fix(memory): flatten multimodal content before provider sync

[erosika]

Problem

Turns with an attached image log this and silently never reach the memory backend:

WARNING agent.memory_manager: Memory provider 'honcho' sync_turn failed: expected string or bytes-like object, got 'list'

Multimodal user messages carry content as a list of typed parts ({type: "text"|"image_url", ...}). _sync_external_memory_for_turn passed that list straight into MemoryManager.sync_all, and providers feed it to regexes — Honcho's sync_turn calls sanitize_context in its synchronous prologue, where re.sub raises the TypeError before the sync thread ever spawns.

The chain:

1. agent/turn_context.py — original_user_message is the raw content (Any), a list for multimodal turns
2. agent/turn_finalizer.py → run_agent.py: _sync_external_memory_for_turn — no string guard (other consumers of original_user_message in turn_context.py do guard with isinstance(..., str))
3. plugins/memory/honcho/__init__.py: sync_turn — sanitize_context(user_content or ""): a non-empty list is truthy, so the list hits re.sub

Other plugins (mem0, hindsight, supermemory, …) would choke on a list the same way.

Fix

Flatten content to plain text at the boundary in _sync_external_memory_for_turn, before sync_all/queue_prefetch_all — one fix covers every provider instead of patching each plugin. New helper flatten_message_content in agent/memory_manager.py (alongside sanitize_context, its main consumer):

• strings pass through, None → ""
• text parts joined with newlines
• images become a [N image(s)] marker so the attachment isn't erased from recall
• turns with no recoverable text (e.g. unknown part types only) are skipped rather than synced empty

Both original_user_message and final_response are guarded — both are typed Any.

Tests

• tests/agent/test_memory_provider.py — unit coverage for flatten_message_content, including a regression case asserting the flattened output survives sanitize_context
• tests/run_agent/test_memory_sync_interrupted.py — boundary coverage: multimodal user message, multimodal response, and no-recoverable-text skip

All 141 tests across the touched suites pass.

[liuhao1024]

Verified flatten_message_content — the multimodal content→string boundary is correct:

1. Type dispatch: None→"", str→passthrough, list→joined text + image markers, fallback str(content).
2. Image counting: pluralization handled ([1 image] vs [2 images]), image-only messages produce marker-only output.
3. Guard in caller: _sync_external_memory_for_turn checks if not (user_text and response_text): return — skips sync when flattening yields empty strings, preventing vacuous memory writes.
4. Content type coverage: text, input_text, output_text, image_url, input_image all handled; unknown types silently skipped.

Test suite covers all branches including the original crash case (sanitize_context(list)). CI green. Clean fix.

[kshitijk4poor]

Thanks for this @erosika — the diagnosis is exactly right: multimodal content reaches the provider sync as a list and crashes sanitize_context's re.sub before the sync thread spawns, and flattening at the _sync_external_memory_for_turn boundary is the right place to fix it once for every provider.

I've opened #44738 as a salvage that cherry-picks your commit verbatim (authorship preserved) and adds one follow-up: the new flatten_message_content turned out to be a near-duplicate of the existing agent/codex_responses_adapter.py:_summarize_user_message_for_log (same dispatch, same part-type sets, identical [N image(s)] marker), which is already imported into turn_finalizer.py — the same file your call site lives in. The follow-up parameterizes that helper with a sep arg (sep="\n" for memory, default sep=" " keeps the logging callers unchanged) and drops the fork, so there's a single source of truth.

Closing this in favor of #44738 — full credit to you for the fix. Happy to fold in any feedback there.