@inetafrica/open-claudia 2.6.4 → 2.6.6

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## v2.6.6
+- **Recall: recent-turns window as judge-gated context.** v2.6.5 fixed quotes, but a bare follow-up like "Push" with no quote still recalled nothing — recall only ever saw the current message. The last ~6 user/assistant turns on this channel (read from the project transcript, current turn excluded) now feed recall the same way the quote does: keyword-matched as origin "context", kept only when the haiku judge confirms relevance, dropped on fail-open. A stale thread can never force-inject packs on keywords alone, but terse follow-ups finally inherit the topic of the conversation.
+
+## v2.6.5
+- **Recall: replied-to quote becomes judge-gated context.** v2.6.4 stripped reply-quotes from matching entirely, which threw away real signal — replying "Push" to a release message recalled nothing. The quoted text now drives keyword matching as origin-"context" candidates that only survive when the haiku relevance judge confirms them (the judge sees the quote to resolve references like "it"/"the app"); on judge failure context-derived candidates are dropped while the user's-own-words baseline is kept, so a quote can never force-inject on keywords alone.
+
 ## v2.6.4
 - **Recall matcher: kill false-positive pack/entity injections.** Three compounding fixes after Sharon/Euvencia/higgsfield got recalled on turns that had nothing to do with them:
   - **Match only the user's own words.** The FTS routers used the full composed prompt — so a Telegram reply-quote fed the entire quoted message (~40 terms) into matching, and any entity named in the quote scored straight past the threshold. The reply-quote block and any 📖 recall-announcement echoes are now stripped before matching (`recallMatchText`); the verbatim prompt still goes to the model untouched.

package/core/recall-filter.js

@@ -10,6 +10,9 @@ const { spawnSubagent } = require("./subagent");
 const FILTER_MODEL = process.env.RECALL_FILTER_MODEL || "haiku";
 const FILTER_TIMEOUT_MS = Number(process.env.RECALL_FILTER_TIMEOUT_MS || 20000);
 const MESSAGE_CLIP_CHARS = 1500;
+// Quote + a window of recent turns: roomier than the message clip so the
+// newest history lines at the tail don't get sliced off.
+const CONTEXT_CLIP_CHARS = 3500;
 
 function enabled() {
   return (process.env.RECALL_FILTER || "on").toLowerCase() !== "off";
@@ -20,21 +23,36 @@ const FILTER_SYSTEM_PROMPT = [
   "Reply with ONLY a JSON array of candidate ids — no prose, no markdown fences.",
 ].join("\n");
 
-// candidates: [{ id, name, description }]. Returns a Set of kept ids,
-// or null to signal fail-open (keep everything).
-async function judgeRelevance(message, candidates) {
+// candidates: [{ id, name, description }]. opts.context: surrounding
+// conversation text (the message the user replied to and/or the last
+// few chat turns) used to disambiguate pronouns and terse follow-ups
+// like "push" or "the app" — context only, not a topic source in its
+// own right. Returns a Set of kept ids, or null to signal fail-open
+// (caller decides what that keeps).
+async function judgeRelevance(message, candidates, opts = {}) {
   if (!enabled() || candidates.length === 0) return null;
   const msg = String(message || "").slice(0, MESSAGE_CLIP_CHARS);
+  const context = String(opts.context || "").slice(0, CONTEXT_CLIP_CHARS);
   const lines = candidates.map((c) => `- ${c.id}: ${c.name}${c.description ? ` — ${c.description}` : ""}`);
   const prompt = [
     "A user sent this message to their assistant:",
     "<<<",
     msg,
     ">>>",
+    ...(context
+      ? [
+          "",
+          "Surrounding conversation (the message it replied to and/or the most recent turns in this chat — context for resolving references like \"the app\", \"it\", or terse follow-ups like \"push\"; judge relevance against what the user's message is about in light of it):",
+          "<<<",
+          context,
+          ">>>",
+        ]
+      : []),
     "",
     "These stored memory notes were keyword-matched as possibly relevant.",
     "Keep only the ones the message is GENUINELY about — the topic, project, or person itself.",
     "Sharing an incidental word with the note (e.g. \"zoom\" matching a video tool, a person's employer matching a project name) is NOT relevance.",
+    "If the conversation context shows the user means a different project than a candidate (e.g. they are discussing app A, and a candidate is the unrelated app B that merely shares words like \"mobile\" or \"gradient\"), drop that candidate.",
     "",
     "Candidates:",
     ...lines,

