hermes - 💡(How to fix) Fix Bug: extract_media() false-positives on example paths in quoted text / code blocks

extract_media() in gateway/platforms/base.py (~line 2577): for match in media_pattern.finditer(content): path = match.group("path").strip() ... media.append((os.path.expanduser(path), has_voice_tag)) The scan is context-blind — it does not skip:

• Fenced code blocks (...)
• Inline code spans (MEDIA:...)
• Blockquotes (> ...)
• Tool output embedded in the response

Fix / Workaround

Problem

extract_media() scans the full response text with a single regex pass (MEDIA_TAG_CLEANUP_RE) without distinguishing live delivery tags from example paths mentioned in prose. This causes false positives when:

1. A skill description, error message, or doc string contains a literal example like inside a code block or quote
2. The agent explains how to use the MEDIA tag in its reply (e.g. "include MEDIA:/path/to/file in your response")
3. A tool returns output that happens to contain a path matching the regex Effect: The matching text is stripped from the user-visible response and the path is added to the media list. validate_media_delivery_path then either rejects it (silent drop) or — if the path happens to exist — delivers an unintended file.

Minimal Reproduction

Ask the agent to explain the MEDIA delivery syntax. Its reply will likely contain something like:

To send an image, include in your response. The backtick-wrapped example matches MEDIA_TAG_CLEANUP_RE, gets stripped from the text, and an attempt is made to deliver /path/to/image.jpg.

Root Cause

extract_media() in gateway/platforms/base.py (~line 2577): for match in media_pattern.finditer(content): path = match.group("path").strip() ... media.append((os.path.expanduser(path), has_voice_tag)) The scan is context-blind — it does not skip:

• Fenced code blocks (...)
• Inline code spans (MEDIA:...)
• Blockquotes (> ...)
• Tool output embedded in the response

Proposed Fix Direction

Before running MEDIA_TAG_CLEANUP_RE, mask content inside fenced code blocks, inline code spans, and blockquotes (replace protected spans with equal-length whitespace to preserve match offsets). This keeps the cleanup substitution correct while eliminating false positives. Happy to submit a PR if the direction looks right — we have a working patch in our deployment.

Environment

• macOS 15, launchd-managed gateway
• Feishu platform adapter
• Triggered by: skills that document MEDIA syntax, tool error messages containing file paths