Skip to content

VALIDATION.md

Latest commit

 

History

History
43 lines (30 loc) · 3.28 KB

VALIDATION.md

File metadata and controls

43 lines (30 loc) · 3.28 KB
Raw

Claude Agent SDK v0.1.0 — validation record

Confirmed on Windows (Python 3.12, Claude Agent SDK 0.2.165)

  • python -m pip check: No broken requirements.
  • python -m pytest -q: 59 passed.
  • examples/native_core_hook_e2e.py: PASS — native hook registration, one deterministic sandbox action, real Core 0.8.1-design-partner, evidence_complete, coverage score 100, signed verification valid.
  • examples/native_core_security_e2e.py: 6/6 PASS — blocked, missing approval, human rejected, human approved, wrong arguments, and pre-execution evidence-write failure; each verified against real Core.
  • examples/native_core_recovery_e2e.py: PASS — action occurred, post-execution evidence failed, downstream reconciliation required; no false completion.
  • examples/offline_core_durability_e2e.py: PASS — SQLite restart, replay denial, one authorized sandbox action with signed Core verification; separate post-execution outage with durable reconciliation and retry denied.
  • python -m build: PASS — isolated-build wheel (.whl) and source archive (.tar.gz).
  • python -m twine check .\dist\*: PASSED for both distributions.
  • Clean wheel installation: PASS — new .venv-wheelcheck environment, dependency resolution, pip check, and from loopgrid_claude_agent import LoopGridClaudeAgent, VERSION printed 0.1.0 LoopGridClaudeAgent.

The real-Core scripts use real SDK hook option construction and deterministic callback inputs; they do not initiate a paid Claude model turn. The simulator does not move money.

Actual Claude model runtime

The diagnostic Claude Agent SDK 0.2.165 session started and returned:

  1. Without runtime authentication: Not logged in · Please run /login.
  2. With an ANTHROPIC_API_KEY: Credit balance is too low.

No model-initiated tool call ran during either attempt. The available evidence points to external API access/billing; it does not prove or disprove live tool-hook compatibility. This additional live-runtime validation remains pending.

When funded API credits are available, from a checkout with LoopGrid Core running:

.\.venv\Scripts\python.exe examples\native_claude_runtime_e2e.py

Pass criteria: exactly one authorized tool action, native pre/post-hook evidence, model-output commitment, observed sandbox receipt, evidence_complete, complete applicable coverage and independent Core verification. Inspect failure diagnostics rather than weakening permission gates.

Public release checks

  • GitHub repository pushed and GitHub Actions green on the exact release commit.
  • Dependencies constrained to the tested Claude SDK 0.2.x line, starting at 0.2.165.
  • PyPI Trusted Publisher configured for loopgridio/loopgrid-claude-agent, release.yml, environment pypi.
  • Wheel/sdist built, Twine metadata passed, and clean-wheel import checked.
  • Publish v0.1.0 as the initial release without claiming the real model-to-MCP scenario has passed.
  • Public PyPI clean installation confirmed after release.

Status: offline and Core-backed validation complete for v0.1.0. Actual Claude model-to-MCP dispatch remains unverified and is recorded above for follow-up. PyPI uses the standard Beta maturity classifier for this initial release. See RELEASE-CHECKLIST.md.