You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Verify original official30 knowledge sections after selected-discovery repair #376
Source repair owner: managed-viewer sidecar, now distinct knowledge lazy-conversion boundary owner, branch fix/knowledge-lazy-conversion-20261001 in /home/ts/wt/openhcs-knowledge-lazy-conversion-20261001; base main6ace57656. Parent owns installed acceptance and the occupied native slot.
Retained installed reproducer: openhcs_get_knowledge_document(document_id=openhcs_official30_benchmark_recipes, section_id=cp-tutorial-pixel-based-classification-openhcs-python, max_chars=50000) returns mcp_transport_failed/TimeoutError at the original10s bound. Parent reports the same MCP PID1658909 remains healthy. No retry, restart, or larger limit authorized. Original launch/stdin/stdout: /home/ts/wt/openhcs-issue-batch-20260929/calibration371-classification372-installed-20261001 (stdin line20; stdout lines1429..1452). Private installed wheel SHA95123985 prefix, parentff52f0e includes371/372; exact wheel digest will be retained in receipt.
Trace selected source through the original KnowledgeBaseService -> import_cellprofiler_pipeline -> PipelineDocumentAuthority owners. Diagnose cold conversion/import/catalog work source-only; do not guess that a healthy transport proves conversion complete. Preserve canonical public pipeline source, errors and existing declaration registries. No duplicate converter, registry, generated-source cache, raw cppipe fallback, or bigger deadline.
Checked open issues and PRs: no matching retrieval owner. Schrodinger PR358 owns cold runtime startup/readiness (ZMQClientService, ZMQExecutionClient and paired ZMQRuntime); these files and process-lifecycle mechanisms are excluded. If the diagnosis reaches an owned seam, coordinate rather than patch it. Classification372 and calibration371 declarations/axis/runtime files also excluded.
Acceptance: original selected-source request through the existing typed capability/service; cold boundary does not prepare unrelated catalog/runtime work; independent declaration/hook case uses unchanged consumers; missing/unsupported conversion remains typed and bounded. One CPU,512MiB combined source RSS,60s per shard, no provider/plugin, native/MCP/UI/Fiji/download/install/heavy lock or parent process interaction. Parent later verifies original installed10s entrypoint; source pass is not installed readiness.
Visible diagnostic draft PR377 at19bea0153 now retains original installed stdin/stdout failure excerpts and byte-exact source logs. Bounded unchanged source converts the actual selected section (10174 chars); both global RegistryService preparation entrypoints were denied and neither was called. Therefore global catalogue preparation is not asserted as the cause. Cold profile locates eager for_module -> LazyDiscoveryDict.values -> whole CellProfiler family discovery (3.283s), including unrelated alignment imports. This is a proven cold-cost boundary, not a reproduction/proof of the installed10s timeout. A red behavioral witness forbids unrelated discovery and reaches original typed official30_conversion_failed. Related issue379 separately records endpoint-versus-local canonical lookup; no common timing cause claimed. No product edits, native/MCP retries, restart or deadline change; parent native slot untouched. Ownership/limits/archive digest/next original-owner seam are in the PR receipt.
Added the independent ExampleHuman installed timeout witness to the existing PR377 receipt (no new owner/feature/product change). Exact retained root: /home/ts/wt/openhcs-issue-batch-20260929/neurite-development-skill383-20261001/output; stdin29 requests examplehuman-openhcs-python/max_chars50000; stdout2704..2726 returns mcp_transport_failed/TimeoutError. The original read remains pending disposition, not replayed. Request/reply slice hashes are recorded without another session snapshot.
Distinct corpus-map section at stdin30 returned all487 installed source lines; that does not qualify the failed generated-source route or biological parity. Captured stderr includes earlier plate-inspection progress and a BigStitcher message; this mixed tail is not a knowledge causal stack. Installed timing cause remains unresolved; no catalogue/Fiji/size diagnosis invented. Same original declaration/conversion owners and MEMB-1/2 constraints apply. Existing WT only, no tests/new cache/install/download/native call/restart/deadline change; frozen paired installation and parent records untouched.
Production source checkpoint PR #377 is now merged to main f2aabe4, paired with actually merged metaclass-registry PR #2/main3294a69. Reviewed original discovery/import/registration template, declaration-owned selection and cooperative new-case behavior; 53 original source controls pass after normal #409 integration and original pinned R0 has no positive deltas. Keeping this issue OPEN for the actual installed knowledge-retrieval/10-second call-boundary acceptance. Current private science installation remains at #408/#411; no package change under its live viewer. Parent owns paired-wheel/native verification at safe closure. Public normal-resolver delivery additionally needs a distinguishable dependency version/minimum; unchanged version0.2.1 does not prove the new API is installed. Schrodinger owns the coordinated release followup.
