@inetafrica/open-claudia 2.6.5 → 2.6.7

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";
@@ -21,14 +24,15 @@ const FILTER_SYSTEM_PROMPT = [
 ].join("\n");
 
 // candidates: [{ id, name, description }]. opts.context: surrounding
-// conversation text (e.g. the message the user replied to) used to
-// disambiguate pronouns like "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).
+// 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, 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:",
@@ -38,7 +42,7 @@ async function judgeRelevance(message, candidates, opts = {}) {
     ...(context
       ? [
           "",
-          "It was a reply to this earlier message (context for resolving references like \"the app\" or \"it\" — judge relevance against what the user's message is about in light of it):",
+          "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,
           ">>>",

package/core/runner.js

@@ -132,6 +132,7 @@ async function buildClaudeArgs(prompt, opts = {}) {
     }
   }
   args.push("--model", settings.model || DEFAULT_CLAUDE_MODEL);
+  if (opts.maxTurns) args.push("--max-turns", String(opts.maxTurns));
   if (settings.effort) args.push("--effort", settings.effort);
   if (settings.budget) args.push("--max-budget-usd", String(settings.budget));
   if (settings.permissionMode) args.push("--permission-mode", settings.permissionMode);
@@ -276,6 +277,11 @@ function compactSeedPrompt(summary, extras = {}) {
     "Treat the following summary as the prior conversation context. Do not repeat it back unless asked.",
     "Continue from this state in future turns.",
     "",
+    // A seed turn that starts working (e.g. resuming a CI watch from the
+    // brief) can background a long-running command, and `claude -p` won't
+    // exit while a background task runs — that wedges the message queue.
+    "RIGHT NOW: reply with ONE short acknowledgement line only. Do NOT use any tools, run any commands, or start/resume any work in this turn — even if the summary, pending tasks, or recent turns suggest obvious next steps. Work resumes with the next user message.",
+    "",
     "Before telling the user you lack context on something they reference:",
     "1. Check the summary below.",
     "2. Read the full archived brief on disk (path given below, when present) — it is the detailed version of this summary.",
@@ -545,10 +551,6 @@ async function compactActiveSession(cwd, opts = {}) {
     if (opts.notify) await send(opts.message || "Context is getting large, compacting first so this stays fast…");
     const summaryRun = await runClaudeCapture(compactSummaryPrompt(), cwd, { resumeSessionId: oldSessionId, skipAutoCompact: true, label: "compact-summary", timeoutMs: COMPACT_SUMMARY_TIMEOUT });
     summary = summaryRun.text || "No prior context was returned by the summarizer.";
-    state[sessionKey] = null;
-    state.sessionUsage = { turns: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUsd: 0, lastInputTokens: 0 };
-    state.isFirstMessage = true;
-    saveState();
   } finally {
     state.isCompacting = false;
     saveState();
@@ -559,9 +561,13 @@ async function compactActiveSession(cwd, opts = {}) {
   // Only seed with the condensed version when the full text actually made it to disk.
   const seedSummary = (condensed && briefPath) ? condensed : (condensed ? `${fullBrief}\n\n${condensed}` : fullBrief);
   const repoFacts = collectRepoStateFacts(cwd);
-  const seedRun = await runClaudeCapture(compactSeedPrompt(seedSummary, { briefPath, condensed: !!(condensed && briefPath), repoFacts, recentTurns }), cwd, { fresh: true, skipAutoCompact: true, label: "compact-seed" });
-  const newSessionId = seedRun.sessionId || state[sessionKey];
-  if (newSessionId) state[sessionKey] = newSessionId;
+  // The old session id stays in place until the seed succeeds: if the seed
+  // dies (timeout, /stop, crash) the conversation keeps resuming the
+  // pre-compaction session instead of being orphaned. maxTurns: 1 hard-stops
+  // a seed that ignores the no-work instruction and reaches for tools.
+  const seedRun = await runClaudeCapture(compactSeedPrompt(seedSummary, { briefPath, condensed: !!(condensed && briefPath), repoFacts, recentTurns }), cwd, { fresh: true, skipAutoCompact: true, label: "compact-seed", timeoutMs: COMPACT_SUMMARY_TIMEOUT, maxTurns: 1 });
+  const newSessionId = seedRun.sessionId || null;
+  state[sessionKey] = newSessionId;
   state.isFirstMessage = false;
   state.lastCompactedAt = Date.now();
   state.sessionUsage = { turns: 0, inputTokens: 0, outputTokens: 0, cacheReadTokens: 0, cacheCreationTokens: 0, costUsd: 0, lastInputTokens: 0 };

package/core/system-prompt.js

@@ -396,10 +396,11 @@ function buildEntityBlock(matches) {
 // 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 we split: the user's own words drive
-// recall as before, and the quoted text is kept separately as context.
-// Context-derived candidates only survive if the LLM relevance judge
-// confirms them — without a working judge they are dropped, so a quote
-// can never force-inject on keywords alone.
+// 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();
 }
@@ -461,6 +462,11 @@ async function promptWithDynamicContext(prompt) {
   lastInjected = { packs: [], entities: [] };
   try {
     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 = [];
@@ -468,18 +474,18 @@ async function promptWithDynamicContext(prompt) {
     try {
       packMatches = mergeMatches(
         packsLib.matchPacks(userText, { limit: 3 }),
-        contextText ? packsLib.matchPacks(contextText, { limit: 3 }) : [],
+        fullContext ? packsLib.matchPacks(fullContext, { limit: 3 }) : [],
         (m) => m.dir,
       );
     } catch (e) {}
     try {
       entityMatches = mergeMatches(
         entitiesLib.matchEntities(userText, { limit: 4 }),
-        contextText ? entitiesLib.matchEntities(contextText, { limit: 4 }) : [],
+        fullContext ? entitiesLib.matchEntities(fullContext, { limit: 4 }) : [],
         (m) => m.slug,
       );
     } catch (e) {}
-    ({ packMatches, entityMatches } = await filterMatches(userText, contextText, packMatches, entityMatches));
+    ({ 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}`;

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": "2.6.5",
+  "version": "2.6.7",
   "description": "Your always-on AI coding assistant — Claude Code, Cursor Agent, and OpenAI Codex via Telegram or Kazee Chat",
   "main": "bot.js",
   "bin": {