seldon

Seldon

Independent plan/spec reviewer for AI coding agents — Multi-judge LLM-as-a-Judge with structured verdicts

Analyzes implementation plans the way Hari Seldon analyzed civilizations — by checking structural assumptions against reality before things go wrong.

Install • Usage • Focus Modes • External Judges • Output Schema

Feed Seldon a plan, spec, or proposal. It reads the document, inspects your workspace for evidence, and returns a verdict — approve, approve_with_changes, or request_major_revision — with a confidence score (0–1) and concrete findings tagged by severity and file references.

Works out of the box as an inline skill for Claude Code. Plug in Codex, OpenAI, or Anthropic as an external judge for true model independence — or run all three and compare verdicts side by side.

Skill

This plugin ships one skill: Seldon, with three pluggable judges.

Seldon (The Reviewer)

Independent plan/spec reviewer with workspace verification.

• Multi-judge LLM-as-a-Judge: Plug in Anthropic (Claude), OpenAI (GPT), or Codex (via the Codex plugin) — or fall back to inline review using the current agent.
• Structured Verdicts: Every response conforms to a strict JSON Schema (approve, approve_with_changes, request_major_revision + confidence + findings).
• Workspace Verification: The inline reviewer and codex judge can traverse your codebase to verify claims; the API runners get every file you pass as arguments.
• Focus Modes: Six pre-baked weighting profiles — balanced, architecture, evaluation, product, operations, safety.
• Severity-Tagged Findings: Each finding lists severity (critical/high/medium/low), why it matters, evidence from the workspace, and path:line references.
• Confidence Score: A numeric 0–1 score from the judge, rendered as a 20-segment visual bar with semantic color labels.
• Schema-Validated Output: A bundled scripts/validate.sh smoke-test harness checks runner output against seldon.schema.json before you trust it.

---

Quick Start

Prerequisites

• Claude Code (CLI) or Claude Desktop — no API keys required for the inline reviewer
• Optional, per external judge:
  • codex → install the Codex plugin (/codex:setup)
  • anthropic → export ANTHROPIC_API_KEY
  • openai → export OPENAI_API_KEY
• For schema validation tooling (optional): python3 with jsonschema (see Testing)

Installation

Option 1: Claude Code Plugin (Recommended)

Install via Claude Code's built-in plugin system:

# Add the marketplace
/plugin marketplace add proyecto26/seldon

# Install the plugin
/plugin install seldon

After installing, the /seldon skill triggers automatically on phrases like "review my plan", "judge this spec", "second opinion on this RFC".

Option 2: CLI Install via skills.sh

# Install all skills
npx skills add proyecto26/seldon

# List available skills
npx skills add proyecto26/seldon --list

This installs to your .claude/skills/ directory.

Option 3: Clone and Copy

git clone https://github.com/proyecto26/seldon.git
cp -r seldon/skills/* .claude/skills/

Option 4: Git Submodule

git submodule add https://github.com/proyecto26/seldon.git .claude/seldon

Reference the skill from .claude/seldon/skills/seldon/.

Option 5: Fork and Customize

1. Fork this repository
2. Customize skills/seldon/SKILL.md for your house style (rubric weights, finding format)
3. Add or modify judge runners in skills/seldon/scripts/
4. Clone your fork into your projects

---

Usage

Trigger phrases that fire the skill:

/seldon my-plan.md
review this plan: docs/migration-plan.md
judge this spec: docs/auth-redesign.md
second opinion on RFC-042

What happens

1. Seldon reads your plan file (and any supporting files you pass)
2. Resolves the judge:
   • auto (default) — probes for codex → ANTHROPIC_API_KEY → OPENAI_API_KEY → falls back to inline
   • explicit — say "judge with codex" / "use the openai judge" / "use anthropic"
3. Inspects the workspace to verify claims — file paths, APIs, dependencies, config, schema
4. Evaluates against a rubric: repo fit, correctness, sequencing, evaluation, safety
5. Returns a structured verdict with a visual confidence bar:

🟡 Confidence  ████████████████░░░░  0.82  (moderate)

Usage Examples