Github Process Agent Issues

Use this skill to automate the manual review-and-fix cadence for GitHub issues tagged Agent in all git repositories discovered under the current working tree.

Scope and triggers

• Use for requests involving:
  • "work through Agent issues"
  • "sweep issues labeled Agent"
  • "claim next Agent issue"
  • "submit PR for solved issue"
  • local repositories under the current folder or projects/<slug>/ subfolders

Shortcut: $github-process-agent-issues

Required companion skills

• $branch
• $agent-browser-relay
• $commit-push
• $pr
• $shell-glob-safe-paths when paths include [], (), *, ?, or spaces.

Preconditions

1. Confirm GitHub CLI authentication with gh auth status.
2. Confirm the active workspace contains cloned repos with remotes.
3. Confirm these companion skills are available: $branch, $agent-browser-relay, $commit-push, $pr.
4. Run with network access and write permission for these repos.
5. Resolve portable Python launcher once before using helper scripts:
   • PYTHON_BIN="$(command -v python3 || command -v python || true)"; [ -n "$PYTHON_BIN" ] || { echo "No Python interpreter found" >&2; exit 1; }
   • Never invoke raw python ....

Core workflow

1. Find open Agent issues across all repos.
2. Process issues one at a time.
3. For each issue:
   • Inspect and claim the issue using gh commands.
   • Create a dedicated branch for the issue using $branch.
   • Implement the fix on that branch.
   • Run all required checks/tests (at minimum local lint/tests/build where available).
   • If any UI/frontend verification is needed, verify via $agent-browser-relay.
   • Commit and push using $commit-push.
   • Create/update PR using $pr.
   • Update issue and PR labels/comments with gh and stop if confidence is low.
4. Repeat until queue is empty.

Mandatory behavior

• All GitHub operations must use the GitHub CLI (gh) or this skill's helper scripts that call gh under the hood.
• Use gh to:
  • list and inspect issues
  • add/remove issue labels
  • post issue comments and status updates
  • check and manage PR state
• Always follow a branch-first approach: do not solve an issue without creating or checking out a dedicated branch first.
• Only progress to the next issue after all required checks pass and PR is created/updated for the current issue.
• Do not bypass human confirmation on sensitive changes.
• Quote any literal path containing shell-glob characters before running shell commands:
  • sed -n '1,120p' 'app/dashboard/[[...path]]/page.tsx'
  • sed -n '1,120p' 'app/research/[slug]/page.tsx'

Script entry points

The skill ships with:

• scripts/agent_issue_sweep.py

Use --help on any command for full flags and defaults.

1) Discover Agent issues

"$PYTHON_BIN" "scripts/agent_issue_sweep.py" discover \
  --root "." \
  --label "Agent"

This command scans every repo under --root and outputs issue rows grouped by repository.

2) Claim one issue for work context

"$PYTHON_BIN" "scripts/agent_issue_sweep.py" claim \
  --root "." \
  --issue owner/repo#123 \
  --base-branch "main"

Use this only to discover the issue and set working context. The workflow for this skill must still explicitly create/confirm the dedicated issue branch with $branch.

3) (Legacy helper for finalize) Finish one issue

"$PYTHON_BIN" "scripts/agent_issue_sweep.py" finalize \
  --root "." \
  --issue owner/repo#123 \
  --result approved \
  --run-tests

This helper should be treated as a compatibility fallback. Primary required path is:

1. branch via $branch
2. verify and test
3. push via $commit-push
4. PR via $pr

4) Required label and comment flow (via gh)

• Add/remove labels with gh issue edit and gh pr edit.
• Post final status on issue and PR with gh issue comment and gh pr comment.
• Apply labels:
  • agent/in-progress while working
  • agent/resolved for approved outcomes
  • agent/blocking for blocked/uncertain outcomes

Mandatory certainty gate

• Default behavior requires full confidence (100) before finalizing.
• If confidence is below 100, pause and do not finalize the PR/issue; request explicit confirmation before continuing.

Human handoff and in-depth checks

• You are still responsible for implementation quality in the branch.
• For human-verification gates (for example CAPTCHA), stop immediately, alert the user, and wait for explicit human confirmation before continuing.