recallnest 2.3.0 → 2.6.0

package/src/mcp-tools-governance.ts

@@ -0,0 +1,475 @@
+import { z } from "zod";
+import { ALIAS_RULES, expandQueryWithAliases, explainAliasExpansion } from "./aliases.js";
+import { buildConflictAuditSummary, clusterConflicts } from "./conflict-advisor.js";
+import { resolveConflictCandidate } from "./conflict-engine.js";
+import { escalateConflicts } from "./conflict-escalation.js";
+import { CONFLICT_ATTENTION_LEVELS, summarizeConflictLifecycle } from "./conflict-lifecycle.js";
+import { formatConflictAudit, formatConflictClusters, formatConflictEscalation, formatConflictList, formatConflictRecord, formatConflictResolution } from "./conflict-output.js";
+import { ConflictStatusSchema } from "./conflict-schema.js";
+import { ConsolidationEngine, formatConsolidationResult } from "./consolidation-engine.js";
+import { archiveDirtyBriefAsset, listDirtyBriefAssets } from "./memory-assets.js";
+import { scanMemoryPromotions, buildPromoteScanDeps, formatPromoteScanResult } from "./memory-promotion.js";
+import { aliasMapFilePath, expandQuery, explainUserAliases, listUserAliases, removeUserAlias, upsertUserAlias } from "./query-expander.js";
+import { recordSkillOutcome } from "./skill-engine.js";
+import { scanForPromotions, formatPromotionResult } from "./skill-promotion.js";
+import { buildWorkflowEvidence, buildWorkflowObservationRecord, inspectWorkflowDashboard, inspectWorkflowHealth } from "./workflow-observation-engine.js";
+import { formatWorkflowEvidencePack, formatWorkflowHealthDashboard, formatWorkflowHealthReport, formatWorkflowObservationSaved } from "./workflow-observation-output.js";
+import { WorkflowObservationOutcomeSchema } from "./workflow-observation-schema.js";
+import type { ToolRegistryDeps } from "./mcp-tool-deps.js";
+import { withLock } from "./distill-lock.js";
+
+export function registerGovernanceTools(deps: ToolRegistryDeps): void {
+  const { registerTool, getComponents, conflictStore, workflowObservationStore, getKGExtractor, getKGStore } = deps;
+
+registerTool(
+  "workflow_observe",
+  "Store an append-only workflow observation for self-evolution. Use this to record whether a continuity primitive or reusable workflow succeeded, failed, was corrected by the user, or was missed entirely. Optionally pass `skillId` to also bump that skill's successCount/failureCount — the canonical way to close the skill feedback loop.",
+  {
+    workflowId: z.string().min(1).max(120).describe("Workflow primitive id, such as resume_context or checkpoint_session"),
+    outcome: WorkflowObservationOutcomeSchema.default("success").describe("success | failure | corrected | missed"),
+    summary: z.string().min(1).max(400).describe("Short description of what happened"),
+    scope: z.string().min(1).max(160).optional().describe("Optional scope such as project:recallnest"),
+    source: z.string().min(1).max(40).default("agent").describe("Source label such as agent, smoke, eval, or manual"),
+    signal: z.string().min(1).max(120).optional().describe("Optional failure/correction signal tag"),
+    task: z.string().min(1).max(240).optional().describe("Optional related task"),
+    tags: z.array(z.string().min(1).max(40)).max(8).default([]).describe("Optional tags"),
+    tools: z.array(z.string().min(1).max(60)).max(6).default([]).describe("Optional tools involved"),
+    skillId: z.string().min(1).max(128).optional().describe("Optional skill id to bump. When present and outcome is success, the skill's successCount increments; otherwise failureCount increments. Missing/corrupt skills are skipped silently (does not block observation write)."),
+    idempotencyKey: z.string().min(1).max(160).optional().describe("Optional stable request key; repeated saves with the same key replace the prior observation"),
+  },
+  async ({ workflowId, outcome, summary, scope, source, signal, task, tags, tools, skillId, idempotencyKey }) => {
+    const record = buildWorkflowObservationRecord({
+      workflowId,
+      outcome,
+      summary,
+      scope,
+      source,
+      signal,
+      task,
+      tags,
+      tools,
+      skillId,
+      idempotencyKey,
+    });
+    const stored = await workflowObservationStore.save(record);
+
+    let skillOutcomeNote = "";
+    if (skillId) {
+      try {
+        const { store } = getComponents();
+        const result = await recordSkillOutcome(store, skillId, outcome);
+        if (result.updated) {
+          skillOutcomeNote = `\nSkill ${skillId.slice(0, 8)} → success=${result.successCount} failure=${result.failureCount}`;
