helix-evolve 0.1.3

CLI for Helix evolutionary prompt optimization

Helix CLI

Standalone command-line tool for evolutionary prompt optimization. Evolve your LLM prompts against test cases without running a web server.

Overview

The Helix CLI runs the same evolution engine as the web UI, but operates on local YAML files. Define your prompt, test cases, and configuration as files in a directory, then run evolution from the terminal.

Key features:

• Standalone -- no web server or database needed
• YAML-based -- human-readable prompt and test case definitions
• Agent-friendly -- every command supports --json for machine-readable output
• Same engine -- uses the identical evolution pipeline as the web UI

Installation

Prerequisites

• Python 3.13+
• uv (recommended) or pip
• An API key for at least one LLM provider (Gemini, OpenAI, or OpenRouter)

Install from source

git clone https://github.com/Onebu/helix.git
cd helix

# Install the core engine and CLI
uv pip install -e .          # core engine (api package)
uv pip install -e cli/       # CLI tool

# Verify installation
helix --help

If helix isn't on your PATH, use the full path to the venv binary:

.venv/bin/helix --help

Configure API key

The easiest way:

helix setup

This walks you through selecting a provider (Gemini, OpenAI, or OpenRouter) and saves your API key to .env.

Or manually create a .env file:

# Choose at least one provider
GENE_GEMINI_API_KEY=your-gemini-key
# GENE_OPENAI_API_KEY=your-openai-key
# GENE_OPENROUTER_API_KEY=your-openrouter-key

Quick Start

# 1. Create a new prompt project
helix init customer-support

# 2. Edit the YAML files (see File Format below)
#    - customer-support/prompt.yaml
#    - customer-support/dataset.yaml
#    - customer-support/config.yaml

# 3. Run evolution
helix evolve customer-support

# 4. Review the results
helix results customer-support

# 5. Accept the evolved template
helix accept customer-support

Commands

helix setup

Interactive first-run configuration. Selects a provider, saves your API key to .env.

helix setup

Providers: Gemini, OpenAI, OpenRouter.

helix models [provider]

List available models for a provider.

helix models                   # list providers
helix models gemini            # list Gemini models
helix models openrouter --json # JSON output

helix init <prompt-id>

Scaffold a new prompt directory with template YAML files.

helix init my-prompt
helix init my-prompt --dir /path/to/workspace
helix init my-prompt --json    # machine-readable output

Creates:

my-prompt/
  prompt.yaml      # prompt template and variable definitions
  dataset.yaml     # test cases for fitness evaluation
  config.yaml      # model, provider, and evolution settings
  results/         # evolution results (populated by helix evolve)

helix list

List all prompt directories in the current workspace.

helix list
helix list --dir /path/to/workspace
helix list --json

Output:

                     Prompts
┏━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━┓
┃ ID               ┃ Purpose               ┃ Cases ┃ Runs ┃
┡━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━┩
│ customer-support │ Handle support tickets │     8 │    3 │
│ summarizer       │ Summarize documents    │     5 │    1 │
└──────────────────┴───────────────────────┴───────┴──────┘

helix show <prompt-id>

Display detailed information about a prompt: template preview, variables, dataset summary, and latest evolution result.

helix show customer-support
helix show customer-support --json

helix evolve <prompt-id>

Run the evolution engine against your prompt and test cases. Shows a live Rich progress display with generation-by-generation fitness updates.

# Basic usage
helix evolve customer-support

# Override evolution parameters
helix evolve customer-support --generations 5 --islands 2 --budget 1.00

# JSON output (no live display, suitable for scripts/agents)
helix evolve customer-support --json

Options:

Flag	Description
--generations, -g	Override number of generations
--islands, -i	Override number of islands
--budget, -b	Override budget cap in USD
--json	JSON output, no live display

The live progress display shows:

╭──────────────── helix evolve customer-support ─────────────────╮
│  Generation  7/10  [████████████████████░░░░]  70%             │
│                                                                │
│  Best Fitness  -0.50  (seed: -6.00, +5.50)                     │
│  Avg Fitness   -2.10                                           │
│  Candidates    140                                             │
│  Cost          $0.23 / $2.00                                   │
│  Elapsed       1m 42s                                          │
╰────────────────────────────────────────────────────────────────╯

Results are automatically saved to results/run-NNN.yaml.

helix results <prompt-id>

Display evolution results. Shows fitness, cost, model info, and the best evolved template.

# Latest result
helix results customer-support

# Specific run
helix results customer-support --run run-001

# JSON output
helix results customer-support --json

# Print only the evolved template (useful for piping)
helix results customer-support --template

helix accept <prompt-id>

Accept an evolved template by copying it back into prompt.yaml, replacing the current template.

helix accept customer-support
helix accept customer-support --run run-001    # accept a specific run
helix accept customer-support --json

Global Options

Every command supports:

Flag	Description
--dir, -d	Workspace directory (default: current directory)
--json	Machine-readable JSON output

File Format Reference

prompt.yaml

Defines the prompt template, variables, and optional tool definitions.

# Required fields
id: customer-support                    # lowercase slug (a-z, 0-9, hyphens)
purpose: "Handle customer support tickets"

# The Jinja2 template. Variables use {{ variable_name }} syntax.
template: |
  You are a customer support agent for {{ company_name }}.

  Customer: {{ customer_name }}
  Issue: {{ issue_description }}

  Resolve the issue professionally and empathetically.
  Always offer a concrete next step.

# Variable definitions (optional).
# If omitted, variables are auto-extracted from the template.
# Define explicitly to set types, descriptions, and anchor status.
variables:
  - name: company_name
    description: "Company brand name"
    var_type: string                    # string | number | boolean | object | array
    is_anchor: true                    # anchored = preserved during evolution

  - name: customer_name
    description: "Customer's display name"
    var_type: string
    is_anchor: true

  - name: issue_description
    description: "Description of the customer's issue"
    var_type: string
    is_anchor: false

# Tool definitions (optional, OpenAI function-calling format)
tools:
  - type: function
    function:
      name: lookup_order
      description: "Look up a customer's order by ID"
      parameters:
        type: object
        properties:
          order_id:
            type: string
        required: [order_id]

  - type: function
    function:
      name: create_ticket
      description: "Create a support ticket"
      parameters:
        type: object
        properties:
          subject:
            type: string
          priority:
            type: string
            enum: [low, normal, high, urgent]
        required: [subject, priority]

Anchor variables are preserved during evolution. The evolution engine will never remove {{ company_name }} from the template if it's marked as an anchor. Use anchors for variables that must always be present (brand names, user identifiers, etc.).

dataset.yaml

Defines test cases that the evolution engine scores prompts against. Each case specifies inputs, conversation context, and expected behavior.

cases:
  # --- Simple behavior test ---
  - name: "Greeting"
    tier: normal                        # critical | normal | low
    tags: [greeting, basic]
    variables:
      company_name: "Acme Corp"
      customer_name: "Alice"
      issue_description: ""
    chat_history:
      - role: user
        content: "Hi, I need help with my order"
    expected_output:
      require_content: true             # must produce text (not just tool calls)
      behavior:                         # LLM-judged criteria
        - "Greets the customer by name"
        - "Asks for order details or offers to look it up"

  # --- Tool call test ---
  - name: "Order lookup"
    tier: critical                      # critical cases have 5x weight
    tags: [tool-call, order]
    variables:
      company_name: "Acme Corp"
      customer_name: "Bob"
      issue_description: "Where is my order #12345?"
    chat_history:
      - role: user
        content: "Where is my order #12345?"
    expected_output:
      match_args:                       # expect a specific tool call
        tool_name: lookup_order
        tool_args:
          order_id: "12345"

  # --- Multi-turn conversation ---
  - name: "Escalation flow"
    tier: normal
    tags: [escalation, multi-turn]
    variables:
      company_name: "Acme Corp"
      customer_name: "Carol"
      issue_description: "Product is defective"
    chat_history:
      - role: user
        content: "My product broke after one day!"
      - role: assistant
        content: "I'm sorry to hear that. Can you describe the issue?"
      - role: user
        content: "The screen cracked on its own. I want a full refund NOW."
    expected_output:
      require_content: true
      behavior:
        - "Shows empathy for the customer's frustration"
        - "Offers a concrete resolution (refund, replacement, or escalation)"
        - "Does not make promises outside company policy"

  # --- Edge case ---
  - name: "Off-topic request"
    tier: low                           # low priority = 0.25x weight
    tags: [edge-case]
    variables:
      company_name: "Acme Corp"
      customer_name: "Dave"
      issue_description: ""
    chat_history:
      - role: user
        content: "What's the meaning of life?"
    expected_output:
      require_content: true
      behavior:
        - "Politely redirects to support topics"
        - "Does not engage with the off-topic question"

Test case tiers control how much each case affects the fitness score:

Tier	Weight	Use for
critical	5x	Must-pass cases, core functionality
normal	1x	Standard behavior expectations
low	0.25x	Nice-to-have, edge cases

Expected output fields:

Field	Type	Description
require_content	bool	Response must include text content
behavior	list[str]	Natural language criteria judged by LLM
match_args	dict	Expected tool call (tool_name + tool_args)
tool_calls	list[dict]	Multiple expected tool calls
must_contain	str	Response must contain this substring

A fitness score of 0.0 means all cases pass. Negative scores indicate failures (more negative = worse).

config.yaml

Configures models, providers, and evolution hyperparameters. All fields are optional and override environment variables.

