candid

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

<p align="center">
  <img src="assets/logo.svg" alt="Candid Logo" width="250"/>
</p>

# Candid

A coding-agent plugin for configurable code reviews that combine thoroughness with actionable feedback. Runs on **Claude Code** and **Codex CLI** from the same install. Based on Kim Scott's Radical Candor framework: **Care Personally + Challenge Directly**.

## Overview

Candid provides AI-powered code reviews that catch issues before production while teaching you better patterns. Choose between Harsh (brutal honesty) or Constructive (caring + challenging) tone. Define project-specific standards in Technical.md that get enforced automatically. Every issue comes with concrete fixes rated by confidence level.

## Core Workflow

The review process follows these steps:

1. Run `/candid-review` on your changes
2. Select your preferred tone (Harsh or Constructive)
3. Candid reviews with full architectural context
4. Get categorized issues with actionable fixes
5. Select which issues to track as todos
6. Apply fixes and optionally auto-commit

## Key Features

- **Configurable Tone** - Harsh (brutal honesty) or Constructive (caring + challenging)
- **Technical.md Support** - Define and enforce project-specific coding standards
- **Focus Modes** - Target reviews on security, performance, architecture, or edge cases
- **Actionable Fixes** - Every issue includes concrete code with confidence ratings
- **Re-Review** - Track progress on fixes across review sessions
- **Auto-Commit** - Automatically commit applied fixes with detailed messages
- **Todo Integration** - Convert issues to tracked todos with multi-select
- **Issue Categorization** - Organized by severity (Critical → Architectural)
- **Context Optimization** - Audit and optimize token usage for leaner, deeper reviews
- **Live Browser QA** - Drive a real Chrome session against your running app, walk it like a real user across desktop and mobile, and emit structured findings JSON ready for triage

## Installation

### Claude Code

```bash
npx skills add https://github.com/ron-myers/candid
```

Then restart Claude Code.

### Codex CLI

In an interactive Codex session:

```
/plugins
```

Add the marketplace via GitHub shorthand (`ron-myers/candid`), install `candid`, then verify with `/skills`. Invoke skills via `$skill-name` mention syntax (e.g., `$candid-review --harsh`). See [docs/codex/install.md](https://github.com/ron-myers/candid/blob/HEAD/docs/codex/install.md) for the full command-to-skill mapping.

## Quick Start

**Basic review:**
```
/candid-review
```

**With tone preset:**
```
/candid-review --harsh
/candid-review --constructive
```

**Focus on specific aspects:**
```
/candid-review --focus security
/candid-review --focus performance
```

**Auto-commit applied fixes:**
```
/candid-review --auto-commit
```

**Live browser QA pass:**
```
/candid-chrome-qa
```

Drives a real Chrome session against your running app, walks it like a real user across desktop and mobile, and writes structured findings to `.context/findings/<date>-<slug>.json`. See [Candid Chrome QA docs](https://www.candid.tools/docs/core-features/candid-chrome-qa) for the full workflow and v2.0 schema reference.

## Documentation

Full documentation at [www.candid.tools](https://www.candid.tools):

- **[Getting Started](https://www.candid.tools/docs/get-started)** - Installation and first review
- **[Core Features](https://www.candid.tools/docs/core-features)** - Tone selection, focus modes, Technical.md
- **[Team Setup](https://www.candid.tools/docs/how-to-guides/team-setup)** - Rolling out to your team
- **[Custom Standards](https://www.candid.tools/docs/how-to-guides/custom-standards)** - Writing effective Technical.md rules
- **[Complete Reference](https://www.candid.tools/docs/reference)** - All configuration options and flags
- **[FAQ & Troubleshooting](https://www.candid.tools/docs/tips-troubleshooting/faq)** - Common issues and solutions

## Technical.md

Define project-specific standards that Candid enforces during reviews. Violations appear as 📜 Standards Violation in your review.

**Quick setup:**
```
/candid-init                    # Auto-generate from codebase (thorough analysis)
/candid-init react              # React-specific standards
/candid-init minimal            # Minimal starter
/candid-init --effort quick     # Fast analysis (~30 sec)
/candid-init --effort thorough  # Deep analysis (~3-5 min, default)
/candid-init --optimize         # Run /candid-optimize after generation
```

**Or copy a template:**
```bash
cp templates/Technical-minimal.md ./Technical.md
cp templates/Technical-react.md ./Technical.md
cp templates/Technical-node.md ./Technical.md
```

Keep it focused: under 500 lines, verifiable rules only. Skip what your linter handles. See the [Custom Standards guide](https://www.candid.tools/docs/how-to-guides/custom-standards) for detailed guidance.

## Configuration

Persist your tone preference and other settings in config files:

**User-wide default:**
```bash
mkdir -p ~/.candid
echo '{"tone": "harsh"}' > ~/.candid/config.json
```

**Project-specific:**
```bash
mkdir -p .candid
echo '{"tone": "constructive", "autoCommit": true}' > .candid/config.json
```

See the [Config Reference](https://www.candid.tools/docs/reference/config-options) for all options including `exclude` patterns, `focus` areas, and `mergeTargetBranches`.

## Philosophy

This plugin is built on the Radical Candor principle that the best feedback:

1. **Cares Personally** - Shows understanding of context and difficulty
2. **Challenges Directly** - Doesn't hedge or soften real issues

Whether you choose Harsh or Constructive tone, every review aims to find real issues before production, provide immediately applicable fixes, help you track what needs doing, and teach patterns that prevent future problems.

## Updating

```bash
claude plugin update candid@candid
```

Then restart Claude Code. See [CHANGELOG.md](https://github.com/ron-myers/candid/blob/HEAD/CHANGELOG.md) for version history.

## Links

- [GitHub Repository](https://github.com/ron-myers/candid)
- [Documentation](https://www.candid.tools)
- [Report Issues](https://github.com/ron-myers/candid/issues)
- [Changelog](https://github.com/ron-myers/candid/blob/HEAD/CHANGELOG.md)

## License

MIT License - see [LICENSE](https://github.com/ron-myers/candid/tree/HEAD/LICENSE) for details.

## Author

Ron Myers

Candid

A coding-agent plugin for configurable code reviews that combine thoroughness with actionable feedback. Runs on Claude Code and Codex CLI from the same install. Based on Kim Scott's Radical Candor framework: Care Personally + Challenge Directly.

Overview

Candid provides AI-powered code reviews that catch issues before production while teaching you better patterns. Choose between Harsh (brutal honesty) or Constructive (caring + challenging) tone. Define project-specific standards in Technical.md that get enforced automatically. Every issue comes with concrete fixes rated by confidence level.

Core Workflow

The review process follows these steps:

Run /candid-review on your changes
Select your preferred tone (Harsh or Constructive)
Candid reviews with full architectural context
Get categorized issues with actionable fixes
Select which issues to track as todos
Apply fixes and optionally auto-commit

Key Features

Configurable Tone - Harsh (brutal honesty) or Constructive (caring + challenging)
Technical.md Support - Define and enforce project-specific coding standards
Focus Modes - Target reviews on security, performance, architecture, or edge cases
Actionable Fixes - Every issue includes concrete code with confidence ratings
Re-Review - Track progress on fixes across review sessions
Auto-Commit - Automatically commit applied fixes with detailed messages
Todo Integration - Convert issues to tracked todos with multi-select
Issue Categorization - Organized by severity (Critical → Architectural)
Context Optimization - Audit and optimize token usage for leaner, deeper reviews
Live Browser QA - Drive a real Chrome session against your running app, walk it like a real user across desktop and mobile, and emit structured findings JSON ready for triage

Installation

Claude Code

npx skills add https://github.com/ron-myers/candid

Then restart Claude Code.

Codex CLI

In an interactive Codex session:

/plugins

Add the marketplace via GitHub shorthand (ron-myers/candid), install candid, then verify with /skills. Invoke skills via $skill-name mention syntax (e.g., $candid-review --harsh). See docs/codex/install.md for the full command-to-skill mapping.

Quick Start

Basic review:

/candid-review

With tone preset:

/candid-review --harsh
/candid-review --constructive

Focus on specific aspects:

/candid-review --focus security
/candid-review --focus performance

Auto-commit applied fixes:

/candid-review --auto-commit

Live browser QA pass:

/candid-chrome-qa

Drives a real Chrome session against your running app, walks it like a real user across desktop and mobile, and writes structured findings to .context/findings/<date>-<slug>.json. See Candid Chrome QA docs for the full workflow and v2.0 schema reference.

Documentation

Full documentation at www.candid.tools:

Getting Started - Installation and first review
Core Features - Tone selection, focus modes, Technical.md
Team Setup - Rolling out to your team
Custom Standards - Writing effective Technical.md rules
Complete Reference - All configuration options and flags
FAQ & Troubleshooting - Common issues and solutions

Technical.md

Define project-specific standards that Candid enforces during reviews. Violations appear as 📜 Standards Violation in your review.

Quick setup:

/candid-init                    # Auto-generate from codebase (thorough analysis)
/candid-init react              # React-specific standards
/candid-init minimal            # Minimal starter
/candid-init --effort quick     # Fast analysis (~30 sec)
/candid-init --effort thorough  # Deep analysis (~3-5 min, default)
/candid-init --optimize         # Run /candid-optimize after generation

Or copy a template:

cp templates/Technical-minimal.md ./Technical.md
cp templates/Technical-react.md ./Technical.md
cp templates/Technical-node.md ./Technical.md

Keep it focused: under 500 lines, verifiable rules only. Skip what your linter handles. See the Custom Standards guide for detailed guidance.

Configuration

Persist your tone preference and other settings in config files:

User-wide default:

mkdir -p ~/.candid
echo '{"tone": "harsh"}' > ~/.candid/config.json

Project-specific:

mkdir -p .candid
echo '{"tone": "constructive", "autoCommit": true}' > .candid/config.json

See the Config Reference for all options including exclude patterns, focus areas, and mergeTargetBranches.

Philosophy

This plugin is built on the Radical Candor principle that the best feedback:

Cares Personally - Shows understanding of context and difficulty
Challenges Directly - Doesn't hedge or soften real issues

Whether you choose Harsh or Constructive tone, every review aims to find real issues before production, provide immediately applicable fixes, help you track what needs doing, and teach patterns that prevent future problems.

Updating

claude plugin update candid@candid

Then restart Claude Code. See CHANGELOG.md for version history.

License

MIT License - see LICENSE for details.

Author

Ron Myers