+        } else {
+          console.warn(`[workflow_observe] skillId=${skillId} update skipped: ${result.reason}`);
+          skillOutcomeNote = `\nSkill ${skillId.slice(0, 8)} → not updated (${result.reason})`;
+        }
+      } catch (err) {
+        console.warn(`[workflow_observe] recordSkillOutcome threw for skillId=${skillId}:`, err);
+        skillOutcomeNote = `\nSkill ${skillId.slice(0, 8)} → update errored, observation still saved`;
+      }
+    }
+
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatWorkflowObservationSaved(stored) + skillOutcomeNote,
+      }],
+    };
+  },
+);
+
+registerTool(
+  "workflow_health",
+  "Inspect workflow observation health: 7d/30d report for one workflow or dashboard of degraded workflows. Read-only. Use when checking if continuity primitives are succeeding or degrading.",
+  {
+    workflowId: z.string().min(1).max(120).optional().describe("Optional workflow primitive id"),
+    scope: z.string().min(1).max(160).optional().describe("Optional scope filter"),
+    limit: z.number().int().min(1).max(30).default(10).describe("Dashboard result limit when workflowId is omitted"),
+  },
+  async ({ workflowId, scope, limit }) => {
+    const text = workflowId
+      ? formatWorkflowHealthReport(await inspectWorkflowHealth(workflowObservationStore, { workflowId, scope }))
+      : formatWorkflowHealthDashboard(await inspectWorkflowDashboard(workflowObservationStore, { scope, limit }), scope);
+    return {
+      content: [{
+        type: "text" as const,
+        text,
+      }],
+    };
+  },
+);
+
+registerTool(
+  "workflow_evidence",
+  "Generate an evidence pack for a workflow primitive with recent issues, top signals, and suggested actions. Read-only. Use when investigating why a workflow is degraded and you need concrete failure examples.",
+  {
+    workflowId: z.string().min(1).max(120).describe("Workflow primitive id, e.g. 'resume_context'"),
+    scope: z.string().min(1).max(160).optional().describe("Scope filter, e.g. 'project:recallnest'"),
+    limit: z.number().int().min(1).max(20).default(5).describe("Max recent issue observations to include"),
+  },
+  async ({ workflowId, scope, limit }) => {
+    const pack = await buildWorkflowEvidence(workflowObservationStore, {
+      workflowId,
+      scope,
+      limit,
+    });
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatWorkflowEvidencePack(pack),
+      }],
+    };
+  },
+);
+
+registerTool(
+  "list_conflicts",
+  "List or inspect conflict candidates where promoted evidence disagrees with existing durable memory. Read-only. Use when reviewing pending conflicts before resolution.",
+  {
+    conflictId: z.string().min(1).max(128).optional().describe("Conflict ID to inspect a single record, e.g. 'c1d2e3f4'"),
+    status: ConflictStatusSchema.optional().describe("Optional status filter"),
+    attention: z.enum(CONFLICT_ATTENTION_LEVELS).optional().describe("Optional lifecycle filter: fresh / aging / stale / escalated / resolved"),
+    canonicalKey: z.string().min(1).max(120).optional().describe("Optional canonical key filter"),
+    groupBy: z.enum(["record", "cluster"]).default("record").describe("Whether to list individual conflicts or grouped clusters"),
+    limit: z.number().int().min(1).max(50).default(20).describe("Max conflicts to list"),
+  },
+  async ({ conflictId, status, attention, canonicalKey, groupBy, limit }) => {
+    if (conflictId) {
+      const record = await conflictStore.getById(conflictId);
+      return {
+        content: [{
+          type: "text" as const,
+          text: record ? formatConflictRecord(record) : `Conflict not found: ${conflictId}`,
+        }],
+      };
+    }
+
+    const records = (await conflictStore.listRecent({ status, canonicalKey, limit: Math.max(limit * 2, limit) }))
+      .filter((record) => !attention || summarizeConflictLifecycle(record).attention === attention)
+      .slice(0, limit);
+    return {
+      content: [{
+        type: "text" as const,
+        text: groupBy === "cluster"
+          ? formatConflictClusters(clusterConflicts(records))
+          : formatConflictList(records),
+      }],
+    };
+  }
+);
+
+registerTool(
+  "resolve_conflict",
+  "Resolve a conflict candidate by keeping existing, accepting incoming, or merging texts. Side effect: updates conflict status and may modify durable memory. Use when list_conflicts shows open conflicts that need a decision.",
+  {
+    conflictId: z.string().min(1).max(128).describe("Conflict ID to resolve"),
