Git-native versioning for AI Skill packages.
sit puts prompts, schemas, golden tests, and release artifacts under semantic version control. It knows what changed, classifies risk, and gates your commits and releases.
pip install sit-toolkitWhen you version a prompt or schema with plain Git, you get +13 -2 lines. You don't know if it's a typo fix or a breaking behavior change. sit does.
$ sit diff v0.3.0..v0.4.0
Skill Diff
Baseline: paper-webpage-builder@0.3.0
Current: paper-webpage-builder@0.4.0
Risk: review-required
Suggested version bump: minor
[prompt]
- PROMPT changed SKILL.md (+13 -2; headings: Core Rule, Workflow)
[script]
- SCRIPT changed scripts/scan_paper.py (review required)
[reference]
- REFERENCE changed references/design_principles.md (+27 -3)
# New package
sit init my-skill && cd my-skill
# Existing project
sit standardize .
# Validate, test, review
sit install-hooks .
sit validate . && sit test .
sit diff HEAD~1..HEAD
sit review HEAD~1..HEAD
sit pr-summary HEAD~1..HEAD
# Release
sit release minor . --bundleLifecycle: sit init, sit standardize, sit onboard, sit doctor
Quality: sit validate, sit test, sit test --run, sit deps check
Diff & Review: sit diff, sit review, sit pr-summary, sit report, sit ci-summary
Release & Safety: sit install-hooks, sit commit, sit release, sit undo
Git passthrough: sit add, sit push, sit pull, sit branch, sit checkout, sit log
sit exposes its capabilities for AI agents via three interfaces:
| Interface | Usage |
|---|---|
| Auto-discovery | sit onboard --agent — one-command setup for Codex, Claude Code, Cursor, etc. |
| Python SDK | from sit.sdk import Sit — direct API calls |
| MCP Server | pip install 'sit-toolkit[mcp]' — 7 tools over stdio |
| LLM Tool-Use | from sit.tool_use import get_tools_openai — OpenAI & Anthropic schemas |
Agent auto-discovery
# Add agent config to an existing skill package
sit onboard --agent ./my-skill
# Or combine with full onboarding
sit onboard --agent ./legacy-projectThis generates .mcp.json (MCP server config) and AGENTS.md (agent rules).
Codex reads AGENTS.md and will run the sit loop after Skill changes:
git status --short, sit validate, sit test, and sit diff HEAD..WORKTREE
for uncommitted working-tree review. Claude Code, Cursor, and other MCP-aware
editors can also discover the sit MCP server through .mcp.json. Restart your
editor after running if it needs to reload project instructions or MCP config.
Python SDK example
from sit.sdk import Sit
s = Sit("./my-skill-package")
s.info() # package metadata
s.validate() # structure checks
s.test() # golden tests
s.diff("./old") # semantic diff
s.pr_summary("./old") # PR summary
s.report(compare="./old") # full reportMCP Server config
pip install 'sit-toolkit[mcp]'
sit-mcp-server{
"mcpServers": {
"sit": { "command": "sit-mcp-server" }
}
}LLM Tool-Use schema
from sit.tool_use import get_tools_openai, get_tools_anthropic
tools = get_tools_openai() # OpenAI format
tools = get_tools_anthropic() # Anthropic formatsit init generates a GitHub Actions workflow that validates, tests, and posts a semantic summary to your PR:
- run: sit validate "$SIT_PACKAGE_DIR"
- run: sit test "$SIT_PACKAGE_DIR"
- run: sit test "$SIT_PACKAGE_DIR" --run
- run: sit ci-summary "$SIT_PACKAGE_DIR" --compare origin/main..HEAD >> "$GITHUB_STEP_SUMMARY"Apache-2.0. See LICENSE.