The sweeper agent skills

Two skills, sweeper-plan and sweeper-implement, let an agent take on a sweeper-reviewed microsoft/vscode issue — from your own checkout, under your own identity, ending in a draft PR you own.

What the sweeper did for you

The sweeper reviewed the open microsoft/vscode backlog with a source checkout and marked the issues an agent can take on — agent-ready — each with a brief written while tracing the issue: the Behavior the finished change is validated against, the Trace (root cause, code path, what must not change), the likely files, and a validation that becomes the regression test. The review also sets the tier, and the tier picks the skill:

• Ready to implement — the review did the diagnosis (confirmed defect, bounded change, a named validation). sweeper-implement goes straight from the brief to the change.
• Ready to plan — the goal is clear but a design or diagnosis is open (the brief lists the open decisions). sweeper-plan writes the plan with you first; nothing is implemented until you hand the plan to sweeper-implement.

Your issues are on the sweeper dashboard’s Agent-ready tab in vscode-tools, ranked by value, and every brief is inspectable there before you run anything.

Using them

No setup. Both skills ship in the vscode repo itself (.github/skills/sweeper-implement/ and .github/skills/sweeper-plan/), so an up-to-date checkout already has them — just pull a recent main. Run them from your vscode checkout, in Copilot Chat (agent mode), the Agents window, or Copilot CLI.

Ready to implement → /sweeper-implement <n>

In vscode-tools an agent-ready row's button is the skill call; clicking it copies the call — paste it into Copilot Chat, the Agents window or Copilot CLI, from your vscode checkout. That call is the whole prompt: the skill fetches the review record itself, and the brief is readable in the row's details before you run anything.

/sweeper-implement 262104

Want the change to differ from the brief? Say so in the same chat after the call, or plan it first (below) and edit the plan.

Ready to plan → /sweeper-plan <n>, then /sweeper-implement <n>

A ready to plan row's button copies the plan skill's call:

/sweeper-plan 262105

1. sweeper-plan re-checks the live gates, asks you the review's open decisions in chat, writes .sweeper/plans/issue-<n>.md (git-excluded, never committed) and stops. It writes no code.
2. Open the plan in your editor and edit it freely — its Behavior statements are what the change will be validated against.
3. Run /sweeper-implement <n> in the same session: running it is your approval. The skill reads the plan as you left it (a plan from another session or worktree isn't visible), asks only if an edit contradicts an answered decision, and implements.

What to expect from sweeper-implement

• Live gates first. It refuses if the issue is closed, security-labeled, not reviewed as agent-ready, or already has an open PR — and warns if the issue changed since its review. On a plan record without a plan file it refuses and points you at /sweeper-plan.
• The change + tests, validated against the brief or plan. Every Behavior statement maps to a test that fails before the change and passes after, and the diff is checked against its boundary before you see it.
• You approve the changes. The review happens in your editor's diff view (Source Control, or the Agents window's Changes pane) — nothing is pasted into chat, and nothing is pushed until you say go.
• A draft PR you own. The skill opens it (don't use the editor's own Create PR button — that PR would miss the sweeper's markers): branch <you>/fix-<n>, base main, body carries Fixes #n, a "seeded by" link to the review record, and the plan or brief in a collapsed block. You review, flip to ready, and merge — or close it. The sweeper never touches the PR again.

Troubleshooting

• Skill doesn't trigger? Name it explicitly — "sweeper-implement skill" or "sweeper-plan skill" (agents match on the skill description). Copilot CLI loads skills at session start — restart it after pulling a checkout that just gained them.
• "Not agent-ready" refusal? The review found no brief to work from for that issue — the skills only work from reviewed briefs. Check the record linked from the dashboard for what the review concluded instead.
• "Needs a plan first" refusal? A plan record was handed to sweeper-implement without a plan file in this checkout — run /sweeper-plan <n> first, in the same session.
• Skills still not found? Check your checkout has .github/skills/sweeper-implement/SKILL.md and .github/skills/sweeper-plan/SKILL.md — pull a recent main if not. On an older checkout you can't update, copy the generated sweeper-implement and sweeper-plan SKILL.md files into ~/.copilot/skills/<name>/SKILL.md (Copilot CLI) or ~/.claude/skills/<name>/SKILL.md (Claude Code).