feat(coding-agent): per-server maxOutputBytes for MCP tool results

[yummybomb]

Problem

pi cuts the model-facing text of every MCP tool result at a fixed 20KB and saves the full text to a file the model is told to read with offset/limit. Without a tool that can read files, the model never sees the middle.

Example: a session started with --no-builtin-tools and only an MCP server's tools. A tool that lists a browser page's site tools returns about 43KB (10.9k tokens) for four tools. The model gets the start and end of that list. In the cut-out middle are the names and refs it needs to call those tools. It cannot open the saved file, so it re-lists the tools, gets the same cut list, and falls back to other tools.

Change

MCP servers accept maxOutputBytes in mcp.json: the bytes of tool-result text the model sees before the middle is cut. The default stays 20480.

{ "mcpServers": { "browser": { "url": "https://example.com/mcp", "maxOutputBytes": 65536 } } }

• validateMcpServerConfig rejects values that are not positive integers.
• createMcpToolDefinition passes the value to convertMcpResult, which passes it to limitMcpContent.
• read_mcp_resource and the resource list tools keep the 20KB default.
• docs/mcp.md lists the key.

Testing

• test/mcp-extension.test.ts: 29 pass. New tests:
  • The config keeps a valid value and rejects 0 and 1.5.
  • A tool definition with maxOutputBytes: 65536 returns 28KB of text whole. With the limit ignored, this test fails.
  • convertMcpResult with maxOutputBytes: 1024 cuts to about 1KB.
• npm run check: biome passes. tsc errors only on the generated model data, which this checkout does not have, and reports nothing in the changed files.

[yummybomb]

Closing: the bench may move off pi.