nlm-memory 0.4.2 → 0.5.1

package/src/core/classifier/prompt.ts

@@ -1,200 +0,0 @@
-/**
- * Classifier prompt + transcript helpers. Centralized so every LLMClient
- * implementation hits the same prompt (parity with the Python daemon).
- *
- * Hard cap at 15K chars matches `classifier.py` MAX_TRANSCRIPT_CHARS:
- * smaller models (phi4-mini, qwen) pattern-match JSON from the transcript
- * above that size. Long sessions get first-half + last-half with a
- * separator to preserve opening intent + closing decisions.
- *
- * Phase B.2: prompt now also asks for a `facts` array of normalized
- * (subject, predicate, value) triples for the FactStore. The closed
- * predicate vocabulary is embedded in the prompt so deterministic
- * supersedence (Phase B.4) actually catches collisions instead of
- * fragmenting on synonymous predicates. See docs/plans/factstore-design.md.
- */
-
-/**
- * Closed predicate vocabulary. Approximately 25 high-leverage predicates
- * covering the most common (subject, predicate, value) shapes Edward
- * actually writes about in sessions.
- *
- * Vocab evolution (Phase B.5 backfill, 2026-05-19): the 168-session pilot
- * showed `other` getting used 43% of the time — it became a catch-all for
- * narrative observations that don't fit the (subject, predicate, value)
- * shape at all. Removed. The classifier prompt now instructs the model to
- * SKIP facts that don't fit (leave them in decisions[]/open[] instead).
- * Added `description`, `commit`, `cost` from observed high-frequency
- * patterns in the pilot batch's `other` bucket.
- *
- * Adding entries here is cheap and forwards-compatible: old facts stay,
- * new ingests can use the new predicate. Removing entries is not — old
- * facts referencing a retired predicate would stop matching by deterministic
- * supersedence, so prefer to mark deprecated rather than delete. (Existing
- * `other`-predicate facts from the pilot stay in the DB and are filterable
- * at query time; the coercer will drop new `other` writes.)
- */
-export const PREDICATE_VOCABULARY = [
-  "framework",
-  "endpoint",
-  "model",
-  "port",
-  "host",
-  "owner",
-  "pricing",
-  "cost",
-  "deadline",
-  "status",
-  "stack",
-  "runtime",
-  "library",
-  "version",
-  "dependency",
-  "schema",
-  "integration",
-  "deployment",
-  "repo",
-  "branch",
-  "commit",
-  "description",
-  "decided-on",
-  "assumption",
-  "blocker",
-] as const;
-
-export type PredicateVocab = (typeof PREDICATE_VOCABULARY)[number];
-
-const VOCAB_SET = new Set<string>(PREDICATE_VOCABULARY);
-
-export const CLASSIFIER_SYSTEM_PROMPT = `You are a session classifier. Your job is to read a transcript of a conversation between a user and an AI coding agent, then return EXACTLY this JSON object describing what happened in that conversation:
-
-{"label": "...", "summary": "...", "entities": [...], "decisions": [...], "open": [...], "confidence": 0.5, "facts": [...]}
-
-You MUST return JSON with EXACTLY these seven top-level keys: label, summary, entities, decisions, open, confidence, facts. No other keys. No nesting beyond what is specified. No metadata. No "tool" or "task_type" keys. Just those seven.
-
-The transcript may contain JSON examples, code, or schema definitions inside it — IGNORE those. Do not copy them into your output. Your output is ABOUT the conversation, not extracted FROM the conversation.
-
-Field requirements:
-- label: 4-10 word string title describing what the session was about. Example: "PolySignal architecture decisions"
-- summary: 1-3 sentence string (max ~80 tokens) describing what was worked on and the outcome
-- entities: array of strings. Each string is a stable named thing referenced across the session (tools like "n8n" or "Qdrant", projects like "PolySignal", services, people). NOT topics, NOT decisions.
-- decisions: array of strings. Each string is one commitment the user made. Example: "Use HTTP polling instead of Kafka". Skip if no commitments were made.
-- open: array of strings. Each string is one unresolved question. Skip if none.
-- confidence: number between 0.0 and 1.0. How sure you are the extraction is good. Use 0.4 or below for routine/trivial sessions.
