Truthound 3.3.1

This patch ships clearer local MCP tool descriptions and JSON Schema field explanations. The tool IDs remain check_file, profile_file and scan_pii. No validation capability, directory grant, output disclosure or network transport is added. check_file checks nulls/duplicate rows; profiling observes types and aggregate statistics; PII findings remain sampled estimates.

Installation and publication

Published on 2026-11-02 from Core source 80f8092d30918c825154e3ff8e5ed3ca4c328b17. Complete same-source verification and PyPI publication succeeded. PyPI 3.3.1 and the active/latest Official MCP Registry record were independently observed. Canonical wheel/sdist hashes match PyPI; the GitHub v3.3.1 release retains those artifacts.

python -m pip install "truthound[mcp]==3.3.1"

Local plugin 0.1.7 pins published Core 3.3.1 with a frozen lock. The existing separate Core plugin is 0.2.1. Core artifact and later plugin-repin source commits are recorded separately. The Styxis public catalog still contains only the Local truthound plugin; no product or vendor directory listing is added.

Verification and limits

Release gates retain complete same-revision tests, native compatibility probes, English/Korean documentation and exact wheel/sdist provenance. Core metadata improvements alone do not establish Agent selection accuracy. Record actual client/model observations separately; no universal discovery guarantee is made. Installation grants no data access and raw values are not returned by the Local summary tools. Paths, column names and aggregates can reach the host model.

Actual Agent selection readiness was attempted through existing Codex/Claude CLI logins. Codex rejected model requests with model_not_found; Claude organization policy disabled subscription access. Observed decisions are zero and accuracy/false-selection rates remain unmeasured. These availability failures are not tool-choice errors. No new key/account or policy bypass was attempted.

The initial macOS/Python 3.13 CI attempt reported a native AI proposal validation failure. It did not reproduce locally; its root cause remains unconfirmed. A bounded same-source retry passed the existing complete gate before publication. The distribution verification record retains the first failure ZIP and retry evidence. No tests, deadlines or permission checks were relaxed.