Process PR Reviews

This workflow supports both CodeRabbit and Codex PR review signals.

CodeRabbit Reviews

“Real review feedback” is strictly defined as:

• inline review comments
• not a review summary
• not a nitpick

Nitpick comments from CodeRabbit must always be ignored to avoid unnecessary noise.

There is no need to analyze the comment content itself.

Data sources

The CodeRabbit workflow only needs these two sources:

1. PR review comments

   bashCopy

   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments
   

   This is the authoritative source for real inline comments.

2. PR reviews

   bashCopy

   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews
   

   This is only used to identify and exclude review summaries / nitpick summaries. It is not used to extract the final result.

Do not treat these as primary sources:

• gh pr view ...
• gh api repos/<owner>/<repo>/issues/<pr_number>/comments

Reason: they are not the authoritative source for actionable inline comments.

Workflow

0. Optional: fetch review thread IDs early if resolve/dismiss may be needed

If the user may ask you to resolve review conversations after triaging them, fetch review thread IDs as soon as you know the PR number:

gh api graphql -f query='query { repository(owner: "<owner>", name: "<repo>") { pullRequest(number: <pr_number>) { reviewThreads(first: 100) { nodes { id isResolved comments(first: 20) { nodes { databaseId path line author { login } body } } } } } } }'

This is not a primary source for actionable review feedback. It is only for mapping inline comments to resolvable thread IDs.

Recommendation:

• If the user only wants to view/list/count review feedback, this step is optional.
• If the user may want to resolve conversations, doing this early is usually more convenient because you can map comment databaseId / path / line to thread IDs in one pass.

1. Fetch inline comments

gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments

Only keep records that satisfy all of the following:

• user.login is coderabbitai[bot] or coderabbitai
• in_reply_to_id == null (only top-level inline comments, not replies)

This is the candidate set.

2. Fetch reviews to exclude review summaries / nitpicks

gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews

Identify CodeRabbit review summaries. Common characteristics include:

• Actionable comments posted: N
• Nitpick comments
• long summary text

These review-level contents are not the final result. They are only used to help determine:

• which items are summaries
• which nitpicks should not be counted as actionable inline comments

Filtering rule

The final goal of the CodeRabbit workflow is always:

Top-level inline comments left by CodeRabbit in pulls/<pr_number>/comments that are neither nitpicks nor summaries

Important:

• Treat CodeRabbit nitpicks as non-actionable by default.
• Do not include nitpicks in counts, summaries, or resolution queues unless the user explicitly asks for nitpicks.

In practice, do the following:

1. Get CodeRabbit top-level inline comments from pulls/<pr_number>/comments
2. Use pulls/<pr_number>/reviews to determine whether the PR contains nitpick summaries
3. In the output, keep only the inline comments you confirm are actionable

Large output handling

If the output of gh api --paginate ... is too large and gets truncated:

1. Record the tool output file path
2. Do not manually read through the entire large JSON blob
3. Hand it off to @explorer to extract:
   • CodeRabbit-authored comments
   • the number of top-level inline comments
   • each comment’s path / line / body

Resolving review conversations

Resolution policy

When triaging review feedback, apply this rule:

• If a comment will not be fixed, you may resolve the conversation after triage.
• If a comment will be fixed, do not resolve it first — make the code change first, then resolve the conversation afterward.

In short:

• won't fix / no code change → triage, then resolve
• will fix / code change required → fix first, then resolve

If the user asks to resolve a CodeRabbit review conversation:

1. Identify the target inline comment from the actionable comment list.
2. Map that comment to its review thread ID via reviewThreads GraphQL data.
   • Match using databaseId when possible.
   • If needed, fall back to path + line + author login.
3. Resolve the thread with GraphQL:

gh api graphql -f query='mutation { resolveReviewThread(input: {threadId: "<thread_id>"}) { thread { isResolved } } }'

Notes:

• Resolve the thread, not the individual comment.
• pulls/<pr_number>/comments remains the source of truth for identifying actionable inline comments.
• reviewThreads is only for thread-level operations such as resolving conversations.

