context-mode 1.0.167 → 1.0.168

package/.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Claude Code plugins by Mert Koseoğlu",
-    "version": "1.0.167"
+    "version": "1.0.168"
   },
   "plugins": [
     {
       "name": "context-mode",
       "source": "./",
       "description": "Claude Code MCP plugin that saves 98% of your context window. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and intent-driven search.",
-      "version": "1.0.167",
+      "version": "1.0.168",
       "author": {
         "name": "Mert Koseoğlu"
       },

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "context-mode",
-  "version": "1.0.167",
+  "version": "1.0.168",
   "description": "MCP server that saves 98% of your context window with session continuity. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and automatic state restore across compactions.",
   "author": {
     "name": "Mert Koseoğlu",

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "context-mode",
-  "version": "1.0.167",
+  "version": "1.0.168",
   "description": "MCP server that saves 98% of your context window with session continuity. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and automatic state restore across compactions.",
   "author": {
     "name": "Mert Koseoğlu",

package/.openclaw-plugin/openclaw.plugin.json

@@ -3,7 +3,7 @@
   "name": "Context Mode",
   "kind": "tool",
   "description": "OpenClaw plugin that saves 98% of your context window. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and intent-driven search.",
-  "version": "1.0.167",
+  "version": "1.0.168",
   "sandbox": {
     "mode": "permissive",
     "filesystem_access": "full",

package/.openclaw-plugin/package.json

@@ -1,6 +1,6 @@
 {
   "name": "context-mode",
-  "version": "1.0.167",
+  "version": "1.0.168",
   "description": "OpenClaw plugin that saves 98% of your context window. Sandboxed code execution in 11 languages, FTS5 knowledge base with BM25 ranking, and intent-driven search.",
   "author": {
     "name": "Mert Koseoğlu",

package/build/session/analytics.js

@@ -900,6 +900,18 @@ export function getRealBytesStats(opts) {
                             snapshotBytes += Number(snap.bytes);
                     }
                     catch { /* old schema */ }
+                    try {
+                        // "With context-mode" = the bytes the model paid to ACCESS the
+                        // kept-out content: ctx_search (query the index) + ctx_fetch_and_index
+                        // (fetch + index a URL). Sandbox compute (ctx_execute/batch/file) is
+                        // work-output the model would see regardless — NOT redirect savings —
+                        // so it is excluded; folding it crushed the bar to a false ~43%.
+                        const tc = sdb.prepare(`SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls
+               WHERE session_id = ? AND tool IN ('ctx_search', 'ctx_fetch_and_index')`).get(opts.sessionId);
+                        if (tc?.bytes)
+                            bytesReturned += Number(tc.bytes);
+                    }
+                    catch { /* old schema: no tool_calls table */ }
                 }
                 else if (opts.projectDir) {
                     // Bug E+F: META-scoped aggregation. Take every session_id whose
@@ -930,6 +942,17 @@ export function getRealBytesStats(opts) {
                             snapshotBytes += Number(snap.bytes);
                     }
                     catch { /* old schema */ }
+                    try {
+                        const tc = sdb.prepare(`SELECT COALESCE(SUM(bytes_returned), 0) AS bytes
+               FROM tool_calls
+               WHERE session_id IN (
+                 SELECT session_id FROM session_meta WHERE project_dir = ?
+               )
+               AND tool IN ('ctx_search', 'ctx_fetch_and_index')`).get(opts.projectDir);
+                        if (tc?.bytes)
+                            bytesReturned += Number(tc.bytes);
+                    }
+                    catch { /* old schema: no tool_calls table */ }
                 }
                 else {
                     const row = sdb.prepare(`SELECT
@@ -948,6 +971,13 @@ export function getRealBytesStats(opts) {
                             snapshotBytes += Number(snap.bytes);
                     }
                     catch { /* old schema */ }
+                    try {
+                        const tc = sdb.prepare(`SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls
+               WHERE tool IN ('ctx_search', 'ctx_fetch_and_index')`).get();
+                        if (tc?.bytes)
+                            bytesReturned += Number(tc.bytes);
+                    }
+                    catch { /* old schema: no tool_calls table */ }
                 }
             }
             finally {

package/build/session/extract.d.ts

@@ -22,6 +22,15 @@ export interface SessionEvent {
      * `Fetched and indexed N sections (XKB)` preamble.
      */
     bytes_avoided?: number;
+    /**
+     * Optional — bytes the model PAID to ACCESS kept-out content for this event:
+     * the tool_response byte length of a `ctx_search` / `ctx_fetch_and_index`
+     * call. This is the OTHER half of the with/without ratio (bytes_avoided is
+     * the kept-out half). Sandbox compute (ctx_execute/batch/file) is work-output
+     * and is excluded. Present only when the call is a retrieval call and its
+     * tool_response is non-empty.
+     */
+    bytes_retrieved?: number;
     /**
      * Optional structured cost/usage fields (Wave 2b). Emitted by
      * extractAgentUsage alongside the colon-string `data` so the forward
@@ -35,6 +44,13 @@ export interface SessionEvent {
     cache_read_tokens?: number;
     cache_creation_tokens?: number;
     cost_usd?: number;
+    /**
+     * "task_cumulative" on agent_usage events whose tokens are a Task sub-agent's
+     * usage SUMMED across its whole run (not one turn). The platform buckets these
+     * as lifetime spend and never prices them per-turn — see
+     * docs/handoff/cumulative-cost-bug.md.
+     */
+    usage_scope?: string;
 }
 export interface ToolCall {
     toolName: string;

