Motivation

• The HTTP/1.1 parser could return a raw_header that included bytes read past the header terminator (\r\n\r\n), and the relay path forwarded those overflow bytes upstream before additional L7 parsing, enabling request smuggling/pipelining to bypass per-request policy checks.

Description

• Change parse_http_request to read one byte at a time (read_u8) and stop exactly when the buffer ends_with(b"\r\n\r\n"), ensuring raw_header contains only the current request header block.
• Remove the previous multi-byte temporary buffer read path and the behavior that returned overflow bytes inside raw_header.
• Adjust header-end detection to use ends_with and retain the existing parse_body_length framing logic.
• Add a regression test parse_http_request_does_not_overread_next_request that writes two requests in one write and verifies they are parsed as two separate requests with no overflow included in the first raw_header.

Testing

• Added unit/regression test: parse_http_request_does_not_overread_next_request (async tokio::test) to verify the parser does not consume bytes belonging to a subsequent pipelined request; the test was added to crates/openshell-sandbox/src/l7/rest.rs.
• Attempted mise run pre-commit in the environment, but it could not complete due to remote tooling/version resolution failures in this sandbox.
• Attempted cargo test for the crate but could not run because cargo is not available in the execution environment, so the new test was not executed here.

Codex Task