# Pragma Review

> Fetch and analyse Pragma's AI review for a GitLab MR. Summarises high-value findings and critiques the review quality. Usage: /pragma-review [mr_number]

- Skill: `ccronca/pragma-review` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ccronca/pragma-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ccronca/pragma-review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: ccronca (https://skillmd.com/u/ccronca)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/ccronca/pragma-review

---


# Pragma Review

Fetch and analyse Pragma's AI-generated review for a GitLab MR.

**Input:** `$ARGUMENTS` — optional MR number (integer). If omitted, detected from context.

## Instructions

1. **Determine the MR number:**
   - If `$ARGUMENTS` contains an integer, use it directly as the MR number.
   - Otherwise, run `glab mr view` and parse the `IID:` field from the output.
   - If both fail, ask the user: "Which MR number would you like to review?"

2. **Ensure the working tree reflects the MR source branch:**
   - Run `glab mr view <MR_NUMBER>` and parse the `Source Branch:` field → `<mr_branch>`.
   - Run `git rev-parse --abbrev-ref HEAD` → `<current_branch>`.
   - If `<current_branch>` equals `<mr_branch>`, continue to step 3.
   - Otherwise:
     - Record `<current_branch>` so it can be restored later.
     - Inform the user: "Switching to `<mr_branch>` to analyse MR code accurately."
     - Run `glab mr checkout <MR_NUMBER>` to switch to the MR branch.
     - Set a flag `BRANCH_SWITCHED=true`.

3. **Determine the repository name:**
   - Run `glab mr view <MR_NUMBER>` (reuse output from step 2) and parse the `Project:` or
     `Web URL` field to extract the repository name (e.g. `pdm-db`).
   - Use just the short repository name (last path segment), not the full group path.

4. **Find the most recent Pragma review for this MR:**
   - Call `mcp__pragma__list_reviews` with `repository` set to the repo name from step 3.
   - Filter returned entries where the filename contains `_mr<MR_NUMBER>_`.
   - If multiple matches exist, select the one with the latest timestamp (the timestamp
     appears in the filename as `YYYYMMDD_HHMMSS`).
   - If no matching review is found, output:
     > "No Pragma review found for MR !<number> in repository <repo>."
     restore the original branch (step 8) and stop.

5. **Fetch the full review content:**
   - Call `mcp__pragma__get_review` with the filename selected in step 4.

6. **Output Section B — High-value findings:**

   Present a prioritised list of findings from the Pragma review that meet ALL of:
   - Blocking, critical, or high-severity (explicitly stated in Pragma's output, or
     clearly impactful based on context)
   - Non-obvious (not something a linter, type-checker, or compiler would catch)
   - Actionable (has a specific file/line reference or a concrete recommendation)

   Format:
   ```
   ## High-Value Findings (Pragma MR !<number>)

   1. [CRITICAL/HIGH] <finding summary> — `<file>:<line>`
      Rationale: <one sentence on why this is worth acting on>

   2. ...
   ```

   If no findings meet the criteria, state: "No high-value findings identified."

7. **Output Section C — Meta-critique:**

   Assess the quality of the Pragma review itself across four dimensions:

   ```
   ## Pragma Review Quality Assessment

   **Coverage:** <Did it address the key areas of the diff? What did it focus on?>
   **Accuracy:** <Are the findings grounded in the actual code? Any hallucinations?>
   **False positives:** <List any findings that appear incorrect or unsubstantiated>
   **Gaps:** <Important aspects of the diff that Pragma missed entirely>
   ```

8. **Restore original branch (if switched):**
   - If `BRANCH_SWITCHED=true`, run `git checkout <current_branch>`.
   - Inform the user: "Restored branch to `<current_branch>`."

## Notes

- Be concise. Section B should rarely exceed 5 items.
- Section C should be factual, not speculative — only flag gaps you can identify from
  the diff or codebase context.
- Do not repeat the full Pragma review verbatim.