GPT-5 Reasoner

Two-phase prompt optimization tool for GPT-5: optimize metadata with Claude, then execute with GPT-5.

Features

• Two-phase architecture: Optimizer (configurable model) analyzes file metadata, executor (GPT-5) processes full content
• Directory support: Automatically discover and include files from directories with filtering
• Dual interfaces: CLI and MCP (Model Context Protocol) support via universal-tool framework
• Smart file handling: Binary detection, UTF-8 validation, path normalization
• Configurable traversal: Control recursion, hidden files, extensions, and file limits
• Type safety: Strongly-typed Rust with comprehensive test coverage (40 tests)

Installation

cargo install gpt5_reasoner

Or build from source:

cd gpt5_reasoner
just build

Usage

CLI

Basic usage with files:

gpt5_reasoner run \
  --prompt "Analyze these files" \
  --prompt-type reasoning \
  --files-json examples/test_files.json

Using directories:

gpt5_reasoner run \
  --prompt "Analyze the codebase" \
  --prompt-type plan \
  --files-json examples/empty_files.json \
  --directories-json examples/directories.json

MCP Server

Run as MCP server:

gpt5_reasoner mcp

Environment Variables

• OPENROUTER_API_KEY (required): API key for OpenRouter
• OPTIMIZER_MODEL (optional): Override default optimizer model
  • Default: anthropic/claude-sonnet-4.5
  • Precedence: parameter > env var > default
• RUST_LOG (optional): Control logging level (e.g., gpt5_reasoner=debug)

Directory Support

DirectoryMeta Structure

{
  "directory_path": "src",
  "description": "Source code files",
  "extensions": ["rs", "toml"],
  "recursive": true,
  "include_hidden": false,
  "max_files": 1000
}

Fields

• directory_path (required): Path to directory (relative or absolute)
• description (required): Description inherited by all files from this directory
• extensions (optional): File extensions to include (case-insensitive, accepts "rs" or ".rs")
  • null or []: Include all files
• recursive (optional, default: false): Traverse subdirectories
• include_hidden (optional, default: false): Include hidden files and directories
  • When false, prunes entire hidden directories from traversal
• max_files (optional, default: 1000): Maximum files per directory (safety cap)

Behavior Notes

• Binary files: Automatically skipped with debug logs
• Extension matching: Case-insensitive, handles both "rs" and ".rs" formats
• Hidden directories: When include_hidden=false, hidden directories are pruned entirely (not just files)
• Path normalization: All paths converted to absolute without resolving symlinks
• Symlinks: Not followed (follow_links=false)
• Deduplication: Files appearing in both files and directories are automatically deduplicated
• Token limit: Final guard enforces 250k token limit on complete prompt
• Multi-dot extensions: foo.rs.bk matches "bk", not "rs"
• Platform: Unix/Linux dotfile-based (Windows hidden attribute not checked)

Examples

Recursive Rust files only

[
  {
    "directory_path": "gpt5_reasoner/src",
    "description": "GPT-5 reasoner implementation",
    "extensions": ["rs"],
    "recursive": true,
    "include_hidden": false,
    "max_files": 1000
  }
]

All files including hidden

[
  {
    "directory_path": "config",
    "description": "Configuration files",
    "recursive": false,
    "include_hidden": true
  }
]

Multiple directories

[
  {
    "directory_path": "src",
    "description": "Source code",
    "extensions": ["rs"],
    "recursive": true,
    "include_hidden": false
  },
  {
    "directory_path": "tests",
    "description": "Test files",
    "extensions": ["rs"],
    "recursive": true,
    "include_hidden": false
  }
]

Model Configuration

Optimizer Model

The optimizer analyzes file metadata to determine which files to include and how to structure the final prompt.

Default: anthropic/claude-sonnet-4.5

Override via:

1. Function parameter (MCP/library usage)
2. OPTIMIZER_MODEL environment variable
3. Falls back to default

Executor Model

The executor processes the full file content with the optimized prompt.

Fixed: openai/gpt-5 (high reasoning effort)

This model is not configurable and always uses reasoning_effort: high.

Reasoning Effort

• Models containing "gpt-5" or "gpt-oss" get reasoning_effort set
• Anthropic models do not support this parameter (correctly omitted)

Architecture

User Prompt + File Metadata
         ↓
    Optimizer (Claude)
         ↓
   Optimized Prompt Template
         ↓
  File Content Injection
         ↓
   Executor (GPT-5 high reasoning)
         ↓
      Final Output

Two-Phase Benefits

1. Cost efficiency: Optimizer sees only metadata (filenames, descriptions), not full content
2. Token optimization: Executor prompt is pre-filtered to relevant files only
3. Flexibility: Easy to swap optimizer models without changing executor

Development

# Run checks (fmt + clippy)
just check

# Run tests
just test

# Build
just build

# Output mode variants
OUTPUT_MODE=normal just test    # Full cargo output
OUTPUT_MODE=verbose just test   # Verbose output

Breaking Changes from 0.1.0

• Default optimizer model changed from openai/gpt-5 to anthropic/claude-sonnet-4.5
• gpt5_reasoner_impl signature now includes directories parameter
• optimize_and_execute MCP function now accepts optional directories parameter

License

MIT - See LICENSE