Current-main audit: selected declaration-local discovery and the metaclass release dependency are already implemented. Current OpenHCS requires metaclass-registry>=0.2.2, so the old unpublished-dependency concern is obsolete. Later #499 installed acceptance retrieved cp-tutorial-translocation-start-openhcs-python through the real native CPPipe importer/document renderer: complete 8011-character source, no errors, 1.954 seconds within the normal10-second observation. That establishes working installed knowledge conversion, but not the original pixel-classification or ExampleHuman sections. Remaining scope is those specific original sections, not general packaging/discovery or an assumed current timeout. Keep catalog/kernel compilation with its current owner; do not relocate warmup or widen observation deadlines. Original receipt: docs/validation/knowledge-native-source-499-20261003.rst and runtime505/mcp.stdin.
changed the title [-]Selected official30 knowledge conversion exceeds the MCP observation deadline[/-][+]Verify original official30 knowledge sections after selected-discovery repair[/+]on Oct 6, 2026
Fixed on main by PR377 (f2aabe4) and metaclass-registry declaration-local discovery; current dependency minimum is >=0.2.2. The selected knowledge conversion no longer prepares the unrelated global catalog. Original source controls passed, and later actual installed knowledge conversion returned the complete translocation source in 1.954 seconds under the normal 10-second observation. Closing the implemented discovery/timeout cause. The exact original pixel-classification/ExampleHuman calls have not been remeasured, so this is not a claim of measured latency for every section.
Activity
test: retain knowledge discovery and custom lookup source witnesses (#…
trissim commentedon Oct 1, 2026
Visible diagnostic draft PR377 at19bea0153 now retains original installed stdin/stdout failure excerpts and byte-exact source logs. Bounded unchanged source converts the actual selected section (10174 chars); both global RegistryService preparation entrypoints were denied and neither was called. Therefore global catalogue preparation is not asserted as the cause. Cold profile locates eager for_module -> LazyDiscoveryDict.values -> whole CellProfiler family discovery (3.283s), including unrelated alignment imports. This is a proven cold-cost boundary, not a reproduction/proof of the installed10s timeout. A red behavioral witness forbids unrelated discovery and reaches original typed official30_conversion_failed. Related issue379 separately records endpoint-versus-local canonical lookup; no common timing cause claimed. No product edits, native/MCP retries, restart or deadline change; parent native slot untouched. Ownership/limits/archive digest/next original-owner seam are in the PR receipt.
trissim commentedon Oct 1, 2026
Added the independent ExampleHuman installed timeout witness to the existing PR377 receipt (no new owner/feature/product change). Exact retained root: /home/ts/wt/openhcs-issue-batch-20260929/neurite-development-skill383-20261001/output; stdin29 requests examplehuman-openhcs-python/max_chars50000; stdout2704..2726 returns mcp_transport_failed/TimeoutError. The original read remains pending disposition, not replayed. Request/reply slice hashes are recorded without another session snapshot.
Distinct corpus-map section at stdin30 returned all487 installed source lines; that does not qualify the failed generated-source route or biological parity. Captured stderr includes earlier plate-inspection progress and a BigStitcher message; this mixed tail is not a knowledge causal stack. Installed timing cause remains unresolved; no catalogue/Fiji/size diagnosis invented. Same original declaration/conversion owners and MEMB-1/2 constraints apply. Existing WT only, no tests/new cache/install/download/native call/restart/deadline change; frozen paired installation and parent records untouched.
docs: retain second installed knowledge timeout witness (#376)
Merge pull request #377 from OpenHCSDev/fix/knowledge-lazy-conversion…
trissim commentedon Oct 2, 2026
Production source checkpoint PR #377 is now merged to main f2aabe4, paired with actually merged metaclass-registry PR #2/main3294a69. Reviewed original discovery/import/registration template, declaration-owned selection and cooperative new-case behavior; 53 original source controls pass after normal #409 integration and original pinned R0 has no positive deltas. Keeping this issue OPEN for the actual installed knowledge-retrieval/10-second call-boundary acceptance. Current private science installation remains at #408/#411; no package change under its live viewer. Parent owns paired-wheel/native verification at safe closure. Public normal-resolver delivery additionally needs a distinguishable dependency version/minimum; unchanged version0.2.1 does not prove the new API is installed. Schrodinger owns the coordinated release followup.
Merge pull request #414 from OpenHCSDev/fix/selected-discovery-minimu…
trissim commentedon Oct 6, 2026
Current-main audit: selected declaration-local discovery and the metaclass release dependency are already implemented. Current OpenHCS requires metaclass-registry>=0.2.2, so the old unpublished-dependency concern is obsolete. Later #499 installed acceptance retrieved cp-tutorial-translocation-start-openhcs-python through the real native CPPipe importer/document renderer: complete 8011-character source, no errors, 1.954 seconds within the normal10-second observation. That establishes working installed knowledge conversion, but not the original pixel-classification or ExampleHuman sections. Remaining scope is those specific original sections, not general packaging/discovery or an assumed current timeout. Keep catalog/kernel compilation with its current owner; do not relocate warmup or widen observation deadlines. Original receipt: docs/validation/knowledge-native-source-499-20261003.rst and runtime505/mcp.stdin.
[-]Selected official30 knowledge conversion exceeds the MCP observation deadline[/-][+]Verify original official30 knowledge sections after selected-discovery repair[/+]trissim commentedon Oct 6, 2026
Fixed on main by PR377 (f2aabe4) and metaclass-registry declaration-local discovery; current dependency minimum is >=0.2.2. The selected knowledge conversion no longer prepares the unrelated global catalog. Original source controls passed, and later actual installed knowledge conversion returned the complete translocation source in 1.954 seconds under the normal 10-second observation. Closing the implemented discovery/timeout cause. The exact original pixel-classification/ExampleHuman calls have not been remeasured, so this is not a claim of measured latency for every section.