Add consistency-reviewer learning agent knowledge base

[nhorton]

Summary

This PR establishes the knowledge base and learning infrastructure for the consistency-reviewer agent, a specialized reviewer for stylistic consistency and agentic process coherence across the DeepWork codebase.

Key Changes

• Core Knowledge Base (.deepwork/learning-agents/consistency-reviewer/core-knowledge.md): Comprehensive reference covering:

  • Consistency principles and the three review domains (job definitions, Python code, agentic process coherence)
  • Job YAML conventions including field ordering, input/output patterns, and critical validation rules
  • Step instruction file structure and writing style guidelines
  • Python source code conventions (type hints, dataclasses, docstrings, error handling, logging, path handling)
  • Agentic process coherence checks (data flow integrity, dependency correctness, review coverage, prompt coherence, granularity, naming)
  • Decision frameworks for severity assessment and when to flag issues

• Topic Guides: Three focused reference documents for specific domains:

  • agentic-process-data-flow.md: DAG structure, dependency rules, workflow ordering, common data flow issues
  • job-yml-schema-and-validation.md: Schema location, field requirements, validation rules, common gotchas
  • step-instruction-quality.md: Section structure, writing style checklist, quality criteria alignment, anti-patterns

• Agent Registration (.claude/agents/consistency-reviewer.md): Agent manifest with description and references to core knowledge and topics

• Learning Infrastructure (.deepwork/learning-agents/consistency-reviewer/additional_learning_guidelines/):

  • issue_identification.md: Guidance on what issues matter most and what to ignore
  • issue_investigation.md: Common root causes and investigation heuristics
  • learning_from_issues.md: Preferences for how learnings are incorporated (topics vs. learnings, core-knowledge updates)
  • .gitkeep files for learnings and topics directories

Notable Implementation Details

• The core knowledge is structured around three distinct but related domains (job YAML, Python code, agentic processes) with explicit guidance on when conventions from one domain should not apply to another
• Severity assessment framework distinguishes between critical runtime failures, high-impact inconsistencies, medium pattern deviations, and low-level style issues
• Learning guidelines emphasize preferring topic updates over incident learnings for convention changes, helping the knowledge base stay current as the codebase evolves
• Comprehensive decision framework helps the agent distinguish between issues worth flagging and minor deviations that are acceptable

https://claude.ai/code/session_01K9TQPhkeDMzn3MSyTVYzr3