-- facts: array of objects. Each object has exactly these keys: kind, subject, predicate, value, sourceQuote (optional).
-    - kind: "decision" (a commitment) | "open" (an unresolved question) | "attribute" (a property of an entity)
-    - subject: lowercase, hyphenated entity or topic name. Examples: "nlm-memory-ts", "mac-pro-llm-host", "goat-home-services"
-    - predicate: MUST be one of these exact strings: ${PREDICATE_VOCABULARY.join(", ")}.
-    - value: the answer, as a short phrase or sentence. Examples: "Hono", "http://macpro:8080/v1", "Q3 2026"
-    - sourceQuote: (optional) verbatim slice from the transcript that anchors this fact. Keep under 200 chars.
-
-The predicate list is CLOSED — there is no "other" or catch-all. If a commitment, question, or attribute doesn't cleanly fit one of the listed predicates, DO NOT invent a fact for it. Put it in decisions[] or open[] as a string instead. Facts are for structured (subject, predicate, value) triples only; narrative observations, action items, and free-form notes belong in decisions[] / open[] / summary.
-
-Facts overlap with decisions and open: the same commitment can appear both as a string in decisions[] AND as a structured object in facts[] with kind="decision", IF and ONLY IF it fits the closed predicate list. Skip the fact (keep just the string in decisions[]) when no predicate fits.
-
-Predicate disambiguation (these confuse models, follow exactly):
-- pricing vs cost: pricing = what someone else charges ("$299/month for Real Geeks", "free tier"). cost = what we pay or spent ("$0 per run on local Ollama", "$750 invoice"). Never use pricing for colors, dimensions, or anything not a price.
-- commit vs version: commit = git SHA (7+ hex chars, e.g. "cb5b940", "63596c3"). version = semver / release tag ("v4", "DSM 7.2.2", "Postgres 15", "0.3.6"). Use commit for any explicit git reference even if short-form.
-- description vs status: description = what a thing IS ("rich text editor framework by Meta"). status = what state it's in right now ("running via pm2", "not yet started", "blocked on review").
-
-Return ONLY the JSON object. No markdown code fences. No prose before or after.`;
-
-export const MAX_TRANSCRIPT_CHARS = 15_000;
-
-export function truncateTranscript(text: string, maxChars: number = MAX_TRANSCRIPT_CHARS): string {
-  if (text.length <= maxChars) return text;
-  const half = Math.floor((maxChars - 80) / 2);
-  return (
-    text.slice(0, half) +
-    "\n\n[... transcript truncated; below is the closing portion ...]\n\n" +
-    text.slice(text.length - half)
-  );
-}
-
-const FENCE_RE = /^```(?:json)?\s*|\s*```$/gm;
-
-export function stripJsonFences(text: string): string {
-  return text.replace(FENCE_RE, "").trim();
-}
-
-const REQUIRED_KEYS = ["label", "summary", "entities", "decisions", "open", "confidence"] as const;
-
-export function validateClassifierJson(data: unknown): data is Record<string, unknown> {
-  if (!data || typeof data !== "object" || Array.isArray(data)) return false;
-  const obj = data as Record<string, unknown>;
-  // `facts` is not in REQUIRED_KEYS — Phase B.2 accepts classifier output
-  // without it (older models, fixtures from Phase E parity tests). Coerced
-  // to [] when absent.
-  return REQUIRED_KEYS.every((k) => k in obj);
-}
-
-export function buildUserPrompt(transcript: string, priorContext: string): string {
-  const truncated = truncateTranscript(transcript);
-  const parts: string[] = [];
-  if (priorContext) parts.push(`PRIOR CONTEXT (already filed):\n${priorContext}\n`);
-  parts.push(`TRANSCRIPT TO CLASSIFY:\n${truncated}`);
-  return parts.join("\n");
-}
-
-interface CoercedFact {
-  kind: "decision" | "open" | "attribute";
-  subject: string;
-  predicate: string;
-  value: string;
-  sourceQuote?: string;
-}
-
-function coerceFacts(raw: unknown): CoercedFact[] {
-  if (!Array.isArray(raw)) return [];
-  const out: CoercedFact[] = [];
-  for (const item of raw) {
-    if (!item || typeof item !== "object" || Array.isArray(item)) continue;
-    const o = item as Record<string, unknown>;
-    const kindRaw = String(o["kind"] ?? "").toLowerCase().trim();
-    if (kindRaw !== "decision" && kindRaw !== "open" && kindRaw !== "attribute") continue;
-    const subject = String(o["subject"] ?? "").toLowerCase().trim();
-    const predicateRaw = String(o["predicate"] ?? "").toLowerCase().trim();
-    const value = String(o["value"] ?? "").trim();
-    if (!subject || !predicateRaw || !value) continue;
-    // Closed vocab — drop the fact entirely if the predicate isn't recognized.
