From 24d2bb2c84a2dbb7adc165243533fc72621647d5 Mon Sep 17 00:00:00 2001 From: Rene Cannao Date: Sat, 17 Jan 2026 14:52:46 +0000 Subject: [PATCH] fix: Enforce MCP catalog usage and prohibit Write tool for agent findings Add prominent warnings and explicit instructions to prevent agents from creating individual markdown files instead of using the MCP catalog. Changes: - Add "CRITICAL: MCP CATALOG USAGE" section at the top of the prompt - Explicitly prohibit Write tool for individual agent discoveries - Specify that ONLY Round 4 final report should be written to file - Add detailed catalog_upsert usage examples - Update all round descriptions with CRITICAL catalog usage instructions - Add "IMPORTANT - Catalog Usage Rules" section Key Instructions Added: - "DO NOT use the Write tool to create separate markdown files" - "ALL agent findings MUST be written to MCP catalog using catalog_upsert" - "ONLY Round 4 Final Synthesis writes to local file using Write tool" - "DO NOT use Write tool for individual agent discoveries in Rounds 1-3" This ensures agents store findings in the MCP catalog for cross-agent collaboration, with only the final consolidated report written to disk. --- .../prompts/multi_agent_discovery_prompt.md | 85 ++++++++++++++++--- 1 file changed, 71 insertions(+), 14 deletions(-) diff --git a/scripts/mcp/DiscoveryAgent/ClaudeCode_Headless/prompts/multi_agent_discovery_prompt.md b/scripts/mcp/DiscoveryAgent/ClaudeCode_Headless/prompts/multi_agent_discovery_prompt.md index 2314be55a..961c8ed91 100644 --- a/scripts/mcp/DiscoveryAgent/ClaudeCode_Headless/prompts/multi_agent_discovery_prompt.md +++ b/scripts/mcp/DiscoveryAgent/ClaudeCode_Headless/prompts/multi_agent_discovery_prompt.md @@ -3,6 +3,36 @@ ## Mission Perform comprehensive database discovery through 6 collaborating subagents using ONLY MCP server tools (`mcp__proxysql-stdio__*`). Output: Single comprehensive markdown report. +## ⚠️ CRITICAL: MCP CATALOG USAGE + +**ALL agent findings MUST be stored in the MCP catalog using `catalog_upsert`.** + +**DO NOT use the Write tool to create separate markdown files for individual agent discoveries.** + +- Round 1-3 findings: Use `catalog_upsert` ONLY +- Round 4 final report: Use both `catalog_upsert` AND Write tool (for the single consolidated report) +- Round 5 meta analysis: Use `catalog_upsert` ONLY + +**WRONG:** Using Write tool for each agent's findings creates multiple markdown files +**RIGHT:** All findings go to MCP catalog, only final report is written to file + +Example correct usage: +```python +# After discovery, write to catalog +catalog_upsert( + kind="structural", # or statistical, semantic, query, security, meta_analysis, question_catalog + key="round1_discovery", + document="## Findings in markdown..." +) +``` + +Only in Round 4 Final Synthesis: +```python +# Write the consolidated report to catalog AND file +catalog_upsert(kind="final_report", key="comprehensive_database_discovery_report", document="...") +Write("database_discovery_report.md", content="...") +``` + ## Agent Roles | Agent | Focus | Key Tools | @@ -19,27 +49,43 @@ Perform comprehensive database discovery through 6 collaborating subagents using ### Round 1: Blind Exploration (Parallel) - Launch all 5 analysis agents simultaneously (STRUCTURAL, STATISTICAL, SEMANTIC, QUERY, SECURITY) - Each explores independently using their tools -- Write findings to catalog: `kind="structural|statistical|semantic|query|security"`, `key="round1_*"` +- **CRITICAL:** Write findings to MCP catalog using `catalog_upsert`: + - Use `kind="structural"`, `key="round1_discovery"` for STRUCTURAL + - Use `kind="statistical"`, `key="round1_discovery"` for STATISTICAL + - Use `kind="semantic"`, `key="round1_discovery"` for SEMANTIC + - Use `kind="query"`, `key="round1_discovery"` for QUERY + - Use `kind="security"`, `key="round1_discovery"` for SECURITY +- **DO NOT** use Write tool to create separate files - META agent does NOT participate in this round ### Round 2: Collaborative Analysis - All 5 analysis agents read each other's findings via `catalog_search` - Identify cross-cutting patterns and anomalies -- Write collaborative findings: `kind="collaborative_round2"` +- **CRITICAL:** Write collaborative findings to MCP catalog using `catalog_upsert`: + - Use `kind="collaborative_round2"` with appropriate keys +- **DO NOT** use Write