snapshot block).
|
||||
// Exactly one is populated per turn; the grading, tolerance, and exemptions are
|
||||
// identical, so a number the model invents is flagged the same way on both.
|
||||
const authoritativePayloads =
|
||||
// v1.32.9 (Coach Guard II / G2+G3) — grade the prose against the typed
|
||||
// Grounding Ledger rather than a per-turn magnitude bag. Activation is
|
||||
// unchanged from Guard I: the ledger is graded ONLY when THIS turn delivered
|
||||
// fresh figures the model was told to ground against (a present tool result
|
||||
// / pinned workout, or the full snapshot on the no-tools path). The
|
||||
// cross-turn, memory, schedule, reference, and guided sources only WIDEN an
|
||||
// active grading — they never ACTIVATE it, so a snapshot figure the model
|
||||
// cited on a no-tool turn is still left alone (the v1.32.1 regression guard
|
||||
// holds). Assistant prose is never a ledger source (D3).
|
||||
const activatingPayloads =
|
||||
toolResultPayloads.length > 0
|
||||
? toolResultPayloads
|
||||
: noToolsSnapshotPayloads;
|
||||
if (!outbound.block && authoritativePayloads.length > 0) {
|
||||
const unverified = findUnverifiedCoachNumbers(
|
||||
replyText,
|
||||
authoritativePayloads,
|
||||
locale,
|
||||
);
|
||||
if (unverified.length > 0) {
|
||||
const { prose: corrected, stripped } = stripUnverifiedNumbers(
|
||||
let groundedFigures: number[] = [];
|
||||
if (activatingPayloads.length > 0) {
|
||||
const ledger = buildGroundingLedger({
|
||||
toolPayloads: activatingPayloads,
|
||||
priorToolFigures,
|
||||
priorUserMessages,
|
||||
// The self-context (goals / about-me) rides the system prompt on both
|
||||
// paths — the ledger's `memory` source.
|
||||
memoryTexts: aboutMe ? [aboutMe] : [],
|
||||
scheduleDoses,
|
||||
// Reference grounding rides the prompt only on the no-tools
|
||||
// full-snapshot path; the guided block rides both paths.
|
||||
referenceGrounding:
|
||||
!toolMode && includeFullSnapshot ? snapshot.referenceGrounding : null,
|
||||
guidedBlock: turnContext.guidedBlock,
|
||||
});
|
||||
// Persist THIS turn's tool figures (bare magnitudes) so a LATER turn can
|
||||
// recall them via the ledger — D3: the tool trace, never the prose.
|
||||
groundedFigures = figuresForPersistence(ledger);
|
||||
if (!outbound.block) {
|
||||
const unverified = findUnverifiedCoachNumbersInLedger(
|
||||
replyText,
|
||||
unverified,
|
||||
ledger,
|
||||
locale,
|
||||
);
|
||||
replyText = corrected;
|
||||
annotate({
|
||||
action: { name: "coach.prose.number_unverified" },
|
||||
meta: {
|
||||
flagged: unverified.length,
|
||||
stripped,
|
||||
// No raw values — just the count + truncated tokens for ops triage.
|
||||
tokens: unverified.slice(0, 6).map((u) => u.source),
|
||||
promptVersion: PROMPT_VERSION,
|
||||
},
|
||||
});
|
||||
if (unverified.length > 0) {
|
||||
const { prose: corrected, stripped } = stripUnverifiedNumbers(
|
||||
replyText,
|
||||
unverified,
|
||||
);
|
||||
replyText = corrected;
|
||||
annotate({
|
||||
action: { name: "coach.prose.number_unverified" },
|
||||
meta: {
|
||||
flagged: unverified.length,
|
||||
stripped,
|
||||
// No raw values — just the count + truncated tokens for ops triage.
|
||||
tokens: unverified.slice(0, 6).map((u) => u.source),
|
||||
promptVersion: PROMPT_VERSION,
|
||||
},
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1049,6 +1106,12 @@ async function handleChatRequest(request: NextRequest): Promise<Response> {
|
||||
})),
|
||||
}
|
||||
: {}),
|
||||
// v1.32.9 (Coach Guard II / G2) — persist THIS turn's tool figures so a
|
||||
// later turn's Grounding Ledger can recall them (D3-safe: tool trace, not
|
||||
// prose). Dropped on a blocked turn (its figures rode the discarded reply).
|
||||
...(!outbound.block && groundedFigures.length > 0
|
||||
? { groundedFigures }
|
||||
: {}),
|
||||
};
|
||||
if (sentinel.malformed) {
|
||||
// Graceful degrade: log so ops can spot a provider whose
|
||||
|
||||
Reference in new issue
Block a user