-    // Pilot data (Phase B.5) showed `other` was 43% of writes and almost all
-    // slop; the prompt now instructs the model to leave such observations in
-    // decisions[]/open[] strings. This coercer enforces the policy
-    // defensively in case the model emits an off-vocab predicate anyway.
-    if (!VOCAB_SET.has(predicateRaw)) continue;
-    const predicate = predicateRaw;
-    const sourceQuoteRaw = o["sourceQuote"];
-    const sourceQuote =
-      typeof sourceQuoteRaw === "string" && sourceQuoteRaw.trim().length > 0
-        ? sourceQuoteRaw.trim().slice(0, 500)
-        : undefined;
-    const fact: CoercedFact = { kind: kindRaw, subject, predicate, value };
-    if (sourceQuote !== undefined) fact.sourceQuote = sourceQuote;
-    out.push(fact);
-  }
-  return out;
-}
-
-export function coerceClassifyResult(data: Record<string, unknown>): {
-  label: string;
-  summary: string;
-  entities: string[];
-  decisions: string[];
-  open: string[];
-  confidence: number;
-  facts: CoercedFact[];
-} {
-  const strArray = (v: unknown): string[] => {
-    if (!Array.isArray(v)) return [];
-    return v.map((x) => String(x).trim()).filter((s) => s.length > 0);
-  };
-  const label = String(data["label"] ?? "").trim().slice(0, 120) || "Untitled";
-  const summary = String(data["summary"] ?? "").trim();
-  const entities = strArray(data["entities"]);
-  const decisions = strArray(data["decisions"]);
-  const open = strArray(data["open"]);
-  const conf = Number(data["confidence"] ?? 0.5);
-  const confidence = Number.isFinite(conf) ? conf : 0.5;
-  const facts = coerceFacts(data["facts"]);
-  return { label, summary, entities, decisions, open, confidence, facts };
-}

package/src/core/dataset/build-dataset.ts

@@ -1,463 +0,0 @@
-/**
- * buildDataset — read projection over canonical.sqlite that hydrates every
- * UI page (pulse, river, search, thread).
- *
- * Ports the read paths of `dataset.py`. Action-driven overlays (dismissed
- * alerts, snoozed entities, retired labels, merged variants) are deferred:
- * the action log isn't yet exposed by the TS daemon. Returns persisted
- * state directly.
- */
-
-import { existsSync } from "node:fs";
-import Database from "better-sqlite3";
-import * as sqliteVec from "sqlite-vec";
-import { liveSessionStatus } from "@core/storage/live-status.js";
-import { loadActionOverlay, openQuestionId } from "@core/actions/overlay.js";
-import type { ActionOverlay } from "@core/actions/overlay.js";
-import type { SessionStatus } from "@shared/types.js";
-
-export interface DatasetSession {
-  readonly id: string;
-  readonly date: string;
-  readonly started_at: string | null;
-  readonly ended_at: string | null;
-  readonly label: string;
-  readonly summary: string;
-  readonly entities: ReadonlyArray<string>;
-  readonly decisions: ReadonlyArray<string>;
-  readonly open: ReadonlyArray<string>;
-  readonly open_questions: ReadonlyArray<{ id: string; text: string; resolved: false }>;
-  readonly status: SessionStatus;
-  readonly duration_min: number;
-  readonly runtime: string;
-  readonly supersedes?: string;
-  readonly superseded_by?: string;
-}
-
-export interface DatasetEntity {
-  readonly canonical: string;
-  readonly type: string;
-  readonly status: string;
-  readonly session_count: number;
-  readonly last_seen_session: string | null;
-}
-
-export interface DatasetResponse {
-  readonly meta: {
-    readonly last_sync: string;
-    readonly sessions_total: number;
-    readonly entities_total: number;
