Iterating PR Bot Fixes
Version 0.2.0. One standalone skill for taking an identified GitHub pull request through bounded review, scoped repairs, verified integration, and eligible task cleanup.
The skill tracks explicit outcomes from independent review and configured GitHub reviewers, investigates root causes across affected code, and keeps findings within the PR's intended scope. The default budget is ten review rounds. A source returning clean retires from further manual requests; a separate final independent check covers later changes without silently re-requesting retired providers.
Copilot's attributable cannot-run response can be recorded as explicit unavailable, with no review coverage from that attempt. This never substitutes
for independent review or overrides an explicit Copilot requirement. CI,
required approvals, branch protection, and merge queues remain separate gates.
CI triage follows the first meaningful failure at the actual tested revision, separates product defects from provider/environment and baseline failures, and bounds authorized retries. Revision refresh distinguishes head, test-merge, and merge-group evidence, handles late feedback, and resumes post-merge cleanup without repeating a verified merge.
Install and use
In a Codex client supporting marketplace plugins:
codex plugin marketplace add sealad886/sealad886-codex-marketplace --ref main
codex plugin add iterating-pr-bot-fixes@sealad886-codex-marketplace
Start a fresh task after installation. Other clients can load the portable
Agent Plugins 1.0 package using its root manifest and skills/ directory.
Example requests:
- "Use $iterating-pr-bot-fixes:iterating-pr-bot-fixes to drive this identified PR through scoped fixes, review, gated merge, and eligible cleanup."
- "Review and repair this PR, but do not push or merge."
- "Resume the PR's existing review ledger; verify current head before continuing."
- "Diagnose this PR's failed check before editing code; identify the tested SHA, failure cause, current merge gate, and whether a bounded rerun is authorized."
- "The base advanced and the queue rebuilt. Refresh current integration evidence while preserving retired clean reviewers."
- "This PR merged, but branch cleanup failed. Verify the merge and resume only the cleanup still authorized and safe."
An explicit full-lifecycle task supplies bounded authority for its necessary commits, ordinary pushes, PR/review actions, gated merge, and eligible branch cleanup. Review-only/no-push/no-merge instructions narrow it. Merely installing, reading, quoting, or automatically loading the skill grants no action. New worktrees require explicit authorization. Release/deployment, destructive Git recovery, settings changes, and new source-export destinations stay outside it.
Requirements and included references
Needs repository access, an available independent reviewer context, project validation tools, and supported authenticated GitHub capabilities for provider actions. GitHub CLI is optional; provider tools/API may be used. No MCP server, runtime dependency, mandatory code indexer, or companion plugin is included.
- Review sources and Copilot policy
- Pattern and scope analysis
- Selective commit safety
- CI diagnosis and bounded retries
- Revision, queue, and interrupted-task refresh
- Conditional integration and cleanup
- Resumable review ledger
Use current provider documentation and installed capabilities. A successful request, green CI, silence, or queue entry never proves review/merge completion. Reports distinguish exact reviewed SHAs, source unavailability, actual merged destination, and any preserved/blocked cleanup.
Maintainer validation
From a checkout of the marketplace repository, run its source checker and distribution checker:
python3 scripts/check_plugin.py plugins/iterating-pr-bot-fixes --layout source
python3 scripts/check_distribution_bundle.py plugins/iterating-pr-bot-fixes
Also validate JSON/YAML, contained links/artwork, exactly one skill, portable and Codex metadata parity, and fresh-client discovery. These checks validate package structure; actual provider requests, agent behavior, and desktop rendering need their own evidence. See security boundaries and MIT license.