Skip to content

VALIDATION.md

Latest commit

 

History

History
68 lines (51 loc) · 2.48 KB

VALIDATION.md

File metadata and controls

68 lines (51 loc) · 2.48 KB
Raw

Validation record — RC3

Target matrix:

  • loopgrid-haystack: 0.1.0 RC3
  • Haystack: haystack-ai==3.3.0
  • LoopGrid Python SDK: loopgrid==0.8.0
  • LoopGrid Core: 0.8.1-design-partner
  • Python: 3.10–3.14

Required local checks

python -m pytest -v
python -m compileall -q src examples tests
python -m build
python -m twine check dist\*

Real Core release gate

Both deterministic E2Es must finish with:

lifecycle = evidence_complete
coverage.score = 100
coverage.complete = true
verification.valid = true
verification.failures = []

examples/human_approval_e2e.py must additionally prove:

human_approved < tool_requested < tool_executed

No public-release claim should be made from RC3 until these checks have actually passed.

RC2 regression focus retained

RC1 exposed a real Haystack tracing-composition edge case: the public haystack.tracing.tracer object is a process-level facade and cannot safely be captured as a delegate before replacing the enabled tracer. Doing so can recurse into the newly installed LoopGrid tracer.

RC2 removed automatic/global-facade delegation. Coexistence remains supported only through an explicitly supplied concrete Haystack tracer instance. The real-runtime test disables any prior tracing state, installs LoopGrid explicitly, and restores disabled tracing afterward.

RC3 hardening focus

RC3 does not add product scope. It closes release-quality edge cases identified during the pre-package audit:

  • tool event identity includes the Haystack Agent step + per-step ordinal, so a legitimate later call is not collapsed merely because a provider reuses the same call ID or omits IDs;
  • result correlation is scoped to the current tool step instead of searching all earlier results;
  • explicit record_human_review() and record_outcome() drain queued framework evidence first, preserving lifecycle order under slow transport;
  • native ConfirmationHook rejection regression proves rejected calls produce neither tool_requested nor tool_executed evidence;
  • native ConfirmationHook modification regression proves LoopGrid records the surviving modified parameters that Haystack actually executes;
  • a real Agent.run_async() runtime regression exercises async hooks and an async tool;
  • README tracing language now matches the RC2 concrete-delegate rule.

The full local suite is expected to contain 29 tests after this hardening pass. Real-Core E2Es are rerun only after all local tests and compile checks pass.