Spec validation before loop start

[adamdaum]

Summary

The loop starts executing immediately without verifying that the PRD, app_spec, and feature_list are well-formed and consistent. A malformed spec can burn through expensive AI iterations before anyone notices.

Current Behavior

• ralph.sh passes the coding prompt to Claude without validating inputs
• If feature_list.json has invalid JSON, missing fields, or no features, the loop may error mid-execution or produce nonsensical results
• No check that app_spec.txt exists or has the expected sections
• No check that referenced tech stack components are compatible

Proposal

Add a validation step to ralph.sh (or as a standalone validate.sh / CLI command) that runs before the loop:

Checks to implement:

1. feature_list.json

   • Valid JSON
   • All required fields present (id, priority, category, description, steps, passes)
   • No duplicate feature IDs
   • No gaps in priority ordering
   • depends_on references exist (if dependency resolution is implemented)

2. app_spec.txt

   • File exists and is non-empty
   • Contains expected sections (Tech Stack, Data Models, API Routes, etc.)

3. PRD.md

   • File exists
   • Contains expected sections

4. Cross-file consistency

   • Features in feature_list reference concepts described in app_spec
   • Tech stack in app_spec matches PRD requirements

Output:

Validating specs...
  ✓ feature_list.json: 10 features, valid schema
  ✓ app_spec.txt: all sections present
  ✓ PRD.md: exists, 2,400 words
  ⚠ Warning: feature API-003 references "Redis" but app_spec tech stack does not include it
  ✓ Validation passed with 1 warning

Considerations

• Validation should be non-blocking by default (warnings) with a --strict flag for hard failures
• Could be integrated into init.sh as well, running after spec generation
• Keep validation fast — it should add seconds, not minutes