Codex Plan Validation (via claude-delegator)

When to use

• complexity

  :

  complex

• taskType

  :

  feature

  or

  refactor

• context.md

  exists or was updated

Procedure

Step 1: Check MCP Availability (CRITICAL - Do This First)

Before any validation work, verify Codex MCP is available:

// Try a simple MCP call to check availability
try {
  mcp__codex__codex({
    prompt: "ping",
    sandbox: "read-only",
    cwd: process.cwd()
  })
  // If successful, MCP is available
} catch (error) {
  // MCP not available - proceed with Claude fallback
}

MCP Unavailable Conditions:

• Tool not found / not registered
• "quota exceeded", "rate limit", "API error", "unavailable"
• Connection timeout
• Any error response

Step 2-6: Validation Process

2. Collect the path to context.md (default:

   {tasksRoot}/{feature-name}/context.md

   ) and read its content

3. Build delegation prompt using the 7-section format below

4. If MCP is available (from Step 1):

   • Call

     mcp__codex__codex

     (include Plan Reviewer instructions in developer-instructions)
   • If successful, proceed to step 6

5. If MCP is unavailable (from Step 1):

   • Claude directly performs the plan review following the Plan Reviewer guidelines below
   • Add note:

     "codex-fallback: Claude performed review directly (MCP unavailable)"

   • Follow the same MUST DO / MUST NOT DO criteria

6. Summarize critical/warning/suggestion items and decide pass/fail

7. Per

   .claude/docs/guidelines/document-memory-policy.md

   : Store full review in

   archives/review-v{n}.md

   , keep only short summary in

   context.md

Delegation Format

Use the 7-section format:

TASK: Review implementation plan at [context.md path] for completeness and clarity.

EXPECTED OUTCOME: APPROVE/REJECT verdict with specific feedback.

CONTEXT:
- Plan to review: [content of context.md]
- Goals: [what the plan is trying to achieve]
- Constraints: [project constraints]

MUST DO:
- Evaluate all 4 criteria (Clarity, Verifiability, Completeness, Big Picture)
- Simulate actually doing the work to find gaps
- Provide specific improvements if rejecting

MUST NOT DO:
- Rubber-stamp without real analysis
- Provide vague feedback
- Approve plans with critical gaps

OUTPUT FORMAT:
[APPROVE / REJECT]
Justification: [explanation]
Summary: [4-criteria assessment]
[If REJECT: Top 3-5 improvements needed]

Tool Call (When MCP Available)

mcp__codex__codex({
  prompt: "[7-section delegation prompt with full context]",
  "developer-instructions": "[contents of plan-reviewer.md]",
  sandbox: "read-only",  // Advisory mode
  cwd: "[current working directory]"
})

Claude Fallback (When MCP Unavailable)

When MCP is not available, Claude performs the validation directly:

1. Apply the same 7-section format as a self-review checklist
2. Evaluate all 4 criteria:
   • Clarity: Are the goals and steps clearly defined?
   • Verifiability: Can success be measured objectively?
   • Completeness: Are all necessary steps included?
   • Big Picture: Does it align with overall architecture?
3. Output in the same format: APPROVE/REJECT with justification
4. Add note indicating fallback mode was used

Output (patch)

notes:
  - "codex-plan: [APPROVE/REJECT], warnings=[count]"
  # If fallback was used:
  - "codex-fallback: Claude performed review directly (MCP unavailable)"