@blockrun/franklin 3.15.60 → 3.15.62

package/dist/agent/loop.js

@@ -1054,6 +1054,15 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
             // llm.ts if `tools` ended up empty, so it's safe to attach here.
             const callToolChoice = forceToolChoiceNextRound;
             forceToolChoiceNextRound = null;
+            // Wall-clock start of the model call. Used by the recordUsage call
+            // a few hundred lines below so franklin-stats.json captures real
+            // latency. Verified 2026-05-05: `franklin stats` reported
+            // `avgLat=0.0s` for every model across 5300+ requests because the
+            // agent-loop callsite always passed 0 for latencyMs (proxy path
+            // already measured correctly). `franklin insights` couldn't surface
+            // "this model is consistently slow" or "fallback was faster" until
+            // this was fixed.
+            const llmCallStartedAt = Date.now();
             try {
                 const result = await client.complete({
                     model: resolvedModel,
@@ -1354,7 +1363,8 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
             // franklin-debug.log; `franklin insights` was therefore useless
             // for spotting a hot routing chain.
             const costEstimate = estimateCost(resolvedModel, inputTokens, usage.outputTokens, 1);
-            recordUsage(resolvedModel, inputTokens, usage.outputTokens, costEstimate, 0, turnFailedModels.size > 0);
+            const llmLatencyMs = Date.now() - llmCallStartedAt;
+            recordUsage(resolvedModel, inputTokens, usage.outputTokens, costEstimate, llmLatencyMs, turnFailedModels.size > 0);
             // ── Circuit breakers: prevent infinite-loop wallet drain ──
             // Per-turn $-cap was removed in v3.11.0 — runaway loops are caught by
             // MAX_TOOL_CALLS_PER_TURN (25) and MAX_TINY_RESPONSES (2) above; the

package/dist/tools/imagegen.js

@@ -230,6 +230,12 @@ function buildExecute(deps) {
         // minutes; text-to-image keeps the original 60s.
         const timeoutMs = referenceImage ? 180_000 : 60_000;
         const timeout = setTimeout(() => controller.abort(), timeoutMs);
+        // Wall-clock start of the paid call, including 402 retry + (optional)
+        // 202 polling. Used by recordUsage below so franklin-stats.json
+        // populates avgLatencyMs for image models. Mirrors the agent-loop fix
+        // in 3.15.61 — same pattern, same reason: insights couldn't surface
+        // "Seedance is slower than grok" while every media call recorded 0.
+        const callStartedAt = Date.now();
         try {
             // First request — will get 402
             let response = await fetch(endpoint, {
@@ -362,11 +368,12 @@ function buildExecute(deps) {
             // insights panel under-reported total spend and never surfaced
             // image-generation models in its "top models" list. Fire-and-forget —
             // stats write must not fail a user-visible generation.
+            const latencyMs = Date.now() - callStartedAt;
             void (async () => {
                 try {
                     const m = await findModel(imageModel);
                     const estCost = m ? estimateCostUsd(m, { quantity: 1 }) : 0;
-                    recordUsage(imageModel, 0, 0, estCost, 0);
+                    recordUsage(imageModel, 0, 0, estCost, latencyMs);
                 }
                 catch { /* ignore stats errors */ }
             })();

