v0.7.1a1 - P1 Alpha

What this Alpha adds

• IntentBench v2 with 100 bilingual and mixed-language development cases.
• Blinded challenge-set import, execution, and reporting tools.
• Same-model with-skill versus without-skill experiment protocol and scorer.
• Project-scoped corrections, Interpretation Gate continuity, and 1 / 第一个 / option-ID resolution.
• Recipient-aware local preview generation with disclosure boundaries.
• Invocation receipts, operator-driven Codex execution traces, privacy-bounded feedback, and real-user trial tooling.
• Fail-closed handling for vague external actions and tighter action-bound approvals.

Verification

• 245 local tests passed.
• Full release gate passed.
• Clean-wheel doctor, onboarding, MCP import, and IntentBench v1/v2 checks passed.
• Creator-shadow leakage: 0.
• Default-user contamination: 0.
• Secret/private-path findings: 0.
• GitHub Actions passed for Windows, macOS, and Linux on Python 3.10 and 3.12.
• Browser smoke, package workflow, and CodeQL passed.

Alpha status

This is a P1 Alpha pre-release. The trial protocol and privacy-bounded evidence tools are ready, but independent evidence from 3-5 real users is still pending. The same-model A/B protocol is prepared; this release does not claim an independent model-performance result. Automatic host interception remains experimental until verified by each supported host.

See the README for installation, onboarding, evaluation, and uninstall instructions.