@blockrun/franklin 3.15.58 → 3.15.60

package/dist/agent/loop.d.ts

@@ -20,6 +20,15 @@ export declare function looksLikeGatewayErrorAsText(parts: ContentPart[]): {
     match: boolean;
     message: string;
 };
+/**
+ * Walk a Dialogue and replace large `image.source.data` (base64) blocks
+ * inside `tool_result.content` arrays with a tiny placeholder. The
+ * accompanying text block already names the file path so the model on
+ * resume can re-Read it if it needs to see the image again. Returns a
+ * shallow clone so the in-memory history (used for the rest of the
+ * current turn) keeps the full image data.
+ */
+export declare function stripLargeImageData(message: Dialogue): Dialogue;
 /**
  * Identify models known to hallucinate tool calls (invented names, literal
  * `[TOOLCALL]` / `<tool_call>` text in answers) — they need the explicit

package/dist/agent/loop.js

@@ -288,6 +288,64 @@ function getBackoffDelay(attempt, maxDelayMs = 32_000) {
     const jitter = base * 0.25 * (Math.random() * 2 - 1); // ±25%
     return Math.max(500, Math.round(base + jitter));
 }
+/**
+ * Threshold for stripping inline base64 image data on session-disk
+ * writes. Mirrors `streaming-executor.ts:PERSIST_THRESHOLD` so a Read of
+ * a small icon (favicon-sized PNG, ~3 KB base64) round-trips through
+ * resume intact, while a Read of a screenshot or generated artwork
+ * (typically 200 KB+ base64) gets path-stubbed.
+ */
+const SESSION_IMAGE_STRIP_THRESHOLD = 50_000;
+/**
+ * Walk a Dialogue and replace large `image.source.data` (base64) blocks
+ * inside `tool_result.content` arrays with a tiny placeholder. The
+ * accompanying text block already names the file path so the model on
+ * resume can re-Read it if it needs to see the image again. Returns a
+ * shallow clone so the in-memory history (used for the rest of the
+ * current turn) keeps the full image data.
+ */
+export function stripLargeImageData(message) {
+    if (!Array.isArray(message.content))
+        return message;
+    let mutated = false;
+    // Cast through `unknown` because Dialogue's content union doesn't expose
+    // the tool_result shape with image blocks at the type level — they flow
+    // in via the loop's outcome-building path. Runtime structure is what
+    // matters here; we only mutate when we positively identify the shape.
+    const newContent = message.content.map((part) => {
+        if (typeof part === 'object' &&
+            part !== null &&
+            part.type === 'tool_result' &&
+            Array.isArray(part.content)) {
+            const tr = part;
+            let inner = tr.content;
+            let innerMutated = false;
+            const cleaned = inner.map((block) => {
+                if (block &&
+                    typeof block === 'object' &&
+                    block.type === 'image' &&
+                    block.source?.type === 'base64' &&
+                    (block.source.data?.length ?? 0) > SESSION_IMAGE_STRIP_THRESHOLD) {
+                    innerMutated = true;
+                    const sz = (block.source.data ?? '').length;
+                    return {
+                        type: 'text',
+                        text: `<image stripped from session log: ${(sz / 1024).toFixed(1)} KB base64. ` +
+                            `See accompanying text block for the source path; re-Read to inline again.>`,
+                    };
+                }
+                return block;
+            });
+            if (innerMutated) {
+                mutated = true;
+                inner = cleaned;
+                return { ...tr, content: inner };
+            }
+        }
+        return part;
+    });
+    return mutated ? { ...message, content: newContent } : message;
+}
 /**
  * Format the user-facing "switching model" line. Includes the resolved
  * concrete model in parentheses when the user-facing alias (e.g.
@@ -483,7 +541,16 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
         });
     };
     const persistSessionMessage = (message) => {
-        appendToSession(sessionId, message);
+        // Strip large base64 image bytes before writing to session jsonl. The
+        // tool_result wrap at line ~1788 inlines image data so vision models
+        // can see it during the live turn — but PNG bytes can be ~600 KB
+        // each, and the inline content bypasses persistLargeResult (which
+        // only checks `result.output.length`). Verified 2026-05-05: a single
+        // Read of `/tmp/mamba_hd_p9.png` produced an 850 KB session jsonl
+        // line; a 5-turn session with multiple .png reads grew to 12 MB.
+        // The model already saw the bytes in this turn's in-memory history,
+        // so disk only needs the path reference for resume.
+        appendToSession(sessionId, stripLargeImageData(message));
         persistSessionMeta();
     };
     pruneOldSessions(sessionId); // Cleanup old sessions on start, protect current

