howhard

How hard should the agent think?

Read the documentation.

howhard chooses the model and reasoning effort for a task delegated to an AI agent. An LLM reads the actual task, current project and writable semantic memory, learning a task-specific capability-fit band over the full supplied inventory, then forecasting clean/caught/escaped outcomes, tokens and time only for that band. One recorded outcome updates both the threshold reader and predictive support for memories.

Selection minimizes expected cost to an accepted result, including same-cell retries after loud failures. Its risk cap applies to defects among accepted results; these estimates are not a calibrated failure guarantee. Without a usable LLM forecast, routing stops explicitly.

Quickstart

You need Python 3.11+, a POSIX shell, Git and a current Codex, Claude Code or Antigravity client with native plugin support.

Install the 3.0 backend from the tagged GitHub release in your project directory:

python3 -m venv .venv-howhard
. .venv-howhard/bin/activate
python -m pip install "howhard @ git+https://github.com/xu-hao/howhard.git@v3.0.0"
howhard --version

Expect howhard 3.0.0. Keep the environment active and start your agent from the same shell so it can find howhard. An existing GUI session may need the absolute path to .venv-howhard/bin/howhard in your project.

Install the native plugin for your host:

# Codex
codex plugin marketplace add xu-hao/howhard
codex plugin add howhard@howhard

# Claude Code
claude plugin marketplace add xu-hao/howhard
claude plugin install howhard@howhard

# Antigravity
agy plugin install https://github.com/xu-hao/howhard
agy plugin list

The plugin supplies the howhard routing and learning skill and the attribute-failure investigation skill. Backend and plugin updates are separate; the native installation guide covers scopes and lifecycle commands.

Ask your coordinating agent to use howhard on a real, authorized task before starting a worker:

Use howhard to choose the model and reasoning effort for this task. Obtain an independent baseline, consult memory, execute with supported host controls, then record the actual outcome and measured usage and update memory.

Invoke the skill explicitly with $howhard or /skills in Codex, /howhard:howhard in Claude Code, or /howhard in Antigravity. To require it before delegation, add the coordinating-agent instruction to your project instructions. Plugin installation alone does not enforce worker launches.

The complete quickstart walks through decision fields, feedback, usage and memory learning. The host must support the chosen controls and a fresh context for the independent baseline; otherwise routing is unavailable. Unknown usage and prices stay unknown. Record actual executions rather than invented outcomes.

Inspect local state and observed-action summaries with:

howhard paths --json
howhard memory list --json
howhard report --json

The model and storage contracts, semantic-memory guide and context contract describe the CLI/API and evidence boundaries. Evaluate routing quality with genuine project work. External routing applications can configure explicitly selected inventories and use the Python or calling-agent interfaces to select models and share learning across providers in the same project store. Contributors use isolated worktrees and the documentation-only exemption.