@dzhechkov/harness-core 0.3.142 → 0.3.144

package/src/compounding.ts

@@ -106,6 +106,47 @@ export interface UsageEvent {
   readonly queryTruncated?: boolean;
 }
 
+export interface ReplayInstance {
+  /** Stable per-PROMPT key — one prompt is one instance, however many lessons it injected. */
+  readonly id: string;
+  readonly query: string;
+  /** Exactly the lesson texts the apply leg injected for this prompt — the WARM arm's only delta. */
+  readonly lessons: readonly string[];
+  /**
+   * Pre-registered slice label (e.g. `task` / `conversational`). `null` until a human assigns it —
+   * nothing here invents a classification, because a class assigned AFTER outcomes are known is
+   * not a pre-registration.
+   */
+  readonly class: string | null;
+}
+
+/**
+ * The ONE definition of "a replayable pair": the readiness gate below COUNTS these and
+ * `dz epoch-replay --emit` EMITS these. A second copy would let readiness say 12 while the runner
+ * emits 9, silently — the drift class this repo keeps catching.
+ *
+ * Rules: a prompt with no query cannot be replayed; a TRUNCATED query is a prefix, not the prompt;
+ * one prompt = one instance (the hook writes one row per injected hit, up to 3 per prompt).
+ */
+export function replayableInstances(
+  usage: readonly UsageEvent[],
+  lessonText: ReadonlyMap<string, string> = new Map(),
+): ReplayInstance[] {
+  const byKey = new Map<string, { query: string; lessons: string[] }>();
+  for (const u of usage) {
+    if (typeof u.query !== 'string' || u.query.trim() === '') continue;
+    if (u.queryTruncated === true) continue; // a prefix is not the prompt
+    const key = u.eventId ?? `${u.runId ?? ''}|${u.ts}|${u.query}`;
+    const entry = byKey.get(key) ?? { query: u.query, lessons: [] };
+    const text = lessonText.get(u.dzId);
+    if (typeof text === 'string' && text.trim() !== '' && !entry.lessons.includes(text)) {
+      entry.lessons.push(text);
+    }
+    byKey.set(key, entry);
+  }
+  return [...byKey.entries()].map(([id, e]) => ({ id, query: e.query, lessons: e.lessons, class: null }));
+}
+
 export interface GuardEvent {
   readonly ts: string;
   readonly verdict: string;
@@ -238,14 +279,10 @@ export function assembleCompoundingReport(facts: CompoundingFacts): CompoundingR
 
   // 3. Replay readiness — cold-vs-warm needs (query -> injected lesson) pairs. They were never
   //    recorded before 2026-07-28, so this gate REPORTS accrual instead of faking a verdict.
-  const replayKeys = new Set<string>();
-  for (const u of usage) {
-    if (typeof u.query !== 'string' || u.query.trim() === '') continue;
-    if (u.queryTruncated === true) continue; // a prefix is not the prompt
-    // one prompt = one pair, however many hits it injected
-    replayKeys.add(u.eventId ?? `${u.runId ?? ''}|${u.ts}|${u.query}`);
-  }
-  const replayablePairs = replayKeys.size;
+  //    The rule ("no query / truncated query / one prompt = one pair") has exactly ONE definition,
+  //    in epoch-replay.ts, and the RUNNER emits precisely what this gate counts — a second copy
+  //    would let readiness say 12 while the runner emits 9, silently.
+  const replayablePairs = replayableInstances(usage).length;
   const replay: ReplayReadiness = {
     replayablePairs,
     minNeeded: MIN_SAMPLES_PER_ARM,