Skip to content

VALIDATION.md

Latest commit

 

History

History
31 lines (23 loc) · 3.69 KB

VALIDATION.md

File metadata and controls

31 lines (23 loc) · 3.69 KB
Raw

Validation record and limitations

This document distinguishes real user-observed tests from tests still required for the new published launcher arrangement.

Completed on Windows (October 11, 2026)

  • Previous uv run --locked plugin: 50 Python tests passed; strict Claude Code plugin validation and genuine uv.lock check passed.
  • Windows Claude Code 2.1.292: original plugin MCP connected with six LoopGrid tools in a clean disposable project.
  • Original plugin + LoopGrid Core 0.8.1: recorded synthetic proposed refund decisions; retry with same idempotency key replayed without duplicating evidence.
  • Fresh Core workspace ws_6f03674a26654418: synthetic five-event Core sandbox policy allow/deny, outcome evidence and signed workspace verification passed.
  • Live Claude Code tool hook: local SQLite showed Bash | result_recorded; the corresponding Core decision dec_009b1bdc33d64086a85eb8a4 verified valid=true, three checked decision events, no failures (10 workspace ledger events).
  • Exported signed evidence ZIP: checksum matched; loopgrid-verify==0.1.0 reported [OK] VERIFIED, bundle integrity attested. The signer key ID was supplied from Core; independent real-world signer identity was not established through separately trusted public-key material.
  • A separate historical default workspace had a known ledger hash-chain failure at event 89. Never claim that workspace passed integrity checks; isolated test workspace was valid.

Published Claude-only launcher tests

  • loopgrid-claude-launcher==0.1.0 was independently published to PyPI from loopgridio/loopgrid-claude-launcher using GitHub OIDC trusted publishing.
  • User installed the published package with uvx --from loopgrid-claude-launcher==0.1.0 loopgrid-claude-launcher; Windows MCP stdio discovery returned all six LoopGrid tools.
  • Security smoke test on the previously built launcher reported inherited AWS and machine-wide LoopGrid service keys filtered, explicitly selected service key forwarded, and remote non-HTTPS Core rejected.
  • Launcher repository GitHub CI was green on Ubuntu, Windows, and macOS (Python 3.12).

Current Claude plugin release checkpoint (October 11, 2026)

  • Completed: Genuine uv.lock generated and checked on Windows (38 resolution entries; the published loopgrid-claude-launcher==0.1.0 and loopgrid-mcp==0.1.0 packages are represented).
  • Completed: Clean Windows Claude Code /mcp smoke test using the pinned published uvx plugin showed a connected plugin:loopgrid:loopgrid server and all six tools. Strict manifest validation passed.
  • Completed: GitHub main @ 8b04b09 CI passed 49 unit tests on Ubuntu and Windows, Python 3.10 and 3.12. A prior CI run failed because the diagnostic test assumed uv existed on bare runners; the test was corrected and CI passed.
  • Prepared: A narrow scanner-review patch changes only hook environment selection, a few security test fixtures and submission documentation. It passes 50 local tests. Re-run all GitHub CI jobs and Claude manifest validation after committing.
  • Captured Anthropic scan: main @ 8b04b09 structurally passed but still reported 4 policy holds and 4 field warnings. The next source scan is needed to determine which, if any, clear. The uvx transitive dependency review remains a substantive human-review question.
  • Before submission: verify the form's compliance attestation for the three declared Python hook commands, confirm actual Core data retention matches the privacy description and obtain an Anthropic reviewer interpretation if necessary.

The plugin does not execute a real refund or other financial action. Tool-result receipts are not independently observed downstream business outcomes.