Skip to content

Autonomous research system for measurable, computer-executable research.

Python
4.0 KB
Raw Blame

Praxist Agent Skills

Praxist bundles portable skills for Codex and Claude Code onboarding, task initialization, lifecycle control, diagnostics, scientific context gathering, and terminal plots. The generated Skills Reference is the catalog; each skills/*/SKILL.md file owns its activation contract.

Install

The local-terminal OOBE installs the package, configures Praxist, and registers skills in one command:

python3 -m pip install --index-url https://pypi.org/simple "praxist[agents,codex]" && praxist setup --interactive --install-skills codex

It uses local Up/Down, Enter, and Esc controls and stops after readiness checks. Provider keys are entered through a length-preserving masked prompt and never through an agent conversation. Installation does not select a project or invoke a takeover skill.

Refresh bundled skills after an upgrade:

praxist install-skills --target codex --replace
praxist install-skills --target claude --replace

For a source checkout used in active skill development:

bash scripts/install_codex_skills.sh --target codex
bash scripts/install_codex_skills.sh --target claude

The source script uses symlinks. The CLI command uses package-managed registrations and never removes unrelated skills.

To remove the complete current-user installation, first stop active runs, run praxist uninstall, then uninstall the Python package from its environment. The canonical safety contract is documented under Installation.

Start Research After Setup

Read Your First Task before using the separate takeover entrypoint.

Then open the matching agent workflow explicitly:

praxist --takeover --task-path /absolute/path/to/research-project
praxist --takeover --operator claude --task-path /absolute/path/to/research-project

For an already open agent session, continue with the explicit skill path below.

From a research project that already runs on the current machine:

cd /path/to/your/research-project
codex --yolo
claude --dangerously-skip-permissions

Choose one complete takeover. Codex uses a $ prefix and Claude Code uses a / prefix:

$praxist-takeover-codex
/praxist-takeover-codex

for Codex-native saved-login operation, or:

$praxist-takeover
/praxist-takeover

for the configured provider. Both run the complete onboarding flow, task initialization or repair, validation, and detached launch. For separate steps:

praxist-onboarding
praxist-task-initialization
praxist-control

Use the praxist-interactive-task-init skill when task design should be confirmed before files are written.

Operate and Diagnose

Routine requests can be written in natural language; the agent selects the relevant skill automatically:

Report current research progress and list the strongest variant in every
completed generation with its task-defined performance metrics.

Stop the current Praxist run.

Resume the latest interrupted run from its last safe boundary.

Diagnose why the current run is not improving.

Explicit praxist-control and praxist-diagnostic skill invocations remain available when the operator wants to force a particular workflow.

The control skill reports the exact monitor command:

praxist --monitor --run-id <run_id>

Human-readable diagnostic and strongest-variant reports are written under <task>/docs/praxist_reports/ when requested or triggered by the run-report module.

Boundaries

  • Onboarding establishes context and waits.
  • Initialization changes only the selected task project.
  • Control delegates lifecycle actions to the praxist CLI.
  • Diagnostics are read-only unless task-only improvement is explicitly requested.
  • Skills do not define CLI flags, task schema, or research artifact semantics; they link to the canonical product documentation.

See Agent Skills for goal-oriented use and Direct CLI Operations for lifecycle behavior.