package/dist/tools/videogen.js

@@ -44,7 +44,7 @@ function estimateVideoCostUsd(durationSeconds = DEFAULT_DURATION) {
 function buildExecute(deps) {
     return async function execute(input, ctx) {
         const rawInput = input;
-        const { output_path, model, image_url, duration_seconds, contentId } = rawInput;
+        const { output_path, model, image_url, duration_seconds, contentId, aspect_ratio } = rawInput;
         if (!rawInput.prompt)
             return { output: 'Error: prompt is required', isError: true };
         // Resolve image_url before sending. The gateway requires a URL (http(s)
@@ -156,6 +156,12 @@ function buildExecute(deps) {
             prompt: chosenPrompt,
             ...(resolvedImageUrl ? { image_url: resolvedImageUrl } : {}),
             ...(duration_seconds ? { duration_seconds } : {}),
+            // aspect_ratio passes through to the gateway. Models that support it
+            // (newer Seedance / grok variants) honor it; models that ignore it
+            // produce their default size. If the gateway rejects an unknown
+            // value, the 400 body surfaces via 3.15.45 diagnostic so the agent
+            // can drop the param and retry.
+            ...(aspect_ratio ? { aspect_ratio } : {}),
         });
         const headers = {
             'Content-Type': 'application/json',
@@ -456,7 +462,15 @@ export function createVideoGenCapability(deps = {}) {
                 "xai/grok-imagine-video bills $0.05/s (8s default ≈ $0.42). Generation " +
                 "takes ~20–60s. ALWAYS confirm with the user before calling — videos " +
                 "are expensive and slow. Pass contentId to attach to a Content piece " +
-                "(budget is checked before paying; asset is recorded on success).",
+                "(budget is checked before paying; asset is recorded on success). " +
+                "PLATFORM TARGETING: when the user says they'll post to X / Twitter, " +
+                "set aspect_ratio: '16:9' AND plan a follow-up `ffmpeg -vf scale=1280:720` " +
+                "step — X rejects videos under 720p with 'aspect ratio too small'. " +
+                "TikTok / Reels / Shorts: aspect_ratio '9:16'. Instagram Square: '1:1'. " +
+                "MODERATION: bytedance/seedance-* refuses photorealistic human faces " +
+                "(`InputImageSensitiveContentDetected.PrivacyInformation`); when the " +
+                "seed image has a real-looking person, use xai/grok-imagine-video " +
+                "instead, or regenerate the keyframe in a more stylized style first.",
             input_schema: {
                 type: 'object',
                 properties: {
@@ -471,6 +485,15 @@ export function createVideoGenCapability(deps = {}) {
                     },
                     image_url: { type: 'string', description: 'Optional seed image (image-to-video). Accepts http(s) URL, data: URI, or local file path — local paths get inlined as base64 data URIs automatically.' },
                     duration_seconds: { type: 'number', description: 'Duration billed for. Default depends on model (8s for grok-imagine-video).' },
+                    aspect_ratio: {
+                        type: 'string',
+                        description: 'Optional aspect ratio hint passed to the model. Common values: ' +
+                            '"16:9" (landscape — X/Twitter, YouTube, TikTok-landscape), ' +
+                            '"9:16" (vertical — TikTok, Reels, Shorts), ' +
+                            '"1:1" (square — Instagram feed). Models that don\'t support the ' +
+                            'param ignore it; if the gateway 400s on an unknown value, the ' +
+                            'error body surfaces — drop the param and retry.',
+                    },
                     contentId: { type: 'string', description: 'Optional Content id to attach and budget against.' },
                 },
                 required: ['prompt'],

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.58",
+  "version": "3.15.60",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {