mine-implementation-review

name	mine-implementation-review
description	Use when mine-orchestrate Phase 3 runs the post-execution quality gate. Reviews changed files against design doc and task files using a Sonnet subagent.
user-invocable	false

Implementation Review

Post-execution quality gate. After /mine-orchestrate finishes, this reviews the full implementation against the original design doc and task files. Uses a Sonnet subagent for structured review across 7 categories.

Arguments

$ARGUMENTS — path to a feature directory (design/specs/NNN-feature/) or a design.md file. If empty, find the most recently modified design/specs/*/design.md and confirm before proceeding.

---

Phase 1: Read the Design Doc, Task Files, and Changed Files

Locate the feature directory

If $ARGUMENTS points to a design/specs/NNN-*/ directory, use it directly.

If $ARGUMENTS points to a design.md file, the feature directory is one level up.

If $ARGUMENTS is empty:

Glob: design/specs/*/design.md

Sort by modification time, take the most recent. The feature directory is one level up. Confirm:

AskUserQuestion:
  question: "Found feature at <feature_dir>. Review its implementation?"
  header: "Confirm feature"
  multiSelect: false
  options:
    - label: "Yes — review it"
    - label: "No — let me specify the path"
      description: "Tell me the correct feature directory and I'll use that"

Read design doc and task files

Read <feature_dir>/design.md in full.

Read all <feature_dir>/tasks/T*.md files in order. If no task files exist, proceed with design-only review and note this in the summary.

Collect changed files

Run git-branch-diff-files to find which files were changed since the branch diverged from its base. This handles base-branch detection automatically (closest remote branch, then default branch fallback):

git-branch-diff-files

If still empty (e.g., no commits yet), fall back to:

git diff --name-only HEAD~1

Read each changed file. If the list is large (more than 15 files), prioritize files referenced in the task Prompt sections first.

---

Phase 2: Dispatch Reviewer Subagent

Review output path

Run get-skill-tmpdir mine-impl-review and use <dir>/review.md for the review output.

Read reviewer prompt

Read ${CLAUDE_CONFIG_DIR:-~/.claude}/skills/mine-implementation-review/reviewer-prompt.md.

Launch review subagent

Launch a general-purpose subagent with model: sonnet. Pass this prompt (fill in bracketed values):

You are reviewing a completed caliper v2 feature implementation.

## Design doc
<full design.md content>

## Task files
<full content of each T*.md in order, separated by "--- T<NN> ---" headers>

## Changed files
<for each changed file: filename header + full content>

## Your instructions
<full reviewer-prompt.md content>

Write your complete structured review to: <temp file path>

The subagent will write the review to the temp file.

---

Phase 3: Present Findings

Read the temp file. Format the results clearly:

1. Checklist results — one line per item: N. <name>: PASS|WARN|FAIL — note
2. Verdict — PASS, FAIL, or ABANDON (bold, prominent)
3. Summary — 2-3 sentences from the subagent
4. Blocking issues — if verdict is FAIL or ABANDON
5. Suggestions — non-blocking notes, if any

Phase 3 is the final output. The caller (mine-orchestrate) handles all gate logic, status updates, and next-step decisions. This skill does not prompt the user or update design.md.