[hijzy]

No description provided.

[Memtensor-AI]

🤖 Open Code Review

Target: PR #726
Task: 351be3e046a117e0
Base: 1.2.401
Head: 1.2.401

🔍 OpenCodeReview found 5 issue(s) in this PR.

---

1. Memory/src/service/read-model/panel-read.ts (L410-L412)

rerelyUsedSkills now issues two extra queries that can each pull 10k full rows (listRecallEvents uses SELECT *, so query, request_json and dropped_json are parsed for every row), and then discards everything outside the 7-day window in JS. Because LIMIT is applied before the date filter, a busy instance silently truncates the window and reports a low count.

Pass the oldest date key (dates[0]) as a SQL-level lower bound (and add created_at indexes) so the database can bound the range instead of relying on a row cap.

---

2. Memory/src/storage/repositories.ts (L3107-L3111)

This read is used for an aggregate count, but SELECT * hydrates the whole RecallEventRecord (query text, request payload, dropped list) for up to 10k rows. recall_events only has idx_recall_events_session_created (leading session_id), so ORDER BY created_at DESC without a session filter requires a full scan plus sort.

For the panel path, consider a narrow aggregate read that projects only created_at and hit_memory_ids_json, plus a created_at index. If this generic method is kept, an optional since clause would let callers push the time window into SQL.

---

3. Memory/src/service/read-model/panel-read.ts (L712)

Summing both signals double-counts a single use: whenever the skill-invoke path also records a recall event (retrievalMode: "skill_invoke"), the same skill contributes once through invoked and again through recalled. It also treats a skill that merely appeared in memory_search results — never applied — as a "use".

When the label is "skill uses", consider counting skill trials only, or deduplicate per skill per turn/session instead of adding the two totals.

---

4. Memory/src/service/read-model/panel.ts (L103)

panelSourceForStatsRow never reads id; making it required on MemoryStatsRow forces this caller to pass a field the helper ignores. Narrowing the parameter keeps the projection honest while the aggregate read model still gets id where it is actually needed.

💡 Suggested Change

Before:

    id: memory.id,

After:

function panelSourceForStatsRow(memory: Omit<MemoryStatsRow, "id">): string {

---

5. Memory/src/service/read-model/panel-read.ts (L410-L413)

The if (input.activeSkillIds.size === 0) return 0; guard inside panelActivatedSkillUseCount cannot save any work: both queries are evaluated as arguments before the helper is entered, so an installation with zero activated skills still pays for two 10k-row reads on every analytics refresh. Short-circuit before the calls (e.g. only build the recentlyUsedSkills input when activeSkillIds.size > 0, otherwise pass empty arrays) so the guard actually avoids the queries.

💡 Suggested Change

Before:

    const recentlyUsedSkills = panelActivatedSkillUseCount({
      activeSkillIds,
      recallEvents: this.deps.repos.runtime.listRecallEvents({ limit: PANEL_USAGE_EVENT_LIMIT }),
      skillTrials: this.deps.repos.runtime.listSkillTrials({ limit: PANEL_USAGE_EVENT_LIMIT }),

After:

    const recentlyUsedSkills = activeSkillIds.size === 0
      ? 0
      : panelActivatedSkillUseCount({
          activeSkillIds,
          recallEvents: this.deps.repos.runtime.listRecallEvents({ limit: PANEL_USAGE_EVENT_LIMIT }),
          skillTrials: this.deps.repos.runtime.listSkillTrials({ limit: PANEL_USAGE_EVENT_LIMIT }),
          dates,
          timeZone
        });

Generated by cloud-assistant via Open Code Review.