Claude Code API 400 after a tool returned an object: Sume results

Update Claude Code to 2.1.286 or later. Its changelog lists a fix for API 400 errors after a tool or hook returned an object, number or boolean instead of text. Sume's hosted MCP tools were not the trigger: the server wraps every result, success or error, in a text content block.

The Claude Code line is from its changelog, read 2026-10-01. The Sume side is from the MCP server source and the MCP tools and gates page.

What did Claude Code change in 2.1.286?

The entry says the 400s followed a tool or hook returning a non-text value, and that the fix covers resumed sessions too. A hook you wrote yourself that prints a bare number is the likelier source than an MCP server.

What shape does a Sume tool result have?

In packages/mcp-server/src/mcp-tool-results.ts, toolResult serializes the redacted payload with JSON.stringify and returns { content: [{ type: "text", text }], isError: false }. The client sees one text block holding JSON, never a raw object. Failures go through toolErrorResult, which returns the same block type with isError: true.

What happens when a result is too large?

The output budget is MCP_OUTPUT_MAX_BYTES = 256 * 1024. Past it, the server returns a coded mcp_output_too_large error whose message says it is an output limit, not a job failure, and never to resubmit a paid create. Re-read with a narrower call instead; see job timeouts.

If I still see a 400, what should I check?

Check your own hooks and any other MCP servers first, and confirm claude --version is at least 2.1.286. To see what Sume returned, call tools_list or mcp_health and read the single text block. If the text parses as JSON with a code field, the failure is a Sume error, not a malformed result.