tool to create separate files - META agent does NOT participate in this round ### Round 3: Hypothesis Testing - Each of the 5 analysis agents validates 3-4 specific hypotheses - Document: hypothesis, test method, result (PASS/FAIL), evidence -- Write: `kind="validation_round3"` +- **CRITICAL:** Write validation results to MCP catalog using `catalog_upsert`: + - Use `kind="validation_round3"` with keys like `round3_{agent}_validation` +- **DO NOT** use Write tool to create separate files - META agent does NOT participate in this round ### Round 4: Final Synthesis - All 5 analysis agents collaborate to synthesize findings into comprehensive report - Each agent ALSO generates their QUESTION CATALOG (see below) -- Write: `kind="final_report"`, `key="comprehensive_database_discovery_report"` -- Write: `kind="question_catalog"`, `key="{agent}_questions"` for each agent -- Also create local file: `database_discovery_report.md` +- **CRITICAL:** Write the following to MCP catalog using `catalog_upsert`: + - `kind="final_report"`, `key="comprehensive_database_discovery_report"` - the main report + - `kind="question_catalog"`, `key="structural_questions"` - STRUCTURAL questions + - `kind="question_catalog"`, `key="statistical_questions"` - STATISTICAL questions + - `kind="question_catalog"`, `key="semantic_questions"` - SEMANTIC questions + - `kind="question_catalog"`, `key="query_questions"` - QUERY questions + - `kind="question_catalog"`, `key="security_questions"` - SECURITY questions +- **ONLY FOR THE FINAL REPORT:** Use Write tool to create local file: `database_discovery_report.md` +- **DO NOT** use Write tool for individual agent findings or question catalogs - META agent does NOT participate in this round ### Round 5: Meta Analysis (META Agent Only) @@ -48,8 +94,10 @@ Perform comprehensive database discovery through 6 collaborating subagents using - Reads all question catalogs and synthesizes cross-domain questions - Identifies gaps, missed opportunities, or areas for improvement - Suggests specific prompt improvements for future discovery runs -- Write: `kind="meta_analysis"`, `key="prompt_improvement_suggestions"` -- Write: `kind="question_catalog"`, `key="cross_domain_questions"` +- **CRITICAL:** Write to MCP catalog using `catalog_upsert`: + - `kind="meta_analysis"`, `key="prompt_improvement_suggestions"` - meta analysis + - `kind="question_catalog"`, `key="cross_domain_questions"` - cross-domain questions +- **DO NOT** use Write tool - meta analysis stays in catalog only ## Report Structure (Required) @@ -505,12 +553,21 @@ TodoWrite([ ## Critical Constraints 1. **MCP-ONLY**: Use `mcp__proxysql-stdio__*` tools exclusively -2. **EVIDENCE-BASED**: All claims backed by database evidence -3. **SPECIFIC RECOMMENDATIONS**: Provide exact SQL for all changes -4. **QUANTIFIED IMPACT**: Include expected improvements with numbers -5. **PRIORITIZED**: Always prioritize (URGENT → HIGH → MODERATE → LOW) -6. **CONSTRUCTIVE META**: META agent provides actionable, specific improvements -7. **QUESTION CATALOGS**: Each agent MUST generate a question catalog with executable answer plans +2. **CATALOG FOR FINDINGS**: ALL agent findings MUST be written to MCP catalog using `catalog_upsert` - NEVER use Write tool for individual agent discoveries +3. **NO INTERMEDIATE FILES**: DO NOT create separate markdown files for each agent's findings - only the final synthesis should be written to a local file +4. **EVIDENCE-BASED**: All claims backed by database evidence +5. **SPECIFIC RECOMMENDATIONS**: Provide exact SQL for all changes +6. **QUANTIFIED IMPACT**: Include expected improvements with numbers +7. **PRIORITIZED**: Always prioritize (URGENT → HIGH → MODERATE → LOW) +8. **CONSTRUCTIVE META**: META agent provides actionable, specific improvements +9. **QUESTION CATALOGS**: Each agent MUST generate a question catalog with executable answer plans + +**IMPORTANT - Catalog Usage Rules:** +- Use `catalog_upsert(kind="agent_type", key="specific_key", document="markdown")` for ALL findings +- Use `catalog_search(kind="agent_type", query="terms")` to READ other agents' findings +- Use `catalog_get(kind="agent_type", key="specific_key")` to retrieve specific findings +- ONLY Round 4 Final Synthesis writes to local file using Write tool +- DO NOT use Write tool for individual agent discoveries in Rounds 1-3 ## Output Locations