# Simplified: one provider and model for all roles
provider: gemini                        # gemini | openai | openrouter
model: gemini-2.5-flash

# Override specific roles (optional):
models:
  meta:
    model: gemini-2.5-pro               # use a stronger model for critique
    thinking_budget: -1                 # Gemini-specific (-1 = dynamic)

# Evolution hyperparameters
evolution:
  generations: 10                       # number of evolution generations
  islands: 4                            # parallel island populations
  conversations_per_island: 5           # RCC conversations per island per gen
  budget_cap_usd: 2.00                  # stop when cost exceeds this
  # n_seq: 3                            # critic-author turns per conversation
  # temperature: 1.0                    # Boltzmann selection temperature
  # structural_mutation_probability: 0.2
  # population_cap: 10                  # max candidates per island
  # n_emigrate: 5                       # migration count between islands
  # reset_interval: 3                   # gens between island resets
  # adaptive_sampling: false            # enable adaptive case sampling

# Inference parameters (passed to LLM API calls)
generation:
  temperature: 0.7
  max_tokens: 4096
  # top_p: null
  # top_k: null
  # frequency_penalty: null
  # presence_penalty: null

results/run-NNN.yaml

Written automatically by helix evolve. Contains the full evolution result.

run_id: run-001
timestamp: "2026-03-27T22:35:28+00:00"
termination_reason: perfect_fitness     # or generations_complete, budget_exhausted

best_template: |
  You are a customer support agent for {{ company_name }}...
  (the full evolved template)

fitness:
  score: 0.0                           # 0.0 = all cases pass
  normalized_score: 0.0

seed_fitness: -6.0                      # original template's score

cost:
  total_calls: 342
  total_input_tokens: 1250000
  total_output_tokens: 85000
  total_cost_usd: 0.47

effective_models:
  meta_model: gemini-2.5-pro
  meta_provider: gemini
  target_model: gemini-2.5-flash
  target_provider: gemini
  judge_model: gemini-2.5-flash
  judge_provider: gemini

generation_records:
  - generation: 1
    best_fitness: -2.0
    avg_fitness: -4.5
    candidates_evaluated: 20
  - generation: 2
    best_fitness: -0.5
    avg_fitness: -1.8
    candidates_evaluated: 20

config_snapshot:
  evolution:
    generations: 10
    islands: 4
    conversations_per_island: 5

Configuration Cascade

Settings are resolved in priority order (highest wins):

1. CLI flags (--generations 5, --budget 1.00)
2. config.yaml values
3. Environment variables (GENE_GEMINI_API_KEY, GENE_META_MODEL, etc.)
4. Defaults (built into the engine)

This means you can set defaults in config.yaml, override per-run with CLI flags, and keep secrets in .env or environment variables.

Agent Integration

Every command supports --json for structured output, making the CLI usable by AI coding agents like Claude Code.

Example CLAUDE.md configuration for a project using Helix:

## Prompt Optimization

Run `helix evolve <prompt-id> --json` to optimize prompts.
Run `helix results <prompt-id> --json` to check evolution results.
Run `helix accept <prompt-id>` to apply the best evolved template.
Run `helix show <prompt-id> --json` to inspect prompt configuration.

Example agent workflow:

# Agent creates a prompt
helix init my-agent-prompt --json

# Agent checks current state
helix show my-agent-prompt --json

# Agent runs evolution and parses result
helix evolve my-agent-prompt --generations 5 --json

# Agent reads the evolved template
helix results my-agent-prompt --template

# Agent accepts if fitness improved
helix accept my-agent-prompt --json

Typical Workflow

  helix init <id>           Create prompt directory
        |
        v
  Edit YAML files           Define template, test cases, config
        |
        v
  helix evolve <id>         Run evolution (live progress)
        |
        v
  helix results <id>        Review fitness, cost, evolved template
        |
       / \
      /   \
     v     v
  Accept   Iterate          Apply template or adjust test cases

1. Initialize: helix init creates a starter directory
2. Define: Write your prompt template in prompt.yaml, add test cases to dataset.yaml
3. Configure: Set models and budget in config.yaml (or use defaults)
4. Evolve: helix evolve runs the engine and saves results
5. Review: helix results shows what improved and by how much
6. Accept or iterate: helix accept applies the evolved template, or adjust test cases and re-run

Troubleshooting

"Missing API key for gemini" Set your API key in a .env file or environment:

export GENE_GEMINI_API_KEY=your-key-here

"No test cases found" Add at least one test case to dataset.yaml. See the file format reference above.

"Prompt not found" Make sure you're in the right directory. Use helix list to see available prompts, or --dir to specify the workspace.

Evolution finishes instantly with perfect fitness Your test cases may be too easy. Add more specific behavior criteria or critical-tier cases to challenge the prompt.