package/core/system-prompt.js

@@ -395,15 +395,33 @@ function buildEntityBlock(matches) {
 // router.js wraps Telegram reply-quotes in "Replying to …:\n---\n…\n---",
 // and quoting a 📖 recall announcement echoes every entity named in it.
 // Matching recall over that text re-injects whatever was quoted (names
-// score 2 — instant threshold), so recall only sees the user's own words.
-function recallMatchText(prompt) {
-  let text = String(prompt || "");
-  text = text.replace(/^Replying to [^\n]*:\n---\n[\s\S]*?\n---\n+/, "");
-  text = text.split("\n").filter((l) => !l.includes("📖")).join("\n");
-  return text.trim();
+// score 2 — instant threshold). So we split: the user's own words drive
+// recall as before, while the quoted text and the last few transcript
+// turns on this channel are kept separately as context. Context-derived
+// candidates only survive if the LLM relevance judge confirms them —
+// without a working judge they are dropped, so neither a quote nor a
+// stale thread can force-inject on keywords alone.
+function stripRecallAnnouncements(text) {
+  return String(text || "").split("\n").filter((l) => !l.includes("📖")).join("\n").trim();
 }
 
-async function filterMatches(matchText, packMatches, entityMatches) {
+function recallMatchParts(prompt) {
+  const text = String(prompt || "");
+  const m = text.match(/^Replying to [^\n]*:\n---\n([\s\S]*?)\n---\n+/);
+  return {
+    userText: stripRecallAnnouncements(m ? text.slice(m[0].length) : text),
+    contextText: stripRecallAnnouncements(m ? m[1] : ""),
+  };
+}
+
+// packMatches/entityMatches carry an `origin` of "user" or "context".
+// Fail-open keeps user-text candidates (the keyword baseline) but drops
+// context-only ones — those are only trustworthy with a working judge.
+async function filterMatches(userText, contextText, packMatches, entityMatches) {
+  const failOpen = () => ({
+    packMatches: packMatches.filter((m) => m.origin === "user"),
+    entityMatches: entityMatches.filter((m) => m.origin === "user"),
+  });
   if (packMatches.length === 0 && entityMatches.length === 0) {
     return { packMatches, entityMatches };
   }
@@ -419,26 +437,57 @@ async function filterMatches(matchText, packMatches, entityMatches) {
       const e = entitiesLib.readEntity(m.slug);
       candidates.push({ id: `entity:${m.slug}`, name: m.name, description: e?.description || "" });
     }
-    const kept = await require("./recall-filter").judgeRelevance(matchText, candidates);
-    if (!kept) return { packMatches, entityMatches }; // fail-open
+    const kept = await require("./recall-filter").judgeRelevance(userText, candidates, { context: contextText });
+    if (!kept) return failOpen();
     return {
       packMatches: packMatches.filter((m) => kept.has(`pack:${m.dir}`)),
       entityMatches: entityMatches.filter((m) => kept.has(`entity:${m.slug}`)),
     };
   } catch (e) {
-    return { packMatches, entityMatches };
+    return failOpen();
+  }
+}
+
+function mergeMatches(primary, secondary, keyOf) {
+  const out = primary.map((m) => ({ ...m, origin: "user" }));
+  const seen = new Set(primary.map(keyOf));
+  for (const m of secondary) {
+    if (seen.has(keyOf(m))) continue;
+    out.push({ ...m, origin: "context" });
   }
+  return out;
 }
 
