@kage-core/kage-graph-mcp 2.5.5 → 2.5.7

package/dist/cli.js

@@ -15,7 +15,6 @@ const CORE_USAGE = `Kage — code-grounded memory for coding agents
 Core commands:
   kage install [--project <dir>]             one-shot: init + index + auto-wire detected agents
   kage scan --project <dir>                  60-second truth report on any repo (zero setup)
-  kage demo                                  watch the reject/withhold loop run in a sandbox
   kage init --project <dir>                  create repo memory (.agent_memory only)
   kage index --project <dir> [--full]        build/refresh code graph + indexes
   kage recall "<query>" --project <dir>      grounded recall from repo memory
@@ -1406,6 +1405,10 @@ async function main() {
                 `~${(0, kernel_js_1.formatTokenCount)(summary.all_time.replay_tokens)} all time ` +
                 `(discovery cost of served memories vs their compressed read cost)`);
         }
+        const usdOverridden = Number.isFinite(Number(process.env.KAGE_USD_PER_MTOK)) && Number(process.env.KAGE_USD_PER_MTOK) > 0;
+        console.log(`\nDollars estimated at $${kernel_js_1.VALUE_DOLLARS_PER_MILLION_TOKENS}/1M input tokens ` +
+            `(${usdOverridden ? "via KAGE_USD_PER_MTOK" : "Sonnet-class default — set KAGE_USD_PER_MTOK for your model"}). ` +
+            `Ledger: .agent_memory/reports/value.json`);
         return;
     }
     if (command === "file-context") {

package/dist/kernel.js

@@ -33,7 +33,7 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.PACKET_MERGE_DRIVER_CONFIG = exports.PACKET_MERGE_ATTRIBUTE_LINE = exports.AUTO_DISTILL_SIGNAL_THRESHOLD = exports.AUTO_DISTILL_TAG = exports.SETUP_AGENTS = exports.MEMORY_TYPES = exports.PACKET_SCHEMA_VERSION = void 0;
+exports.PACKET_MERGE_DRIVER_CONFIG = exports.PACKET_MERGE_ATTRIBUTE_LINE = exports.AUTO_DISTILL_SIGNAL_THRESHOLD = exports.AUTO_DISTILL_TAG = exports.VALUE_DOLLARS_PER_MILLION_TOKENS = exports.SETUP_AGENTS = exports.MEMORY_TYPES = exports.PACKET_SCHEMA_VERSION = void 0;
 exports.memoryRoot = memoryRoot;
 exports.packetsDir = packetsDir;
 exports.pendingDir = pendingDir;
@@ -1017,8 +1017,14 @@ function recordRecallAccess(projectDir, results) {
 // ---------------------------------------------------------------------------
 const VALUE_LEDGER_SCHEMA_VERSION = 1;
 const VALUE_LEDGER_EVENT_CAP = 5000;
-// Rough Sonnet-class input price used for the dollar estimate: $15 per 1M tokens.
-const VALUE_DOLLARS_PER_MILLION_TOKENS = 15;
+// Input price used for the dollar estimate. Default: Sonnet-class ~$3 per 1M input tokens (the
+// typical coding-agent tier) — deliberately conservative so the savings figure never overstates.
+// Override with KAGE_USD_PER_MTOK to match your model (e.g. 15 for Opus, 2.5 for GPT-4o, 0.8 for
+// Haiku). The old default ($15, Opus pricing) overstated savings ~5x for most users.
+exports.VALUE_DOLLARS_PER_MILLION_TOKENS = (() => {
+    const raw = Number(process.env.KAGE_USD_PER_MTOK);
+    return Number.isFinite(raw) && raw > 0 ? raw : 3;
+})();
 function valueLedgerPath(projectDir) {
     return (0, node_path_1.join)(reportsDir(projectDir), "value.json");
 }
@@ -1112,7 +1118,7 @@ function recordValueEvent(projectDir, event) {
     recordValueEvents(projectDir, [event]);
 }
 function estimatedTokenDollars(tokensSaved) {
-    return Number(((tokensSaved / 1_000_000) * VALUE_DOLLARS_PER_MILLION_TOKENS).toFixed(2));
+    return Number(((tokensSaved / 1_000_000) * exports.VALUE_DOLLARS_PER_MILLION_TOKENS).toFixed(2));
 }
 function summarizeValueWindow(events, cutoff) {
     const window = { tokens_saved: 0, replay_tokens: 0, stale_withheld: 0, stale_caught: 0, recalls: 0, caller_answers: 0 };
@@ -7015,10 +7021,11 @@ 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 +8019,14 @@ 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 +8036,21 @@ 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 +15633,17 @@ 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 +15656,7 @@ 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

@@ -1,7 +1,7 @@
 {
   "name": "@kage-core/kage-graph-mcp",
-  "version": "2.5.5",
-  "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.",
+  "version": "2.5.7",
+  "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, no account, no API key.",
   "main": "dist/index.js",
   "files": [
     "dist/**/*.js",