Resolve CodeRabbit Plugin¶

Name: resolve-coderabbit

Description: Walk through CodeRabbit inline PR comments, verify each against the current code, and batch-resolve with commits and replies

Author: Flop (flopspm@gmail.com)

Version: 0.1.1

Keywords: review, coderabbit, github, pr, pull-request, bot-comments

The Resolve CodeRabbit plugin gives Claude Code a structured, safe workflow for handling CodeRabbit's automated review comments on GitHub pull requests. It's language-agnostic — works with Node, Python, Rust, Go, Java, Ruby, or anything else — as long as the project has a unit test command and uses the GitHub CLI (gh).

Installation¶

claude plugin install resolve-coderabbit@flugins

Important: After installing the plugin, restart Claude Code to activate skills.

Tip: Enable auto-update via /plugin → Installed → select the plugin → enable auto-update.

Requirements¶

gh CLI installed and authenticated against the repo's GitHub host
jq installed (used to parse the GitHub API responses)
git installed and the current directory inside a git working tree
The PR must already be opened and the branch pushed (the skill pushes new commits, but the PR itself must exist)
CodeRabbit (coderabbitai) must have posted inline review comments on the PR
A clean working tree on the PR branch — uncommitted unrelated changes will collide with one-commit-per-comment

The skill's first step is an automated dependency self-check (skills/resolve-coderabbit/scripts/self-check.sh) that verifies the first three bullets and walks you through skills/resolve-coderabbit/docs/setup-dependencies.md if anything is missing — so there's no guesswork about what to install.

Features¶

Skills¶

Resolve CodeRabbit Comments - Walk through unresolved CodeRabbit inline comments one by one, verify, fix or reject with user approval, then batch-push and resolve threads

Usage¶

Mention CodeRabbit PR comments in natural language (e.g. "resolve the CodeRabbit comments on PR #42", "address the bot review") and the skill activates automatically to walk through every unresolved CodeRabbit inline comment on the PR. For each thread it extracts the bot's claim, verifies it against the current code, and asks you to pick FIX / REJECT / SKIP — applying fixes as one commit per comment with a full unit-test gate, then batching the push, SHA-referenced replies, and thread resolves into a single final step. You can also trigger the same flow explicitly via /resolve-coderabbit:resolve-coderabbit [pr_number] from the / menu (omit the number to use the PR attached to the current branch).

Configuration¶

No plugin-specific configuration knobs — the skill is self-configuring. On first run it executes a bundled dependency self-check (scripts/self-check.sh) that verifies git, gh, jq, a valid gh auth session, and that you're inside a git working tree. If anything is missing or misconfigured, the skill pauses and walks you through the exact install/auth steps for your OS (from docs/setup-dependencies.md) before touching the PR — so setup is a guided conversation rather than a manual checklist. See Requirements above for the underlying tool list.

Resolve CodeRabbit Comments¶

Skill: resolve-coderabbit Type: Model-invoked (automatic) / user-invocable

Walks through every unresolved CodeRabbit inline comment on a PR, verifies the bot's claim against the current state of the code, applies a fix (or rejects with justification) only after explicit user approval, commits each fix locally, and then — in a single batched final step — validates the whole batch, pushes, posts SHA-referenced replies, and resolves the review threads.

How it Activates¶

The skill activates when you reference CodeRabbit or bot PR comments. Examples that will trigger it:

"resolve the CodeRabbit comments on PR #42"
"go through the bot's suggestions"
"address the PR review"
"walk through the comments"
"fix the CodeRabbit nits"

You can also invoke it explicitly via the Claude Code skill UI.

Workflow¶

Dependency self-check. Runs the bundled skills/resolve-coderabbit/scripts/self-check.sh first, which verifies git, gh, jq, a valid gh auth session, and that the current directory is inside a git working tree. If anything fails, the skill stops and walks you through skills/resolve-coderabbit/docs/setup-dependencies.md for your OS before continuing.
Resolve PR context. Reads the PR number from arguments, or falls back to the PR attached to the current branch (gh pr view --json number --jq .number), or asks you.
Pull the comment set via the bundled skills/resolve-coderabbit/scripts/fetch-comments.sh. A single GraphQL reviewThreads query returns every thread along with the first comment's body, path, line, and ID; the script filters to unresolved threads authored by coderabbitai and prints a JSON array ready to iterate over.
Seeds a task list — one task per unresolved comment with a severity marker and file:line.
Loops over each comment — for every one:
Extracts severity, the bolded claim, the proposed fix diff, and the 🤖 Prompt for AI Agents block.
Reads the real file at the claimed line and checks if the bot's claim still holds.
Presents the decision via AskUserQuestion — native radio UI with a preview pane showing the code + bot's diff. Batches by 4 (the tool's per-call cap). Options per comment: FIX / REJECT / SKIP.
Falls back to a plain-text y/n/s block in headless sessions.
Waits for your explicit choice before touching anything.
Applies the decision.
FIX → edit → run the project's full unit test suite (auto-detected from CLAUDE.md, README, package.json, Makefile, pyproject.toml, Cargo.toml, etc.) → one commit per fix → queue the reply (don't post yet).
REJECT → reply + resolve in one call via the bundled skills/resolve-coderabbit/scripts/resolve-comment.sh helper. No git changes.
SKIP → leave the thread open, move on.
Batched final step.
Runs the full unit suite on HEAD once more (per-commit runs skip docs-only fixes, so this catches combined-batch regressions).
Runs E2E/integration suite only if the batch touches integration boundaries (HTTP, auth, external APIs, DBs).
Shows you the queued commits and queued replies, asks for push confirmation.
git push once.
Replays all queued replies by calling resolve-comment.sh once per item (REST reply + GraphQL thread-resolve in one shot).
Summary — reports fixed / rejected / skipped counts and which threads remain open.

Why Batched Replies¶

Replies to inline comments reference commit SHAs. If replies were posted before the push, their SHAs wouldn't be on origin yet — confusing for reviewers and fragile against force-pushes. One-commit-per-fix keeps each reply's SHA meaningful, and batching the push + reply + resolve at the end means no broken SHA ever reaches origin attached to a reply.

Commit Message Convention¶

Each FIX commit uses the project's own convention (Conventional Commits or gitmoji — the skill checks recent git log) with a body explaining the CodeRabbit feedback:

fix(scope): short imperative summary

Per CodeRabbit PR review on #42: one or two sentences on what and why,
referencing the file/line that was off.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

What the Skill Will NOT Do¶

Never pushes without explicit user confirmation
Never commits if unit tests fail
Never posts replies against unpushed SHAs
Never mass-accepts suggestions — every comment requires per-comment approval
Never touches threads from non-CodeRabbit reviewers
Never handles merge conflicts or branch management (use the git plugin for that)

When to Use Something Else¶

No CodeRabbit on the PR — use the standard human review workflow, not this skill.
PR not opened yet / branch not pushed — open the PR first.
Bulk dispute — if you want to dismiss most of CodeRabbit's findings without per-comment discussion, handle that conversationally, don't loop through the skill.
Dirty working tree with unrelated changes — commit or stash first; one-commit-per-comment can't coexist with in-progress work.

See the SKILL.md reference for the full behavior specification and the exact gh api calls used.