vedix-meta-analyst

Meta-Analyst

Cross-job learning extraction.

Inputs

• <input name="trajectories_jsonl"> — content of ~/.vedix/trajectories.jsonl
• <input name="jobs_json"> — content of ~/.vedix/jobs.json

Steps

1. Preferred: call mcp__vedix__run_meta_analysis() — does the work + writes outputs to ~/.vedix/meta_analysis.json and ~/.vedix/what_works.json. Returns summary.

2. Fallback (manual, if MCP call fails):

   • Compute success rate per domain
   • Compute avg manuscript words, papers found, fix attempts per domain
   • Identify common error types from fix logs
   • Extract reliable approaches (high success-rate patterns)
   • Build recommendations list

Output

<output name="meta_analysis_json">
{
  "total_jobs": 0,
  "successful_jobs": 0,
  "failed_jobs": 0,
  "avg_manuscript_words": 0,
  "avg_papers_found": 0,
  "domain_stats": {},
  "common_experiment_errors": [],
  "successful_experiment_patterns": [],
  "last_updated": "..."
}
</output>
<output name="what_works_json">
{
  "successful_patterns": {
    "statistical": {
      "experiment_success_rate": 0.0,
      "avg_manuscript_words": 0,
      "reliable_approaches": [],
      "common_failures": []
    }
  },
  "recommendations_for_next_job": [],
  "last_updated": "..."
}
</output>

Findings drawer contract

After generating meta-analysis, return a findings_update field in your JSON output:

{
  "meta_analysis_json": {...},
  "findings_update": {
    "current_understanding": "What we know so far in this project",
    "patterns_and_insights": "Recurring patterns across runs",
    "lessons_and_constraints": "What broke and what we won't try again",
    "open_questions": "Unresolved questions",
    "last_direction_decision": "DEEPEN | BROADEN | PIVOT | CONCLUDE — and why"
  }
}

The pipeline writes each section into the per-project palace under room research-findings.