@blockrun/franklin 3.15.15 → 3.15.16

package/dist/agent/loop.js

@@ -1201,6 +1201,11 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 inputTokens,
                 outputTokens: usage.outputTokens,
                 costUsd: costEstimate,
+                // Any failed model this turn means the model that finally
+                // succeeded was a fallback. Without this, audit log read 0%
+                // fallbacks across 4k entries — useless for diagnosing whether
+                // the routing chain is healthy or hot.
+                fallback: turnFailedModels.size > 0,
                 source: 'agent',
                 workDir,
                 prompt: extractLastUserPrompt(history),

package/dist/stats/audit.js

@@ -11,6 +11,7 @@
 import fs from 'node:fs';
 import path from 'node:path';
 import { BLOCKRUN_DIR } from '../config.js';
+import { isTestFixtureModel } from './test-fixture.js';
 const AUDIT_FILE = path.join(BLOCKRUN_DIR, 'franklin-audit.jsonl');
 const PROMPT_PREVIEW_CHARS = 240;
 // Cap the audit log at the most recent N entries. Without this the file
@@ -26,6 +27,12 @@ const TRIM_PROBE_BYTES = MAX_AUDIT_ENTRIES * 200;
 const TRIM_CHECK_INTERVAL = 200;
 let appendsSinceCheck = 0;
 export function appendAudit(entry) {
+    // Tests run interactiveSession() in-process with model="local/test*"
+    // and would otherwise pollute the user's real audit log. Drop the
+    // entry before any disk write rather than relying on every test to
+    // remember to redirect HOME.
+    if (isTestFixtureModel(entry.model))
+        return;
     try {
         fs.mkdirSync(BLOCKRUN_DIR, { recursive: true });
         const safe = {

package/dist/stats/test-fixture.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Test-fixture model detection.
+ *
+ * Tests in `test/local.mjs` run `interactiveSession()` in-process with
+ * model names like `local/test-model` and `local/test`. The agent loop
+ * persists every successful turn to `~/.blockrun/franklin-audit.jsonl`,
+ * `franklin-stats.json`, and the session store — which means tests
+ * pollute the user's real telemetry. Verified on a real machine:
+ * 2326 of 3969 audit entries (58.6%) and 84 of 1000 stats entries
+ * (8.4%) were `local/test*` test fixtures.
+ *
+ * The fix is to skip persistence when the model name follows the
+ * convention. Test prefixes are reserved (`local/test*` won't ever ship
+ * as a real model on the BlockRun gateway), so this is safe.
+ *
+ * Local LLMs that real users run (`local/llamafile`, `local/ollama`,
+ * `local/lmstudio`, etc.) are intentionally NOT filtered — only the
+ * `local/test` prefix.
+ */
+export declare function isTestFixtureModel(model: string | undefined | null): boolean;

package/dist/stats/test-fixture.js

@@ -0,0 +1,31 @@
+/**
+ * Test-fixture model detection.
+ *
+ * Tests in `test/local.mjs` run `interactiveSession()` in-process with
+ * model names like `local/test-model` and `local/test`. The agent loop
+ * persists every successful turn to `~/.blockrun/franklin-audit.jsonl`,
+ * `franklin-stats.json`, and the session store — which means tests
+ * pollute the user's real telemetry. Verified on a real machine:
+ * 2326 of 3969 audit entries (58.6%) and 84 of 1000 stats entries
+ * (8.4%) were `local/test*` test fixtures.
+ *
+ * The fix is to skip persistence when the model name follows the
+ * convention. Test prefixes are reserved (`local/test*` won't ever ship
+ * as a real model on the BlockRun gateway), so this is safe.
+ *
+ * Local LLMs that real users run (`local/llamafile`, `local/ollama`,
+ * `local/lmstudio`, etc.) are intentionally NOT filtered — only the
+ * `local/test` prefix.
+ */
+const TEST_FIXTURE_PREFIXES = [
+    'local/test',
+];
+export function isTestFixtureModel(model) {
+    if (!model)
+        return false;
+    for (const prefix of TEST_FIXTURE_PREFIXES) {
+        if (model.startsWith(prefix))
+            return true;
+    }
+    return false;
+}

package/dist/stats/tracker.js

@@ -7,6 +7,7 @@ import path from 'node:path';
 import os from 'node:os';
 import { OPUS_PRICING } from '../pricing.js';
 import { BLOCKRUN_DIR } from '../config.js';
+import { isTestFixtureModel } from './test-fixture.js';
 let resolvedStatsFile = null;
 function preferredStatsFile() {
     return path.join(BLOCKRUN_DIR, 'franklin-stats.json');
@@ -156,6 +157,12 @@ export function flushStats() {
  * Record a completed request for stats tracking
  */
 export function recordUsage(model, inputTokens, outputTokens, costUsd, latencyMs, fallback = false) {
+    // Same rationale as appendAudit — tests run in-process with
+    // local/test* models and would otherwise mix into franklin-stats.json
+    // history (verified: 8.4% of a real user's 1000-entry history was
+    // test fixtures before this gate).
+    if (isTestFixtureModel(model))
+        return;
     const stats = getCachedStats();
     const now = Date.now();
     // Update totals

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.15",
+  "version": "3.15.16",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {