Coder Agents integration

Coder Agents support is beta.

tokenjuice install coder-agents writes a reusable workspace skill to .agents/skills/tokenjuice/SKILL.md at the current git/project root. Coder Agents discovers skills from .agents/skills/ and loads their full instructions on demand, so tokenjuice uses that native skill surface instead of a shell hook.

Install

tokenjuice install coder-agents
tokenjuice doctor coder-agents
tokenjuice uninstall coder-agents

By default tokenjuice resolves the current git root and writes <git-root>/.agents/skills/tokenjuice/SKILL.md. Set CODER_AGENTS_PROJECT_DIR=/path/to/project to target a specific project directory during tests or scripted installs.

Behavior

• The skill frontmatter uses name: tokenjuice, matching the skill directory.
• The skill tells Coder Agents to prefer tokenjuice wrap -- <command> for terminal commands likely to produce long output.
• The skill tells Coder Agents to treat compacted output as authoritative.
• The only documented escape hatch is tokenjuice wrap --raw -- <command>.
• Existing Coder skill content is backed up without clobbering older backups; uninstall restores that exact pre-existing file when possible.

Current beta caveat

Coder Agents workspace skills are prompt/tool guidance, not command hooks. This integration does not register MCP tools, change Coder templates, or rewrite shell output; it gives Coder Agents a discoverable workspace skill for deciding when to invoke tokenjuice around noisy terminal commands.