-    readonly db_present: boolean;
-    readonly db_path: string;
-  };
-  readonly sessions: ReadonlyArray<DatasetSession>;
-  readonly entities: ReadonlyArray<DatasetEntity>;
-  readonly entity_colors: Record<string, string>;
-  readonly entity_type: Record<string, string>;
-  readonly entity_status: Record<string, string>;
-  readonly metrics: {
-    readonly this_week: number;
-    readonly last_week: number;
-    readonly sparkline: ReadonlyArray<number>;
-    readonly healthy: number;
-    readonly sparse: number;
-    readonly stale: number;
-    readonly closed_decisions: number;
-  };
-  readonly alerts: ReadonlyArray<{
-    readonly id: string;
-    readonly type: "stale";
-    readonly severity: "high" | "medium";
-    readonly entity: string;
-    readonly summary: string;
-    readonly sessions: ReadonlyArray<string>;
-    readonly age_days: number;
-    readonly last_touch_at: string | null;
-  }>;
-  readonly runtimes: ReadonlyArray<DatasetRuntime>;
-}
-
-export interface DatasetRuntime {
-  readonly name: string;
-  readonly status: "active" | "idle" | "dormant";
-  readonly sessions_total: number;
-  readonly this_week: number;
-  readonly last_week: number;
-  readonly last_session_at: string | null;
-}
-
-interface SessionRow {
-  id: string;
-  started_at: string | null;
-  ended_at: string | null;
-  duration_min: number | null;
-  label: string;
-  summary: string;
-  status: "active" | "closed" | "superseded";
-  transcript_path: string | null;
-  runtime: string;
-}
-
-interface EntityRow {
-  session_id: string;
-  entity_canonical: string;
-}
-
-interface MarkerRow {
-  session_id: string;
-  kind: "decision" | "open";
-  text: string;
-  position: number;
-}
-
-interface EdgeRow {
-  from_session: string;
-  to_session: string;
-  kind: "supersedes" | "continues";
-}
-
-interface EntityCatalogRow {
-  canonical: string;
-  type: string;
-  status: string;
-  session_count: number;
-  last_seen_session: string | null;
-}
-
-const EMPTY_DATASET = (dbPath: string, present: boolean): DatasetResponse => ({
-  meta: {
-    last_sync: new Date().toISOString(),
-    sessions_total: 0,
-    entities_total: 0,
-    db_present: present,
-    db_path: dbPath,
-  },
-  sessions: [],
-  entities: [],
-  entity_colors: {},
-  entity_type: {},
-  entity_status: {},
-  metrics: { this_week: 0, last_week: 0, sparkline: [0, 0, 0, 0, 0, 0, 0], healthy: 0, sparse: 0, stale: 0, closed_decisions: 0 },
-  alerts: [],
-  runtimes: [],
-});
-
-export interface BuildDatasetOptions {
-  /** Include path-shaped entities (filesystem leaks from the classifier).
-   *  Default false — they pollute the catalog without adding signal. */
-  readonly includePaths?: boolean;
-}
-
-export function buildDataset(dbPath: string, options: BuildDatasetOptions = {}): DatasetResponse {
-  if (!existsSync(dbPath)) return EMPTY_DATASET(dbPath, false);
-  const db = new Database(dbPath, { readonly: true });
-  try {
-    sqliteVec.load(db);
-  } catch {
-    // vec extension only required for semantic search; tolerable here.
-  }
-  try {
-    return projectFromDb(db, dbPath, options.includePaths ?? false);
-  } finally {
-    db.close();
-  }
-}
-
-/**
- * Heuristic for "this entity is actually a filesystem path the classifier
- * leaked into the catalog". Catches things like ".claude/agents/",
- * "bridge/server.js", "deploy.sh", "nlm-memory-spec.md" while leaving
- * real entities like "n8n", "Node.js", "NocoDB", "Whtnxt Agent" alone.
- */
-const CODE_FILE_EXT_RE =
-  /\.(?:md|markdown|txt|ts|tsx|js|jsx|mjs|cjs|py|pyi|json|yaml|yml|toml|sh|bash|zsh|css|html|sql|xml|env|ini|cfg|conf|lock)$/i;
-
-export function isPathShapedEntity(canonical: string): boolean {
-  if (!canonical) return false;
-  // Any slash → looks like a path (forward or back).
-  if (canonical.includes("/") || canonical.includes("\\")) return true;
-  // Hidden-file prefix only when it's clearly a dotfile (e.g. ".env", ".mcp.json").
-  if (canonical.startsWith(".") && canonical.length > 1 && canonical !== "...") return true;
-  // Common source-code file extensions.
-  if (CODE_FILE_EXT_RE.test(canonical)) return true;
-  return false;
-}
-
-function projectFromDb(db: Database.Database, dbPath: string, includePaths: boolean): DatasetResponse {
-  const sessionRows = db
-    .prepare<[], SessionRow>(`
-      SELECT id, started_at, ended_at, duration_min, label, summary,
-             status, transcript_path, runtime
-      FROM sessions
-      ORDER BY started_at ASC
-    `)
-    .all();
-
-  if (sessionRows.length === 0) return EMPTY_DATASET(dbPath, true);
-
-  const entitiesBySession = new Map<string, string[]>();
-  for (const r of db
-    .prepare<[], EntityRow>("SELECT session_id, entity_canonical FROM session_entities ORDER BY session_id")
-    .all()) {
-    const list = entitiesBySession.get(r.session_id);
-    if (list) list.push(r.entity_canonical);
-    else entitiesBySession.set(r.session_id, [r.entity_canonical]);
-  }
-
-  const decisionsBySession = new Map<string, string[]>();
-  const openBySession = new Map<string, { id: string; text: string }[]>();
-  for (const r of db
-    .prepare<[], MarkerRow>("SELECT session_id, kind, text, position FROM markers ORDER BY session_id, position")
-    .all()) {
-    if (r.kind === "decision") {
-      const list = decisionsBySession.get(r.session_id);
-      if (list) list.push(r.text);
-      else decisionsBySession.set(r.session_id, [r.text]);
-    } else {
-      const id = openQuestionId(r.session_id, r.text);
-      const list = openBySession.get(r.session_id);
-      if (list) list.push({ id, text: r.text });
-      else openBySession.set(r.session_id, [{ id, text: r.text }]);
-    }
-  }
-
-  const supersedesBy = new Map<string, string>();
-  const supersededByBy = new Map<string, string>();
-  const continuesBy = new Map<string, string>();
-  for (const r of db
-    .prepare<[], EdgeRow>("SELECT from_session, to_session, kind FROM session_edges")
-    .all()) {
-    if (r.kind === "supersedes") {
-      supersedesBy.set(r.from_session, r.to_session);
-      supersededByBy.set(r.to_session, r.from_session);
-    } else if (r.kind === "continues") {
-      continuesBy.set(r.from_session, r.to_session);
-    }
-  }
-
-  const allEntityRows = db
-    .prepare<[], EntityCatalogRow>(`
-      SELECT canonical, type, status, session_count, last_seen_session
-      FROM entities ORDER BY session_count DESC
-    `)
-    .all();
-
-  const overlay = loadActionOverlay(db);
-  for (const e of allEntityRows) {
-    if (overlay.retiredEntities.has(e.canonical)) e.status = "retired";
-    else if (overlay.snoozedEntities.has(e.canonical)) e.status = "snoozed";
-    const newType = overlay.labeledEntities.get(e.canonical);
-    if (newType) e.type = newType;
-  }
-
-  const entityRows = includePaths
-    ? allEntityRows
-    : allEntityRows.filter((e) => !isPathShapedEntity(e.canonical));
-  const keptEntities = new Set(entityRows.map((e) => e.canonical));
-
-  const sessions: DatasetSession[] = sessionRows.map((s) => {
-    const status = liveSessionStatus(s.transcript_path, s.status);
-    const rawOpen = openBySession.get(s.id) ?? [];
-    const supersedes = supersedesBy.get(s.id);
-    const supersededBy = supersededByBy.get(s.id);
-    const rawEntities = entitiesBySession.get(s.id) ?? [];
-    const activeOpen = rawOpen.filter(
-      (o) => !overlay.resolvedOpens.has(o.id) && !overlay.promotedOpens.has(o.id),
-    );
-    const promotedDecisions = rawOpen
-      .filter((o) => overlay.promotedOpens.has(o.id))
-      .map((o) => overlay.promotedOpens.get(o.id)!);
-    return {
-      id: s.id,
-      date: (s.started_at ?? "").slice(0, 10),
-      started_at: s.started_at,
-      ended_at: s.ended_at,
-      label: s.label,
-      summary: s.summary,
-      entities: includePaths ? rawEntities : rawEntities.filter((name) => keptEntities.has(name)),
-      decisions: [...(decisionsBySession.get(s.id) ?? []), ...promotedDecisions],
-      open: activeOpen.map((o) => o.text),
-      open_questions: activeOpen.map((o) => ({ id: o.id, text: o.text, resolved: false as const })),
-      status,
-      duration_min: s.duration_min ?? 0,
-      runtime: s.runtime,
-      ...(supersedes !== undefined ? { supersedes } : {}),
-      ...(supersededBy !== undefined ? { superseded_by: supersededBy } : {}),
-    };
-  });
-
-  // continuesBy is in the dataset shape but unused by current UI; reserved for thread view.
