Skip to content

Fix source-aware orchestration and native document review - #1931

Merged
Paul Lizer (paullizer) merged 1 commit into
microsoft:Developmentfrom
paullizer:paullizer-deployed-retest-investigation
Nov 13, 2026
Merged

Paul Lizer (paullizer) merged 1 commit into
microsoft:Developmentfrom
paullizer:paullizer-deployed-retest-investigation

Conversation

@paullizer

Copy link
Copy Markdown
Collaborator

Why

The deployed retest exposed plans that looked valid but could not execute against the selected file schemas or native reader formats. Some failures were reported as missing named results, while correct native titles could be truncated by Markdown table parsing.

Approach

  • Validate complete native header assertions during planning when authoritative schema information is available, and preserve specific semantic rejection rules for bounded correction. Keep projections separate from source-schema assertions and calculations separate from preview retrieval.
  • Make document Gather references and report sections source-aware. Allow one known incompatible-section correction before source reading, without extending the existing call budget or replaying failed reads/effects.
  • Read source-bound PDF page annotation contents and supported OBJR bindings, preserving reply references and absent/empty/inaccessible states with bounded text and records.
  • Preserve native values when serializing table cells, stop at Markdown block interruptions, and retain inline HTML in valid data rows. Version native answer checking as v5 and native pagination as 1.1.
  • Improve coverage and retained-result refusal diagnostics without weakening guards. Preserve semantic document understanding, deterministic calculation, and mixed interpretation-plus-computation workflows. Advance the application to 0.261.439 and update related documentation/regressions.

Validation

Functional runs used network blocking, PYTHONDONTWRITEBYTECODE=1, and a session-local pytest cache. Repository-relative commands below omit only that cache location.

  • python .\scripts\check_cosmos_response_boundaries.py --run-contracts - 1,565 passed, 1,104 reviewed owners, 430 doubles, 0 review gaps.
  • python -m pytest -q -p test_support.cosmos_network_guard --tb=short --disable-warnings functional_tests\test_orchestration_planner_failure_diagnostics.py functional_tests\test_orchestration_planner_contract_recovery.py functional_tests\test_orchestration_services.py functional_tests\test_orchestration_document_action_binding.py functional_tests\test_orchestration_relational_results.py functional_tests\test_native_document_fidelity.py functional_tests\test_native_audit_reliability.py functional_tests\test_tabular_relational_calculations.py functional_tests\test_orchestration_failure_telemetry.py - 272 passed.
  • python -m pytest -q -p test_support.cosmos_network_guard --tb=short --disable-warnings functional_tests\test_native_audit_answer_delivery.py functional_tests\test_native_audit_reliability.py functional_tests\test_orchestration_search_result_integrity.py - 116 passed after final serialization refinements.
  • python -m pytest -q --tb=short --disable-warnings ui_tests\test_v2_chat_table_layout.py - 11 passed in the actual V2 renderer, including mobile/desktop title cells and immediately following headings.
  • npm --prefix .\application\v2_ui run build -- --outDir ..\..\ui_tests\artifacts\orchestration-plan-editor - passed.
  • python .\functional_tests\test_docs_app_surface_coverage.py - 7/7 passed; committed surface inventory remains current.
  • python .\functional_tests\test_docs_site_quality.py - 6/6 passed.
  • Existing XSS/access-control inspectors against the actual added application lines - 0 issues. git --no-pager diff --check - passed.

The first full contract run caught a long-input regression from verbose guidance. Shortening the guidance restored complete excerpt delivery without truncating inputs or increasing limits; the full gate rerun passed.

Boundaries and remaining gaps

  • The historical TRE-09 retained-result refusal is not claimed resolved. Its guard is unchanged; precise lower-code and hashed-alias diagnostics were added.
  • The exact lower rules of 13 historical generic planning rejections and per-request immutable historical serving provenance remain unverified.
  • No deployment, live model Send/Retry/replay, source repair, or live acceptance retest occurred. Original retest exports remain unchanged; their acceptance statuses were not rewritten.
  • Release notes remain unchanged at the user's request. No new settings, routes, browser runtime libraries, or deployer changes.

Details: docs/explanation/fixes/DEPLOYED_RETEST_TOOL_CONTRACT_FIX.md.

Validate native source schemas and preserve actionable planning diagnostics. Read source-bound PDF annotations, retain saved-result safeguards, and serialize native values without corrupting Markdown block boundaries. Add regression coverage and documentation; advance the application to 0.261.439.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@paullizer
Paul Lizer (paullizer) merged commit c158f26 into microsoft:Development Nov 13, 2026
13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant