Pr Sweep logo

Pr Sweep

OrganizationPopular
TanStack
pr-sweep

Sweep open (or listed) PRs with up to 100 parallel agents: security-scan outside contributors, rebase onto main when behind (push --force-with-lease), approve pending first-time-contributor CI when relevant, optionally rebase in-house PRs, and report who should review. Supports full, changed-only, behind-only, and conflict-only scopes for cheap daily runs. Use when the user runs /pr-sweep (or /pr-inbound-sweep), or asks to "sweep PRs", "sweep inbound PRs", "security-check outside PRs", "rebase outsider PRs", "rebase our PRs", "approve waiting CI on PRs", "daily PR sweep", or "prep external PRs for review".

Overview

PublisherTanStack
Repositoryai
Skill namepr-sweep
Stars
3.2K
Forks
361
Bundled files
1
LicenseMIT
Links
  • Markdown instructions

    A SKILL.md file the model loads on demand, so it only costs tokens when a request actually matches.

  • Works with any LLM

    AI skills are plain Markdown, not provider-specific code, so this works with GPT, Claude, Gemini, Grok, or a local model.

  • 1 bundled files

    Scripts, templates, and references the model can read while it works. Files are read-only and never executed.

  • Open source

    Published by TanStack on GitHub. Read the source before you install it.

Installation

Install the Pr Sweep AI skill in TypingMind to use it with any LLM, or drop it into another agent that reads SKILL.md.

1

Install in TypingMind

TypingMind installs a skill straight from its GitHub folder — it reads SKILL.md, bundles the resource files, and stores the result locally.

  1. Open the app and go to Plugins → Skills.
  2. Choose "Install from GitHub".
  3. Paste the skill folder URL below and confirm.
    https://github.com/TanStack/ai/tree/main/.agents/skills/pr-sweep
  4. Enable the skill in any chat where you want it available.
Plugins → Skills → Add skill → From GitHub URL, then paste the folder URL and press Continue.
2

Install in another agent

Any agent that reads the Agent Skills format can use this skill — copy the folder into that agent's skills directory.

Claude Code — .claude/skills
git clone --depth 1 https://github.com/TanStack/ai.git /tmp/ai
mkdir -p .claude/skills
cp -r /tmp/ai/.agents/skills/pr-sweep .claude/skills/pr-sweep
Restart Claude Code after copying so it picks up the new skill.

Use it in TypingMind

Enable Pr Sweep in any TypingMind chat and the model takes it from there. Its name and description sit in the system prompt, and the moment a request matches, the model loads the full instructions itself — you never invoke it by hand, and it costs no tokens until it is actually used.

The model loads Pr Sweep on its own as soon as a request matches it.

Works with any AI model

AI skills are plain Markdown instructions rather than provider-specific code, so Pr Sweep is not tied to the model it was written for. Install it once in TypingMind and use it with GPT-5, Claude, Gemini, Grok, DeepSeek, Mistral, Llama, or a local model you run yourself — all on your own API keys.

  • Loaded only when it is needed

    The system prompt carries just the name and description. The instructions are fetched on the first matching request, so an idle skill costs nothing.

  • Switch models mid-chat

    Because the skill is instructions rather than code, changing model does not break it — the next model reads the same SKILL.md.

Skill instructions

This is the SKILL.md content the model loads. Read it before installing — a skill is instructions your model will follow.

PR Sweep

Prep PRs for review. Fan out one subagent per PR (cap 100). Default is dry-run (report only). Mutating steps require --apply (or the user saying "apply" / "go ahead").

Args

InvocationBehavior
/pr-sweepAll open non-draft PRs (full audit)
/pr-sweep 12 34 56Only those PR numbers
/pr-sweep --applyFull set, then rebase and push (--force-with-lease)
/pr-sweep --apply 12 34Rebase and push listed PRs only
/pr-sweep --outside-onlyAction target = outside authors only (default for mutations)
/pr-sweep --include-in-houseAlso rebase/update in-house branches (still never merge)
/pr-sweep --behindOnly PRs behind base / BEHIND / not up to date
/pr-sweep --conflictsOnly CONFLICTING / DIRTY / dirty merge state
/pr-sweep --changedOnly PRs changed since last snapshot (or updatedAt within 24h if no snapshot)
/pr-sweep --dailyRecommended daily recipe: --changed ∪ --behind ∪ --conflicts ∪ new outside PRs; security-scan new outside; lighter pass on the rest
/pr-sweep --apply --daily --include-in-houseDaily apply: prep outside + rebase ours when behind/conflicting

Combine freely: --apply --daily --include-in-house. Explicit PR numbers always win over filters.

Modes

  • dry-run (default): fetch, classify, security-scan, decide relevance / rebase need / CI need / assign recommendation. No push, no CI approve, no comments.
  • apply: after dry-run logic, rebase + git push --force-with-lease + CI approve for PRs that pass security and are marked actionable, in the same turn. --apply on the invocation is consent — do not wait for a second yes. Still never merge a PR and never comment on a PR. A local rebase with no push is a failed apply. Record results in SWEEP-*.md only.

Daily routine (recommended)

Option A — Grok /loop (same machine, session-scoped, expires ~7d)

text
/loop 1d /pr-sweep --apply --daily --include-in-house

Or dry-run every morning and apply only when you say go:

text
/loop 1d /pr-sweep --daily --include-in-house

/loop intervals: Nm / Nh / Nd (min 60s). Cancel with scheduler_list → scheduler_delete <id>.

Option B — Manual weekday

text
/pr-sweep --daily --include-in-house          # dry-run first
/pr-sweep --apply --daily --include-in-house  # after skimming plan

What --daily processes

Build the action set as the union of:

  1. New outside PRs — open outside non-draft not present in the previous snapshot (full security scan).
  2. Changed — updatedAt newer than last snapshot sweptAt, or head SHA changed vs snapshot.
  3. Behind — mergeStateStatus is BEHIND or not up to date with default branch.
  4. Conflicts — mergeable == CONFLICTING or mergeStateStatus in DIRTY, BLOCKED with dirty indicators.
  5. CI waiting approval — outside PRs with first-time-contributor gate (cheap; no full diff if already in snapshot as security: clean).

Skip from action set (still note counts in report):

  • Drafts (unless listed explicitly)
  • security: alert from prior snapshot until human clears
  • PRs marked blocked-conflicts in the last 24h with no updatedAt change (avoid thrashing)
  • Bot version/release PRs (changeset-release/*, pure Renovate) unless --include-bots

Cost target: daily should touch tens, not all open history. Full /pr-sweep remains the weekly deep scan.

Snapshot (enables --changed / --daily)

Path: .agent/pr-sweep/snapshot.json

json
{
  "repo": "TanStack/ai",
  "sweptAt": "2026-08-10T18:00:00Z",
  "defaultBranch": "main",
  "defaultBranchSha": "abc…",
  "prs": {
    "1069": {
      "author": "mikemikimike",
      "outside": true,
      "headSha": "def…",
      "updatedAt": "2026-08-10T04:01:40Z",
      "security": "clean",
      "mergeable": "MERGEABLE",
      "mergeStateStatus": "UNSTABLE",
      "lastAction": "approve-ci",
      "lastActionAt": "2026-08-10T17:30:00Z"
    }
  }
}

Write/update after every run (dry-run or apply). Diff against this file for --changed. If missing, treat all open PRs as new for one full pass, then write the snapshot.

Prerequisites

bash
gh auth status
gh repo view --json nameWithOwner,defaultBranchRef,owner

Stop if not authenticated. Confirm you are in the target repo (or pass owner/repo if the user named one).

Safety (hard rules)

  1. Never merge a PR.
  2. Never git push --force. Only git push --force-with-lease.
  3. Never mutate on a security finding (security: "alert"). Report and stop that PR.
  4. Default action target = outside only. In-house rebases require --include-in-house (or explicit PR numbers that happen to be in-house).
  5. In-house security: skip malware fan-out unless the PR touches scripts/, .github/workflows/, lockfiles, or install lifecycle — then light scan only.
  6. Worktrees only for apply-mode git ops (spawn_subagent with isolation: "worktree"). Do not checkout foreign branches in the main workspace.
  7. Cap concurrent apply subagents at 8 (worktrees + API). Read-only fan-out may go up to 100.
  8. If rebase conflicts cannot be resolved cleanly in under ~10 minutes of agent work, abort, leave a note in the report, do not push partial state.
  9. --apply / "apply" / "go ahead" / "yes" after a dry-run is consent. Write the apply plan into the report and mutate immediately. Do not stop for a second yes, including when more than 5 PRs would be mutated. Daily /loop … --apply likewise fires without re-prompting.
  10. Prefer rebase for both outside and in-house. Use merge-from-base only when the agent JSON says so (many merge commits, prior merge strategy).
  11. Push is part of rebase. After a clean rebase (or merge-from-base), run git push --force-with-lease onto the PR head remote, then verify gh pr view N --json headRefOid changed. Do not mark the PR done until the remote moved or you recorded push-403.
  12. Never comment on a PR. GitHub already notifies the author on push. A comment from the maintainer reads as a review ping / waiting-on-author. Record rebase/CI results in .agent/pr-sweep/SWEEP-*.md only. No gh pr comment, no review comments, no issue comments.

Procedure

1. Resolve repo + "us" set

bash
OWNER_REPO=$(gh repo view --json nameWithOwner --jq '.nameWithOwner')
OWNER=$(echo "$OWNER_REPO" | cut -d/ -f1)
REPO=$(echo "$OWNER_REPO" | cut -d/ -f2)
DEFAULT_BRANCH=$(gh repo view --json defaultBranchRef --jq '.defaultBranchRef.name')
DEFAULT_SHA=$(git rev-parse origin/$DEFAULT_BRANCH 2>/dev/null || gh api repos/$OWNER/$REPO/commits/$DEFAULT_BRANCH --jq .sha)

Build in-house logins (US):

  1. Org members (if owner is an org):
    bash
    gh api orgs/$OWNER/members --paginate --jq '.[].login' 2>/dev/null
  2. Repo collaborators with admin or maintain or push:
    bash
    gh api repos/$OWNER/$REPO/collaborators --paginate --jq '.[] | select(.permissions.admin or .permissions.maintain or .permissions.push) | .login'
  3. Logins in CODEOWNERS (and resolve teams when cheap).
  4. Bots: dependabot[bot], renovate[bot], github-actions[bot], copilot-swe-agent[bot], etc. → in-house.

Author is outside if login ∉ US.

2. List target PRs

bash
gh pr list --state open --limit 200 \
  --json number,title,url,author,isDraft,baseRefName,headRefName,headRepository,headRepositoryOwner,isCrossRepository,mergeable,mergeStateStatus,reviewDecision,statusCheckRollup,labels,additions,deletions,changedFiles,createdAt,updatedAt,assignees,headRefOid

(REST fallback if GraphQL 502s.)

Apply filters in order:

  1. Explicit numbers → use only those.
  2. Drop drafts from action set (mention under skipped).
  3. If --conflicts → keep only conflicting/dirty.
  4. If --behind → keep only behind/not up to date.
  5. If --changed → keep only snapshot-diffed changes (or 24h updatedAt if no snapshot).
  6. If --daily → union of new-outside ∪ changed ∪ behind ∪ conflicts ∪ waiting-approval (see Daily).
  7. Else → all open non-draft.

If count > 100, process 100 most-recently-updated; list the rest under "Skipped (over budget)".

3. Fan-out (read phase)

Spawn up to 100 parallel general-purpose subagents. One PR per agent.

Cheap path (daily / already-clean outside): if snapshot has security: clean and head SHA unchanged and only behind/conflicts flag flipped, skip full gh pr diff malware scan — re-fetch mergeability + checks only.

Each agent returns one JSON object only:

json
{
  "number": 123,
  "title": "...",
  "url": "https://github.com/...",
  "author": "login",
  "outside": true,
  "draft": false,
  "security": "clean|alert|review",
  "securityReasons": ["..."],
  "relevant": true,
  "relevanceReason": "<=120 chars",
  "behindBase": true,
  "mergeable": "MERGEABLE|CONFLICTING|UNKNOWN",
  "rebasePlan": "none|rebase|merge-from-base|blocked-conflicts|blocked-security|n/a-skip",
  "ci": {
    "overall": "passing|failing|pending|waiting-approval|none",
    "needsWorkflowApproval": false,
    "failedChecks": [],
    "pendingChecks": []
  },
  "assignForReview": true,
  "assignTo": ["login-or-team"],
  "priority": "P0|P1|P2|P3",
  "actionsPlanned": [
    "security-alert",
    "rebase",
    "push-force-with-lease",
    "approve-ci",
    "none"
  ],
  "blockers": "<=120 chars or empty",
  "summary": "<=160 chars"
}
Per-PR agent instructions (embed fully)
You are auditing GitHub PR <URL> in <OWNER/REPO> for pr-sweep.
Mode: read-only. Do not push, comment, approve, or merge.

1) Fetch:
   gh pr view <N> --json title,body,author,isDraft,baseRefName,headRefName,headRepositoryOwner,isCrossRepository,mergeable,mergeStateStatus,reviewDecision,statusCheckRollup,labels,files,additions,deletions,commits,url,assignees,headRefOid
   gh pr checks <N> 2>/dev/null || true
   # Full diff only if: outside AND (new OR security not yet clean in snapshot OR headSha changed)
   gh pr diff <N>   # when required

2) outside: true if author.login not in: <US_LOGINS_CSV>

3) Security:
   - Outside: always for new/changed head; use checklist references/security-checklist.md (next to this SKILL.md)
   - In-house: clean by default unless scripts/CI/lockfile/install lifecycle touched
   security: alert | review | clean

4) relevant + relevanceReason

5) behindBase + rebasePlan:
   - none | rebase | merge-from-base | blocked-conflicts | blocked-security | n/a-skip
   - In-house with --include-in-house: rebasePlan is rebase/merge-from-base/blocked-conflicts (not n/a-skip)
   - In-house without include: rebasePlan n/a-skip

6) CI: overall + needsWorkflowApproval + failed/pending checks

7) assignForReview, assignTo, priority, actionsPlanned

Return ONLY the JSON object on one line.

4. Aggregate dry-run report

Write .agent/pr-sweep/SWEEP-YYYY-MM-DD.md (create dirs). Structure:

markdown
# PR Sweep — <OWNER/REPO> — <date> — mode: dry-run|apply — scope: full|daily|behind|conflicts|changed

## Security alerts

## Needs human eyes

## Outside PRs — recommended actions

## In-house PRs — recommended actions # when --include-in-house

## Skipped

## Apply plan

Also update snapshot.json (even on dry-run) with current head SHAs and merge state.

Print a short chat summary: alert count, would-rebase count (outside / in-house), CI approvals, top assign list (max 5), report path.

If mode is dry-run, stop here (unless user then says apply).

5. Apply mode

--apply is consent. Write the Apply plan into the report and mutate in this turn. Do not wait for another yes.

Done for a rebase target = default-branch is an ancestor of the remote head SHA (push landed) or result: push-403 is recorded. Local-only rebase = failed apply.

Process PRs that are actionable:

AuthorSecurityFlagMutate?
outsidecleandefaultyes (rebase, approve-ci)
outsidealert/reviewanyno (report only)
in-houseclean--include-in-houseyes (rebase only; CI approve usually N/A)
in-house—no flagno
5a. Security gate

If security != "clean" → skip mutations.

5b. Rebase / update base (worktree subagent)

spawn_subagent with isolation: "worktree", max 8 concurrent. Embed the push + verify steps in the child prompt (completion criterion: remote SHA changed).

bash
BEFORE=$(gh pr view <N> --json headRefOid --jq .headRefOid)
gh pr checkout <N>
git fetch origin <DEFAULT_BRANCH>
git rebase origin/<DEFAULT_BRANCH>
# only if agent said merge-from-base:
# git merge origin/<DEFAULT_BRANCH>
# conflicts: resolve only clear non-overlapping/import/lockfile/generated cases.
# else: git rebase --abort; result=blocked-conflicts; do not push.

git push --force-with-lease
# Forks: push to the upstream `gh pr checkout` set (often not origin).
# If no upstream: git push --force-with-lease <fork-remote> HEAD:<headRefName>

AFTER=$(gh pr view <N> --json headRefOid --jq .headRefOid)
# MUST: AFTER != BEFORE. If equal, the push did not land — not done.

Org-fork 403 (maintainerCanModify false): record push-403, do not retry loops. Still never git push --force.

Apply child prompt (embed fully)
You are applying pr-sweep to GitHub PR https://github.com/OWNER/REPO/pull/N.
Mode: APPLY in a worktree. Never merge. Never git push --force (lease only). Never comment on the PR.

DONE only when the PR's remote head SHA changed, or you return result=push-403 or blocked-conflicts.
A local rebase with no push is a FAILED apply. Do not stop after rebase.

1. BEFORE=$(gh pr view N --json headRefOid --jq .headRefOid)
2. gh pr checkout N
3. git fetch origin DEFAULT_BRANCH
4. git rebase origin/DEFAULT_BRANCH
   Conflicts: resolve only trivial non-overlapping import/lockfile/generated cases.
   Else git rebase --abort and return blocked-conflicts (no push).
5. git push --force-with-lease
   Forks: push the upstream gh pr checkout configured (often not origin).
   No upstream: git push --force-with-lease <fork-remote> HEAD:<headRefName>
6. AFTER=$(gh pr view N --json headRefOid --jq .headRefOid)
   If AFTER == BEFORE and rebase was not already-current: push did not land — not done.

Return ONLY JSON:
{"number":N,"pushed":true|false,"newHeadSha":"...","result":"rebased-pushed|already-current|blocked-conflicts|push-403|error","blockers":"","summary":""}
5c. Approve waiting workflows (outside, relevant, clean)
bash
gh api -X POST repos/$OWNER/$REPO/actions/runs/<RUN_ID>/approve

If 403/404 → note manual approval needed. Do not re-run failing CI unless asked.

6. Final report + assignment recommendations

Update the markdown report with Results: mutated, still blocked, assign-for-review table.

Do not gh pr edit --add-assignee unless the user said "assign them".

Offer once (opt-in): publish report as secret gist:

bash
gh gist create .agent/pr-sweep/SWEEP-YYYY-MM-DD.md --desc "PR sweep — <OWNER/REPO> — <date>"

Secret is the default — never --public unless asked.

Chat closer: ≤5 lines — alerts, actions taken, top PRs to review, report path, gist URL if created.

Orchestrator tips

  • Prefer one read fan-out, then a smaller apply fan-out only for real mutations.
  • Daily first: load snapshot → filter → only then fan out. Avoid 60-agent full scans every day.
  • In-house rebases share the same force-with-lease rules; branches on the origin repo push to origin.
  • Org-fork push 403s (maintainerCanModify false / org policy): report partial; do not retry loops.
  • Cost: <20 PRs → main thread fine; no need for 100 agents.
  • Does not replace triage-github (backlog ranking) or pr-babysit (ongoing CI/comment fixing).

Reference

Bundled files

The model reads these on demand while the skill is loaded. They are exposed as readable files and are never executed.

Frequently asked questions

What does the Pr Sweep AI skill do?

Sweep open (or listed) PRs with up to 100 parallel agents: security-scan outside contributors, rebase onto main when behind (push --force-with-lease), approve pending first-time-contributor CI when relevant, optionally rebase in-house PRs, and report who should review. Supports full, changed-only, behind-only, and conflict-only scopes for cheap daily runs. Use when the user runs /pr-sweep (or /pr-inbound-sweep), or asks to "sweep PRs", "sweep inbound PRs", "security-check outside PRs", "rebase outsider PRs", "rebase our PRs", "approve waiting CI on PRs", "daily PR sweep", or "prep external...

Why use Pr Sweep on TypingMind?

Because you install it once and use it with any model. Pr Sweep is plain Markdown rather than provider-specific code, so the same skill runs on GPT-5, Claude, Gemini, Grok, or a local model — and you can switch model mid-chat without it breaking. TypingMind runs on your own API keys, so you pay providers directly instead of a per-seat subscription, and your skills and chats stay in your own storage.

How do I install Pr Sweep in TypingMind?

Open Plugins → Skills → Install from GitHub in TypingMind and paste https://github.com/TanStack/ai/tree/main/.agents/skills/pr-sweep. TypingMind reads its SKILL.md and bundles its files and installs it as a skill you can enable per chat.

Which AI models can use Pr Sweep?

Any model you connect in TypingMind. AI skills are plain Markdown instructions rather than provider-specific code, so GPT, Claude, Gemini, Grok, and local models can all load this skill when a request matches it.

How many AI models can I use with Pr Sweep?

As many as you like. As long as a model supports skills, you can use Pr Sweep with it — GPT, Claude, Gemini, Grok, DeepSeek, Mistral, Llama and more — all on TypingMind with your own API keys.

Is the Pr Sweep AI skill free?

Yes. It is published on GitHub by TanStack under the MIT license. You only pay your own AI provider for the tokens you use.

What are AI skills?

An AI skill is a reusable instruction bundle that teaches an AI model how to do one specific task. It follows the open Agent Skills format: a SKILL.md file with a name and description, plus any scripts, templates or reference files the model may need. The model reads the instructions only when your request matches the skill, so an installed skill costs nothing until it is used.

How are AI skills different from plugins or MCP servers?

A plugin or MCP server gives a model new tools to call — code that runs somewhere and returns a result. An AI skill gives the model knowledge and process instead: how to approach a task, which steps to follow, what good output looks like. Skills are plain Markdown, so they need no server, no API key and no runtime, and they work with any model.

View all

Set up your own AI workspace now

Get notified about new features and future giveaways by subscribing to our newsletter 👇