fix(mcp): cap search_graph default limit to 200 (was 500K)

[amitmynd]

Problem

search_graph with a broad file_pattern (e.g. **/ads/**) returns output that exceeds LLM tool result limits. In my case, querying a 12K-node project returned 3,358,396 characters, which Claude Code couldn't process:

search_graph (MCP)(project: "Users-amitbhalla-Downloads-myndintegrated", file_pattern: "**/ads/**")

Error: result (3,58,396 characters) exceeds maximum allowed tokens.

Root cause

The default limit for the regex/file_pattern search path uses MCP_HALF_SEC_US (500,000), which appears to be a microsecond timeout constant that was accidentally used as a row count. When no explicit limit is passed, the query returns up to 500K rows, serializing all of them into the JSON response.

• src/mcp/mcp.c:1344: cbm_mcp_get_int_arg(args, "limit", MCP_HALF_SEC_US) (500,000)
• src/store/store.c:2304: fallback to ST_HALF_SEC (also 500,000)
• Tool description says "Default: unlimited"

Fix

Changed the default limit to 200 in both the MCP handler and the store layer. Updated the tool description to reflect this.

Before: 1,287 matching nodes, all returned, 3.3M characters output
After: 1,287 total (reported in total field), 200 returned, 57K characters output, has_more: true

Pagination with offset and limit continues to work as expected for callers that need more results.

Changes

File	Change
src/mcp/mcp.c	Default limit 500,000 -> 200 in handle_search_graph
src/store/store.c	Fallback limit ST_HALF_SEC -> 200 in cbm_store_search
src/mcp/mcp.c	Tool description: "Default: unlimited" -> "Default: 200"

Test plan

• Build succeeds with make -f Makefile.cbm cbm
• search_graph with broad file_pattern returns 200 results (57K chars) instead of all matches (3.3M chars)
• Response includes total: 1287, has_more: true for pagination
• list_projects and --help work correctly
• Explicit limit param still overrides the default

🤖 Generated with Claude Code