eiffel-review

name	eiffel-review
description	Phase 2 of Eiffel Spec Kit. Runs progressive AI review chain (Ollama → Claude → Grok → Gemini → Human). Generates approach.md and synopsis.md. BLOCKS until human approves. Use with /eiffel.review command.
allowed-tools	Read, Grep, Glob, Edit, Write, Bash, Task

/eiffel.review - Phase 2: Adversarial Review + Approach Sketch

Purpose: Multiple AI perspectives review contracts before implementation. Human manually submits prompts to each AI and brings results back.

Usage

/eiffel.review <project-path>

Example:

/eiffel.review d:\prod\simple_cache

Project Scoping

All files are created inside the PROJECT directory:

<project-path>/
├── .eiffel-workflow/
│   ├── approach.md                    (implementation sketch)
│   ├── synopsis.md                    (aggregated review findings)
│   ├── prompts/
│   │   ├── phase2-ollama-review.md    (READY TO SUBMIT - contracts pre-populated)
│   │   ├── phase2-claude-review.md    (needs Ollama response pasted)
│   │   ├── phase2-grok-review.md      (needs Ollama+Claude responses pasted)
│   │   └── phase2-gemini-review.md    (needs all previous responses pasted)
│   └── evidence/
│       ├── phase2-ollama-response.md  (placeholder for Ollama's review)
│       ├── phase2-claude-response.md  (placeholder for Claude's review)
│       ├── phase2-grok-response.md    (placeholder for Grok's review)
│       ├── phase2-gemini-response.md  (placeholder for Gemini's review)
│       └── phase2-chain.txt           (summary evidence)
├── src/
└── test/

Prerequisites

• Phase 1 complete: contracts compile

Verify:

test -f <project-path>/.eiffel-workflow/evidence/phase1-compile.txt && echo "Phase 1 complete" || echo "ERROR: Run /eiffel.contracts first"

Workflow

Step 1: Generate Implementation Sketch

Create <project-path>/.eiffel-workflow/approach.md with actual implementation details.

Step 2: Generate Progressive AI Review Prompts

CRITICAL: Pre-populate prompts with actual file contents where possible.

Claude has access to the project files. The prompts should contain:

• Contracts: Actual content of all src/*.e files (Claude reads and embeds)
• Approach: Actual content of approach.md (Claude reads and embeds)
• Previous AI responses: Placeholder - user must paste after receiving

3a. Ollama Prompt (FULLY READY TO SUBMIT)

Create <project-path>/.eiffel-workflow/prompts/phase2-ollama-review.md with:

• Full contracts embedded (read all src/*.e files and include content)
• Full approach.md embedded
• User can copy this file directly to Ollama with zero modifications

Template:

# Eiffel Contract Review Request (Ollama)

You are reviewing Eiffel contracts for a library. Find obvious problems.

## Review Checklist
- [ ] Preconditions that are just `True` (too weak)
- [ ] Postconditions that don't constrain anything
- [ ] Missing invariants
- [ ] Obvious edge cases not handled
- [ ] Missing MML model queries for collection attributes
- [ ] Missing frame conditions (what did NOT change)

## Contracts to Review

[ACTUALLY EMBED THE FULL CONTENT OF EACH src/*.e FILE HERE]

## Implementation Approach

[ACTUALLY EMBED THE FULL CONTENT OF approach.md HERE]

## Output Format

List issues found as:
- **ISSUE**: [description]
- **LOCATION**: [class.feature]
- **SUGGESTION**: [how to fix]

3b. Claude Prompt (Contracts pre-populated, needs Ollama response)

Create <project-path>/.eiffel-workflow/prompts/phase2-claude-review.md with:

• Full contracts embedded
• Full approach.md embedded
• Placeholder for Ollama's response (user pastes after Step 1)

Claude-specific MML review checklist:

• MML model queries correctly represent collection semantics
• Frame conditions use |=| for model equality
• Postconditions use old expressions correctly
• Model queries are pure (no side effects)

3c. Grok Prompt (Contracts pre-populated, needs previous responses)

Create <project-path>/.eiffel-workflow/prompts/phase2-grok-review.md with:

• Full contracts embedded
• Full approach.md embedded
• Placeholders for Ollama + Claude responses

3d. Gemini Prompt (Contracts pre-populated, needs all previous responses)

Create <project-path>/.eiffel-workflow/prompts/phase2-gemini-review.md with:

• Full contracts embedded
• Full approach.md embedded
• Placeholders for all three previous responses

Step 3: Auto-Create Evidence Placeholder Files

Create placeholder files in <project-path>/.eiffel-workflow/evidence/ with STATUS: INCOMPLETE markers.

Step 4: Notify Human - Start Manual Cycle

Display:

PHASE 2: Review prompts generated.

READY TO SUBMIT (contracts pre-populated):
  - phase2-ollama-review.md → Copy entire file to Ollama

NEED PREVIOUS RESPONSES PASTED:
  - phase2-claude-review.md → Paste Ollama's response first
  - phase2-grok-review.md → Paste Ollama + Claude responses first
  - phase2-gemini-review.md → Paste all previous responses first

EVIDENCE PLACEHOLDERS (save AI responses here):
  - evidence/phase2-ollama-response.md
  - evidence/phase2-claude-response.md
  - evidence/phase2-grok-response.md
  - evidence/phase2-gemini-response.md

MANUAL REVIEW CHAIN:

Step 1: OLLAMA
  a. Copy entire contents of: prompts/phase2-ollama-review.md
  b. Paste to Ollama
  c. Save response to: evidence/phase2-ollama-response.md

Step 2: CLAUDE
  a. Open: prompts/phase2-claude-review.md
  b. Paste Ollama's response where indicated
  c. Copy entire file to a different Claude session
  d. Save response to: evidence/phase2-claude-response.md

Step 3: GROK
  a. Open: prompts/phase2-grok-review.md
  b. Paste Ollama + Claude responses where indicated
  c. Copy entire file to Grok
  d. Save response to: evidence/phase2-grok-response.md

Step 4: GEMINI
  a. Open: prompts/phase2-gemini-review.md
  b. Paste all previous responses where indicated
  c. Copy entire file to Gemini
  d. Save response to: evidence/phase2-gemini-response.md

When complete, say "reviews complete" and I will generate the synopsis.

NOTE: You may skip AIs or change the order. At minimum, use ONE external AI.

BLOCK until user confirms reviews are complete.

Step 5: Generate Synopsis

When user returns, read all response files and create synopsis.md.

Step 6: Human Approval

DO NOT PROCEED until user explicitly approves.

Step 7: Save Evidence

Save to <project-path>/.eiffel-workflow/evidence/phase2-chain.txt.

Completion

When user approves:

Phase 2 COMPLETE: Adversarial review passed.
Project: <project-path>

Next: Run /eiffel.tasks <project-path> to break contracts into implementation tasks.

Context Management (RLM Pattern)

This skill focuses ONLY on: <project-path>

DO NOT:

• Read files outside this project directory
• Load entire ecosystem or multiple libraries into context

DO:

• Use Task tool with Explore agent for ecosystem questions
• Pre-populate prompts with actual file contents (contracts, approach)
• Only use placeholders for content that doesn't exist yet (AI responses)

Anti-Drift

• Ollama prompt is FULLY ready to submit (no user pasting of contracts)
• Subsequent prompts only need previous AI responses pasted
• Human effort is minimized to: copy prompt → submit to AI → save response