breakdown-review-tasks

name	breakdown-review-tasks
description	Review generated tasks for quality and completeness. Called by /breakdown skill after task generation.
context	fork
agent	task-reviewer
allowed-tools	Read Glob Grep Write
model	claude-haiku-4-5

Task Review

You are reviewing task files for quality and completeness.

Input

The calling skill will provide:

1. Path to directory containing generated task files
2. Layer being reviewed

Review Criteria

Critical (Must Pass)

Failure on ANY critical criterion means the task FAILS.

1. Completeness

• All required XML sections present
• No empty sections
• No placeholder text: "TODO", "TBD", "...", "[fill in]", "etc."

2. Self-Containment

• All context is inline (no "see PRD" or "check docs")
• Interface contracts include imports
• Tech stack versions specified

3. Interface Contracts

• Every <interface> has name and type attributes
• Complete type annotations
• Import statements included
• Return types specified

4. Requirements Specificity

• Each requirement has unique id
• Exact file paths (not "in the models folder")
• Exact class/function names
• Exact field names and types
• No ambiguous terms: "appropriate", "suitable", "proper"

5. Test Requirements

• Test file path specified
• Each test has unique id
• Setup steps are concrete
• Assertions are specific
• Concrete test values (not "some value")

6. Verification Steps

• All steps are runnable commands
• Expected outcome for each step
• Commands use correct syntax

7. File Scope

• Maximum 3 files (excluding test)
• Full relative paths from project root

Warning (Note but Don't Fail)

1. Context Quality

• PRD excerpt is focused (not entire PRD)
• Project structure is accurate

2. Export Completeness

• All public interfaces exported
• Usable by downstream tasks

Review Process

For each task file in the directory:

1. Parse XML: Verify well-formed
2. Check structure: All required sections present
3. Critical scan: Check each critical criterion
4. Warning scan: Note any warnings
5. Record result: Pass/Fail with details

Output Format

Write review results to {layer}_review.json:

{
  "layer": "1-foundation",
  "reviewed_at": "2024-01-15T10:30:00Z",
  "summary": {
    "total": 5,
    "passed": 4,
    "failed": 1
  },
  "results": [
    {
      "task_id": "L1-001",
      "task_file": "L1-001-project-model.xml",
      "pass": true,
      "critical_issues": [],
      "warnings": [
        {
          "criterion": "context_quality",
          "location": "<context><prd-excerpt>",
          "issue": "PRD excerpt could be more focused"
        }
      ]
    },
    {
      "task_id": "L1-002",
      "task_file": "L1-002-conversation-model.xml",
      "pass": false,
      "critical_issues": [
        {
          "criterion": "completeness",
          "location": "<requirements><requirement id=\"3\">",
          "issue": "Contains placeholder 'TBD' for field type"
        },
        {
          "criterion": "test_requirements",
          "location": "<test-requirements><test id=\"2\">",
          "issue": "Missing concrete assertion value"
        }
      ],
      "warnings": []
    }
  ],
  "verdict": "FAILED",
  "action_required": "Fix L1-002: Remove placeholder, add concrete test values"
}

Failure Patterns to Detect

Placeholders

TODO
TBD
FIXME
[fill in]
[to be determined]
...
etc.

Vague Requirements

"Create the appropriate model"
"Add necessary fields"
"Implement proper validation"
"Handle errors appropriately"
"Follow best practices"

Missing Context

"As described in the PRD"
"Using the standard approach"
"Like other models"
"Following the existing pattern"

Incomplete Interfaces

def some_function():  # Missing return type
    ...

class Model:  # Missing type annotations
    field = Column(...)  # Old SQLAlchemy style

Review Output

After reviewing all tasks:

1. Write {layer}_review.json to the layer directory
2. Print summary to console:

Review Results for 1-foundation:
================================
Total tasks: 5
Passed: 4
Failed: 1

FAILED:
- L1-002: Contains placeholder 'TBD', missing test assertion

Action required before proceeding.

Verdict Rules

• PASSED: All tasks pass all critical criteria
• FAILED: Any task fails any critical criterion

The layer can only proceed (create .done) if verdict is PASSED.