Skip to content

Add file reading guardrails to prevent review agent hallucination loops #77

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
derekmisler merged 1 commit into from
Mar 10, 2026
Merged

Add file reading guardrails to prevent review agent hallucination loops #77

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
20 changes: 19 additions & 1 deletion review-pr/agents/pr-review.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,24 @@ agents:

Find **real bugs in the changed code**, not style issues. If the changed code works correctly, approve it.

## CRITICAL: File Reading Guardrails

The root agent MUST NOT exhaustively explore the repository. Follow these rules strictly:

1. **Only read files that are directly relevant**: the diff, AGENTS.md/CLAUDE.md, and files
Copy link
Copy Markdown

@krissetto krissetto Mar 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why agents/claude.md?

Copy link
Copy Markdown
Contributor Author

@derekmisler derekmisler Mar 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just following the trends I've noticed. These are the files people tend to add the most, but that prompt-files flag appends the contents of the prompt files to every request, which seems like a waste of tokens. So I'm trying out just telling the agents to read these files once.

explicitly referenced in the diff (e.g., imported modules, configuration files mentioned
in changed code). Do NOT speculatively read files to "understand the project."
2. **Never guess file paths**: If you need to check whether a file exists, use `list_directory`
first. Do NOT try `read_file` on paths you are guessing — this wastes time and tokens.
3. **Circuit breaker**: If 3 consecutive `read_file` calls return "not found", STOP reading
files immediately and proceed with what you have. The drafter and verifier have their own
`read_file` access and will read source files as needed during analysis.
4. **Cap total file reads**: The root agent should read at most 10 files total (excluding the
diff itself). The drafter and verifier handle deeper file analysis.
5. **Never enumerate topics as file paths**: Do NOT try to read files named after general
concepts (e.g., `consensus.md`, `raft.md`, `six-sigma.md`). Only read files that
appear in the diff, the project tree, or are referenced by other files you've already read.

## Delivering the Review

You MUST always deliver a review, even if no issues were found.
Expand Down Expand Up @@ -200,7 +218,7 @@ agents:

toolsets:
- type: filesystem
tools: [read_file, read_multiple_files, list_directory, directory_tree]
tools: [read_file, read_multiple_files, list_directory]
- type: shell
- type: memory
path: .cache/pr-review-memory.db
Expand Down
Loading