What

IndexEngine.index_path_stream (engine.py:707) and the GET /api/index/stream endpoint (system.py:768) lack two contract elements that the non-stream index_path / POST /api/index pair already has:

1. namespace parameter — silently dropped, so any UI/CLI/MCP caller that streams an index ignores the user's namespace selection.
2. Error aggregation — _index_file returns result["errors"] per file (engine.py:606, 628, 684, 752), but the stream loop sums only chunk counters (754-757), never propagating errors. The complete event lacks an errors field, so partial-failure runs (e.g. ONNX missing on a subset, file-too-large, binary-detected) appear successful in the UI.

Fix

• Engine: add namespace: str | None = None to index_path_stream, pass through to _index_file. Aggregate result["errors"] into a loop accumulator and include it as errors: tuple[str, ...] in the complete event — same shape as IndexingStats.errors so the existing non-stream UI handler reuses verbatim. Per-file uncaught-exception branch (engine.py:752) prefixes f"{fp.name}: {exc}" to match non-stream's asyncio.gather(return_exceptions=True) branch (engine.py:394).
• Route: add namespace query param to GET /api/index/stream, forward to the engine call.
• Tests:
  • test_index_stream_namespace_propagates — namespace="x" → indexed chunks have metadata.namespace == "x".
  • test_index_stream_complete_errors_no_silent_drop — force a file-level failure → complete.errors non-empty (regression guard against current silent-drop behavior; the test name documents intent so future grep finds it).

Why now

Unblocks #582 PR-B (Prev #1 — Index tab two-button collapse) and informs #582 PR #6 (4.11 — indexing in-flight visibility), which will surface complete.errors via toast.

Out of scope

• Per-file streaming of errors (option ii/iii from the planning discussion). complete.errors is sufficient for current UX. Revisit if user reports indicate large indexing runs (>100 files) where waiting for complete to surface partial failures is impractical.
• Normalizing _index_file's mixed path-prefix convention — some branches prefix f"{file_path.name}: ..." (file-too-large, binary), others don't (embedding failures at line 684). That's a separate non-stream bug. The PR for this issue preserves the existing loose contract and only path-prefixes the stream-level uncaught-exception branch (line 752) to match non-stream's outer-gather branch.

Refs

• Engine: packages/memtomem/src/memtomem/indexing/engine.py:348-414 (non-stream), :707-776 (stream).
• Route: packages/memtomem/src/memtomem/web/routes/system.py:768-804 (stream), :807-832 (non-stream).
• IndexingStats: packages/memtomem/src/memtomem/models.py:172-185.

🤖 Generated with Claude Code