Summary

Version 2026.2.14 introduced a regression that causes constant "Media failed: ENOENT" errors when tool results contain the text "MEDIA:" (e.g., in session transcripts, documentation, or code examples).

Environment

• OpenClaw version: 2026.2.14
• Previous working version: 2026.2.12
• OS: Linux (Ubuntu 24.04)
• Channel: WhatsApp

Root Cause

New code in dist/pi-embedded-8DITBEle.js scans tool result content for MEDIA: tokens without a startsWith("MEDIA:") guard.

Affected Code (v2026.2.14)

File: dist/pi-embedded-8DITBEle.js (lines ~411-420)

for (const item of content) {
    if (entry.type === "text" && typeof entry.text === "string") {
        MEDIA_TOKEN_RE.lastIndex = 0;
        let match;
        while ((match = MEDIA_TOKEN_RE.exec(entry.text)) !== null) {
            const p = match[1]?.replace(/^[`"'[{(]+/, "").replace(/[`"'\]})\\,]+$/, "").trim();
            if (p && p.length <= 4096) paths.push(p);
        }
    }
}

The Problem

• Uses regex /\bMEDIA:\s*?([^\n]+)?/gi which matches MEDIA: anywhere in text (not just line-start)
• Processes tool results (like memory_search returning session transcripts)
• Session transcripts contain instructional text with examples: "MEDIA:https://example.com/image.jpg"
• Regex matches those examples → system attempts to fetch them → ENOENT errors

Why v2026.2.12 Doesn't Have This

This code doesn't exist in v2026.2.12:

$ grep -r "MEDIA_TOKEN_RE.lastIndex" v2026.2.12/dist/*.js
(no output)

$ grep -r "MEDIA_TOKEN_RE.lastIndex" v2026.2.14/dist/*.js
v2026.2.14/dist/pi-embedded-8DITBEle.js:            MEDIA_TOKEN_RE.lastIndex = 0;

Comparison

v2026.2.12 (working):

• Only splitMediaFromOutput parses MEDIA: directives (with startsWith guard) ✅
• Tool results are not scanned for media tokens
• No false positives

v2026.2.14 (broken):

• splitMediaFromOutput still has guard for assistant output ✅
• NEW: pi-embedded scans tool results WITHOUT guard ❌
• Result: example URLs in tool output trigger fetch attempts

Reproduction Steps

1. Use memory_search tool configured with sources: ["memory", "sessions"]
2. Have session transcripts containing the text MEDIA: (e.g., from previous agent instructions or documentation)
3. Run any query that returns those transcripts
4. Observer ENOENT errors for random text fragments

Symptoms

Constant error messages like:

⚠️ Media failed: ENOENT: no such file or directory, open ')) return nextPayload;'
⚠️ Media failed: ENOENT: no such file or directory, open '^>]+>(\s*\([^)]*\))?$/i;'
⚠️ Media failed: ENOENT: no such file or directory, open 'tokens from text content blocks (all OpenClaw tools).'
⚠️ Media failed: Failed to fetch media from https://example.com/image.jpg

Examples of text fragments being treated as file paths:

• Code snippets: ')) return nextPayload;'
• Regex patterns: ^>]+>(\s*\([^)]*\))?$/i;'
• Documentation text: tokens from text content blocks
• Example URLs from instructional text

Impact

• High: Error spam disrupts normal operations
• Tool results containing documentation/examples become unusable
• Session transcripts trigger the bug every time memory_search is used
• Users must rollback to v2026.2.12 to avoid the issue

Fix Suggestion

Add a startsWith("MEDIA:") guard to the new code in pi-embedded, similar to how splitMediaFromOutput handles it:

for (const item of content) {
    if (entry.type === "text" && typeof entry.text === "string") {
        // Split into lines and check each line
        const lines = entry.text.split('\n');
        for (const line of lines) {
            // Guard: only process lines that start with MEDIA:
            if (!line.trimStart().startsWith("MEDIA:")) {
                continue;
            }
            
            MEDIA_TOKEN_RE.lastIndex = 0;
            let match;
            while ((match = MEDIA_TOKEN_RE.exec(line)) !== null) {
                const p = match[1]?.replace(/^[`"'[{(]+/, "").replace(/[`"'\]})\\,]+$/, "").trim();
                if (p && p.length <= 4096) paths.push(p);
            }
        }
    }
}

Alternatively, filter tool results before scanning them, or add metadata to distinguish between agent-generated directives and tool result content.

Workaround

Rollback to v2026.2.12:

npm install -g [email protected] --force
systemctl --user restart openclaw-gateway

Related Code

For reference, the working splitMediaFromOutput function has this guard (line ~1606 in deliver-Dlw-4HTg.js):

if (!line.trimStart().startsWith("MEDIA:")) {
    keptLines.push(line);
    lineOffset += line.length + 1;
    continue;
}

This same pattern should be applied to the new tool result scanning code.