I built MRSF because Markdown review workflows don’t persist well across edits.
Inline comments clutter documents. GitHub PR comments disappear once merged. And AI agents don’t have a structured way to attach feedback to Markdown that survives refactors.
MRSF is a JSON/YAML sidecar format for Markdown reviews. It includes:
- Anchored comments (line + span + selected text)
- Deterministic re-anchoring after edits
- A JSON Schema for validation
- A CLI (validate, reanchor, status)
- An MCP server so LLM agents can read/write review comments programmatically
The core idea is simple: keep review metadata outside the Markdown file, but robust enough that comments survive document evolution.
I’d especially appreciate feedback on:
- The re-anchoring strategy
- Whether this overlaps too much with LSP diagnostics
- Whether this would be useful outside AI/agent workflows
- Any obvious flaws in the schema design
I was thinking about something like that a lot. The anchoring is the hard part. One direction i was thinking in: store the commit sha within the anchor. this could allow "reanchor" based on changes in some sense. For get-colibri.com (project i am working on) I am currently considering inline html comments ai <-> human communication.
Thanks for the feedback. Yea, anchoring is tough and right now it's content based rather than commit-diff based. I've been thinking about adding an additional strategy for commit-aware anchoring, for instance using the CLI "mrsf reanchor --strategy git-diff" or something.
Inline comments clutter documents. GitHub PR comments disappear once merged. And AI agents don’t have a structured way to attach feedback to Markdown that survives refactors.
MRSF is a JSON/YAML sidecar format for Markdown reviews. It includes:
- Anchored comments (line + span + selected text) - Deterministic re-anchoring after edits - A JSON Schema for validation - A CLI (validate, reanchor, status) - An MCP server so LLM agents can read/write review comments programmatically
The core idea is simple: keep review metadata outside the Markdown file, but robust enough that comments survive document evolution.
I’d especially appreciate feedback on:
- The re-anchoring strategy - Whether this overlaps too much with LSP diagnostics - Whether this would be useful outside AI/agent workflows - Any obvious flaws in the schema design
Repo: https://github.com/wictorwilen/MRSF/