@heixui/dsh-session-suite 0.4.4 → 0.5.0

package/client.js

@@ -270,6 +270,8 @@ window.__ModuleLoader__.load({
         'memory.extractionTurns': '{n} turn(s) ended',
         'memory.extractionRegistered': 'hook registered',
         'memory.extractionNotRegistered': 'HOOK NOT REGISTERED',
+        'memory.extractionNoTrace': 'no extraction trace recorded by this Host',
+        'memory.diagTitle': 'Extraction diagnostics:',
         'memory.accept': 'Keep',
         'memory.reject': 'Discard',
         'memory.accepted': 'Kept',
@@ -516,6 +518,8 @@ window.__ModuleLoader__.load({
         'memory.extractionTurns': '结束 {n} 轮',
         'memory.extractionRegistered': '钩子已注册',
         'memory.extractionNotRegistered': '钩子未注册',
+        'memory.extractionNoTrace': '本次运行没有记录到提取痕迹',
+        'memory.diagTitle': '提取诊断：',
         'memory.accept': '保留',
         'memory.reject': '丢弃',
         'memory.accepted': '已保留',
@@ -1026,6 +1030,23 @@ window.__ModuleLoader__.load({
   overflow-y: auto;
   overscroll-behavior: contain;
 }
+/* The extraction diagnostic. Deliberately larger than the meta text around it and set
+   on its own tinted strip: it exists to answer "why is the queue empty", and the first
+   version was 11px grey text inside a bordered box that the user reported not seeing
+   at all. A diagnostic that cannot be found is not a diagnostic. */
+.dss-extraction-diag {
+  display: flex; flex-wrap: wrap; gap: 6px 16px; align-items: baseline;
+  margin-top: 8px; padding: 8px 12px;
+  border-radius: 8px;
+  background: var(--dsw-alias-bg-layer-2, #f2f3f5);
+  font-size: 12.5px; line-height: 1.5;
+  color: var(--dsw-alias-label-secondary, #4a4a4a);
+  font-variant-numeric: tabular-nums;
+}
+.dss-diag-title {
+  font-weight: 600;
+  color: var(--dsw-alias-label-primary, #1f1f1f);
+}
 /* The summary charts. Two narrow panels for composition and health, one wide panel
    for tags: composition and health are short lists, tags can run long. Bars are divs,
    not SVG, so they inherit the theme and stay legible in both palettes. */
@@ -3511,21 +3532,40 @@ window.__ModuleLoader__.load({
               // `registered` is the third state: the hook was never installed at all
               // (the optional registration threw), which is a different bug from
               // "installed but never called".
-              extraction
-                ? h(
-                    'div',
-                    { className: 'dss-row-meta' },
-                    h('span', null, t('memory.extractionEvents', { n: extraction.eventsSeen ?? 0 })),
-                    h('span', null, t('memory.extractionTurns', { n: extraction.turnEndsSeen ?? 0 })),
-                    h(
+              //
+              // This block renders even when `extraction` is null, and says so. The
+              // previous version was gated on `extraction` being present — so a Host
+              // that never recorded a trace (registration failed, or the status
+              // endpoint reads a different engine) showed nothing at all, which is
+              // exactly the case the diagnostic existed to expose.
+              //
+              // Rendered at a readable size rather than in the row-meta micro type:
+              // it is a diagnostic, and a diagnostic nobody can find is the same as
+              // no diagnostic. The 11px meta line was reported as "I don't see it".
+              h(
+                'div',
+                { className: 'dss-extraction-diag' },
+                h('span', { className: 'dss-diag-title' }, t('memory.diagTitle')),
+                h(
+                  'span',
+                  null,
+                  extraction
+                    ? t('memory.extractionEvents', { n: extraction.eventsSeen ?? 0 })
+                    : t('memory.extractionNoTrace'),
+                ),
+                extraction
+                  ? h('span', null, t('memory.extractionTurns', { n: extraction.turnEndsSeen ?? 0 }))
+                  : null,
+                extraction
+                  ? h(
                       'span',
                       null,
                       extraction.registered
                         ? t('memory.extractionRegistered')
                         : t('memory.extractionNotRegistered'),
-                    ),
-                  )
-                : null,
+                    )
+                  : null,
+              ),
             )
 
       // The list gets its own scroll box.

package/cordis.patch.yml

@@ -53,6 +53,14 @@
         # the budget would claim six memories are relevant when the evidence
         # says otherwise.
         recallLimit: 6
+        # How much of the conversation before the current message is folded into the
+        # retrieval query. A follow-up like "那它为什么慢？" has no referent for "它", so
+        # searching those characters alone cannot find the memory that answers it.
+        # The previous user turns are appended BEFORE the current question, so the
+        # question's own terms still dominate the score. 0 disables the behaviour and
+        # searches the current message exactly as before.
+        contextQueryChars: 160
+        contextQueryTurns: 3
         # How many of the most important memories ride along regardless of the
         # question. Without a resident layer, an ordinary follow-up like "继续"
         # shares no vocabulary with any memory and a strict relevance filter

package/lib/memory-host.js

@@ -175,7 +175,13 @@ export function installMemory({ ctx, root, config, log }) {
             log(`explicit memory stored (${instruction.scope})`)
           }
 
-          const question = lastUserText(messages)
+          // The question, plus enough of the conversation before it to resolve a
+          // short follow-up. See contextualQuery: a bare 「那它为什么慢？」 has no
+          // referent for 「它」 and cannot match the memory that answers it.
+          const question = contextualQuery(messages, {
+            chars: typeof config?.contextQueryChars === 'number' ? config.contextQueryChars : undefined,
+            turns: typeof config?.contextQueryTurns === 'number' ? config.contextQueryTurns : undefined,
+          })
           const text = await buildRecallText({ engine, cwd, question, config })
           if (agent) recallByScope.set(agent, { text, at: Date.now() })
           lastRecall = text
@@ -495,6 +501,80 @@ function lastUserText(messages) {
   return ''
 }
 
+/**
+ * The current question plus the tail of the conversation before it.
+ *
+ * Retrieval used to search for the last user message alone, so a follow-up like
+ * 「那它为什么慢？」 searched for those seven characters and nothing else. "它" is a
+ * pronoun with no referent in the query, so the memory that explains the slowness could
+ * not be found no matter how well it was written. That is a real gap, not a tuning
+ * problem.
+ *
+ * The shape here is deliberate:
+ *
+ *   • the current question comes LAST, so it dominates the lexical score (BM25 and the
+ *     coverage route both reward a match in the query; putting context first would let
+ *     an older turn outvote the thing being asked now);
+ *   • only the previous USER turns are included, not assistant replies. Assistant text
+ *     is verbose, and a long reply would drown the query in words the user never typed;
+ *   • the tail is capped in characters rather than turns, because one long message and
+ *     ten short ones should not cost the same;
+ *   • nothing is included when the current question already stands alone (long enough
+ *     and not pronoun-led), so an ordinary self-contained question is searched exactly
+ *     as before.
+ *
+ * This is a heuristic, not a language model. It will not resolve 「那个东西」 back to a
+ * noun three turns ago. It does fix the common case — a short follow-up — and it costs
+ * nothing at query time.
+ *
+ * @param {readonly any[]} messages step messages
+ * @param {{ chars?: number, turns?: number }} [limits] budget
+ * @returns {string} query text
+ */
+function contextualQuery(messages, limits = {}) {
+  const current = lastUserText(messages)
+  if (!current) return ''
+
+  const maxChars = limits.chars ?? 160
+  const maxTurns = limits.turns ?? 3
+
+  // A self-contained question needs no help. The test is NOT length alone: a short
+  // question that names its own subject ("导出的图片为什么断链") is self-contained,
+  // while a long one made of pronouns ("那这个呢，还有刚才说的那个，它到底是怎么回事")
+  // is not. Pronouns are the actual signal — they are what a query cannot resolve.
+  //
+  // An earlier version used `length >= 24`, which was a guess and misclassified a
+  // perfectly clear 20-character question as needing context. Pronoun presence is a
+  // narrower, more defensible rule: it fires exactly when the query contains a word
+  // that refers to something outside itself.
+  const PRONOUNS = /(它|他|她|这个|那个|这些|那些|该|此|上述|前面|刚才|上面|之前那个)/
+  const looksSelfContained = !PRONOUNS.test(current)
+
+  const earlier = []
+  let budget = maxChars
+  let seen = 0
+  for (let index = (messages?.length ?? 0) - 1; index >= 0 && seen < maxTurns; index -= 1) {
+    const message = messages[index]
+    if (message?.role !== 'user') continue
+    const text = contentText(message.content)
+    if (!text) continue
+    seen += 1
+    // The newest user message IS `current`; skip it so it is not duplicated.
+    if (text === current) continue
+    if (budget <= 0) break
+    const slice = text.length > budget ? text.slice(-budget) : text
+    earlier.unshift(slice)
+    budget -= slice.length
+  }
+
+  // Nothing to add, or the question stands on its own: search it as-is.
+  if (earlier.length === 0 || looksSelfContained) return current
+
+  // Context first, question last. The engine tokenises the whole string, and the
+  // question's own terms are what must win.
+  return `${earlier.join(' ')} ${current}`.trim()
+}
+
 /**
  * Ask a small model call what, if anything, is worth remembering from this turn.
  *

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@heixui/dsh-session-suite",
-  "version": "0.4.4",
+  "version": "0.5.0",
   "description": "Session checkpoints for DSH: rewind a conversation with its files and memory, permanent session deletion, and shareable exports",
   "type": "module",
   "main": "./index.js",