crucible

crucible

If the user asks to see the banner, NFO, or splash screen, display the following:

          ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
          ░                                                       ░
          ░     ▄████▄  ██▀███  █    ██ ▄████▄  ██▓ ▄▄▄▄  ██▓   ░
          ░    ▒██▀ ▀█ ▓██ ▒ ██▒██  ▓██▒██▀ ▀█ ▓██▒▓█████▄▓██▒   ░
          ░    ▒▓█    ▄▓██ ░▄█ ▒██  ▒██▒▓█    ▄▒██▒▒██▒ ▄█▒██░   ░
          ░    ▒▓▓▄ ▄██▒██▀▀█▄ ░██  ░██▒▓▓▄ ▄██░██░▒██░█▀ ▒██░   ░
          ░    ▒ ▓███▀ ░██▓ ▒██░ ████▓▒▒ ▓███▀ ░██░░▓█  ▀█░██████░
          ░    ░ ░▒ ▒  ░ ▒▓ ░▒▓░ ▒░▒░▒░░ ░▒ ▒  ░▓  ░▒▓███▀░ ▒░▓  ░
          ░      ░  ▒    ░▒ ░ ▒░ ░ ▒ ▒░  ░  ▒   ▒ ░▒░▒   ░░ ░ ▒  ░
          ░                                                       ░
          ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
          ╔═══════════════════════════════════════════════════════╗
          ║               RELEASE INFORMATION                     ║
          ╠═══════════════════════════════════════════════════════╣
          ║                                                       ║
          ║  Skill.......: crucible                               ║
          ║  Author......: machug          (hughtec.com)          ║
          ║  Version.....: 0.1.0                                  ║
          ║  Released....: 2026                                   ║
          ║  License.....: MIT                                    ║
          ║  Requires....: Python 3.10+, litellm                  ║
          ║                                                       ║
          ╠═══════════════════════════════════════════════════════╣
          ║                PIPELINE OVERVIEW                       ║
          ╠═══════════════════════════════════════════════════════╣
          ║                                                       ║
          ║  gather ──> review ──> challenge ──> synthesize        ║
          ║                                                       ║
          ║  Adversarial project review via multi-LLM panel.      ║
          ║  Each model plays a different adversarial persona.     ║
          ║  N models enter. 1 honest report leaves.              ║
          ║                                                       ║
          ╚═══════════════════════════════════════════════════════╝

Multi-LLM adversarial project review. Gathers project context and sends it to multiple AI models, each reviewing through a different adversarial persona and dimension. Produces a structured report with severity ratings and actionable recommendations.

Important: Claude is an active participant in this review, not just an orchestrator. You (Claude) will provide your own independent assessment alongside the external models, challenge findings you disagree with, and synthesize the final report.

Requirements

• Python 3.10+ with litellm package installed
• API key for at least one provider (set via environment variable)

IMPORTANT: Do NOT install the llm package. This skill uses litellm for API providers.

Supported Providers

Provider	API Key Env Var	Example Models
OpenAI	OPENAI_API_KEY	gpt-5.4, gpt-5.4-pro, o3-pro, o4-mini
Anthropic	ANTHROPIC_API_KEY	claude-opus-4-6, claude-sonnet-4-6
Google	GEMINI_API_KEY	gemini/gemini-3.1-pro-preview, gemini/gemini-2.5-pro
xAI	XAI_API_KEY	xai/grok-4.20-0309-reasoning, xai/grok-4-0709
Azure AI	AZURE_AI_API_KEY	foundry/<deployment-name>
Mistral	MISTRAL_API_KEY	mistral/mistral-large
Groq	GROQ_API_KEY	groq/llama-3.3-70b-versatile
OpenRouter	OPENROUTER_API_KEY	openrouter/openai/gpt-5.2-pro
Deepseek	DEEPSEEK_API_KEY	deepseek/deepseek-chat
ZAI (GLM)	ZAI_API_KEY	zai/glm-5.1, zai/glm-5-turbo
Moonshot	MOONSHOT_API_KEY	moonshot/kimi-k2.5
Codex CLI	(ChatGPT subscription)	codex/gpt-5.3-codex
Gemini CLI	(Google account)	gemini-cli/gemini-3.1-pro-preview

Discover latest models: Run cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py discover-models

Run cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py providers to see which keys are set.

Review Dimensions

Each dimension is a focused lens for project evaluation:

Dimension	Focus
progress	Plan drift, milestone tracking, velocity, scope creep
tech-stack	Framework choices, architecture, dependencies, complexity
code-quality	Type safety, test coverage, dead code, error handling
pipeline	CI/CD health, deploy frequency, build reproducibility
security	Dependency vulns, secret management, OWASP risks
devex	Onboarding friction, README accuracy, project structure

Adversarial Personas

Each model is assigned a different adversarial persona for maximum critique diversity:

Persona	Perspective
investor	Skeptical VC doing due diligence — progress vs burn, PMF evidence
staff-engineer	Seen it all — architecture sustainability, tech debt, scaling risks
devops-lead	Will be paged at 3am — deploy reliability, observability, rollback
qa-skeptic	Zero tolerance — test coverage, edge cases, regression prevention
new-hire	Day one developer — setup friction, documentation accuracy, tribal knowledge

Process

Step 1: Gather Project Context

Automatically collect project context. This is the critical input — the richer the context, the better the review. Gather ALL of the following that exist:

Git & History:

git log --oneline -30
git log --format="%h %s" --since="2 weeks ago"
git diff --stat HEAD~10..HEAD
git branch -a

Project Structure:

ls -la
find . -name "*.json" -maxdepth 2 -not -path "*/node_modules/*" | head -20

Dependencies (read the relevant files):

• package.json / requirements.txt / Cargo.toml / go.mod
• Lock files (check existence, freshness)

CI/CD Configuration (read if they exist):

• .github/workflows/ / Jenkinsfile / .gitlab-ci.yml / vercel.json

Issue Tracker:

• If beads is available: bd list --status=open and bd stats
• If there's a TODO.md or similar: read it

Code Metrics (sample):

• Count files by type: find . -name "*.ts" -not -path "*/node_modules/*" | wc -l
• Check for test files: find . -name "*.test.*" -o -name "*.spec.*" | head -10
• Check for type config: tsconfig.json, mypy.ini, .eslintrc

Build & Deploy:

• Try npm run / yarn / make to see available scripts
• Check for Dockerfile, docker-compose.yml
• Check for deploy scripts or configs

Compile ALL gathered context into a single JSON structure:

{
  "project_name": "...",
  "gathered_at": "ISO timestamp",
  "git": {
    "recent_commits": "...",
    "branches": "...",
    "recent_diff_stats": "..."
  },
  "structure": {
    "top_level_files": "...",
    "file_counts_by_type": {}
  },
  "dependencies": {
    "package_json": "... or null",
    "lock_file_exists": true
  },
  "ci_cd": {
    "github_actions": "... or null",
    "deploy_config": "... or null"
  },
  "issues": {
    "open_count": 0,
    "in_progress": [],
    "recent_closed": []
  },
  "code_quality": {
    "test_files_count": 0,
    "type_config_exists": true,
    "lint_config_exists": true
  }
}

Write this context to a temporary file (e.g., /tmp/crucible-context-{project}.json).

Step 2: Select Models

First, discover available models:

cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py discover-models

If that fails, fall back to:

cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py providers

Present available models to the user using AskUserQuestion with multiSelect. Build options from discover-models output. Recommend at least 3 models for good adversarial coverage.

Also ask which dimensions to review (default: all 6), or let the user focus on specific areas.

Step 3: Run Review Round

Send the context to all selected models in parallel, each assigned a different persona and dimension:

cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py review --models MODEL1,MODEL2,MODEL3 --context /tmp/crucible-context-{project}.json

The script outputs JSON with all reviews. Parse it and present each review to the user.

Claude's Own Review: After receiving model reviews, provide YOUR OWN independent assessment. You have advantages the models don't: you can read the actual code, run commands, and verify claims. Use this to:

• Confirm or challenge model findings with evidence
• Add findings the models missed
• Flag any model hallucinations (e.g., claiming a file exists that doesn't)

Step 4: Cross-Examination (Optional)

Ask the user: "Want to run a cross-examination round where models challenge each other's findings?"

If yes, save the reviews to a file and run:

cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py challenge --models MODEL1,MODEL2 --reviews /tmp/crucible-reviews-{project}.json

Step 5: Synthesize Report

Consolidate all reviews (model + Claude's own) into a structured final report:

# Crucible Report: {Project Name}
**Date:** {date}
**Models:** {list}
**Dimensions reviewed:** {list}

## Executive Summary
{2-3 sentence overall assessment}

## Scorecard

| Dimension | Score | Severity | Key Finding |
|-----------|-------|----------|-------------|
| Progress  | X/10  | ...      | ...         |
| Tech Stack| X/10  | ...      | ...         |
| ...       | ...   | ...      | ...         |

**Overall: X/10**

## Critical Findings
{Findings rated CRITICAL by any reviewer, with cross-examination results}

## Warnings
{Findings rated WARNING}

## Healthy Areas
{What's working well — important for morale}

## Recommended Actions
{Prioritized list of concrete next steps}

## Reviewer Disagreements
{Areas where models disagreed — these often reveal the most interesting insights}

## Cost Summary
{Token usage and cost breakdown from the review}

Write the report to ./crucible-report-{date}.md in the project directory.

Step 6: Export Actions (Optional)

Ask the user if they want to export critical findings and recommendations as issues:

• If beads is available: create beads issues with bd create
• Otherwise: list them as actionable items

CLI Reference

# Run review
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py review --models gpt-5.4,xai/grok-4-0709 --context context.json

# Specific dimensions only
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py review --models gpt-5.4 --context ctx.json --dimensions progress,code-quality

# Cross-examination round
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py challenge --models gpt-5.4,xai/grok-4-0709 --reviews reviews.json

# List providers
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py providers

# Discover models
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py discover-models

# List dimensions
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py dimensions

# List personas
cd ${CLAUDE_PLUGIN_ROOT}/skills/crucible/scripts && python3 crucible.py personas

Key Principles

1. Evidence over opinion — every finding must cite specific files, commits, or configs
2. Actionable over academic — recommendations should be concrete and prioritized
3. Honest over kind — the point is to find problems before they find you
4. Diverse perspectives — different models have different biases; that's a feature
5. Claude verifies — you have the ability to actually read code and run commands; use it to fact-check model claims