Skip to content

fix(review-workflow): prevent context overflow in Expert Code Review workflow #8199

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Evangelink merged 2 commits into from
May 14, 2026
+4 −7
Merged

fix(review-workflow): prevent context overflow in Expert Code Review workflow #8199

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
19ece3a
Initial plan
Copilot May 13, 2026
7cf8214
Fix review-shared.md: prevent primary agent from reading background s…
Copilot May 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 4 additions & 7 deletions .github/workflows/shared/review-shared.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,10 +39,7 @@ Review pull request #${{ github.event.pull_request.number || github.event.issue.
## Instructions

1. Fetch the full diff for the pull request.
2. Call the `expert-reviewer` agent. Make sure to call it as subagent (`task` tool, `agent_type: "general-purpose"`, `model: "claude-opus-4.6"`). And make sure to follow the guidance on subagent calls from within the `expert-reviewer` agent. We expect 2+ levels of agents to be called.
3. Do **not** post comments or reviews yourself, except for the fallback in step 4 if the subagent posts nothing. The subagent will post its own comments using the available safe-output tools:
- **Inline review comments** on specific diff lines via `create_pull_request_review_comment`
- **Design-level concerns** (not tied to a line) via `add_comment`
- **Final review verdict** (COMMENT or REQUEST_CHANGES) via `submit_pull_request_review`
- **Never use APPROVE** — the agent must not count as a PR approval. Use COMMENT for clean reviews.
4. If the subagent does not post anything (e.g. no issues found), this is the only exception to step 3: post a brief fallback review using `submit_pull_request_review` with event `COMMENT` (not `APPROVE`). Do not use `add_comment` for this fallback.
2. Call the `expert-reviewer` agent as a **background** task (`task` tool, `agent_type: "general-purpose"`, `model: "claude-opus-4.6"`, `mode: "background"`). Include the PR number, repository owner/name, and the full diff content in the subagent prompt.
3. **Immediately after launching the background task** — do NOT wait for it to finish and do NOT read its result — call `noop` with a brief status message such as `"Expert-reviewer launched in background for PR #N. It will post the review directly."`. Then stop. The subagent has direct access to the safe-output tools and will post its own review (`create_pull_request_review_comment`, `add_comment`, `submit_pull_request_review`) without any further action from you.

> **Important**: Reading the background agent result would pull its entire conversation (2+ million tokens from spawning 21 dimension sub-agents) into your context, causing a server error. Do not call `read_agent` or any equivalent after calling `noop`.
Loading