agent-tools

agent-tools

My custom skills and subagents for working with Claude Code.

How you drive it (pull-driven by default)

The default posture is pull-driven: nothing runs while you're away. You sit down, run a check-in sweep, and hand-dispatch the work you choose. You decide the unit of work; the dispatcher (or, for a single task, a runner) carries it out.

Typical session:

1. File issues (describe what you want done)
2. Sit down and review state across the project(s)
3. Hand off the work you choose: @dispatcher work the backlog for a scoped batch, or @runner implement #N for a single task
4. Review the resulting PRs

A loop/timer may remind you to run this sweep, but it never runs the sweep itself. Dispatching is a deliberate act you take when you sit down.

Opt-in: work the whole backlog unattended

When a batch of well-specified issues makes it worth it, you can escalate to the fully autonomous loop for that session:

claude --agent dispatcher

In this mode the dispatcher reads your GitHub issue queue, decides execution shape, and fires runners. Runners dispatch workers that branch, edit files, open PRs, watch CI, and merge — the whole pipeline runs unattended until the queue is worked. Turn it on deliberately, per session; it is the escalation, not the standing mode.

Architecture

graph LR
    Human((Human)) -->|files issues| GH[(GitHub Issues)]
    Human -->|claude --agent dispatcher| D[Dispatcher]
    D -->|reads queue| GH
    D -->|dispatches| R[Runner]
    R -->|opens draft PR| GH
    R -->|dispatches| W[Worker]
    W -->|edits files, commits| Repo[(Source)]
    R -->|pushes, watches CI, merges| GH
    R -->|dispatches| Rev[Reviewer]
    Rev -->|approve + merge or request-changes| GH

What's in here

• skills/ -- operational knowledge. Process discipline, git workflow, dispatch patterns, sandbox preflight, release-audit anchoring, and the rest of the patterns I've found load-bearing across projects.
• agents/ -- subagent definitions. runner does one task end-to-end; dispatcher gathers context, decides execution shape, and fires runners (single, parallel, sequential, etc.).

Agents

Agent	What it does	Invoke with
dispatcher	Scopes units of work, decides execution shape, fires runners	@dispatcher work the backlog (in-session) or claude --agent dispatcher (CLI)
runner	Implements one GitHub issue end-to-end (branch, draft PR, CI, merge)	@runner implement #N
reviewer	Reviews a PR: approve+merge, request-changes+draft, or approve+note ordering	@reviewer review #N
worker	Bounded code-change executor; reads context, edits files, commits	(dispatched by runner)
auditor	Read-only audit of a codebase against a rubric; files findings as issues	@auditor audit <domain>, or dispatched for audit+remediate

See agents/README.md for invocation details and when to skip the dispatcher and go straight to the runner.

Feedback loop

graph LR
    Work[Work Sessions] -->|observes gaps| FF[field-feedback\nagent-feedback]
    FF -->|files issues| Q[(Issue Queue)]
    Q -->|triage labels| L[Labeled Queue]
    L -->|dispatcher| Fixes[PRs + Merges]
    Fixes -->|improves| Skills[Skills & Agents]
    Skills -->|better| Work

The field-feedback and agent-feedback skills file GitHub issues automatically when agents encounter problems during dispatch. @dispatcher triage open issues labels and prioritizes them. You then work that queue on your next check-in — hand-dispatching the fixes you choose, or, when the backlog warrants it, opting into the autonomous loop above to let runners work it unattended. Either way the loop closes; what varies is whether you drive each hop or let the queue do it.

How it fits together

• Skills provide operational knowledge loaded at dispatch time
• Agents define roles (dispatcher, runner, worker, reviewer)
• Durable state (GitHub issues, PRs, CLAUDE.md, code) is the substrate -- agents are ephemeral, state persists
• The self-improving loop: use -> observe -> file -> fix -> repeat

Getting started on a new project

1. touch CLAUDE.md -- marks the project for workspace survey.
2. Write Overview, Architecture, and Current Status sections.
3. @dispatcher work the backlog or @runner implement #N.

Install

agent-tools ships as a Claude Code plugin, and the repo is its own marketplace. This works in both the CLI and the desktop app: