fix(plugin): remove misleading recall top score (#4053)

* fix(plugin): report server recall top score

Derive the status snapshot's top score from server-assembled context so /ov no longer reports 0.00 for scored recalls.

Co-Authored-By: Claude Sonnet 4.6 noreply@anthropic.com

* test(plugin): remove redundant recall state tests

* fix(plugin): remove misleading recall top score

---------

Co-authored-by: qin-ctx <qinhaojie.exe@bytedance.com>

examples/claude-code-memory-plugin/README.md

@@ -270,7 +270,7 @@ The plugin renders a one-line status of OpenViking under your Claude Code input
 Examples:
 
 ```text
-OV ✓ │ Fable 5 · ctx 42% │ ↩ 6 mem (0.92) · 50ms   6 memories injected; model + context usage
+OV ✓ │ Fable 5 · ctx 42% │ ↩ 6 mem · 50ms          6 memories injected; model + context usage
 OV ⚠ slow                                  probe missed the 1 s budget (server may be lagging)
 OV ✗ offline                               server unreachable
 OV ⚡ bypass │ Fable 5 · ctx 42%            OPENVIKING_BYPASS_SESSION* matched

examples/claude-code-memory-plugin/README_CN.md

@@ -264,7 +264,7 @@ bypass 命中时所有 hook 直接放行，不联系 OpenViking。
 示例：
 
 ```text
-OV ✓ │ Fable 5 · ctx 42% │ ↩ 6 mem (0.92) · 50ms   注入 6 条记忆；模型 + 上下文占比
+OV ✓ │ Fable 5 · ctx 42% │ ↩ 6 mem · 50ms          注入 6 条记忆；模型 + 上下文占比
 OV ⚠ slow                                  探针超过 1s 预算（服务器可能在抽风）
 OV ✗ offline                               服务器不可达
 OV ⚡ bypass │ Fable 5 · ctx 42%            命中 OPENVIKING_BYPASS_SESSION*

examples/claude-code-memory-plugin/STATUSLINE.md

@@ -42,7 +42,7 @@ The default composition runs left to right, joined by ` │ `. Segments are cond
 |               | `OV ✗ offline`         | red    | probe errored (refused, DNS fail, network down)                                                                       |
 |               | `OV ⚡ bypass`         | yellow | session matched `OPENVIKING_BYPASS_SESSION` or `*_PATTERNS`                                                           |
 | Model · ctx   | `Fable 5 · ctx 42%`    | dim / by usage | always, when CC's statusline payload carries `model` / `context_window`. Registering a custom statusLine replaces CC's native context indicator, so this reproduces it: `<70%` dim, `70–89%` yellow, `≥90%` red (native thresholds). Hide with `OPENVIKING_STATUSLINE_CTX=off`. Shows in bypass mode too — it describes the CC conversation, not OV |
-| Recall        | `↩ 6 mem (0.92) · 50ms` | dim    | last user prompt actually injected memories. `(0.92)` is the top similarity score among picked items; latency is the recall round-trip |
+| Recall        | `↩ 6 mem · 50ms`        | dim    | last user prompt actually injected memories; latency is the recall round-trip                                                    |
 | Capture       | `✎ 573/20k · 2 arch`   | dim    | tokens pending toward the next archive (sawtooth — resets on commit), `2 arch` = archives produced this session       |
 |               | `✎ committed · 2 arch` | dim    | the turn that just finished produced an archive                                                                       |
 |               | `✎ 2 arch`             | dim    | nothing pending, but archives already exist this session                                                              |
@@ -127,7 +127,6 @@ For an assistant adding a segment that consumes existing state. All files live u
 {
   "reason": "ok" | "bypass" | "offline" | "no_results" | "filtered_out" | "short_query" | ...,
   "count": 6,                        // memories actually injected
-  "top_score": 0.92,                 // max score among picked items
   "latency_ms": 180,
   "cc_session_id": "ff875009-...",
   "ts": 1778139288759

examples/claude-code-memory-plugin/scripts/auto-recall.mjs

@@ -399,7 +399,6 @@ async function main() {
       hint_items: 0,
       tokens_used: estimateTokens(endpointBlock),
       tokens_budget: cfg.recallTokenBudget,
-      top_score: 0,
       cc_session_id: sessionId,
       reason: "ok",
     });
@@ -436,14 +435,12 @@ async function main() {
   }
 
   const built = await buildInjectionBlock(picked, effectivePeer.peerId);
-  const topScore = picked.reduce((m, it) => Math.max(m, clampScore(it.score)), 0);
   writeRecallState({
     count: picked.length,
     content_items: built?.contentCount ?? 0,
     hint_items: built?.hintCount ?? 0,
     tokens_used: built?.budgetUsed ?? 0,
     tokens_budget: cfg.recallTokenBudget,
-    top_score: topScore,
     cc_session_id: sessionId,
     reason: "ok",
   });

examples/claude-code-memory-plugin/scripts/ov-status.mjs

@@ -6,7 +6,7 @@
  *   - Server URL + /health probe
  *   - Resolved identity (account/user)
  *   - Last session-start injection (size, age, audit path)
- *   - Last auto-recall (item count, top score, token budget use)
+ *   - Last auto-recall (item count, token budget use)
  *   - Toggle state for the three injection paths
  *   - Auth source — which file/env actually drove url + api_key, mirroring
  *     config.mjs's priority chain (env → ovcli.conf → ov.conf → default)
@@ -94,12 +94,11 @@ async function main() {
   // 3. Last auto-recall
   const recall = readJsonState("last-recall.json");
   if (recall) {
-    const top = typeof recall.top_score === "number" ? recall.top_score.toFixed(2) : "?";
     const used = recall.tokens_used ?? 0;
     const budget = recall.tokens_budget ?? 0;
     console.log(
       `Last auto-recall: ${fmtAge(recall.ts)} — ${recall.count ?? 0} items, ` +
-      `top ${top}, ${used}/${budget} tokens (${recall.reason || "ok"})`,
+      `${used}/${budget} tokens (${recall.reason || "ok"})`,
     );
   } else {
     console.log("Last auto-recall: (none yet)");

examples/claude-code-memory-plugin/scripts/statusline.mjs

@@ -189,16 +189,9 @@ async function main() {
 
   // Recall summary: only meaningful when we actually injected memories this
   // turn. Skip the segment for empty/bypass/no-results reasons to keep the
-  // line tight. The (0.92) trailing parens is the top score among picked
-  // items — quality hint without an extra segment. Token/char count is
-  // omitted: the only number we have is a chars/4 heuristic, which is
-  // misleading enough on CJK text that displaying it does more harm than
-  // good. Count + score + latency convey the relevant signal.
+  // line tight. Count confirms the injection; latency shows its round-trip cost.
   if (recall && recall.reason === "ok" && recall.count > 0) {
-    const top = typeof recall.top_score === "number" && recall.top_score > 0
-      ? ` (${recall.top_score.toFixed(2)})`
-      : "";
-    const seg = `↩ ${recall.count} mem${top}`
+    const seg = `↩ ${recall.count} mem`
       + (typeof recall.latency_ms === "number" ? ` · ${recall.latency_ms}ms` : "");
     parts.push(dim(seg));
   }