hexicodes 0.2.0

Minimal contract-driven coding-agent runtime + CLI

Hexi v0.2.0

Hexi is a minimal, contract-driven (hexagonal) coding-agent runtime and CLI. It runs exactly one agent step per invocation against a local git repository.

Test Drive (5 Minutes)

Run this in any local git repo you can safely modify.

1. Install Hexi:

pip install -e .

Optional OpenRouter support:

pip install -e ".[openrouter]"

2. Initialize Hexi files:

hexi init

3. Onboard provider/model and key:

hexi onboard

When prompted, pick any provider. For OpenRouter providers, install the optional extra and provide OPENROUTER_API_KEY.

4. Verify setup:

hexi doctor

Expected: provider/model printed and Doctor check passed.

5. Run one agent step:

hexi run "Add one tiny test for an existing function and run pytest"

6. Inspect what changed:

hexi diff
tail -n 20 .hexi/runlog.jsonl

7. If you want to switch providers later:

hexi onboard

Re-run onboarding to update .hexi/local.toml.

What it is

• Python package (PyPI distribution): hexicodes
• Core contracts in hexi.core
• Side-effect adapters in hexi.adapters
• One-step execution with structured event logging to .hexi/runlog.jsonl

What it is not

• No daemon, no background workers, no web UI
• No MCP server and no SQLite in v0.2.0
• No multi-agent orchestration

Install

pip install -e .

OpenRouter support (optional)

• HTTP adapter only (openrouter_http provider):

pip install -e ".[openrouter-http]"

• SDK adapter only (openrouter_sdk provider):

pip install -e ".[openrouter-sdk]"

• Both OpenRouter adapters:

pip install -e ".[openrouter]"

Dev/test dependencies:

pip install -e ".[dev]"

CLI

• hexi --help or hexi help : show command help
• hexi --version or hexi version : print installed version
• hexi init : create .hexi/config.toml, .hexi/local.toml, .hexi/runlog.jsonl
• hexi onboard : interactive setup for provider/model and optional local key paste
• hexi new : scaffold a project from built-in Hexi templates (non-interactive by default)
• hexi demo : fancy interactive flow with random/model-generated ideas and template scaffolding
• hexi run "<task>" : execute one agent step and emit structured events
• hexi apply --plan plan.json : execute a validated ActionPlan file directly (debug/replay mode)
• hexi diff : show current git diff
• hexi doctor : verbose diagnostics; use --probe-model for live “What model are you?” check
• hexi plan-check --file plan.json : validate/troubleshoot ActionPlan JSON directly

Documentation (MkDocs + Read the Docs)

Build docs locally:

pip install -e ".[docs]"
mkdocs serve

Read the Docs config is in .readthedocs.yml.

Configuration design choices

Hexi uses layered TOML configuration:

1. .hexi/config.toml (repo defaults)
2. .hexi/local.toml (local machine overrides)
3. Environment variables (recommended for secrets)

For secrets, env vars are preferred. hexi onboard can write keys to .hexi/local.toml for local/testing convenience.

Config shape (.hexi/config.toml)

[model]
provider = "openai_compat" # openrouter_http | openrouter_sdk | openai_compat | anthropic_compat
model = "gpt-4o-mini"

[providers.openrouter_http]
base_url = "https://openrouter.ai/api/v1"
api_style = "openai" # openai | anthropic

[providers.openrouter_sdk]
base_url = "https://openrouter.ai/api/v1"

[providers.openai_compat]
base_url = "https://api.openai.com/v1"

[providers.anthropic_compat]
base_url = "https://api.anthropic.com"

[policy]
allow_commands = ["git status", "git diff", "pytest", "python -m pytest"]
max_diff_chars = 4000
max_file_read_chars = 4000

Local override example (.hexi/local.toml)

[model]
provider = "openrouter_http"
model = "anthropic/claude-sonnet-4-6"

[providers.openrouter_http]
api_style = "anthropic"

[secrets]
openrouter_api_key = "..."

Env vars

• OPENROUTER_API_KEY for openrouter_http and openrouter_sdk
• OPENAI_API_KEY for openai_compat
• ANTHROPIC_API_KEY for anthropic_compat

Packaging

• Distribution name: hexicodes
• Console script: hexi
• Optional extras:
  • openrouter-http
  • openrouter-sdk
  • openrouter
  • docs
  • dev

Included example projects

• examples/todo_refiner : minimal CLI-wrapper agent integration
• examples/embedded_step : direct embedded RunStepService usage
• examples/policy_loop : multi-step user-gated loop using repeated hexi run

Example ActionPlans (hexi apply)

• examples/action_plans/read_only.json : inspect file + emit progress
• examples/action_plans/write_file.json : create a file in-repo
• examples/action_plans/run_tests.json : run pytest -q
• examples/action_plans/mixed_step.json : read + write + run + emit

Try one now:

hexi plan-check --file examples/action_plans/mixed_step.json
hexi apply --plan examples/action_plans/mixed_step.json --task "ActionPlan debug run"

Included Hexi-native templates

• templates/hexi-python-lib : tested library starter with Hexi wiring
• templates/hexi-fastapi-service : FastAPI service starter with Hexi wiring
• templates/hexi-typer-cli : Typer CLI starter with Hexi wiring
• templates/hexi-data-job : data job starter with dry-run and Hexi wiring
• templates/hexi-agent-worker : embedded Hexi runtime starter

Provenance

Made with ❤️ from 🇵🇪. El Perú es clave 🔑.