agent-tools

Skills for coding agents, written for research work: simulation and analysis codebases, publication figures, and the scientific software around them. They run under Claude Code and Codex, which both read a SKILL.md from a skills directory.

Skills

• code-review-scientific: multi-agent review of a GitLab merge request in a scientific codebase.
• figure-polish: conventions and QA for publication figures made with matplotlib and mplpub.
• dream: consolidate past session history into a short set of transferable facts loaded in every session.
• status-report, handoff, pickup: carry session objectives through a report, a saved handoff, and a pickup.

code-review-scientific

Multi-agent review of a GitLab merge request in a scientific codebase. Weighs correctness against a small, readable diff, judges the change from the user's side as well as the developer's, and hunts for the structural simplification rather than stopping at style. Writes the result to a review-<MR-number>.md file, and posts it as a single GitLab comment when asked. Assumes the glab command line tool is authenticated.

figure-polish

Conventions and QA for publication figures made with matplotlib and mplpub. Fixed column widths of 3.4 and 6.6 inches, compact multi-panel layout, layout variants rendered side by side into one contact sheet instead of one guess per round trip. Nothing is reported as finished until the rendered image has been looked at.

dream

Consolidates the session history of every local agent config directory into a short ~/.claude/DREAMED.md of transferable facts: the conventions and procedures that still hold in a project that does not exist yet. Import that file from your ~/.claude/CLAUDE.md with a line @~/.claude/DREAMED.md. Runs only when invoked explicitly, and writes that file plus ~/.claude/DREAMED.sources.md, a ledger of the incidents behind each line that is never imported.

Session skills

Three skills that keep the thread of a session, built around objectives with stable IDs such as A1, A2, B1. status-report reports the objectives, their statuses, and the blockers. handoff writes objectives, current state, and local or HPC job details to handoff.md in the working directory. pickup reads the repository README and that handoff, then restores or establishes the objectives.

Invocation

handoff, pickup, and dream run only when you type their name, and cost no context otherwise. That takes two files: disable-model-invocation: true in the frontmatter for Claude Code, and agents/openai.yaml with policy.allow_implicit_invocation: false for Codex. The other skills keep a trigger description and can also fire on their own.

Examples

• examples/figure-polish: the same figure request, run without and with the figure-polish skill, with both scripts and both rendered figures.
• examples/code-review-scientific: the same dynasor commit, reviewed without and with the code-review-scientific skill, with both reviews.

Install

git clone https://github.com/erikfransson/agent-tools.git
agent-tools/install.sh

install.sh symlinks every skill under skills/ into ~/.claude/skills, and into ~/.codex/skills when ~/.codex exists. Name skills as arguments to link only those. An existing directory with the same name is moved aside as <name>.before-agent-tools rather than overwritten.

Update

git -C agent-tools pull

The symlinks pick the new content up immediately, including in a running session. A skill added to the repo after install needs one more install.sh <name>.

License

MIT, see LICENSE.