@kage-core/kage-graph-mcp 2.5.5 → 2.5.6

package/dist/kernel.js

@@ -7015,10 +7015,11 @@ function gcProject(projectDir, options = {}) {
             skipped.push({ id: packet.id, title: packet.title, reason: "already deprecated" });
             continue;
         }
-        // Serialized transcript / tool-output / file-content dumps carry no durable knowledge
-        // and bloat recall + the graph. Always delete them (deprecating would leave the blob on
-        // disk) — this also reclaims legacy dumps written before the capture-time guard existed.
-        if (isSerializedDumpTitle(packet.title) || isSerializedDumpBody(packet.body)) {
+        // Serialized transcript / tool-output / file-content dumps and ungrounded conversational
+        // chatter (a path-less rant at the assistant) carry no durable knowledge and bloat recall +
+        // the digest. Always delete them (deprecating would leave the blob on disk) — this also
+        // reclaims legacy junk written before the capture-time guards existed.
+        if (isSerializedDumpTitle(packet.title) || isSerializedDumpBody(packet.body) || isUngroundedConversationalCapture(packet)) {
             if (!options.dryRun)
                 (0, node_fs_1.unlinkSync)(path);
             deleted.push({ id: packet.id, title: packet.title });
@@ -8012,6 +8013,14 @@ function hasRepoGroundingSignal(text) {
 // punctuation, or one of a curated set of rhetorical / second-person-at-the-assistant
 // phrases — so it stays clear of normal declarative learnings (which read as statements, not
 // outbursts addressed to "you").
+// Phrases that rant AT the assistant ("why are you...", "it's your job", "don't stop"). These
+// never appear in a curated learning, so they mark chatter even when the outburst name-drops a
+// platform word (github, pr, x, linkedin) that the loose grounding matcher would otherwise
+// treat as a repo reference. This is the override that closes the leak where a frustrated
+// message peppered with nouns slipped through and was captured as approved memory.
+function looksFrustratedAtAssistant(text) {
+    return /\b(why are you|why did you|why would you|why aren'?t you|are you (kidding|serious|even|really)|it'?s your job|that'?s your job|do your job|don'?t stop|stop asking|stop before you|keep going|hurry up|just do it|figure it out yourself|i (already )?told you)\b/i.test(text ?? "");
+}
 function looksLikeRawUserUtterance(text) {
     const t = (text ?? "").trim();
     if (!t)
@@ -8021,18 +8030,21 @@ function looksLikeRawUserUtterance(text) {
     if (/[!?]{2,}/.test(t))
         return true;
     // Rhetorical questions / second-person-imperative frustration directed at the assistant.
-    return /\b(why are you|why did you|why would you|why aren'?t you|are you (kidding|serious|even|really)|it'?s your job|that'?s your job|do your job|don'?t stop|stop asking|stop before you|keep going|hurry up|just do it|figure it out yourself|i (already )?told you)\b/i.test(t);
-}
-// Capture-noise guard for ungrounded chat. A packet trips it only when ALL hold: it cites zero
-// repo paths, its text carries no repo grounding signal (no path/file/identifier/command), and
-// it reads as a raw conversational user utterance. The conjunction is what keeps it safe — a
-// real ungrounded decision or convention names a symbol, file, command, or rule and so is never
-// caught. Such packets are routed to pending (not auto-approved) at capture time and withheld
-// from recall, mirroring how serialized dumps are gated.
+    return looksFrustratedAtAssistant(t);
+}
+// Capture-noise guard for ungrounded chat. A packet trips it when it cites zero repo paths and
+// either (a) it rants at the assistant — which overrides incidental repo-ish words, since the
+// grounding matcher false-positives on bare platform names — or (b) it reads as a raw outburst
+// and carries no repo grounding signal at all. The override + conjunction keep it safe: a real
+// ungrounded decision or convention is declarative (no "why are you.../don't stop") and usually
+// names a symbol, file, command, or rule, so it is never caught. Such packets route to pending
