prepare-review-input

name: prepare-review-input description: Normalize input (changes, folder, files, class, buffer) into structured review-input.json for review agents. argument-hint: "[changes|folder|file|files|buffer] [target]" allowed-tools: Read, Grep, Glob, Bash, Write, mcp__plugin_solid-coder_pipeline__get_output_path user-invocable: false

Prepare Review Input

Input

• SOURCE_TYPE: $ARGUMENTS[0] — one of: changes, folder, file, files, buffer
  • If $ARGUMENTS[0] is not a known type you must stop and show an error message. Unknown Type
  • If NO arguments provided → default to changes (staged + unstaged + untracked)
• OUTPUT_ROOT: $ARGUMENTS[1] — output root; if not provided call mcp__plugin_solid-coder_pipeline__get_output_path with {operation: "review"} to get the default
• CANDIDATE_TAGS: optional list of tags from the orchestrator (extracted from rule.md frontmatters)
• OUTPUT_PATH: {OUTPUT_ROOT}/prepare
• SKILL_ROOT: ${CLAUDE_PLUGIN_ROOT}/skills/prepare-review-input

Workflow

Phase 1: Input Detection & Validation

Create Preparation task list and execute it.

• 1.1 Create output folder — Create OUTPUT_PATH
• 1.2 Detect source type —
  • SOURCE_TYPE not recognized, fail with the message (don't continue):
  • NO arguments → changes (default: review current git changes)
  • is a directory path → folder
  • is a file path → file
  • contains newlines or is after explicit buffer keyword → buffer
• 1.3 Validate target — Confirm file/folder is accessible (skip for changes type)
• 1.4 Read output schema — Read output.schema.json from this skill's directory

Phase 2: Loading content

• 2.0.1 Load content
  • if mode is Changes then load content using Phase 2.1
  • if mode is Folder then load content using Phase 2.2
  • if mode is File then load content using Phase 2.3
  • if mode is Files then load content using Phase 2.5
  • if mode is Buffer then load content using Phase 2.4

Phase 2.1: Changes

• 2.1.1 Run prepare-changes script —

      python3 {SKILL_ROOT}/scripts/prepare-changes.py -o OUTPUT_PATH/review-input.json
  

  • Collects staged + unstaged + untracked changes, parses diffs, writes per-file changed_ranges.
• 2.1.2 Extract units + compute has_changes — Run:

      python3 {SKILL_ROOT}/scripts/extract-units.py OUTPUT_PATH/review-input.json
  

  • Updates the JSON in place: fills files[].units with {name, kind, line_start, line_end, has_changes} and recomputes summary.total_units / summary.changed_units.
  • has_changes is true if any changed_range overlaps the unit's line span, else false. For files with null/empty changed_ranges (untracked), all units are has_changes: true.
  • Do NOT do this work by hand / via inline Python. If extract-units fails, fix the script instead of bypassing it.
• 2.1.3 Match tags — Read OUTPUT_PATH/review-input.json, examine detected_imports + the source files, merge matched tags into matched_tags (see Tag Matching).
• 2.1.4 Write matched_tags back — Update matched_tags in the JSON.

Phase 2.2: Folder

• 2.2.1 Discover .swift files — Glob the folder recursively.
• 2.2.2 Write initial review-input.json — Build files[] with each path and changed_ranges: null. Set source_type: "folder". Then run extract-units.py OUTPUT_PATH/review-input.json to fill units (all has_changes: true).
• 2.2.3 Match tags — See Tag Matching.

Phase 2.3: File

• 2.3.1 Write initial review-input.json — files: [{file_path: <path>, changed_ranges: null, units: []}], source_type: "file". Run extract-units.py to fill units.
• 2.3.2 Match tags — See Tag Matching.

Phase 2.5: Files (explicit file list)

• 2.5.1 Parse file list — JSON array or space-separated paths.
• 2.5.2 Write initial review-input.json — One entry per file with changed_ranges: null. Set source_type: "files". Run extract-units.py.
• 2.5.3 Match tags — See Tag Matching.

Phase 2.4: Buffer

• 2.4.1 Write initial review-input.json — Put the input string in buffer.input, source_type: "buffer". Run extract-units.py to fill buffer.units.
• 2.4.2 Match tags — See Tag Matching.

Phase 3: Assembly & Output

• 3.1 Extract imports (for folder/file/buffer modes) — Scan the loaded source code for import <ModuleName> statements. Collect unique module names into detected_imports array (sorted). For changes mode, the script already handles this.

• 3.2 Write structured output - using defined schema write to OUTPUT_PATH/review-input.json (include detected_imports and matched_tags)

  If validation fails, read the error message, fix the JSON to match the schema, re-write, and re-validate.

Diff Parsing

Parse unified diff to extract changed line ranges per file:

1. Split by ^diff --git a/(.*) b/(.*)$ — use b/ path as file_path
2. Find hunk headers: ^@@ -\d+(?:,\d+)? \+(\d+)(?:,(\d+))? @@
   • Group 1 = new file start line
3. Walk hunk lines from start line:
   • + line (not +++): mark line as changed, increment line counter
   • - line (not ---): skip (deletion), do NOT increment
   • context line: increment line counter
4. Coalesce adjacent changed lines into ranges { "start": N, "end": M }

Unit Identification

Handled by scripts/extract-units.py. The script finds every top-level class, struct, protocol, enum, extension declaration in each Swift source file and emits {name, kind, line_start, line_end, has_changes}. line_end is the line before the next unit (or end-of-file for the last unit).

Do not re-implement this logic in the agent's prompt or via inline Bash heredocs — always call the script.

Tag Matching

If CANDIDATE_TAGS is provided (non-empty list), determine which tags apply to the code being reviewed. You already have the code loaded from Phase 2.

1. Start with detected_imports from the script output (raw import names like SwiftUI, UIKit, Combine)
2. For each CANDIDATE_TAG, check if the code matches it:
   • Check detected_imports — does any import name match the tag (case-insensitive)?
   • Check code patterns — e.g., tag gcd matches DispatchQueue usage, tag structured-concurrency matches async/await/Task { usage, tag combine matches Publisher/Subscriber usage
   • Use your judgment for semantic tags that don't map directly to imports
3. Merge all matched tags into the matched_tags array (lowercase, deduplicated)

If CANDIDATE_TAGS is empty or not provided, set matched_tags to an empty array.

Constraints

• Do NOT modify any source files
• Do NOT write buffer input to a separate file — keep it in the JSON
• Do NOT resolve any issues with the workflow, if anything fails:
  • file not found
  • will consume too many tokens > 200k You fail with a message