Follow-up to #463 (chunker reader). Two coupled writer-side changes plus the e2e round-trip the chunker enabled.

Summary

• Writer canonical form — append_entry now emits > tags: ["a", "b"] (explicit > prefix on every line, JSON-quoted) instead of the lazy-continuation form tags: ['a', 'b'] (Python repr()). Chunker reader (PR feat(chunking): promote per-entry blockquote tags to ChunkMetadata.tags #463) already accepts both shapes, so old files still parse.
• mem_edit header preservation — new replace_chunk_body helper keeps the heading + > created: / > tags: blockquote intact when the caller supplies body-only new_content. Prefix new_content with ## to bypass preservation and revert to full-replacement (pre-RFC semantic).
• mem_search(tag_filter=...) round-trip — multi-agent e2e now asserts mem_search(tag_filter=f"shared-from={src}") returns the share copy. Closes the gap discovered in PR test(multi-agent): end-to-end integration coverage #462.

Why two writer changes in one PR

The reviewer flagged on #463 that mem_edit's existing replace_lines call wasn't a new regression but became a much louder footgun once the chunker strips the blockquote from chunk content. Post-PR-A, a caller who reads chunk.content and feeds it back through mem_edit would supply body-only and silently erase the metadata header. Fixing that in the same PR as the canonical writer keeps the related contract change coherent.

What changed

• packages/memtomem/src/memtomem/tools/memory_writer.py
  • append_entry switches to > tags: {json.dumps(list(tags))}.
  • New _find_body_start_index(chunk_lines) walks a chunk's line range and returns the index where the body starts, after heading + leading blockquote group (with lazy-continuation tolerance for legacy on-disk files).
  • New replace_chunk_body(file_path, start_line, end_line, new_content) preserves heading + leading blockquote unless new_content starts with ## (full replace).
• packages/memtomem/src/memtomem/server/tools/memory_crud.py
  • mem_edit calls replace_chunk_body instead of replace_lines.
• packages/memtomem/tests/test_memory_writer.py — existing tag-format pin updated to canonical form.
• New tests/test_memory_writer_tag_format.py — 5 cases pinning the canonical output (explicit > prefix, JSON serialization, round-trip via json.loads, empty-tag omission, > created: prefix preserved).
• New tests/test_mem_edit_header_preservation.py — 6 cases (canonical / legacy lazy / no-header / explicit ## full-replace / non-first sub-chunk of oversized section / trailing-newline round-trip).
• tests/test_chunking_blockquote_tags.py — adds block-list-shape case (> tags:\n> - a\n> - b). Writer never emits this, but the shared _parse_tags_value helper supports it and we lock that in.
• tests/test_multi_agent_integration.py test_share_copies_chunk_with_audit_tag — adds forward-direction tag_filter round-trip assertion.

Test plan

• Full suite: uv run pytest -m "not ollama" → 2402 passed, 0 failed (PR-A baseline 2390 + 12 new).
• uv run ruff check && uv run ruff format --check (full repo).
• uv run mypy on changed source files — clean.

Out of scope (still tracked)

Three remaining follow-ups noted in #463 review, all in planning/mem-add-tags-blockquote-promote-rfc.md §Follow-ups:

• _split_section line-number drift after blockquote strip (small follow-up PR; mem_edit / UI only, search unaffected).
• add_entries_batch broadcast-tag cleanup (memory_crud.py:392–401); orthogonal cleanup PR.
• chunk_links table to permanently close the shared-from=<old-uuid> audit-chain gap on reindex UUID churn — separate RFC mem-agent-share-chunk-links-rfc.md.

🤖 Generated with Claude Code