+// (not auto-approved) at capture time and are withheld from recall, like serialized dumps.
 function isUngroundedConversationalCapture(packet) {
     if (packet.paths && packet.paths.length > 0)
         return false;
     const text = `${packet.title ?? ""}\n${packet.body ?? ""}`;
+    if (looksFrustratedAtAssistant(text))
+        return true;
     if (hasRepoGroundingSignal(text))
         return false;
     return looksLikeRawUserUtterance(text);
@@ -15615,20 +15627,17 @@ function kageResume(projectDir) {
         updated_at: packetRecency(packet),
         age: humanPacketAge(packetRecency(packet)),
     }));
-    const hasSessionContent = Boolean(lastSession || lastChangeMemory || pendingAutoDistilled || reconciliation.unresolved_count);
-    const hasContent = hasSessionContent || recentMemory.length > 0;
+    // A new session starts on a fresh task, so SessionStart does NOT replay last session's files,
+    // commands, or a recency-ranked memory list — that was pre-task noise (it fires before the
+    // first prompt, so it can't be task-targeted) and the surface that leaked raw command dumps and
+    // junk packets. Task-relevant memory is pulled on the first prompt via prompt-context; recall
+    // at the moment a file is read via file-context. SessionStart carries only always-on curated
+    // repo facts (the pinned block, below) plus a few actionable open-thread pointers.
+    const hasSessionContent = Boolean(lastChangeMemory || pendingAutoDistilled || reconciliation.unresolved_count);
+    const hasContent = hasSessionContent;
     const lines = [];
-    if (hasContent) {
-        lines.push("# Previously (Kage)");
-        if (lastSession) {
-            lines.push(`Last session ${lastSession.session_id} (${lastSession.observations} observation${lastSession.observations === 1 ? "" : "s"}, ended ${lastSession.last_at}).`);
-            if (lastSession.paths.length)
-                lines.push(`Worked on: ${lastSession.paths.join(", ")}`);
-            if (lastSession.commands.length)
-                lines.push(`Commands: ${lastSession.commands.join("; ")}`);
-            if (lastSession.distilled_titles.length)
-                lines.push(`Learned: ${lastSession.distilled_titles.join("; ")}`);
-        }
+    if (hasSessionContent) {
+        lines.push("# Open threads (Kage)");
         if (lastChangeMemory) {
             lines.push(`Change memory: ${lastChangeMemory.title} — ${lastChangeMemory.summary}`);
         }
@@ -15641,22 +15650,7 @@ function kageResume(projectDir) {
                 lines.push(`  - ${item.packet_id}: ${item.title}`);
         }
     }
-    // Compact timeline index: one line per recent packet, full detail (summary)
-    // only for the newest few, hard-capped to the resume token budget.
-    const block = lines.slice(0, 15);
-    if (recentPackets.length) {
-        block.push("", "## Recent memory");
-        recentPackets.forEach((packet, position) => {
-            const entry = `[${packet.id.slice(0, 12)}] ${packet.type} ${packet.title} (${humanPacketAge(packetRecency(packet))})`;
-            const candidate = [entry];
-            if (position < RESUME_TIMELINE_DETAILED && packet.summary) {
-                const summary = packet.summary.replace(/\s+/g, " ").trim();
-                candidate.push(`    ${summary.length > 160 ? `${summary.slice(0, 157)}...` : summary}`);
-            }
-            if (estimateTokens([...block, ...candidate].join("\n")) <= RESUME_CONTEXT_TOKEN_BUDGET)
-                block.push(...candidate);
-        });
-    }
+    const block = lines.slice(0, 12);
     // Lead the SessionStart injection with the team's pinned, always-on repo memory (the
     // curated high-signal facts), not just the recent-timeline digest — parity with recall's
     // context block, so a new session starts already holding the key knowledge, not only a

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kage-core/kage-graph-mcp",
-  "version": "2.5.5",
+  "version": "2.5.6",
   "description": "Team memory for coding agents: captures the decisions, runbooks, and bug fixes that get lost, verified against your code and shared via git. MCP server, zero deps, no account.",
   "main": "dist/index.js",
   "files": [