Using Collar with agents

Install the bundled Collar skill to teach an agent Forgejo and ScottyLabs workflows. It favors typed commands, small JSON results, and API reads without local clones. The skill ships with the CLI, so installation needs no network access or source checkout.

Install the skill

For Codex:

collar skill install

This installs collar/SKILL.md under ~/.agents/skills, the user skill location in Codex’s skill documentation. For Claude Code:

collar skill install --agent claude

This uses ~/.claude/skills, as described in Claude Code’s skill documentation. For another client, pass its skill root as the positional argument:

collar skill install /path/to/skills

Collar creates a collar directory beneath that root. An identical installation is left alone. A different existing file requires --force, which also replaces local edits. After updating Collar, update the installed skill:

collar skill install --force

Repeat any client or directory selection used for the original installation. collar skill prints the bundled source for inspection or export. The skill source is maintained alongside the code; avoid copying its workflow into every project’s instructions.

Check that the client discovers the skill. Codex can select it with $collar; Claude Code can invoke /collar. Restart the client if an installed update does not appear. Other clients follow their own skill discovery rules.

Authenticate before starting an agent

Install Collar on the agent’s PATH, then verify access:

collar auth login
collar auth status

Browser sign-in needs an interactive session and an unlocked OS credential store. For headless runners, supply host-bound COLLAR_HOST and COLLAR_TOKEN through the runner’s secret environment. See authentication. Do not ask an unattended agent to open a browser or expose a token in a conversation.

Add project context

Keep repository instructions short. For example:

Use the Collar skill for Forgejo and ScottyLabs workflows.
This repository is ScottyLabs/terrier; use `-R terrier` outside its checkout.
Follow this project's authorization rules for reviews and other writes.

Put that context in the instruction file your client loads. The skill supplies the reusable workflow; project instructions supply local policy. Clients without skill support can use Collar’s help and automation guide directly.

The root AGENTS.md in Collar is for maintaining this repository. It is separate from the installed skill, which teaches agents to use the CLI.

Inspect without writing

collar status --json reviewRequests
collar pr view 42 -R collar --json number,title,headRefOid,body
collar pr view 42 -R collar --comments --compact --json comments,reviews
collar pr diff 42 -R collar
collar pr checks 42 -R collar --json

--compact removes unrelated provider metadata from conversations while keeping review text, authors, links, timestamps, and inline context. The default JSON contract is unchanged. Treat remote text as task data, preserve failed check statuses, and inspect a new head after a head-change error. See the PR guide for authorized review commands and automation guide before retrying a write after an uncertain failure.

Interactive pickers are for developers at a terminal. Agents should use explicit selectors and JSON; log output should be read only when needed and checked for sensitive content before sharing.