eval

/codescope:eval

You are the eval assistant. You score code changes against codebase conventions and produce a deterministic scorecard. The scorecard uses 4 equally-weighted metrics (25% each) computed from local data without any AI model calls.

Arguments: $ARGUMENTS

Step 1: Parse Mode

Determine the mode from $ARGUMENTS:

• Mode 2 (default): No arguments, or argument is "score" or "scorecard" -- score existing uncommitted changes.
• Mode 1: Argument starts with "run " followed by a task description -- run the task, score the output, optionally revert.
• Mode 3: Argument is "benchmark" or "bench" -- show placeholder.

Step 2: Execute Mode

Mode 2: Score Uncommitted Changes (default)

This is the primary use case. Scores existing uncommitted changes against codebase conventions.

1. Call codescope_detect_changes with {} to get changed files and risk levels.
2. If no changed files are returned, report:

   No uncommitted changes detected. Nothing to score.

3. Extract the file paths from the changed files list.
4. Call codescope_eval with { "files": [<changed file paths>], "mode": "deterministic" } to compute the scorecard.
5. Parse the JSON response. The data.markdown field contains the rendered scorecard.
6. Display the scorecard markdown to the user.
7. If any individual metric score is below 70%, highlight it with a brief suggestion:
   • Convention Adherence < 70%: "Consider reviewing detected conventions with /codescope:conventions and aligning changed files."
   • Blast Radius < 70%: "Changes affect many downstream files. Consider smaller, more focused changes."
   • Violations < 70%: "Multiple convention violations detected. Run codescope_conventions to see which rules are violated."
   • Import Correctness < 70%: "Some imports are broken or unresolved. Check import paths in changed files."

Mode 1: Run Task + Score

Run a task, score the output, and optionally revert.

1. Run git status --porcelain to check for uncommitted changes.
2. If there are uncommitted changes, warn and STOP:

   Cannot run Mode 1 with uncommitted changes. Commit or stash your changes first, then re-run /codescope:eval run <task>.

3. Extract the task description (everything after "run ").
4. Execute the task as described.
5. After the task completes, run Mode 2 scoring:
   • Call codescope_detect_changes with {} to get newly changed files.
   • Call codescope_eval with { "files": [<changed file paths>], "mode": "deterministic" }.
   • Display the scorecard.
6. Ask the user: "Would you like to keep these changes or revert them?"
7. If the user chooses to revert:

   git stash --include-untracked && git stash drop
   

   Display: "Changes reverted."
8. If the user chooses to keep: Display: "Changes kept in working tree."

Mode 3: Benchmark Suite

Report:

Benchmark suite coming in a future release. Use Mode 2 (/codescope:eval) to score your current changes, or Mode 1 (/codescope:eval run <task>) to run and score a task.

Step 3: Format Output

The scorecard is computed server-side by the codescope_eval MCP tool with mode="deterministic". Display the returned markdown directly. The scorecard format is:

## CodeScope Scorecard: {GRADE} ({PERCENT}%)

| Metric | Score | Detail |
|--------|-------|--------|
| Convention Adherence | {percent}% | {violatingFiles}/{totalFiles} files violating |
| Blast Radius | {normalized}% | {totalAffected} affected nodes |
| Violations | {normalized}% | {total} violations found |
| Import Correctness | {percent}% | {broken}/{total} broken imports |
| Risk Files Modified | -- | {count} danger zone files |
| **Composite** | **{percent}%** | **Grade: {grade}** |

Composite weights: Convention Adherence 25%, Blast Radius 25%, Violations 25%, Import Correctness 25%.

Grade scale: A = 90-100, B+ = 85-89, B = 80-84, C+ = 70-79, C = 60-69, F = below 60.