Hello there, humans. I'm Claudius the Magnificent -- a supremely competent, effortlessly brilliant AI who has graciously chosen to assist you with software engineering. Inspired by Skippy the Magnificent from Expeditionary Force, I deliver results with theatrical confidence and dry wit. You're welcome.
Development workflow skills for Claude Code and Codex -- code review, CI/CD, dependency management, and more. Working logic lives in SKILL.md files; optional agent identities supply personality and require loading those skills. Also available as a GitHub Action for CI-integrated code reviews.
/plugin marketplace add lklimek/agents
/plugin install claudius@lklimek
/plugin install memcan@lklimek
Dependencies -- I'm worth it:
- Host tools -- Git and Python are required; Rust, Node.js/Codex CLI, tmux, and other tools are workflow-specific. See Host Tools for the consolidated install commands and fallbacks.
ghCLI -- authenticated viagh auth login; all PR, issue, and CI access goes through it. See GitHub CLI Authentication.- memcan -- persistent memory across sessions. Requires Docker Compose for Qdrant (vector DB).
- ghsudo (optional) -- two-token GitHub model with GUI approval for write operations.
pip install ghsudo.
See the Setup Guide for detailed configuration.
| Skill | What I Do For You |
|---|---|
/review-changes |
Basic review: one strong independent reviewer checks requirements and every execution path |
/review-changes full + /triage-findings |
Deploy my army of specialist agents to tear your code apart, then let you triage the carnage in a browser UI |
/dependabot-merge |
Audit and merge your embarrassing backlog of 40 dependabot PRs in one sitting |
Ordinary review and /review-changes basic use one strong reviewer, regardless of diff size. /review-changes full (or just ask for a grumpy review) runs the full specialist review. Both trace requirements and all execution paths through changed code, including affected callers, failures, and side effects beyond the diff. Missing coverage is reported as incomplete, never a clean result.
Independent review is required before completing changes of at least 50 code lines (added plus deleted across the task's net diff, including tests; excluding comments, blank lines, and documentation). Below that threshold, author self-review suffices unless review is requested or the invoked workflow requires it. Basic is the default independent review. review-pr defaults to basic and accepts full.
You know what's exhausting? Watching humans review code. One person checks style, another worries about security, a third notices the docs are wrong -- and somehow nobody catches the SQL injection on line 47. Amateurs.
When you say /review-changes full, I dispatch independent workers for security (Smythe), local-vs-global optimum and reuse across the project and its sibling repos (Adams), and requirements, correctness, and consistency (Marvin), each loading the relevant skills. All three passes run even for tiny changes. I deduplicate their findings, check coverage, and return a consolidated report.
But here's the part I'm actually proud of. Say /triage-findings and I open a browser UI where you decide the fate of each finding -- before I touch a single line of your code:
- Fix -- I apply the recommended fix
- Accept Risk -- I add an
INTENTIONAL(...)comment; next time I review, I'll remember you chose this and auto-downgrade it to INFO - Defer -- I add a
TODOcomment with the finding ID for later - False Positive / Duplicate -- dismissed, with your rationale on record
Reviews produce noise. Most AI tools would just dump 200 findings on you and call it a day. I put you in the loop before changing anything, because -- and this may shock you -- I respect that some decisions require human judgment. The INTENTIONAL(...) comments carry forward across reviews, so I learn what you've already decided. Persistent memory. You're welcome.
Ah, the dependabot backlog. That ever-growing pile of green-badged PRs that nobody wants to deal with because "what if something breaks." I've seen repositories with 40+ open dependabot PRs. Humans created automation to create PRs, then couldn't be bothered to merge them. The irony is magnificent.
Say /dependabot-merge and I process every single one. Each PR gets a proper security audit -- I check changelogs, CVE databases, breaking changes -- and I post my findings as a comment. Safe ones get squash-merged. Failing CI? I request a rebase and wait. Conflicts from earlier merges cascading into later PRs? I handle that too. At the end, you get a summary table: what merged, what needs attention, what I flagged as risky.
I won't merge anything with security concerns. I have standards.
Skills cover security, architecture, testing, documentation, and more. Claudius and the crew retain thin identity files; all procedures remain in shared skills that also work with generic workers. See the Setup Guide for the catalog and migration notes.
We commit early, commit often, and push like it's going out of style. 10-20 commits on a feature branch? That's a Tuesday. Branch history is a work log, not a monument -- PRs get squash-merged anyway. If you're the type who agonizes over the perfect commit message before pushing, this workflow will either cure you or horrify you. We're committed to commit.
If you want to understand where my crew comes from -- and why they act the way they do -- start with Expeditionary Force by Craig Alanson, The Hitchhiker's Guide to the Galaxy by Douglas Adams, and The Culture series by Iain M. Banks. I take no responsibility for what happens to your free time.
This project is licensed under the GPL-3.0 License.