Codex Reviews

“Real review feedback” is strictly defined as:

• inline review comments
• not the review summary card

There is no need to analyze the comment content itself.

Important behavior differences from CodeRabbit

• Codex review is silent while running.
• Unlike CodeRabbit, Codex does not expose an in-progress PR check for review status.
• While Codex is reviewing, the PR main conversation thread gets an eyes reaction from chatgpt-codex-connector[bot].
• If Codex finds no issues, it may leave no actionable inline comments and instead react to the PR main conversation thread with +1 (thumbs up).
• If Codex finds issues, it may create inline review comments in pulls/<pr_number>/comments and a review summary card in pulls/<pr_number>/reviews.

Data sources

The Codex workflow uses these sources:

1. PR review comments

   bashCopy

   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments
   

   This is the authoritative source for actionable inline Codex comments.

2. PR reviews

   bashCopy

   gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews
   

   This is used to identify the Codex review summary card such as ### 💡 Codex Review. It is not the primary source of actionable inline feedback.

3. Issue comment reactions on the PR main thread

   First fetch the PR issue node / comments if needed:

   bashCopy

   gh api repos/<owner>/<repo>/issues/<pr_number>/comments
   gh api repos/<owner>/<repo>/issues/<pr_number>/reactions
   

   Use reactions on the PR main conversation thread only to detect Codex review state:

   • eyes from chatgpt-codex-connector[bot] → Codex review appears to be in progress
   • +1 from chatgpt-codex-connector[bot] on the PR main thread → Codex reviewed and found no issues

   These reactions are status signals, not actionable review feedback.

Do not treat these as primary sources for actionable comments:

• gh pr view ...
• gh api repos/<owner>/<repo>/issues/<pr_number>/comments
• gh api repos/<owner>/<repo>/issues/<pr_number>/reactions

Reason: actionable Codex review feedback still lives in PR review comments, not in the PR issue timeline.

Workflow

1. Fetch inline comments

gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/comments

Only keep records that satisfy all of the following:

• user.login is chatgpt-codex-connector[bot]
• in_reply_to_id == null (only top-level inline comments, not replies)

This is the candidate set of actionable Codex comments.

2. Fetch reviews to identify the summary card

gh api --paginate repos/<owner>/<repo>/pulls/<pr_number>/reviews

Identify Codex review summaries. Common characteristics include:

• ### 💡 Codex Review
• explanatory text such as Here are some automated review suggestions for this pull request.
• “About Codex in GitHub” help text

These review-level contents are not the final result. They are only used to understand whether Codex posted a review summary.

3. Optionally inspect PR main-thread reactions for status

If the user asks whether Codex is still reviewing, or whether Codex finished with no findings, inspect reactions on the PR main thread.

Interpret them as follows:

• eyes by chatgpt-codex-connector[bot] → likely still reviewing / review in progress
• +1 by chatgpt-codex-connector[bot] with no Codex inline comments → likely completed with no findings

Do not count these reactions as review comments.

Filtering rule

The final goal of the Codex workflow is always:

Top-level inline comments left by Codex in pulls/<pr_number>/comments

In practice, do the following:

1. Get Codex top-level inline comments from pulls/<pr_number>/comments
2. Use pulls/<pr_number>/reviews only to recognize the summary card
3. If there are no Codex inline comments, optionally inspect PR main-thread reactions to distinguish:
   • still reviewing (eyes)
   • reviewed with no findings (+1)
   • no observable Codex activity

Large output handling

If any gh api --paginate ... output is too large and gets truncated:

1. Record the tool output file path
2. Do not manually read through the entire large JSON blob
3. Hand it off to @explorer to extract:
   • Codex-authored inline comments
   • the number of top-level inline comments
   • each comment’s path / line / body