Repository navigation
MCP plate viewer cannot stream ordinary TIFF/Z-stack raw images through BioFormats #201
Description
Owner: @trissim. Reproducer (fresh-source MCP, 2026-09-29): create four 8x8 TIFF planes named A01_s001_w1_z001_t001.tif through z004 under a source plate (the fixture in tests/unit/test_streaming_service.py::test_3d_point_archive_reopens_with_native_z_domain_and_features creates them). Call openhcs_stream_plate_files_to_viewer with plate_path=source, kind=image, file_paths=[the four absolute TIFF paths], and an isolated Napari viewer port. Auto detection selects BioFormatsHandler. It fails with StorageResolutionError trying to load a nonexistent escaped-and-suffixed address such as A01%5Fs001%5Fw1%5Fz001%5Ft001.tif_s001_w1_z001_t001.tif. A single four-plane image.ome.tif also fails explicit-file resolution with PlateFileRecordAmbiguityError because several inventory records share the same physical path; querying kind=image, limit=1 then streaming constructs image.ome.tif_s001_w1_z001_t001.tif and fails loading.
Expected: MCP can stream the exact source image(s) represented by the inventory and display raw planes beside a source-bound ROI ZIP. Preserve per-plane identity and reject genuinely ambiguous selections; do not infer source identity from the ROI filename. Acceptance: focused tests for both physical-file cases plus a fresh MCP/UI viewer reopen on an isolated display, with raw-only and raw+ROI state inspected. This blocks full live visual QA for PR #198; its native ROI-only stream already succeeds.
Activity
trissim commentedon Sep 29, 2026
Fix is on draft PR #198 at 39a6d0c. The direct plate-stream route now carries inventory-owned SourcePixelRef through the viewer request and resolves exact backend plane addresses (rather than loading physical container paths as plane addresses). Focused ordinary-TIFF and four-plane OME tests pass; fresh-source MCP streamed both to isolated :91/5791, and viewer readback reports OME Z 1-4. Issue remains open pending same-coordinate raw/result/combined QA and reviewed installation.
trissim commentedon Sep 29, 2026
Fixed by merged PR #198 (main e95995e). Focused ordinary-TIFF and OME-plane tests pass; fresh installed MCP streamed four OME Z planes and a persisted source-bound point ROI into the isolated Napari viewer. Raw-only/result-only/combined same-coordinate captures and point feature selection were inspected. This validates transport/alignment for the synthetic fixture, not biological blind-analysis acceptance. The current Codex MCP child is from the older checkout and requires client reconnect to use the installed main launcher.