Bug: Server Return Type

[MartelleV]

The code-review-graph server is returning a prompt where messages[0] is a raw dict instead of a proper Message object (or plain string), which Claude Code's MCP client rejects.

I think the MCP server's prompt template for debug_issue is constructing messages like:

{"role": "user", "content": "..."}  # raw dict

Hence it returns:

McpError: MCP error 0: Error rendering prompt 'debug_issue': messages[0] must be Message or str, got dict. Use Message({'role': 'user', 'content': '## Rules for Token-Efficient Graph Usage\n1. ALWAYS call get_minimal_context first with a task description.\n2. Use detail_level="minimal" on all tool calls unless the minimal output is insufficient.\n3. Only escalate to detail_level="standard" or "verbose" for the specific entities that need deeper inspection.\n4. Never request more than 3 tool calls per turn unless absolutely necessary.\n5. Prefer targeted queries (query_graph with a specific symbol) over broad scans (list_communities with full members).\n6. When reviewing changes: detect_changes(detail_level="minimal") → only expand on high-risk items.\n\n## Debug Workflow\n1. Call get_minimal_context(task="debug: Error").\n2. Call semantic_search_nodes(query=<keywords from description>, detail_level="minimal", limit=5).\n3. For the top 1-2 results, call query_graph(pattern="callers_of", target=<name>, detail_level="minimal").\n4. If the issue involves execution flow: call get_flow(name=<relevant flow>) for the single most relevant flow.\n5. Only call get_review_context or get_impact_radius if you need to trace the blast radius of a specific change.'}) to wrap the value.

My prompt was to use /code-review-graph:debug_issue prompt with Claude Code v2.1.119 on macOS. All models of Claude (Sonnet, Opus, Haiku), regardless of effort, got this error. Code-review-graph is installed via pipx, v2.3.3.