Skip to content

Developer Skills

ZodiacCore publishes skills for developers building services with ZodiacCore. These files are project assistance for coding agents, not runtime Python dependencies. Install them into the service project where the agent will work.

Together these skills, zodiac check, and the generated AGENTS.md form a quality loop: zodiac-docs checks that you write against the version-matched docs first, zodiac-core-integration-summary audits adoption (and drives zodiac check for mechanical wiring), and AGENTS.md carries the judgment rules the machine cannot check.

Available Skills

Skill Purpose
zodiac-docs Look up version-matched docs, source, and tests before answering usage questions or implementing APIs, routes, schemas, CLI, templates, migrations, and troubleshooting.
zodiac-core-integration-summary Audit a service and produce a ZodiacCore adoption matrix. Mechanical wiring evidence comes from zodiac check --format json.

zodiac-core-integration-summary must not restate CLI rule catalogs. It runs zodiac check for anti-patterns, then judges pagination, cache, configuration, schemas, and ➖/❓. If the user asks to fix mechanical ❌ cells, apply each error rule's change and rerun the CLI. A green check report certifies mechanical wiring only. Import-layer anti-patterns are in the CLI; thin routers, DI, session ownership, and lifecycle judgment stay in AGENTS.md.

Use zodiac-docs to explain an API or version-specific behavior, and when adding or changing routes, schemas, envelopes, pagination, or exceptions.

Install Into A Project

Agents discover project skills from .agents/skills/. Packaged ZodiacCore skills live in the installed zodiac-core wheel. Link them from the version you actually installed:

uv add --dev "zodiac-core[zodiac]"
uv run zodiac skills install

zodiac new prints this after uv sync --extra dev so the links follow the project's locked zodiac-core, not the CLI that generated the files. Run it from the service root or a subdirectory; it walks up to pyproject.toml.

Defaults to Codex and links .agents/skills/zodiac-* to the matching folder in the package. Use --agent for Claude (.claude/skills), Cursor (.cursor/skills), Copilot (.github/skills), Gemini (.gemini/skills), or --agent all. Unix uses a directory symlink; Windows uses a directory junction. The command gitignores the packaged zodiac-* directories. Re-run after uv sync; existing links to packaged zodiac-* skills are retargeted to the current install. Use --force only to replace a copied directory. Do not copy or commit those directories.

Remove the packaged links with uv run zodiac skills uninstall (same --agent and project-root walk-up). Other skills in the directory stay. Use --force only to delete a copied zodiac-* directory.

Start a new agent session in that project, or reload the session if the client requires it, then ask for the skill by name:

$zodiac-core-integration-summary

Generated projects document uv run zodiac skills install (Codex by default; pass --agent for other clients) and uv run zodiac check in AGENTS.md and README.md. The CLI is the source of truth for mechanical wiring whether or not the skill is installed.

When To Use

Use zodiac-core-integration-summary when you want an adoption matrix, 对接表, or to fix wiring the CLI already reported. Run:

uv run zodiac check --format json

If the project does not install the CLI extra, or the installed zodiac-core predates zodiac check, stop and prompt the user to upgrade zodiac-core and add zodiac-core[zodiac] to dev dependencies, then rerun. Do not run the latest CLI from uvx against an older project.

Map rules[] into the matrix columns, then fill pagination, cache, configuration, and schemas by reading the code. Apply change only when the user asked to fix mechanical errors.

Use zodiac-docs for questions about ZodiacCore APIs, configuration, CLI, generated templates, upgrades, and runtime behavior, and before implementing those APIs in a service. It resolves the local source revision or the downstream project's locked ZodiacCore version, then uses matching documentation, source, and tests as evidence. Do not implement from memory; read the skill's source-map rows first.

Both skills are read-only by default. They should not modify code unless you explicitly ask the agent to implement or fix something. Fixes for mechanical wiring must follow the CLI change field.