gspdev-eval-changes

1. Change intent achieved? Did the PR/commit land what it set out to do?
2. Skill intent coherent? Does the skill body deliver on its description? (Catches drift between SKILL.md frontmatter description: and the body — a separate failure mode from the change-intent question.)
3. Substance preserved or improved? For modifications: did the change avoid sacrificing what the agent needs? For additions: does the new content fit the skill's existing contracts (constraints, output specs, quality bars)?

Examples of changes this covers:

• Adding a new step to a methodology
• Adding a new sibling reference file
• Modifying a rule, constraint, or rubric
• Changing the agent's role or execution mode
• Trimming verbose prose
• Extracting content to siblings
• Renaming inputs/outputs
• Adding a new SKILL.md (skill intent check from scratch)

A change that achieves its goal (Dim 1 = PASS) can still drift the skill away from its stated purpose (Dim 2 = FAIL) if it adds behavior the description doesn't reflect, or removes a step the description promised. A coherent intent + description (Dims 1-2 = PASS) can still break an output spec (Dim 6 = FAIL). All three gates matter.

Spawns one general-purpose evaluator agent per modified file in parallel. Each evaluator compares BEFORE (if any) and AFTER against the parent SKILL.md using a fixed 8-dimension rubric, returns PASS / CONCERNS / FAIL with quoted findings and surgical recommendations.

Different from siblings:

• /gspdev-audit — static checks across the whole repo (contracts, installer, version sync). Read-only, no eval.
• /gspdev-housekeeping — drift detection (count mismatches, stale references). Mechanical.
• /gspdev-prompt-audit — semantic analysis of all skill prompts (dead weight, contradictions). Whole-repo, slow.

This skill is targeted: it evaluates the quality of a specific change before it merges.

Input: Optional commit range (default: main..HEAD + uncommitted) Output: Aggregated verdict table + per-file findings; optional inline restoration of CONCERNS-flagged content Agent: general-purpose (one per evaluated file, dispatched in parallel)

Step 0: Resolve scope

Determine which files to evaluate:

1. If $ARGUMENTS contains a commit range (e.g., main..HEAD, HEAD~3..HEAD, feature-branch..main), use it directly
2. Otherwise default to main..HEAD plus uncommitted working-tree changes

RANGE="${ARGUMENTS:-main..HEAD}"

# Committed changes in range
COMMITTED=$(git diff --name-only "$RANGE" 2>/dev/null | grep -E '^(gsp|dev)/(skills|agents)/.*\.md$' || true)

# Uncommitted working-tree changes
UNCOMMITTED=$(git diff --name-only HEAD 2>/dev/null | grep -E '^(gsp|dev)/(skills|agents)/.*\.md$' || true)

# Combine, dedupe
FILES=$(printf "%s\n%s\n" "$COMMITTED" "$UNCOMMITTED" | sort -u | grep -v '^$')