| name | verify-config |
| description | Audits CLAUDE.md, REASONIX.md, DEEPSEEK.md, and .ai/ reference files against actual codebase patterns and hard requirements. Use to detect configuration drift and ensure documentation stays in sync. |
Verify Configuration Skill
Audits project documentation against actual codebase conventions and enforces hard requirements
(task execution protocol, subagent delegation, multi-assistant sync integrity, session management).
Steps
1. Read current documentation
- Read
CLAUDE.md
- Read
REASONIX.md (if exists)
- Read
DEEPSEEK.md (if exists)
- Read
.ai/reference/critical-rules.md
- Read
.ai/reference/task-execution.md
- Read
.ai/reference/session-management.md
- Read
.ai/reference/templates/session-handoff.md.txt
- Read
.ai/session-context.md
- Read
.ai/patterns/cqrs-patterns.md
- Read
.claude/settings.json
2. Codebase pattern audit (existing)
- Pick a representative domain from
{ApplicationName}.Domain.* as the reference implementation
- Read 2-3 command handlers (Create, Update, Delete)
- Read 2-3 query handlers (GetById, GetList)
- Read 2-3 Model records from
Models/
- Read 2-3 Response records
- Check
Extensions/ServiceCollectionExtensions.cs
- For each documented convention, verify it matches actual code
- Categorize: Correct / Drift / Missing / Stale
3. Hard requirement checks
3a. Task execution protocol
.ai/progress/ folder exists
.ai/completed/ folder exists
.ai/plans/ folder exists
CLAUDE.md Task Execution section explicitly says progress files are MANDATORY
REASONIX.md (if exists) Task Execution section explicitly says progress files are MANDATORY
DEEPSEEK.md (if exists) Task Execution section explicitly says progress files are MANDATORY
.ai/reference/task-execution.md is tool-agnostic (no EnterPlanMode / mv references)
3b. Subagent delegation
CLAUDE.md has a HARD RULE subagent delegation section naming what the main conversation may/may not do
REASONIX.md (if exists) has the same HARD RULE section
DEEPSEEK.md (if exists) has the same HARD RULE section
- Agent tables in all files include a Model column
- Model names use
deepseek-v4-pro/deepseek-v4-flash (not sonnet/haiku)
- Model mapping line exists:
deepseek-v4-pro ↔ sonnet · deepseek-v4-flash ↔ haiku
3c. Multi-assistant sync integrity
DEEPSEEK.md (if exists) exists alongside CLAUDE.md and REASONIX.md
.claude/skills/ entries are thin wrappers (contain !cat .ai/skills/.../SKILL.md``)
.github/skills/ entries are full content copies (do NOT contain !cat ...``)
.reasonix/skills/ (if exists) skill-derived entries are thin wrappers (reference .ai/skills/)
.reasonix/skills/ agent-derived entries have runAs: subagent and reference .ai/agents/
- No stale dirs in
.reasonix/skills/ (all have an .ai/ source)
.claude/settings.json has ./.reasonix in additionalDirectories
.claude/settings.json sync hooks mention .reasonix
.claude/settings.json has a DEEPSEEK.md sync hook triggering on .ai/reference/critical-rules.md, .ai/reference/task-execution.md, .ai/agents/**, and CLAUDE.md
- Sync scripts (
sync-skills.py, sync-agents.py) push to .reasonix/skills/
.gitignore includes CLAUDE.md.deepseek-backup
3d. Session management
.ai/reference/session-management.md has generic [assistant name] Written By (not hardcoded to one assistant)
.ai/reference/session-management.md lists all assistants that share the session context
.ai/reference/templates/session-handoff.md.txt includes all assistants in the pick-lists
.ai/session-context.md has no stale [Claude Code | GitHub Copilot]-only references
3e. DEEPSEEK.md content integrity
DEEPSEEK.md inlines the subagent delegation table (same 8 agent types as CLAUDE.md)
DEEPSEEK.md inlines the most critical coding rules as direct content (not cross-references to .ai/reference/critical-rules.md)
DEEPSEEK.md inlines the task execution protocol (plan → progress file → complete cycle)
DEEPSEEK.md is self-contained — a DeepSeek model can follow it without resolving nested file references
DEEPSEEK.md structure is flatter than CLAUDE.md (fewer levels of indirection, fewer cross-references)
DEEPSEEK.md is ~180-220 lines (concise enough for DeepSeek context but comprehensive)
DEEPSEEK.md references .ai/ folder only as an appendix for deep dives, not as required reading
DEEPSEEK.md acknowledges it runs inside Claude Code's runtime (tools available: Agent, Skill, Read, Write, Edit, Bash, Glob, Grep, EnterPlanMode, TaskCreate, etc.)
4. .ai/ folder structure
- Verify
.claude/settings.json has plansDirectory pointing to local .ai/plans
- Verify
.ai/progress/ and .ai/plans/ folders exist
- Check that no project-specific config leaked to global
~/.claude/ or ~/.reasonix/
5. Present findings
- Show a summary table of all checks with their status
- For each failure, show the specific gap and how to fix it
- Categorize:
- Correct — passes
- Missing — required file/section doesn't exist
- Stale — references wrong assistant names or model names
- Drift — content contradicts the canonical pattern
- Wait for user approval before making changes
Output Format
## Configuration Audit Report
### Summary
- Conventions checked: N
- Correct: N
- Failures: N
### Hard Requirements
| # | Check | Status | Details |
|---|-------|--------|---------|
| 1 | Progress files mandatory in CLAUDE.md | PASS/FAIL | ... |
| 2 | Subagent delegation HARD RULE in REASONIX.md | PASS/FAIL | ... |
| 3 | Agent table has Model column | PASS/FAIL | ... |
| 4 | Model names are deepseek-v4-pro/flash | PASS/FAIL | ... |
| 5 | Model mapping documented | PASS/FAIL | ... |
| 6 | .claude/settings.json includes ./.reasonix | PASS/FAIL | ... |
| 7 | .reasonix/skills/ thin wrappers | PASS/FAIL | ... |
| 8 | .reasonix/skills/ agent skills runAs: subagent | PASS/FAIL | ... |
| 9 | Session management is assistant-agnostic | PASS/FAIL | ... |
| 10 | Handoff template includes all assistants | PASS/FAIL | ... |
| ... | ... | ... | ... |
### Codebase Pattern Audit
| # | Convention | Status | Details |
|---|-----------|--------|---------|
| 1 | Data access style | Correct/Drift/Missing/Stale | ... |
### Recommended Fixes
1. ...
2. ...