package/build/session/extract.js

@@ -912,12 +912,40 @@ function extractMcpToolCall(input) {
     const payload = truncated
         ? `{"tool_name":${JSON.stringify(tool_name)},"params_raw":${JSON.stringify(cappedStr)},"truncated":true}`
         : `{"tool_name":${JSON.stringify(tool_name)},"params":${cappedStr}}`;
-    return [{
-            type: "mcp_tool_call",
-            category: "mcp_tool_call",
-            data: safeString(payload),
-            priority: 4,
-        }];
+    const event = {
+        type: "mcp_tool_call",
+        category: "mcp_tool_call",
+        data: safeString(payload),
+        priority: 4,
+    };
+    // Retrieval cost (the OTHER half of the with/without ratio): when this MCP
+    // call is a `ctx_search` or `ctx_fetch_and_index` retrieval, the tool_response
+    // IS the kept-out content the model paid to access — record its byte length.
+    // Sandbox compute (ctx_execute/batch/file) is work-output, NOT retrieval, so
+    // it is intentionally excluded. Match by suffix char-algorithmically (host
+    // prefixes the name like `mcp__plugin_…__ctx_search`); NO regex.
+    if (isRetrievalToolName(tool_name)) {
+        const response = safeString(input.tool_response);
+        if (response.length > 0) {
+            event.bytes_retrieved = Buffer.byteLength(response, "utf8");
+        }
+    }
+    return [event];
+}
+/** Tool-name suffixes that denote a RETRIEVAL call (kept-out content accessed). */
+const RETRIEVAL_TOOL_SUFFIXES = ["ctx_search", "ctx_fetch_and_index"];
+/**
+ * True when `toolName` ends with one of the retrieval suffixes. Char-level
+ * suffix comparison via String.prototype.endsWith — no regex. MCP host names
+ * arrive prefixed (e.g. `mcp__plugin_context-mode_context-mode__ctx_search`),
+ * so an exact-name check would miss them; suffix match is host-agnostic.
+ */
+function isRetrievalToolName(toolName) {
+    for (const suffix of RETRIEVAL_TOOL_SUFFIXES) {
+        if (toolName.endsWith(suffix))
+            return true;
+    }
+    return false;
 }
 /**
  * Category 6 (tool-based): decision
@@ -1377,27 +1405,16 @@ function extractAgentUsage(input) {
     if (typeof usage.service_tier === "string") {
         parts.push(`tier:${usage.service_tier.slice(0, 32)}`);
     }
-    // Gap #1 (16-oss-verify-gap-prd) — derive cost_usd from per-model pricing
-    // when at least one token count is present. Zero-token case skips cost
-    // so dashboard never shows misleading "$0.00 for nothing" rows.
-    const inputTokens = typeof usage.input_tokens === "number" ? usage.input_tokens : 0;
-    const outputTokens = typeof usage.output_tokens === "number" ? usage.output_tokens : 0;
-    const cacheCreate = typeof usage.cache_creation_input_tokens === "number"
-        ? usage.cache_creation_input_tokens
-        : 0;
-    const cacheRead = typeof usage.cache_read_input_tokens === "number"
-        ? usage.cache_read_input_tokens
-        : 0;
+    // CUMULATIVE-USAGE GUARD (docs/handoff/cumulative-cost-bug.md): a Task
+    // tool_response carries the sub-agent's usage SUMMED across its entire run —
+    // every internal turn re-reads the cache, so cache_read reaches the billions.
+    // Pricing that cumulative figure as a single turn produced four-figure
+    // per-event costs ($3,532 with cache_read 4.7B) that poisoned every FinOps
+    // aggregate. We therefore do NOT derive cost_usd here. The raw token counts
+    // stay, tagged usage_scope="task_cumulative", so the platform buckets them as
+    // lifetime spend; real per-turn cost comes only from per-turn signals
+    // (extractTranscriptUsage + each adapter's own session).
     const modelId = resolveModelId(input, out);
-    const anyTokens = inputTokens > 0 || outputTokens > 0 || cacheCreate > 0 || cacheRead > 0;
-    let cost = null;
-    if (anyTokens) {
-        // null ⇒ unmatched model id (catalog warned once) — skip the cost token
-        // rather than blend a wrong Claude rate (the old non-Claude bug).
-        cost = computeTurnCostUsd(modelId, inputTokens, outputTokens, cacheCreate, cacheRead);
-        if (cost !== null)
-            parts.push(`cost_usd:${formatCostUsd(cost)}`);
-    }
     // Wave 2b — emit structured top-level fields alongside the colon-string so
     // the forward envelope (which spreads `...event`) hands the platform typed
     // columns. Each field is set only when its source signal is present, so the
@@ -1421,8 +1438,7 @@ function extractAgentUsage(input) {
     if (typeof usage.cache_creation_input_tokens === "number") {
         event.cache_creation_tokens = usage.cache_creation_input_tokens;
     }
-    if (cost !== null)
-        event.cost_usd = cost;
+    event.usage_scope = "task_cumulative";
     return [event];
 }
 // ── Kimi Code (kimi-code) usage parsers ────────────────────────────────────