Skip to content

About

Agent skills for Claude Code: council (plan, build and review a task across Claude, Codex and OpenCode in Herdr), commit, pr and triage-reviews

Topics

Resources

Stars

6 stars

Watchers

0 watching

Forks

Latest commit

 

History

9 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

agent-skills

Skills I use every day with Claude Code. Each one is a folder under skills/ with a SKILL.md.

npx skills add digitaltracer/agent-skills   # asks which skills and which agents

Or copy a folder into ~/.claude/skills/. You start each one with its slash command (/council, /commit, /pr, /triage-reviews, /triage-comment); the model never picks those up on its own. fix-ci is the exception: the model also uses it when CI fails.

council

Runs a task past a council of coding agents in tabs you can watch: Claude, Codex and OpenCode, two to four of them, each on the model you name.

  • plan debates a design across the agents, from a brief you approve, to a plan with evidence another agent can carry out.
  • implement has one agent build the plan on the current branch, a commit per step.
  • review has the agents review the branch's diff against the plan. You get the findings that checked out, and the implementer applies the ones you pick, or all of them with --autofix.
  • run does all of it in order. resume <task> picks up an interrupted run.
/council plan <task> using claude with opus and codex
/council implement --plan .agents/tasks/<task>/plan.md --worktree
/council review using claude and codex
/council run <task> --autofix using claude and codex

Your session leads: it writes the brief, checks every claim against the code, settles disagreements and writes the results. It never writes code itself.

Needs Herdr (the app, or hc from herdr-core), jq, python3, and the CLIs you put on the council. The scripts are called from ~/.claude/skills/council/, so install it for Claude Code.

commit

Conventional commits from what you staged. It scans for credentials on every commit (vendor token prefixes, private keys, connection strings, .env files), and --no-verify doesn't skip that. If the staged changes mix concerns it proposes a split. --atomic works from the whole working tree and makes one commit per concern after you confirm. It never pushes and never adds AI attribution.

pr

Opens a pull request with gh. The description is written for people: what the change does and why, never file or function names. It picks labels the repo already has, suggests reviewers from CODEOWNERS, adds checklist items only when the diff calls for them (migrations, new env vars, docs), and shows you everything before creating it. The body template is mine; change it to your team's.

triage-reviews

Fetches every unresolved review comment on the branch's PR, from bots and people, and checks each claim against the code before believing it. You get a table: valid, false positive, stale (fixed by a later commit) or nit, each with its proof. On approval it makes the smallest fix each finding needs. /loop 10m /triage-reviews --fix keeps a PR clean while bots keep posting.

triage-comment

The same check for one comment you paste, or a link to it. It rules out stale comments first (a later commit already changed those lines), then cosmetic ones and ones a single read settles. Only a correctness claim that survives that goes to two subagents: one tries to build a concrete failure from the code, the other tries to name the guard that prevents it. If the first can't build one, that's the verdict. You get one verdict with the deciding file:line, and a fix on approval.

fix-ci

Diagnoses a failing GitHub Actions run from its URL, a PR number, or the current branch. It reads the failing step's log, then looks for an answer before inventing one: another workflow in the repo that already solves the problem, or what changed since the last green run. The fix stays small. In a monorepo it checks sibling services for the same defect. After you allow a push, it watches the run until it passes.

Licence

MIT, see LICENSE.

About

Agent skills for Claude Code: council (plan, build and review a task across Claude, Codex and OpenCode in Herdr), commit, pr and triage-reviews

Topics

Resources

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages