Persona Review

日本語

Evidence-based code reviews in a character's voice, with optional expressions. Codex records the technical assessment first, then changes its wording while preserving findings, severity, evidence, and uncertainty. A blunt persona must not invent defects; clean code does not need a criticism.

Development preview: 0.4.0-rc.1 is not released. The helper currently requires Python 3.10+. The shell and PowerShell scripts launch Python. A Bash alternative is under discussion and is not implemented.

Requirements

A signed-in Codex environment supporting local plugins, skills, and local file execution. The manifest declares requires_local_executor; this is not a cloud-only plugin. Availability depends on the Codex version and account.
Python 3.10+ for configuration, validation, and image management. Its standard library is sufficient: no pip packages, Node.js, jq, HTTP/MCP server, or separate helper API key. JavaScript development tools do not imply Python is installed.
Codex's image_gen and a registered reference image for new expressions. If generation, saving, or display fails, the review finishes with an explanation. No images are bundled.
A connected GitHub reader for GitHub URLs, or local files to review.

Platform	Helper	Optional launcher
Windows	Python 3.10+	PowerShell 5.1+; checks `py -3`, `python3`, then `python`
macOS	Python 3.10+	POSIX `sh`; checks `python3`, then `python`
Linux	Python 3.10+	POSIX `sh`; checks `python3`, then `python`

Launchers never install a runtime or change execution policy. See runtime setup.

Install a reviewed local package

Obtain a reviewed persona-review-marketplace-<version>.zip from the maintainer. This preview has no published release. Private source access requires permission; maintainers can build the archive.

Expand the ZIP into a new directory. It includes the catalog and correct plugin layout, so no JSON editing is needed.
Use your normal signed-in Codex. Check codex --version, codex plugin --help, and codex login status. If the command is missing, follow official CLI setup or your installed CLI's launch instructions. Do not hard-code internal desktop app paths or copy credentials from another user.

From the extracted persona-review-marketplace directory, run:

codex plugin marketplace add .
codex plugin add persona-review@persona-review-local
codex plugin list --json

Start a new Codex session. Confirm persona-review:review and persona-review:review-config are available in its skill catalog, then ask:

Use persona-review:review-config to show my current settings.

Normal use should not require an absolute skill path. If the skills are missing, follow plugin discovery checks.

Windows extraction example, from the folder containing the ZIP:

Expand-Archive -LiteralPath .\persona-review-marketplace-0.4.0-rc.1.zip -DestinationPath .\persona-review-preview
Set-Location .\persona-review-preview\persona-review-marketplace

macOS/Linux, using an existing unzip command or a graphical archive tool:

unzip persona-review-marketplace-0.4.0-rc.1.zip -d persona-review-preview
cd persona-review-preview/persona-review-marketplace

These CLI commands were available in 0.163.0-alpha.10.2; check local help for other versions. Local marketplaces are separate from public directory submission. The single-plugin ZIP is not a marketplace directory. Installing a local package does not publish it.

Use it

Use persona-review:review-config to register Mina with a calm, direct voice. Use my attached image as the reference and choose hybrid image mode.

Ask Mina to review this project's changes, focusing on security and correctness.

Keep Mina's appearance but use the registered direct persona for this review.

Appearance and voice are independent. Scope is auto / diff / full / file. Focus accepts correctness, security, performance, architecture, maintainability, and testing. Focus never hides serious findings outside it. Unread code and tests that were not run stay unverified.

Image mode	Behavior
`always`	Generate and save a new expression for each review.
`cached_only`	Use valid saved images only; never generate missing ones.
`hybrid`	Reuse a matching expression; generate and save it when missing.
`pre_generate`	Prepare happy, neutral, concerned, and stern; reuse valid images and retry missing ones.

Image identity includes the character, appearance, and reference-image hash. Changing only the voice reuses images. Images and provenance are saved outside the plugin; old images remain. cached_only can use valid saved images without image generation. When no usable image exists, the output explains why and includes the technical review.

Try the helper

From the plugin directory, use sh scripts/persona-review.sh on macOS/Linux or & .\scripts\persona-review.ps1 in Windows PowerShell. A verified Python executable can also run scripts/persona_review.py directly.

sh scripts/persona-review.sh --data-dir "../demo data" show
sh scripts/persona-review.sh --data-dir "../demo data" register-persona calm examples/persona.json
sh scripts/persona-review.sh --data-dir "../demo data" register mina examples/character-v2.json
sh scripts/persona-review.sh --data-dir "../demo data" set character_id mina
sh scripts/persona-review.sh --data-dir "../demo data" resolve
sh scripts/persona-review.sh validate-review examples/review.json

Put --data-dir before the subcommand: --data-dir PATH resolve, not resolve --data-dir PATH. Use a new temporary directory for experiments. Saving commands write to the selected location.

Defaults are %LOCALAPPDATA%/persona-review on Windows, or $XDG_DATA_HOME/persona-review / ~/.local/share/persona-review on macOS/Linux. PERSONA_REVIEW_HOME and --data-dir can select another trusted location. Project overrides are disabled by default. See configuration and migration and image workflow.

Reading v1 configuration is nonmutating. A write or migrate makes an exact backup before saving v2. Old A means always; old B means pre_generate and can now generate missing expressions. Choose cached_only to forbid generation. Do not write migrated configuration with an older helper.

Validation and limits

Offline CI requires no model/API credentials. The matrix is configured for Windows/macOS/Linux with Python 3.10 and 3.13, including launcher and archive tests. Check results for the specific commit; adding a matrix is not evidence that it ran. Baseline v0.3.0 passed 82 tests in each Windows/Linux job.

Independent Windows checks of v0.3.0 confirmed two clean plugin installations and fresh-data reviews with generated images. Authenticated model execution was separate and used explicit installed skill paths. This does not prove fully unaided installation, Desktop final-image rendering, or non-Windows Codex integration. See recorded validation scope.

From a source checkout (tests and build tools are not in the runtime ZIP):

python -m unittest discover -s tests -v
python tests/review_quality.py check tests/fixtures/review_quality/reference-results.json
python tools/build_dist.py --output dist
python tools/build_dist.py --output dist --check

Quality fixtures are hand-authored contracts, not real LLM results. Neither schema validation nor the small live evaluation establishes general accuracy. Image checks cover headers, size, hashes, and provenance; the skill must inspect image contents. The helper never runs reviewed code or calls an LLM/image API. See quality evaluation and design.

License

Apache-2.0. User images, reviewed code, settings, credentials, and generated review data are not bundled. Do not commit them.