+    resolution: z.enum(["accept_incoming", "keep_existing", "merge"]).describe("How to resolve the conflict"),
+    mergedText: z.string().min(1).max(2000).optional().describe("Optional merged text override when resolution is merge"),
+    notes: z.string().min(1).max(320).optional().describe("Optional operator notes"),
+  },
+  async ({ conflictId, resolution, mergedText, notes }) => {
+    const { store, embedder } = getComponents();
+    const result = await resolveConflictCandidate({
+      store,
+      embedder,
+      conflictStore,
+    }, {
+      conflictId,
+      resolution,
+      ...(mergedText ? { mergedText } : {}),
+      notes,
+    });
+
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatConflictResolution(result),
+      }],
+    };
+  }
+);
+
+registerTool(
+  "audit_conflicts",
+  "Generate a conflict audit summary showing priority clusters by staleness and escalation level. Read-only. Use when triaging which conflict clusters to resolve first.",
+  {
+    status: ConflictStatusSchema.optional().describe("Optional status filter"),
+    canonicalKey: z.string().min(1).max(120).optional().describe("Optional canonical key filter"),
+    limit: z.number().int().min(1).max(500).default(100).describe("How many conflict records to scan"),
+    top: z.number().int().min(1).max(20).default(5).describe("How many priority clusters to show"),
+  },
+  async ({ status, canonicalKey, limit, top }) => {
+    const records = await conflictStore.listRecent({
+      status,
+      canonicalKey,
+      limit,
+    });
+    const summary = buildConflictAuditSummary(records, top);
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatConflictAudit(summary),
+      }],
+    };
+  }
+);
+
+registerTool(
+  "escalate_conflicts",
+  "Preview or apply conflict aging policy to mark stale conflicts for operator review. Side effect: when apply=true, persists escalation metadata. Use when conflicts have aged past their attention threshold.",
+  {
+    attention: z.enum(["stale", "escalated"]).default("stale").describe("Only consider stale or escalated conflicts"),
+    canonicalKey: z.string().min(1).max(120).optional().describe("Optional canonical key filter"),
+    limit: z.number().int().min(1).max(500).default(100).describe("How many open conflicts to scan"),
+    top: z.number().int().min(1).max(20).default(10).describe("How many eligible conflicts to include"),
+    apply: z.boolean().default(false).describe("When false, preview only. When true, persist escalation metadata."),
+    notes: z.string().min(1).max(320).optional().describe("Optional operator note when applying escalation"),
+  },
+  async ({ attention, canonicalKey, limit, top, apply, notes }) => {
+    const result = await escalateConflicts({
+      conflictStore,
+    }, {
+      attention,
+      canonicalKey,
+      limit,
+      top,
+      apply,
+      notes,
+    });
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatConflictEscalation(result),
+      }],
+    };
+  }
+);
+
+registerTool(
+  "list_dirty_briefs",
+  "List memory briefs generated before current cleanup rules that may need re-indexing. Read-only. Use when auditing brief quality or before running clean_dirty_briefs.",
+  {},
+  async () => {
+    const rows = listDirtyBriefAssets();
+    if (rows.length === 0) {
+      return { content: [{ type: "text" as const, text: "No dirty briefs found." }] };
+    }
+    const lines = [
+      "Brief ID  Title  Scope  Reasons",
+      "--------  -----  -----  ----------------------------------------",
+      ...rows.map(row => `${row.id.slice(0, 8)}  ${row.title}  [${row.scope}]  ${row.reasons.join("; ")}`),
+    ];
+    return { content: [{ type: "text" as const, text: lines.join("\n") }] };
+  }
+);
+
+registerTool(
+  "clean_dirty_briefs",
+  "Archive dirty briefs and remove their indexed asset entries. Side effect: when apply=true, moves briefs to archive and deletes index rows. Use when list_dirty_briefs shows stale briefs that need cleanup.",
+  {
+    apply: z.boolean().default(false).describe("When false, preview only (no writes). When true, archive briefs and delete indexed rows."),
+  },
+  async ({ apply }) => {
+    const { store } = getComponents();
+    const rows = listDirtyBriefAssets();
+    if (rows.length === 0) {
+      return { content: [{ type: "text" as const, text: "No dirty briefs found." }] };
+    }
+
+    if (!apply) {
+      const preview = rows.map(row => `${row.id.slice(0, 8)}  ${row.title}  [${row.scope}]  ${row.reasons.join("; ")}`).join("\n");
+      return {
+        content: [{
+          type: "text" as const,
