OrodruinLabs/nazgul:patch

Nazgul Patch

Examples

• /nazgul:patch "fix CORS headers on API" — Quick patch with single reviewer
• /nazgul:patch --no-review "update README badges" — Skip review, just implement
• /nazgul:patch --discuss "add rate limiting" — Discuss gray areas first, then patch
• /nazgul:patch — Interactive mode, prompts for description

Arguments

$ARGUMENTS

Current State

• Config: !cat nazgul/config.json 2>/dev/null | head -3 || echo "NOT_INITIALIZED"
• Existing patches: !ls nazgul/tasks/patches/PATCH-*.md 2>/dev/null | wc -l || echo "0"
• Active reviewers: !jq -r '.agents.reviewers // [] | join(", ")' nazgul/config.json 2>/dev/null || echo "none"

Instructions

Pre-flight

1. Check if nazgul/config.json exists. If not: "Nazgul not initialized. Run /nazgul:init first." and STOP.
2. Parse $ARGUMENTS for flags:
   • --no-review → skip review step
   • --discuss → run gray area discussion before planning
   • Remaining text → patch description
3. If no description after parsing, ask interactively: "What do you want to patch?"

Display Banner

Output per references/ui-brand.md:

─── ◈ NAZGUL ▸ PATCHING ────────────────────────────────

Step 1: Create Patch Task

1. Scan nazgul/tasks/patches/ for next available PATCH-NNN number (create directory if needed)
2. Create nazgul/tasks/patches/PATCH-NNN.md:

# PATCH-NNN: [description]

## Metadata
- **Status**: IN_PROGRESS
- **Created**: [ISO 8601 timestamp]
- **Source**: /nazgul:patch
- **Flags**: [--no-review, --discuss, or none]

## Description
[full description from user]

## Subtasks
[Generated by planner in Step 3]

## Implementation Log
[Filled during execution]

Step 2: Discuss (if --discuss flag)

Only if --discuss was passed:

1. Analyze the patch description for gray areas based on domain:
   • Code changes → scope, backwards compatibility, edge cases
   • Config changes → environments affected, rollback plan
   • Bug fixes → root cause, regression test needed
2. Ask the user about each gray area (one at a time)
3. Lock decisions into the patch task manifest under ## Decisions

Step 3: Plan

Use the planner agent to generate 1-3 subtasks for the patch:

• Dispatch planner with a simplified prompt: "Generate 1-3 small subtasks for this patch. No research, no docs. Just implementation steps."
• Planner reads the patch description (and decisions if --discuss was used)
• Subtasks are written into the patch manifest under ## Subtasks

Step 4: Implement

Use the implementer agent to execute each subtask:

• Dispatch implementer with the patch manifest
• Implementer follows existing patterns (reads CLAUDE.md, follows conventions)
• Each subtask gets an atomic git commit with prefix: patch(PATCH-NNN):
• Update patch manifest with implementation log

Step 5: Review (unless --no-review)

If --no-review was NOT passed:

1. Auto-select a single reviewer based on file types changed:
   • .ts/.tsx/.js/.jsx → frontend-reviewer (if available) or code-reviewer
   • .py → code-reviewer
   • .md/.json/.yaml → skip review (docs/config only)
   • Mixed → code-reviewer
2. Run the selected reviewer on the changed files
3. If APPROVED: mark patch DONE
4. If CHANGES_REQUESTED: show feedback, implementer fixes, re-review (max 2 retries)

If --no-review was passed:

• Mark patch DONE immediately after implementation

Step 6: Complete

1. Update patch manifest status to DONE
2. Append to nazgul/plan.md under ## Patches section (create section if it doesn't exist):

   - [x] PATCH-NNN: [description] (sha: [commit])
   

3. Display completion with Next Up block per ui-brand.md:

   ─── ◈ NAZGUL ▸ COMPLETE ✦ ──────────────────────────────
   
   ✦ PATCH-NNN: [description]
     Commits: [N] | Files changed: [N] | Review: [approved/skipped]
   
   ─── ◈ NEXT ─────────────────────────────────────────────
     /nazgul:start to continue main objective
     /nazgul:patch for another quick fix
   ────────────────────────────────────────────────────────
   

Error Handling

• If planner fails to generate subtasks: "Could not plan this patch. Try providing more detail or use /nazgul:start for a full task."
• If implementer fails: Mark patch BLOCKED with error details
• If reviewer rejects after max retries: Mark patch BLOCKED, show consolidated feedback