Skip to content

Warning: blindly using code-review-graph can make model faster but noticeably worse at code understanding #314

@cxc1357

Description

@cxc1357

The current injected instructions push the agent to always use code-review-graph first.

In practice, this improves speed and reduces token usage, but it also makes the model over-trust graph summaries and miss important implementation details in source code. The result is shallower reasoning, more template-like edits, and weaker understanding of edge cases, recovery logic, compatibility code, and tests.

It may be helpful to append guardrails like this to the injected instructions:

Guardrails

  • Use graph tools first to narrow scope, not to replace source reading.
  • Do not make code changes based only on graph summaries.
  • For any non-trivial change, read the actual implementation files and relevant tests before concluding.
  • When changing behavior, database logic, migrations, retries, fallbacks, recovery, or compatibility code, always verify the exact source.
  • If graph and source disagree, trust the source code.
  • In final answers, explain which files/functions/tests were actually read.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions