fix(memory): extend the frontmatter read to the closing fence when the byte cap cuts it off - #3617

AgentScope Version

2.0.12dev (main @ 4acacb9)

Description

The agentic-memory indexer reads only a byte-capped prefix of each memory file before parsing frontmatter (retrieval_max_tokens_per_frontmatter defaults to 256 tokens = 1024 bytes). The parsing regex requires a closing fence, so when the frontmatter block itself outgrows the cap the fence sits outside the prefix, the regex cannot match, and the file's description / type metadata are all silently dropped from the memory index (#3616 has the repro). #3345 already fixed the sibling case where only the trailing newline falls outside the prefix.

Changes — extract the prefix computation into _frontmatter_snippet and extend the read window to the closing fence when one exists beyond the cap:

closing = raw.find(b"\n---", 1)
if closing != -1:
    newline = raw.find(b"\n", closing + 1)
    fence_end = len(raw) if newline == -1 else newline + 1
    snippet_end = max(snippet_end, min(len(raw), fence_end))

The fence sits near the top of the file by definition, so the extension is bounded; raw is already fully in memory (read_file), so nothing extra is read. Behavior matrix:

How to test

Three regression cases added to the existing FrontmatterParseTest in tests/middleware_filesystem_memory_test.py: the over-cap file now yields description + type through the extended window; the unclosed and within-cap shapes are pinned to their current behavior.

python -m pytest tests/middleware_filesystem_memory_test.py
# 13 passed.

Fixes #3616

Checklist

• An issue has been created for this PR
• I have read the CONTRIBUTING.md
• Docstrings are in Google style
• Related documentation has been updated (e.g. links, examples, etc.) in documentation repository
• Code is ready for review