+          text: `Dirty briefs detected: ${rows.length}\n\n${preview}\n\nCall clean_dirty_briefs with apply=true to archive them.`,
+        }],
+      };
+    }
+
+    let archived = 0;
+    let deleted = 0;
+    for (const row of rows) {
+      try {
+        archiveDirtyBriefAsset(row);
+        archived += 1;
+      } catch (err) {
+        console.error("[recallnest] Failed to archive dirty brief:", err instanceof Error ? err.message : String(err));
+      }
+      try {
+        deleted += await store.bulkDelete([row.scope]);
+      } catch (err) {
+        console.error("[recallnest] Failed to delete dirty brief index rows:", err instanceof Error ? err.message : String(err));
+      }
+    }
+
+    return {
+      content: [{
+        type: "text" as const,
+        text: `Dirty briefs: ${rows.length}\nArchived: ${archived}\nIndex rows deleted: ${deleted}`,
+      }],
+    };
+  }
+);
+
+registerTool(
+  "consolidate_memories",
+  "Run semantic consolidation: cluster similar memories, merge near-duplicates, and detect contradictions. Side effect: when apply=true, archives merged entries. Use when a scope has grown large and needs deduplication.",
+  {
+    scope: z.string().min(1).max(160).describe("Scope to consolidate (e.g. project:recallnest)"),
+    clusterThreshold: z.number().min(0.5).max(1.0).default(0.82).describe("Min similarity to form a cluster (default 0.82)"),
+    mergeThreshold: z.number().min(0.5).max(1.0).default(0.92).describe("Min similarity to merge/archive (default 0.92)"),
+    maxEntries: z.number().min(10).max(2000).default(500).describe("Max entries to scan (default 500)"),
+    apply: z.boolean().default(false).describe("When false, preview only (scan + report without archiving). When true, actually merge/archive."),
+  },
+  async ({ scope, clusterThreshold, mergeThreshold, maxEntries, apply }) => {
+    const { store } = getComponents();
+
+    if (!apply) {
+      // Dry-run: use a read-only wrapper that blocks writes
+      const readOnlyStore = {
+        list: store.list.bind(store),
+        getById: store.getById.bind(store),
+        vectorSearch: store.vectorSearch.bind(store),
+        update: async () => null, // no-op in dry-run
+      };
+      const engine = new ConsolidationEngine(readOnlyStore, { clusterThreshold, mergeThreshold, maxEntriesPerRun: maxEntries }, getKGStore());
+      const result = await engine.run(scope);
+      return {
+        content: [{
+          type: "text" as const,
+          text: `[DRY-RUN] ${formatConsolidationResult(result)}\n\nRe-run with apply=true to execute merges.`,
+        }],
+      };
+    }
+
+    // P0-1: serialize applied consolidation per scope across the 11 mcp-server processes.
+    // Skip (not queue) if another process is already consolidating this scope — the
+    // store-write lock already prevents write corruption; this just avoids redundant work.
+    const outcome = await withLock(
+      `consolidate-${scope}`,
+      async () => {
+        const engine = new ConsolidationEngine(store, { clusterThreshold, mergeThreshold, maxEntriesPerRun: maxEntries }, getKGStore());
+        return engine.run(scope);
+      },
+      { onBusy: "skip", expireMs: 600_000 },
+    );
+    if (!outcome.ran) {
+      return {
+        content: [{
+          type: "text" as const,
+          text: `⏭️ consolidate_memories skipped: another process is consolidating scope "${scope}".`,
+        }],
+      };
+    }
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatConsolidationResult(outcome.result),
+      }],
+    };
+  }
+);
+
+registerTool(
+  "scan_skill_promotions",
+  "Scan cases and patterns in a scope for potential promotion to reusable skills. Read-only. Use when you want to discover recurring procedures that deserve formalization as skills.",
+  {
+    scope: z.string().min(1).max(160).describe("Project scope to scan for promotion candidates"),
+    minOccurrences: z.number().min(2).max(20).default(3).describe("Minimum similar cases to trigger a promotion suggestion"),
+  },
+  async ({ scope, minOccurrences }) => {
+    const { store } = getComponents();
+    const result = await scanForPromotions(store, scope, {
+      minCaseOccurrences: minOccurrences,
+    });
+
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatPromotionResult(result),
+      }],
+    };
+  },
+);
+
+registerTool(
+  "manage_alias",
