Rather than limiting search to a predefined parameter space, Praxist runs an evidence-driven research loop that can explore methods, architectures, and strategies. Give it an already runnable project with measurable evaluation, and it produces task-local runs with evaluated evidence and successive research agendas. It operates alongside Codex or Claude Code, with direct CLI operation also available.
For research engineers with runnable projects when the best path is unknown, and plugin authors adding integrations
Praxist keeps scientific truth in explicit external task projects while the package provides task-agnostic protocols, orchestration, storage, replay, and plugins. Each run freezes configuration and writes task-local evidence and lifecycle artifacts rather than modifying the source project.
docs/concepts/architecture.md (Ownership Boundary); README.md (Praxist And The Task Project)
Core contracts
Defines protocols, resolution, storage, replay, budgets, credentials, and extension interfaces.
praxist/core/
Plugin catalog
Supplies replaceable runtimes, providers, workflow stages, tools, and budget policies.
praxist/plugins/
Research launcher
Launches registry-backed research runs and exposes run monitoring commands.
praxist/cli/start.py
Task templates
Provides scaffolding for task-owned evaluators, metrics, and research contracts.
templates/tasks/
Generated
