Teaching the agent

What an agent needs from the manual — the call, the flags that are the reviewer's and not its own, what the exit status means, how to find the line a comment is about, what to do afterwards — is written down once more as a skill, in skills/slop-review/SKILL.md. It is in the Agent Skills format that Claude Code, OpenCode and most other coding agents read, so one file serves them all; what differs is where each one looks for it.

Claude Code installs it as a plugin, straight from this repository:

claude plugin marketplace add https://gitlab.com/c0va23/slop-review.git
claude plugin install slop-review@slop-review

/slop-review then asks for a review, and Claude reaches for the skill by itself when work is ready to be looked at. claude plugin update slop-review brings in whatever the skill says now.

OpenCode reads the same file from ~/.claude/skills/, which is also where Claude Code finds a skill that did not come as a plugin. The skills CLI puts it there, or wherever the agents you name keep theirs:

npx skills add https://gitlab.com/c0va23/slop-review -g

Or copy the directory by hand: it is one file, and ~/.claude/skills/slop-review/SKILL.md is read by both.

The skill is held to the program by a test: every flag it names is one the command takes, the review it shows the agent is rendered by the same code that prints one, the message it tells the agent to stop on is the one that is printed, and the default ranges and the decisions are the ones in the tree. Changing any of those means changing the skill in the same commit.