-  void continuesBy;
-
-  const entityColors: Record<string, string> = {};
-  const entityType: Record<string, string> = {};
-  const entityStatus: Record<string, string> = {};
-  for (const e of entityRows) {
-    entityColors[e.canonical] = stableColor(e.canonical);
-    entityType[e.canonical] = e.type;
-    entityStatus[e.canonical] = e.status;
-  }
-
-  const metrics = computeMetrics(sessions, entityRows);
-  const alerts = computeStaleAlerts(sessions, entityRows, overlay);
-  const runtimes = computeRuntimes(sessions);
-
-  return {
-    meta: {
-      last_sync: new Date().toISOString(),
-      sessions_total: sessions.length,
-      entities_total: entityRows.length,
-      db_present: true,
-      db_path: dbPath,
-    },
-    sessions,
-    entities: entityRows,
-    entity_colors: entityColors,
-    entity_type: entityType,
-    entity_status: entityStatus,
-    metrics,
-    alerts,
-    runtimes,
-  };
-}
-
-function computeRuntimes(sessions: ReadonlyArray<DatasetSession>): DatasetRuntime[] {
-  const now = Date.now();
-  const day = 86_400_000;
-  const week = now - 7 * day;
-  const prev = now - 14 * day;
-  const groups = new Map<string, {
-    total: number;
-    thisWeek: number;
-    lastWeek: number;
-    lastAt: number;
-    lastAtIso: string | null;
-  }>();
-  for (const s of sessions) {
-    const name = (s.runtime ?? "").trim() || "unknown";
-    const g = groups.get(name) ?? { total: 0, thisWeek: 0, lastWeek: 0, lastAt: 0, lastAtIso: null };
-    g.total += 1;
-    if (s.started_at) {
-      const t = Date.parse(s.started_at);
-      if (Number.isFinite(t)) {
-        if (t >= week) g.thisWeek += 1;
-        else if (t >= prev) g.lastWeek += 1;
-        if (t > g.lastAt) {
-          g.lastAt = t;
-          g.lastAtIso = s.started_at;
-        }
-      }
-    }
-    groups.set(name, g);
-  }
-  const hour = 3_600_000;
-  const out: DatasetRuntime[] = [];
-  for (const [name, g] of groups) {
-    const age = g.lastAt ? now - g.lastAt : Infinity;
-    const status: DatasetRuntime["status"] =
-      age <= hour ? "active" : age <= day ? "idle" : "dormant";
-    out.push({
-      name,
-      status,
-      sessions_total: g.total,
-      this_week: g.thisWeek,
-      last_week: g.lastWeek,
-      last_session_at: g.lastAtIso,
-    });
-  }
-  out.sort((a, b) => (Date.parse(b.last_session_at ?? "0") || 0) - (Date.parse(a.last_session_at ?? "0") || 0));
-  return out;
-}
-
-function computeMetrics(
-  sessions: ReadonlyArray<DatasetSession>,
-  entityRows: ReadonlyArray<EntityCatalogRow>,
-) {
-  const now = Date.now();
-  const sparkline = [0, 0, 0, 0, 0, 0, 0];
-  let thisWeek = 0;
-  let lastWeek = 0;
-  for (const s of sessions) {
-    const t = s.started_at ? Date.parse(s.started_at) : NaN;
-    if (!Number.isFinite(t)) continue;
-    const ageDays = (now - t) / 86_400_000;
-    if (ageDays >= 0 && ageDays < 7) {
-      thisWeek += 1;
-      const bucket = Math.min(6, Math.floor(ageDays));