+  "Manage user-level query alias rules (data/alias-map.json). Matching aliases append canonical tokens to the query in the BM25 channel only ('我的桥' → '+telegram-ai-bridge') — they deliberately do NOT alter the vector embedding, so casual aliases cannot cause semantic drift. Actions: add/remove/list, plus explain to debug which rules (builtin full-channel + user BM25-only) fire for a query. Use when a colloquial entity name fails to recall its canonical project/tool.",
+  {
+    action: z.enum(["add", "remove", "list", "explain"]).describe("add: upsert a rule; remove: delete by trigger; list: all rules (builtin + user); explain: show which rules match a query"),
+    trigger: z.string().min(1).max(60).optional().describe("Alias literal to match in queries, case-insensitive inclusion (required for add/remove)"),
+    expansions: z.array(z.string().min(1).max(60)).max(8).optional().describe("Canonical tokens to append (required for add, max 8)"),
+    query: z.string().min(1).max(500).optional().describe("Query to explain (required for explain)"),
+  },
+  async ({ action, trigger, expansions, query }) => {
+    const text = (() => {
+      switch (action) {
+        case "add": {
+          if (!trigger || !expansions?.length) throw new Error("invalid input: manage_alias add 需要 trigger 与 expansions");
+          const result = upsertUserAlias(trigger, expansions);
+          return result.ok
+            ? `${result.action}: "${result.entry.trigger}" → [${result.entry.expansions.join(", ")}](写入 ${aliasMapFilePath()},BM25 通道立即生效)`
+            : `拒绝: ${result.reason}`;
+        }
+        case "remove": {
+          if (!trigger) throw new Error("invalid input: manage_alias remove 需要 trigger");
+          return removeUserAlias(trigger)
+            ? `removed: "${trigger}"`
+            : `未找到 trigger "${trigger}"(只能删用户规则;builtin 规则在 src/aliases.ts)`;
+        }
+        case "list": {
+          const user = listUserAliases();
+          return [
+            `Builtin rules (${ALIAS_RULES.length}, read-only, full-channel vector+BM25):`,
+            ...ALIAS_RULES.map(r => `  /${r.pattern.source}/ → ${r.expansion}${r.comment ? ` (${r.comment})` : ""}`),
+            ``,
+            `User rules (${user.length}, BM25-only, ${aliasMapFilePath()}):`,
+            ...(user.length ? user.map(r => `  "${r.trigger}" → [${r.expansions.join(", ")}]`) : ["  (none)"]),
+          ].join("\n");
+        }
+        case "explain": {
+          if (!query) throw new Error("invalid input: manage_alias explain 需要 query");
+          const builtinMatched = explainAliasExpansion(query);
+          const userMatched = explainUserAliases(query);
+          return [
+            `Query: ${query}`,
+            `Full-channel expanded (builtin): ${expandQueryWithAliases(query) === query ? "(unchanged)" : expandQueryWithAliases(query)}`,
+            `BM25-channel expanded (builtin synonyms + user aliases): ${expandQuery(query) === query ? "(unchanged)" : expandQuery(query)}`,
+            `Matched builtin rules (${builtinMatched.length}):`,
+            ...builtinMatched.map(m => `  /${m.pattern}/ → ${m.expansion}`),
+            `Matched user aliases (${userMatched.length}):`,
+            ...userMatched.map(m => `  "${m.trigger}" → [${m.expansions.join(", ")}]`),
+          ].join("\n");
+        }
+      }
+    })();
+
+    return { content: [{ type: "text" as const, text }] };
+  },
+);
+
+registerTool(
+  "promote_scan",
+  "Scan transcript-downgraded evidence (profile/preferences sunk to events) and auto-promote recurring high-importance facts back to durable memory. Default dry-run (no writes); when dryRun=false it creates durable entries (dedup is idempotent). Use to backfill sparse durable profile/preferences from accumulated transcript evidence.",
+  {
+    scope: z.string().min(1).max(160).describe("Scope to scan, e.g. cc:project:recallnest"),
+    dryRun: z.boolean().default(true).describe("When true (default), only report candidates without writing"),
+    minOccurrences: z.number().int().min(2).max(20).default(3).describe("Min cluster members before promoting"),
+    minImportance: z.number().min(0).max(1).default(0.6).describe("Min average cluster importance to promote"),
+  },
+  async ({ scope, dryRun, minOccurrences, minImportance }) => {
+    const { store, embedder } = getComponents();
+    const kgExtractor = getKGExtractor();
+    const deps = buildPromoteScanDeps({ store, embedder, conflictStore, kgExtractor });
+    const result = await scanMemoryPromotions(deps, scope, { dryRun, minOccurrences, minImportance });
+
+    return {
+      content: [{
+        type: "text" as const,
+        text: formatPromoteScanResult(result),
+      }],
+    };
+  },
+);
+}

