fix: return connector tool errors as tool results instead of throwing

[Ashu (ashudhanda)]

Summary

Closes #427

Every func in createOpenWikiConnectorTools() currently lets errors propagate. A thrown tool error aborts the whole agent run, so the model never sees the error messages that were written for it — e.g. the openwiki_call_mcp_tool guidance to run openwiki_list_mcp_tools first and retry with an exact discovered tool name. A single hallucinated tool name or one expired connector token kills the entire wiki-build instead of letting the agent self-correct or continue with other connectors.

Change

Added a small wrapToolErrors helper in src/connectors/tools.ts and applied it to all seven connector tools via .map(wrapToolErrors) on the returned array, following the approach suggested (and verified with a local patch) in the issue:

function wrapToolErrors<
  T extends { func: (...args: any[]) => Promise<unknown> },
>(tool: T): T {
  const originalFunc = tool.func;
  tool.func = (async (...args: any[]) => {
    try {
      return await originalFunc(...args);
    } catch (error) {
      return `Tool error: ${
        error instanceof Error ? error.message : String(error)
      }`;
    }
  }) as T["func"];
  return tool;
}

Behavior

• Tool failures now return Tool error: <message> as the tool result, matching how LangChain's own ToolNode reports errors, so the agent can retry with corrected input or move on
• Bad tool name → model retries after discovery; expired token → model skips that connector and continues; the run completes with partial results instead of crashing
• No behavior change on the success path; internal helper throws (input validation, MCP-backed checks) are preserved and simply surface through the wrapper

Kept the scope minimal per the issue; making ingestAllConnectors collect per-connector failures individually could be a follow-up.