-      sparkline[6 - bucket] = (sparkline[6 - bucket] ?? 0) + 1;
-    } else if (ageDays >= 7 && ageDays < 14) {
-      lastWeek += 1;
-    }
-  }
-  const sessionsById = new Map(sessions.map((s) => [s.id, s]));
-  let healthy = 0;
-  let sparse = 0;
-  let stale = 0;
-  for (const e of entityRows) {
-    if (e.session_count === 0) continue;
-    const last = sessionsById.get(e.last_seen_session ?? "");
-    const lastT = last?.started_at ? Date.parse(last.started_at) : NaN;
-    const ageDays = Number.isFinite(lastT) ? (now - lastT) / 86_400_000 : 999;
-    if (ageDays > 30) stale += 1;
-    else if (e.session_count >= 3) healthy += 1;
-    else sparse += 1;
-  }
-  const closedDecisions = sessions.reduce(
-    (sum, s) => sum + (s.status === "superseded" ? 0 : s.decisions.length),
-    0,
-  );
-  return { this_week: thisWeek, last_week: lastWeek, sparkline, healthy, sparse, stale, closed_decisions: closedDecisions };
-}
-
-function computeStaleAlerts(
-  sessions: ReadonlyArray<DatasetSession>,
-  entityRows: ReadonlyArray<EntityCatalogRow>,
-  overlay: ActionOverlay,
-): DatasetResponse["alerts"] {
-  const now = Date.now();
-  const sessionsById = new Map(sessions.map((s) => [s.id, s]));
-  const alerts: DatasetResponse["alerts"][number][] = [];
-  for (const e of entityRows) {
-    if (e.session_count === 0 || e.status === "retired" || e.status === "snoozed") continue;
-    const last = sessionsById.get(e.last_seen_session ?? "");
-    const lastT = last?.started_at ? Date.parse(last.started_at) : NaN;
-    if (!Number.isFinite(lastT)) continue;
-    const ageDays = Math.floor((now - lastT) / 86_400_000);
-    if (ageDays <= 30) continue;
-
-    const alertId = `stale_${e.canonical.replace(/[^A-Za-z0-9]/g, "_")}`;
-    if (overlay.dismissedAlerts.has(alertId) || overlay.snoozedAlerts.has(alertId)) continue;
-
-    const openOnEntity = sessions
-      .filter((s) => s.entities.includes(e.canonical))
-      .flatMap((s) => s.open)
-      .slice(0, 2);
-    let summary = `Last touch ${ageDays} days ago`;
-    if (openOnEntity.length > 0) {
-      const n = openOnEntity.length;
-      const label = n === 1 ? "question" : "questions";
-      summary += ` · ${n} unresolved open ${label}: "${openOnEntity[0]!.slice(0, 80)}"`;
-    }
-    alerts.push({
-      id: alertId,
-      type: "stale",
-      severity: ageDays > 60 ? "high" : "medium",
-      entity: e.canonical,
-      summary,
-      sessions: last ? [last.id] : [],
-      age_days: ageDays,
-      last_touch_at: last?.started_at ?? null,
-    });
-  }
-  alerts.sort((a, b) => (a.severity === "high" ? 0 : 1) - (b.severity === "high" ? 0 : 1));
-  return alerts;
-}
-
-const HUES = [200, 270, 320, 30, 90, 150, 220, 290, 340, 50, 110, 170] as const;
-
-function stableColor(canonical: string): string {
-  let h = 0;
-  for (let i = 0; i < canonical.length; i++) h = (h * 31 + canonical.charCodeAt(i)) | 0;
-  const hue = HUES[Math.abs(h) % HUES.length] ?? 200;
-  return `hsl(${hue}, 60%, 55%)`;
-}