 async function promptWithDynamicContext(prompt) {
   lastInjected = { packs: [], entities: [] };
   try {
-    const matchText = recallMatchText(prompt);
+    const { userText, contextText } = recallMatchParts(prompt);
+    let historyText = "";
+    try {
+      historyText = stripRecallAnnouncements(require("./transcripts").recentTranscriptText({ limit: 6 }));
+    } catch (e) {}
+    const fullContext = [contextText, historyText].filter(Boolean).join("\n\n");
+    const packsLib = require("./packs");
+    const entitiesLib = require("./entities");
     let packMatches = [];
     let entityMatches = [];
-    try { packMatches = require("./packs").matchPacks(matchText, { limit: 3 }); } catch (e) {}
-    try { entityMatches = require("./entities").matchEntities(matchText, { limit: 4 }); } catch (e) {}
-    ({ packMatches, entityMatches } = await filterMatches(matchText, packMatches, entityMatches));
+    try {
+      packMatches = mergeMatches(
+        packsLib.matchPacks(userText, { limit: 3 }),
+        fullContext ? packsLib.matchPacks(fullContext, { limit: 3 }) : [],
+        (m) => m.dir,
+      );
+    } catch (e) {}
+    try {
+      entityMatches = mergeMatches(
+        entitiesLib.matchEntities(userText, { limit: 4 }),
+        fullContext ? entitiesLib.matchEntities(fullContext, { limit: 4 }) : [],
+        (m) => m.slug,
+      );
+    } catch (e) {}
+    ({ packMatches, entityMatches } = await filterMatches(userText, fullContext, packMatches, entityMatches));
+    packMatches = packMatches.slice(0, 3);
+    entityMatches = entityMatches.slice(0, 4);
     return `${buildDynamicContextBlock()}${buildPackBlock(packMatches)}${buildEntityBlock(entityMatches)}\n\nCurrent user request:\n${prompt}`;
   } catch (e) {
     return prompt;

package/core/transcripts.js

@@ -58,6 +58,38 @@ function appendProjectTranscript(role, text, metadata = {}, state = currentState
   }
 }
 
+// Last few user/assistant turns on THIS channel, for recall context.
+// The current turn's user message is already appended by the time the
+// prompt is built, so the trailing user entry is dropped — the live
+// message drives recall directly as origin "user".
+function recentTranscriptText({ limit = 6, perEntryChars = 280 } = {}, state = currentState()) {
+  const info = transcriptProjectInfo(state);
+  if (!info) return "";
+  try {
+    if (!fs.existsSync(info.transcriptPath)) return "";
+    const raw = fs.readFileSync(info.transcriptPath, "utf-8").trim();
+    if (!raw) return "";
+    const channel = String(currentChannelId() || "");
+    const entries = [];
+    for (const line of raw.split("\n").slice(-80)) {
+      try {
+        const obj = JSON.parse(line);
+        if (obj.role !== "user" && obj.role !== "assistant") continue;
+        if (channel && obj.chat?.id && String(obj.chat.id) !== channel) continue;
+        if (typeof obj.text !== "string" || !obj.text.trim()) continue;
+        entries.push(obj);
+      } catch (e) {}
+    }
+    if (entries.length > 0 && entries[entries.length - 1].role === "user") entries.pop();
+    return entries
+      .slice(-limit)
+      .map((e) => `${e.role}: ${e.text.replace(/\s+/g, " ").trim().slice(0, perEntryChars)}`)
+      .join("\n");
+  } catch (e) {
+    return "";
+  }
+}
+
 function promptWithTranscriptPointer(prompt, state = currentState()) {
   if (!state.currentSession) return prompt;
   return projectTranscripts.withPointer(prompt, state.currentSession.dir, state.currentSession.name, state.userId);
@@ -72,6 +104,7 @@ module.exports = {
   transcriptProjectInfo,
   transcriptPointerNote,
   appendProjectTranscript,
+  recentTranscriptText,
   promptWithTranscriptPointer,
   stripTranscriptPointerForStorage,
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@inetafrica/open-claudia",
-  "version": "2.6.4",
+  "version": "2.6.6",
   "description": "Your always-on AI coding assistant — Claude Code, Cursor Agent, and OpenAI Codex via Telegram or Kazee Chat",
   "main": "bot.js",
   "bin": {