package/dist/tools/modal.js

@@ -325,7 +325,9 @@ export const modalCreateCapability = {
                 body.cpu = cpuCoerced;
             if (typeof memoryCoerced === 'number')
                 body.memory = memoryCoerced;
+            const callStartedAt = Date.now();
             const res = await postWithPayment(modalEndpoint('create'), body, 'Franklin Modal sandbox create', ctx.abortSignal, 90_000);
+            const latencyMs = Date.now() - callStartedAt;
             if (!res.ok) {
                 const err = res.body.error ? String(res.body.error) : res.raw.slice(0, 300);
                 // Surface the per-field validation issues — usually the
@@ -356,7 +358,7 @@ export const modalCreateCapability = {
             });
             // Stats — surface Modal usage in `franklin insights` like other paid tools.
             try {
-                recordUsage(`modal/${tier}`, 0, 0, price, 0);
+                recordUsage(`modal/${tier}`, 0, 0, price, latencyMs);
             }
             catch { /* ignore */ }
             return {
@@ -448,7 +450,9 @@ export const modalExecCapability = {
             };
             if (coercedTimeout !== undefined)
                 body.timeout = coercedTimeout;
+            const callStartedAt = Date.now();
             const res = await postWithPayment(modalEndpoint('exec'), body, 'Franklin Modal sandbox exec', ctx.abortSignal, Math.max(30_000, ((coercedTimeout ?? 300) + 30) * 1000));
+            const latencyMs = Date.now() - callStartedAt;
             if (!res.ok) {
                 // 400 here usually means the agent built the wrong shape (bad
                 // sandbox_id, malformed command). Dump the full raw body so the
@@ -479,7 +483,7 @@ export const modalExecCapability = {
             const hasAnyOutput = stdout.length > 0 || stderr.length > 0;
             const exitCode = rawExit !== null ? rawExit : (hasAnyOutput ? 0 : -1);
             try {
-                recordUsage('modal/exec', 0, 0, EXEC_PRICE_USD, 0);
+                recordUsage('modal/exec', 0, 0, EXEC_PRICE_USD, latencyMs);
             }
             catch { /* ignore */ }
             const summary = `exit ${exitCode}` + (rawExit === null ? ' (inferred — no exit_code field in response)' : '');
@@ -527,13 +531,15 @@ export const modalStatusCapability = {
         }
         catch { /* ignore */ }
         try {
+            const callStartedAt = Date.now();
             const res = await postWithPayment(modalEndpoint('status'), { sandbox_id }, 'Franklin Modal sandbox status', ctx.abortSignal, 30_000);
+            const latencyMs = Date.now() - callStartedAt;
             if (!res.ok) {
                 const err = res.body.error ? String(res.body.error) : res.raw.slice(0, 300);
                 return { output: `ModalStatus failed (${res.status}): ${err}`, isError: true };
             }
             try {
-                recordUsage('modal/status', 0, 0, STATUS_PRICE_USD, 0);
+                recordUsage('modal/status', 0, 0, STATUS_PRICE_USD, latencyMs);
             }
             catch { /* ignore */ }
             const status = res.body.status || 'unknown';
@@ -573,7 +579,9 @@ export const modalTerminateCapability = {
         }
         catch { /* ignore */ }
         try {
+            const callStartedAt = Date.now();
             const res = await postWithPayment(modalEndpoint('terminate'), { sandbox_id }, 'Franklin Modal sandbox terminate', ctx.abortSignal, 30_000);
+            const latencyMs = Date.now() - callStartedAt;
             // Always remove from tracker — even on failure, retrying is wasteful.
             sessionSandboxTracker.remove(sandbox_id);
             if (!res.ok) {
@@ -585,7 +593,7 @@ export const modalTerminateCapability = {
                 };
             }
             try {
-                recordUsage('modal/terminate', 0, 0, TERMINATE_PRICE_USD, 0);
+                recordUsage('modal/terminate', 0, 0, TERMINATE_PRICE_USD, latencyMs);
             }
             catch { /* ignore */ }
             return { output: `Sandbox \`${sandbox_id}\` terminated.` };

package/dist/tools/videogen.js

@@ -167,6 +167,12 @@ function buildExecute(deps) {
             'Content-Type': 'application/json',
             'User-Agent': `franklin/${VERSION}`,
         };
+        // Wall-clock start of the paid call (submit + poll + download). Fed
+        // to recordUsage below so franklin-stats.json populates avgLatencyMs
+        // for video models. Same fix as 3.15.61 (agent loop) — five
+        // recordUsage callsites in this codebase, three of them were
+        // hardcoding 0.
+        const callStartedAt = Date.now();
         const onAbort = (ctrl) => () => ctrl.abort();
         // Phase 1: submit the job. First POST triggers a 402; we sign and retry.
         // The signed paymentHeaders must be reused on every GET poll — the server
@@ -300,11 +306,12 @@ function buildExecute(deps) {
             // Prefer the live gateway price when the model is in the catalog;
             // fall back to the legacy $0.05/s estimate otherwise. Fire-and-
             // forget — stats write must not fail a user-visible generation.
+            const latencyMs = Date.now() - callStartedAt;
             void (async () => {
                 try {
                     const m = await findModel(videoModel);
                     const estCost = m ? estimateCostUsd(m, { duration_seconds: dur }) : estimateVideoCostUsd(dur);
-                    recordUsage(videoModel, 0, 0, estCost, 0);
+                    recordUsage(videoModel, 0, 0, estCost, latencyMs);
                 }
                 catch { /* ignore stats errors */ }
             })();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.60",
+  "version": "3.15.62",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {