ai-stp：面向 AI 智能体的跨 harness 配置管理 CLI 工具

English · Русский

The command is ai-stp. The distribution on PyPI is ai-stp-cli. The primary consumer is the user's agent: every command returns one JSON envelope. ai-stp does not call a model API and does not require a model key.

One setup, nine kinds, exact versions

A setup belongs to one harness from creation. The closed kinds are instruction, skill, mcp, hook, command, agent, plugin, setting, cli. command is a named slash invocation; cli is a standalone executable. Memory and rules are content of those kinds, not extra kinds. A published version pins exact component versions and is immutable.

CLI assembles, the provider writes

The CLI and the agent select, validate, and bundle. Every command is one JSON envelope. The CLI does not write the harness's native files. The web owns the account and the public catalog; it does not assemble or install a setup. Only the harness provider writes native state.

Trust is origin, version, and consent. Compatibility — graph, target, and policy — must decide before apply. A published version is an immutable digest bound to object version, policy, and device. Sync continues from the last confirmed cursor. Author verification is origin, not a safety verdict on the bytes. author_verified and component_verified are independent.

Local work needs no account. The public catalog is readable anonymously. Google or GitHub sign-in unlocks private objects, sync, publication, devices, and grants.

Install the CLI, then let the agent drive it

uv tool install ai-stp-cli
ai-stp task intents --json

Next commands the agent should read from compact discovery

ai-stp task intents --json
ai-stp task start --intent initialize --idempotency-key initialize-session-01 --json

help --agent --json is the full registry when a shipped intent is not enough. Do not dump it as a prelude.

ai-stp is the executable. ai-stp-cli is the package name. Copying uv tool install ai-stp installs a distribution this project does not publish.

Desktop application

apps/desktop is ai-stp-desktop, a Tauri 2 desktop shell over the same CLI machine contract — catalog, task-engine journeys, targets, and recovery in a native window for Linux, macOS, and Windows. Release bundles embed the CLI as a sidecar, so no separate install is required. Download the latest build from GitHub Releases (bundles are unsigned; SHA256SUMS covers every artifact). Details and the boundary rules live in apps/desktop/README.md and ADR-0222.

Supported harnesses

full — native surface and a provider route. partial — one of those. — — no native surface. The matrix is ai-stp toolchain harness-capabilities.

An unknown harness is undefined. Automatic install is refused.

Current direction: native Rust CLI

The native ai-stp-v2 preview is available for Linux x86-64, Apple Silicon macOS and Windows x86-64 in GitHub Releases. Archives include checksums and build provenance. The binary runs without Python and uses explicit isolated state. Its implemented capabilities, platform requirements and verification instructions live in apps/cli-v2/README.md.

The production ai-stp executable remains the Python CLI until verified state ownership transfer (ADR-0227). Native configuration installation, connected account/sync/publication, automated updates and desktop distribution remain unfinished. Setup systems are ai-stp installation components with separate release repositories (ADR-0228).

main is the protected release branch; work starts from dev and enters it by PR before promotion to main. For a new workstation, follow CONTRIBUTING.md.

Phase status belongs to docs/engineering/implementation-roadmap.md: read that file when you need the current evidence, not a summary here.

Contributing

• CONTRIBUTING.md: how changes enter this repository.
• SECURITY.md: how to report a vulnerability.
• CODE_OF_CONDUCT.md: expectations for participation.
• AGENTS.md: rules for people and agents. Read before any repository change.

Documentation

• docs/index.md: map of product, architecture, contract, engineering, and operations documentation.
• docs/product/vision.md: the problem, users, value, and positioning of ai_stp.
• docs/product/scope.md: required MVP capabilities, harness statuses, and explicit exclusions.
• docs/architecture/overview.md: overall data flow and the boundaries of the local and server environments.
• specs/index.md: versioned requirements that the code must satisfy.
• User docs: English · Русский

License

AGPL-3.0-or-later. Network use of the platform is covered: if ai-stp is offered as a service, the source of the modified version stays available to those users.

The catalog belongs to the guild. Public harness installation components retain their separate repositories and licenses.

Components and setups published by users are independent works licensed by their authors; the platform license does not apply to them.