Changelog

All notable changes to Dual AI Verifier are documented here.

[Unreleased]

Publish the plugin source, usage documentation, project log, and CI checks.
Keep the reviewer result available as report_markdown for Codex's Outputs panel.

A single ai_review MCP tool with Gemini-first provider selection.
Automatic DeepSeek fallback when Gemini has no usable credential or cannot complete a request.
JSON review contract with pass, revise, and blocked verdicts.
Evidence, fixes, missing-test, and open-question fields for each review.
Chinese-friendly Markdown rendering through report_markdown.
Skill workflow that mirrors each review to outputs/ai-review.md.
macOS Keychain credential loading for both providers.
Local unit tests, plugin validation, and shell/Python syntax checks.

Provider credentials are read through inherited file descriptors instead of being copied into command arguments.
DeepSeek credential loading no longer uses a FIFO or a temporary directory.
Gemini credentials are sent in the x-goog-api-key header rather than a URL query string.
Review responses fail closed when required JSON fields are missing or empty.
MCP parameter errors return JSON-RPC errors instead of terminating the server.
Request, summary, context, tests, and combined input limits are enforced.
The default Gemini model was moved from the retired gemini-2.5-flash to gemini-3.1-flash-lite.

This is provider failover, not independent dual-model consensus: DeepSeek is not called when Gemini succeeds.
Codex has no universal post-final-message interception hook; enforcement occurs at the skill workflow boundary.
Review content is sent to the selected third-party provider. Do not include secrets or unnecessary private data in the context.
A missing provider credential or blocked network request yields an explicit unavailable/blocked result and never an invented pass.