pr-operations

The pr-operations skill provides a standardized, automated interface for interacting with GitHub Pull Request review threads and discussion comments. Designed for software engineers and maintainers, this skill ensures that PR feedback loops remain clean, organized, and technically accurate by enforcing the use of erk exec commands over raw API calls. By centralizing operations, it prevents common pitfalls such as leaving threads unresolved or duplicating feedback across review iterations.

This skill is intended for automated agents or developers who need to process large volumes of code review feedback, batch resolve discussion threads, or perform inline code commenting. It includes an intelligent classification model that categorizes feedback by complexity—ranging from local file fixes to complex, cross-cutting architectural changes—to determine whether an action should proceed automatically or require human verification. It also implements a sophisticated deduplication algorithm for inline comments to maintain high signal-to-noise ratios in PR history.

• Perform atomic resolve-and-reply operations for code review threads using erk exec resolve-review-thread and resolve-review-threads.

• Fetch structured data on unresolved threads and PR discussion comments for analysis and reporting.

• Classify feedback into four complexity levels: local fix, multi-location, cross-cutting, and related, to optimize batch processing.

• Automatically post new inline review comments while applying a proximity-based deduplication logic to prevent redundant feedback.

• Distinguish between simple PR discussion replies and formal code review thread resolution to ensure GitHub PR status remains accurate.

• Always prioritize erk exec commands; raw gh api calls are prohibited for thread resolution because they fail to mark threads as resolved.

• Batch processing is encouraged for efficiency; provide input as JSON via stdin for resolve-review-threads.

• Utilize the batch confirmation flow for complex or cross-cutting changes (Batch 3-4) to ensure safety and code quality.

• When posting inline comments, the agent respects a 2-line tolerance window to ensure feedback is mapped correctly to evolving codebases.

• Inputs typically involve PR metadata and comment bodies, while outputs include structured API interaction logs and confirmation of resolved state.