package/src/memory-boundaries.ts

@@ -61,8 +61,29 @@ export interface IngestBoundaryResolution {
   boundary: MemoryBoundaryMetadata;
 }
 
-const TRANSCRIPT_SOURCES = new Set(["cc", "codex", "gemini"]);
-const TRANSCRIPT_SCOPE_PREFIXES = ["cc:", "codex:", "gemini:"];
+// 在栈各端的 transcript 来源与 scope 前缀。
+// ⚠️ 接入新的 AI 端时必须同步这两处 —— 漏掉的端不会被认成 evidence 层，
+// 其会话碎片既不会被降权，也会被 layer admission 当成 durable 结论放行。
+// （kimi / antigravity 就漏了：kimi 2026-07-19 进栈、antigravity 07-29 进栈，
+//   直到 08-01 才发现库里的 `kimi:` scope 一直被当结论对待。）
+//
+// minis（2026-08-12 加，iPhone 上的 Minis agent，经 iCloud 文件信箱回流对话记录）：
+// 这一次是**接入前先补**，不是事后发现——补的动机值得记下来，因为 Minis 比前面几端更危险。
+// 它的 CLI 把 agent 自己的工具调用回传也标成 `type:"user"` / `role:"user"`（实测一场对话
+// 137 条里 69 条 user，其中仅 11 条是真人发言，另外 58 条是 `[Tool result: ...]`）。
+// 所以漏掉它不只是"碎片被当结论"，而是**把 shell 输出提炼成用户说过的话**——
+// 例如 `ssh: connect to host ... timed out` 会变成一条"用户说过"的记忆。
+// 导出侧已在上游过滤（只导真人对话），这里是第二道闸：即使某次导出没过滤干净，
+// 走 ingest 的内容也会被正确降权到 evidence 层，而不是当成 durable 结论。
+const TRANSCRIPT_SOURCES = new Set(["cc", "codex", "gemini", "kimi", "antigravity", "minis"]);
+const TRANSCRIPT_SCOPE_PREFIXES = [
+  "cc:",
+  "codex:",
+  "gemini:",
+  "kimi:",
+  "antigravity:",
+  "minis:",
+];
 
 export function getConflictPolicyForCategory(category: DurableMemoryCategory): MemoryConflictPolicy {
   return category === "events" || category === "cases"
@@ -135,7 +156,9 @@ export function buildDefaultCanonicalKey(params: {
           ? [params.category, slot.type, slot.brand, slot.item]
           : slot.type === "reply-style"
             ? [params.category, slot.type, ...slot.traits]
-            : [params.category, slot.type, slot.preferredTool, "over", slot.avoidedTool],
+            : slot.type === "implicit-usage"
+              ? [params.category, slot.type, slot.subject]
+              : [params.category, slot.type, slot.preferredTool, "over", slot.avoidedTool],
       );
     }
   }

package/src/memory-evolution.ts

@@ -13,6 +13,7 @@
  */
 
 import { parseEmotion, isEmotionScoringEnabled } from "./memory-schema.js";
+import { computeUsageStatus } from "./usage-tracker.js";
 
 // ---------------------------------------------------------------------------
 // Types
@@ -328,6 +329,16 @@ export function computeDecayScore(
   importance: number,
   now?: number,
   metadata?: string,
+): number {
+  return computeDecayScoreInternal(evo, importance, now, metadata, 1);
+}
+
+function computeDecayScoreInternal(
+  evo: EvolutionMetadata,
+  importance: number,
+  now: number | undefined,
+  metadata: string | undefined,
+  frequencyFactor: number,
 ): number {
   const ts = now ?? Date.now();
 
@@ -340,7 +351,7 @@ export function computeDecayScore(
   const recencyBoost = evo.lastAccessedAt
     ? Math.pow(0.5, Math.max(0, (ts - evo.lastAccessedAt) / 86_400_000) / 30) // 30-day half-life for recency
     : 0.5; // never accessed → neutral
-  const frequencyScore = Math.min(1, rawFreq * recencyBoost);
+  const frequencyScore = Math.min(1, rawFreq * recencyBoost) * Math.max(0, Math.min(1, frequencyFactor));
 
   const clampedImportance = Math.max(0, Math.min(1, importance));
 
@@ -363,3 +374,64 @@ export function computeDecayScore(
     IMPORTANCE_WEIGHT_BASE * clampedImportance
   );
 }
+
+function parseMetadataObject(metadata: string | undefined): Record<string, unknown> {
+  if (!metadata) return {};
+  try {
+    const parsed: unknown = JSON.parse(metadata);
+    return parsed !== null && typeof parsed === "object" && !Array.isArray(parsed)
+      ? parsed as Record<string, unknown>
+      : {};
+  } catch {
+    return {};
+  }
+}
+
+function readUsageCounts(meta: Record<string, unknown>): { injection: number; use: number } {
+  const usage = meta.usage;
+  const usageObj = usage && typeof usage === "object" && !Array.isArray(usage)
+    ? usage as Record<string, unknown>
+    : {};
+  return {
+    injection: typeof meta.accessCount === "number" ? meta.accessCount : 0,
+    use: typeof usageObj.useCount === "number" ? usageObj.useCount : 0,
+  };
+}
+
+function isPendingProspectiveReminder(meta: Record<string, unknown>, now: number): boolean {
+  const prospective = meta.prospective;
+  if (!prospective || typeof prospective !== "object" || Array.isArray(prospective)) return false;
+  const obj = prospective as Record<string, unknown>;
+  if (obj.status !== "pending") return false;
+  if (typeof obj.expiresAt !== "string" || obj.expiresAt.length === 0) return true;
+  const expiresAt = Date.parse(obj.expiresAt);
+  return Number.isNaN(expiresAt) || expiresAt >= now;
+}
+
+function isUsageDecayProtected(meta: Record<string, unknown>, importance: number, now: number): boolean {
+  if (importance >= 0.95) return true;
+  if (typeof meta.promotedTo === "string" && meta.promotedTo.trim().length > 0) return true;
+  return isPendingProspectiveReminder(meta, now);
+}
+
+/**
+ * GC-only usage-adjusted decay score.
+ *
+ * Cold means "repeatedly surfaced but never actually cited"; in that case the
+ * frequency term is discounted instead of rewarding unhelpful exposure. This is
+ * intentionally separate from computeDecayScore(), which is still used by
+ * online retrieval ranking.
+ */
+export function computeUsageAdjustedDecayScore(
+  evo: EvolutionMetadata,
+  importance: number,
+  now?: number,
+  metadata?: string,
+): number {
+  const ts = now ?? Date.now();
+  const meta = parseMetadataObject(metadata);
+  const counts = readUsageCounts(meta);
+  const status = computeUsageStatus(counts.injection, counts.use);
+  const factor = status === "cold" && !isUsageDecayProtected(meta, importance, ts) ? 0.2 : 1;
+  return computeDecayScoreInternal(evo, importance, ts, metadata, factor);
+}

package/src/memory-health-rebalance.ts

@@ -125,6 +125,10 @@ function computeTierAccessSignal(
   }
 }
 
+// ⚠️ 约束（2026-07-21 通用Wiki上下文记忆Agent方案借鉴审计 P1-4）：本函数用 accessCount 参与
+// importance 计算（importance←accessSignal，越常访问 importance 越高的马太逻辑）。当前为孤儿代码——
+// 全树仅测试引用、生产侧零 import。接线前须评估与「usage 只进检索先验、绝不抬 importance/事实置信」
+// 立场的冲突：频次抬 importance 会自增强（越常召回越权威），违反该立场。接线须 Alice 明确批准。
 export function computeRebalancedImportance(
   currentImportance: number,
   tier: MemoryTier,