feat(reward): sentinel-based exchange-count bypass for cron episodes

apps/memos-local-plugin/core/config/defaults.ts

@@ -135,6 +135,7 @@ export const DEFAULT_CONFIG: ResolvedConfig = {
       minContentCharsForCompletion: 40,
       toolHeavyRatio: 0.7,
       minAssistantCharsForToolHeavy: 80,
+      cronSentinels: [],
     },
     l2Induction: {
       minSimilarity: 0.65,

apps/memos-local-plugin/core/config/schema.ts

@@ -212,6 +212,14 @@ const AlgorithmSchema = Type.Object({
      * scored normally even if tool calls dominate. Default 80.
      */
     minAssistantCharsForToolHeavy: NumberInRange(80, 0, 10_000),
+    /**
+     * User-turn prefixes that identify cron/scheduled episodes. When
+     * the first user turn starts with any of these, the exchange-count
+     * gate is bypassed. Default: Hermes cron sentinel.
+     */
+    cronSentinels: Type.Array(Type.String(), {
+      default: ["[IMPORTANT: You are running as a scheduled cron job"],
+    }),
   }, { default: {} }),
   l2Induction: Type.Object({
     /** Cosine ≥ this to associate a new trace with an existing L2 policy. */

apps/memos-local-plugin/core/reward/reward.ts

@@ -392,7 +392,7 @@ function looksLikeTrivialContent(text: string): boolean {
 function decideSkipReason(
   snapshot: import("../session/types.js").EpisodeSnapshot,
   traces: readonly TraceRow[],
-  cfg: Pick<RewardConfig, "minExchangesForCompletion" | "minContentCharsForCompletion" | "toolHeavyRatio" | "minAssistantCharsForToolHeavy">,
+  cfg: Pick<RewardConfig, "minExchangesForCompletion" | "minContentCharsForCompletion" | "toolHeavyRatio" | "minAssistantCharsForToolHeavy" | "cronSentinels">,
 ): string | null {
   // Prefer the live snapshot's turn list; fall back to traces when the
   // snapshot came from a SQLite row (no turns materialised).
@@ -443,9 +443,34 @@ function decideSkipReason(
   // 1. Not enough real conversation turns (need at least N user-assistant exchanges)
   const exchanges = Math.min(userTurns, assistantTurns);
   if (exchanges < cfg.minExchangesForCompletion) {
-    return (
-      `对话轮次不足（${exchanges} 轮），需要至少 ${cfg.minExchangesForCompletion} 轮完整的问答交互才能生成摘要。`
-    );
+    // Cron episodes always have exactly 1 user turn (the task prompt) so this
+    // gate always fires for them. Bypass it when the first user content starts
+    // with a known cron sentinel — cron jobs are inherently substantive, and
+    // the content/triviality checks below provide the real substance filter.
+    //
+    // Fallback: if snapshot.turns is empty (recovery path), we also check
+    // snapshot.meta?.initialUserText. This field is set by the pipeline on
+    // episode creation and is generally reliable, but could be absent or
+    // stale in unusual recovery scenarios — if so, the episode falls through
+    // to the old "skip" behavior (no false positives, just a missed score).
+    const sentinels = cfg.cronSentinels ?? [];
+    if (sentinels.length > 0) {
+      const firstUserContent =
+        userContents[0] ??
+        (snapshot.meta?.initialUserText as string | undefined) ??
+        "";
+      if (sentinels.some((s) => firstUserContent.startsWith(s))) {
+        // Cron episode — skip the exchange-count gate, fall through to content checks.
+      } else {
+        return (
+          `对话轮次不足（${exchanges} 轮），需要至少 ${cfg.minExchangesForCompletion} 轮完整的问答交互才能生成摘要。`
+        );
+      }
+    } else {
+      return (
+        `对话轮次不足（${exchanges} 轮），需要至少 ${cfg.minExchangesForCompletion} 轮完整的问答交互才能生成摘要。`
+      );
+    }
   }
 
   // 2. No user messages at all

apps/memos-local-plugin/core/reward/types.ts

@@ -64,6 +64,14 @@ export interface RewardConfig {
    * that the tool-heavy heuristic would otherwise skip. Default 80.
    */
   minAssistantCharsForToolHeavy: number;
+  /**
+   * User-turn prefixes that identify cron/scheduled episodes. When the
+   * first user turn starts with any of these prefixes, the exchange-count
+   * gate (minExchangesForCompletion) is bypassed — cron jobs are
+   * inherently substantive regardless of turn count. Content and
+   * triviality checks still apply. Default: Hermes cron sentinel.
+   */
+  cronSentinels?: string[];
 }
 
 // ─── User feedback inputs ──────────────────────────────────────────────────

apps/memos-local-plugin/tests/unit/reward/reward.integration.test.ts

@@ -376,4 +376,59 @@ describe("reward/integration", () => {
     expect(res.feedbackCount).toBe(2);
     expect(res.rHuman).toBeGreaterThan(0);
   });
+
+  it("cron episode with 1 exchange is not skipped when sentinel matches", async () => {
+    const sid = "s_cron_1";
+    const eid = "ep_cron_1";
+    seedEpisode(handle, eid, sid, ["tr_cron_1"]);
+    seedTrace(handle, "tr_cron_1", eid, sid, {
+      userText:
+        "[IMPORTANT: You are running as a scheduled cron job. Please review recent activity and write a reflection card.",
+      agentText:
+        "Reviewed the last 48 hours of conversation traces and wrote a reflection card to ~/Faye/memory/reflections/2026-05-31.md. Commit succeeded.",
+    });
+
+    const events: RewardEvent[] = [];
+    const bus = createRewardEventBus();
+    bus.onAny((e) => events.push(e));
+
+    const runner = createRewardRunner({
+      tracesRepo: handle.repos.traces,
+      episodesRepo: handle.repos.episodes,
+      feedbackRepo: handle.repos.feedback,
+      llm: fakeLlm({
+        completeJson: {
+          "reward.reward.r_human.v3": {
+            goal_achievement: 0.8,
+            process_quality: 0.8,
+            user_satisfaction: 0.8,
+            label: "success",
+            reason: "cron reflection card written and committed",
+          },
+        },
+      }),
+      bus,
+      cfg: {
+        ...cfg(),
+        minExchangesForCompletion: 2,
+        minContentCharsForCompletion: 40,
+        cronSentinels: ["[IMPORTANT: You are running as a scheduled cron job"],
+      },
+      now: () => NOW,
+    });
+
+    const res = await runner.run({
+      episodeId: eid as unknown as Parameters<typeof runner.run>[0]["episodeId"],
+      feedback: [],
+      trigger: "implicit_fallback",
+    });
+
+    // Must reach the LLM scorer, not be abandoned as trivial.
+    expect(res.humanScore.source).toBe("llm");
+    expect(res.rHuman).toBeGreaterThan(0);
+    expect(events.some((e) => e.kind === "reward.updated")).toBe(true);
+    // Must NOT be heuristic-skipped (reward.scored with source=heuristic means skipped).
+    const scoredEvent = events.find((e) => e.kind === "reward.scored");
+    expect((scoredEvent as { source?: string } | undefined)?.source).not.toBe("heuristic");
+  });
 });