Rebased on current origin/main and reran validation locally.

Validated:

gofmt -w mcp\cmd.go mcp\transport.go mcp\transport_test.go
go test ./mcp
go test -race ./mcp
go test ./...
git diff --check origin/main...HEAD

All passed after the rebase.

Fresh local re-validation on the current checkout:

• go test ./mcp -> passed
• gofmt -l mcp/cmd.go mcp/transport.go mcp/transport_test.go -> no output
• git diff --check origin/main...HEAD -> passed

The GitHub checks are also green as of the latest run.

Rebased this branch onto the latest upstream main and force-pushed the cleaned head.

Validated locally:

• gofmt -l mcp/cmd.go mcp/transport.go mcp/transport_test.go (no output)
• go test ./mcp
• go test -race ./mcp
• go test ./...
• git diff --check origin/main..HEAD

Nice refactor in the stdio framing path. A few notes from reading the diff:

1. Pre-decode bound — Checking len(frame)+len(part) against MaxMessageBytes before validateJSONFrame avoids allocating/decoding oversized payloads. That composes well with other stdio hardening work on malformed envelopes.

2. ReadSlice + ErrBufferFull — The continue loop on bufio.ErrBufferFull is correct for long lines, but it does grow frame without an early cap when maxMessageBytes == 0. Worth confirming that is intentional (unbounded mode preserves prior behavior) and that the capped path rejects before unbounded growth when maxMessageBytes > 0.

3. Limit semantics — TestIOConnReadMaxMessageBytes uses int64(len(input)) without the trailing newline, which matches rejecting only the JSON payload bytes. A one-line doc comment on whether the limit includes \n/\r\n would help server authors pick values.

4. Surface area — Exposing the knob on CommandTransport, StdioTransport, and IOTransport covers the main server entry points I have been exercising in go-sdk transport tests.

I ran a quick read-only review against current main; I have not tested this branch locally. Looks like a solid fix for #984.

Thanks for the careful read. I pushed 880a8de with a small follow-up for the limit semantics you called out.

What changed:

• documented MaxMessageBytes as applying to the JSON-RPC payload bytes, not the NDJSON line ending
• changed readFrame so LF/CRLF are trimmed before the final capped-size check
• kept the bufio.ErrBufferFull path capped when MaxMessageBytes > 0
• added regression coverage for payload-at-limit with LF and CRLF
• added a regression check that MaxMessageBytes == 0 preserves the old unbounded behavior

Validation:

• gofmt -w mcp\cmd.go mcp\transport.go mcp\transport_test.go
• go test ./mcp
• go test -race ./mcp
• go test ./...
• git diff --check

Thanks for the follow-up. The payload-vs-line-ending semantics and the CRLF/LF regression tests match what I had in mind. Looks good from my read.