vllm-test-audit

Plugin vllm-test-audit from marketplace ai-marketplace in TorchedHat/ai-marketplace. Commands read from the repo's manifest — 19 days ago.

Skills (2)

• audit-contract

  Reference-only skill defining the numeric stability audit contract. Contains the 3 inclusion criteria, strong contracts, not-strong-by-default clauses, assertion pattern guidance, and output formats for both phases. Loaded by audit-agent and review-agent as shared knowledge.

• auto-fix

  Remediation playbook for tests classified COINCIDENTALLY_CORRECT by the vLLM numeric-stability audit. Maps each fragile-oracle finding to a concrete fix — tolerance oracle, batch-invariant mode, golden strings, or principled threshold — with recipes, alignment pitfalls, and verification steps. Use w

Subagents (3)

• audit-agent

  Phase 1 agent for the vLLM test oracle auditor. Scopes to a PR, directory, or single test, finds test files, analyzes each for fragile assertions, and produces structured evidence with initial classifications. Use when auditing tests, checking a PR for brittle numeric assumptions, or analyzing test

• coordinator-agent

  Pipeline coordinator for the vLLM test oracle auditor. Orchestrates Phase 1 (audit-agent) and Phase 2 (review-agent) sequentially, verifying outputs between phases. Use when running the full audit pipeline end-to-end on a directory or PR. - audit-contract

• review-agent

  Phase 2 agent for the vLLM test oracle auditor. Adversarially verifies Phase 1 classifications by challenging each criterion rating. Loads the same audit contract as Phase 1 to ensure consistent clause references. Must run in a separate Claude invocation from Phase 1. - audit-contract

Hooks (2)

• SessionStart

  cp "${CLAUDE_PLUGIN_ROOT}/vllm_test_audit/output_object.py" "${CLAUDE_PROJECT_DIR}/output_object.py" 2>/dev/null || true

• SessionEnd

  rm -f "${CLAUDE_PROJECT_DIR}/output_object.py" 2>/dev/null || true

Similar plugins