workbench-dev-team

workbench-dev-team

A Claude Code plugin that runs a three-agent development pipeline against a GitHub project board. Items flow from triage → development → review without you touching them. Part of the claude-workbench marketplace.

What it does

You work out of a GitHub project board. New issues land with no acceptance criteria. Well-refined ones sit in Ready. Work-in-progress has open draft PRs. PRs waiting on review pile up.

This plugin runs three agents on a local 20-minute clock to move items through the pipeline for you:

• Inspector Lestrade (Sonnet) — triage. Reads items in the Inbox lane, writes acceptance criteria, scores WSJF, moves them to Backlog for your review. The WSJF write also lands two GitHub-native issue attributes The Index derives server-side: the issue Type (PBI) and an issue-level Priority (Urgent/High/Medium/Low) mapped from the WSJF — org-repo-only and best-effort. Also runs blocker + consolidation sweeps: after a repo gets fresh triage work, he re-reads all of its open issues and (1) marks blocked-by dependencies (native GitHub issue dependencies, additive only) so blocked items stay out of Dr. Watson's queue, and (2) consolidates follow-ups — folds expand-from comments into the issue they target and merges unmistakable near-duplicate follow-ups into the earliest anchor (native duplicate-close, high bar, ambiguous clusters flagged not closed) so the backlog stops sprawling.
• Dr. Watson (Opus, $10/run cap) — development. Has two modes: The Index mode (Dispatch-driven, picks the top Ready/In Progress item, clones the repo, writes code and tests against AC, opens a PR, moves to In Review) and Direct mode (invocable as a sub-agent from Claude Code or Cowork for ad-hoc dev work — no The Index calls, just runs the /develop skill in a sub-agent context). Both modes follow the /develop skill for the actual coding.
• Sherlock Holmes (Opus parent, Sonnet lenses, $7/run cap) — code review. Reviews open PRs, approves or requests changes. Escalates to you after 3 change rounds — but your input resets that count: comment, review, or weigh in on the PR and the window restarts from your last word, so an escalated PR you've decided on gets a fresh review instead of bouncing straight back. Reviews fan out across blind, read-only lens sub-agents (AC conformance, correctness, security, test honesty), with every blocker adversarially verified before it lands; only the parent writes, so there's still exactly one App-signed verdict. Falls back to a single inline pass when the fan-out is unavailable. Findings route by locality: anything actionable in the code a PR touched blocks (request changes, however minor), and a hard correctness/security/test defect blocks wherever it lives — while a soft observation about untouched code becomes a tracked follow-up issue, never dropped. The precise routing rule lives in Holmes's canonical review contract (agents/holmes.md, §4e). On approval Holmes routes each follow-up himself — expanding the earliest related open issue in place (a comment Lestrade folds into its acceptance criteria) rather than opening a near-duplicate, and opening a new anchor issue via create_issue (App-signed as Holmes, board-added and PBI-typed) only when nothing related exists, so follow-ups expand the original instead of multiplying. When a follow-up is one sighting of an invariant that should hold across a whole class of call-sites (a containment guard, a null-check, a helper every caller owes), he sweeps the tree for every violating site and tracks them as one umbrella issue with a checkbox per site — closing the class in a single PR rather than minting a fresh single-site issue every review, the treadmill that otherwise turns one finding into an endless #A → #B → #C chain. On a change request there's no round-trip to spin out: Watson is already fixing the blockers, so he implements every non-blocking follow-up in the same PR — no extra issues, no churn. Follow-ups become tracked issues only on approval; on a bounce they're built straight into the PR.

A fourth component — Dispatch — is the local scheduled task that polls the board every 20 minutes and fires the right agent for each pending item. Dispatch is the only thing that's scheduled; the three agents run as dispatched subprocesses.

Install

/plugin marketplace add mike-bronner/claude-workbench
/plugin install workbench-dev-team@claude-workbench

That installs the agents, the Dispatch prompt, and the bundled skills (see below). Nothing is scheduled yet.

Bundled skills

The plugin ships three skills: