[he-yufeng]

Summary

Fixes #2707.

The streamable HTTP client was closing SSE responses as soon as a terminal JSON-RPC response arrived. That leaves the underlying HTTP/1.1 connection partially drained, so the next request can pay the keepalive penalty described in the issue.

This change keeps reading the POST SSE response, resumption GET response, and resumed GET stream to EOF after the terminal response has been delivered. Terminal responses still stop further message handling and still avoid reconnecting; the only behavior change is that the transport no longer aborts the response early on the normal completion path.

To verify

• uv run pytest tests\shared\test_streamable_http.py -k "drains_after_terminal_response"
• uv run pytest tests\shared\test_streamable_http.py::test_streamable_http_client_tool_invocation tests\shared\test_streamable_http.py::test_streamable_http_client_resumption
• uv run ruff check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
• uv run ruff format --check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
• git diff --check -- src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py

[he-yufeng]

Updated in bb834da.

You were right about the exception path. I moved response_complete outside the try scopes and now both paths return immediately if the terminal response was already delivered before a drain error:

• _handle_sse_response(): no reconnect after terminal POST SSE response + post-terminal drain error.
• _handle_reconnection(): no recursive retry after resumed stream delivers the terminal response + post-terminal drain error, so the response is not forwarded twice.

Added focused regressions for both probes:

• test_sse_response_does_not_reconnect_after_terminal_then_drain_error
• test_reconnection_does_not_retry_after_terminal_then_drain_error

Validation run locally:

PYTHONUTF8=1 PYTHONPATH=src uv run --no-sync pytest tests\shared\test_streamable_http.py -q -k "terminal_then_drain_error or terminal_response or failed_resume" --basetemp .tmp\pytest-2750-20260604b -p no:cacheprovider
PYTHONUTF8=1 PYTHONPATH=src uv run --no-sync pytest tests\shared\test_streamable_http.py -q --basetemp .tmp\pytest-2750-20260604-full -p no:cacheprovider
PYTHONUTF8=1 uv run --no-sync ruff check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
PYTHONUTF8=1 uv run --no-sync ruff format --check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
PYTHONUTF8=1 uv run --no-sync pyright src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
git diff --check

Results: the focused subset passed, the full tests/shared/test_streamable_http.py file passed (66 passed), ruff/format/pyright/diff-check passed.

I also tried mypy on the touched files, but on this Windows checkout it reports existing project-wide issues while following imports (missing win32/jsonschema/requests stubs, os.killpg/SIGKILL on Windows stubs, existing session/test typing errors). I did not treat that as caused by this patch.

[he-yufeng]

Addressed the latest review and rebased this branch on current main.

Changes in the follow-up commit:

• keep response_complete outside the SSE iteration try-block so terminal responses remain terminal even if draining raises afterwards;
• return from the POST SSE drain exception path once a terminal response has already been delivered;
• return from the resumed GET exception path once a terminal response has already been delivered, avoiding duplicate delivery/reconnect after a completed response;
• add regression coverage for both _handle_sse_response() and _handle_reconnection() when the stream yields a terminal response and then raises during drain.

Validation on Windows:

• uv run --no-sync pytest tests\shared\test_streamable_http.py -q -k "terminal_then_drain_error or terminal_response or failed_resume" --basetemp .tmp\pytest-2750-rebase -p no:cacheprovider -> 5 passed
• uv run --no-sync pytest tests\shared\test_streamable_http.py -q --basetemp .tmp\pytest-2750-rebase-full -p no:cacheprovider -> 66 passed
• uv run --no-sync ruff check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py -> passed
• uv run --no-sync ruff format --check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py -> passed
• uv run --no-sync pyright src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py -> 0 errors
• git diff --check -> passed

[he-yufeng]

Addressed the coverage gap on the non-terminal SSE error path.

What changed:

• added a focused test that covers the pre-terminal stream error path and verifies it still reconnects with the last SSE id.

Validation on head 58bf5a5:

• uv run --no-sync pytest tests\shared\test_streamable_http.py -q --basetemp .tmp\pytest-2750-coverage-full -p no:cacheprovider -> 67 passed
• uv run --no-sync ruff check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py tests\interaction\transports\test_hosting_resume.py -> passed
• uv run --no-sync ruff format --check src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py tests\interaction\transports\test_hosting_resume.py -> passed
• uv run --no-sync pyright src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py tests\interaction\transports\test_hosting_resume.py -> passed
• git diff --check -> passed

[he-yufeng]

Follow-up after the CI strict-no-cover failure:

• The new regression test now covers the SSE drain exception path, so I removed the stale # pragma: no cover from that line.
• Re-ran the focused streamable HTTP suite and static checks locally.

Verified:

uv run --no-sync pytest tests\shared\test_streamable_http.py -q --basetemp .tmp\pytest-2750-nocover -p no:cacheprovider
uv run --no-sync ruff check --no-cache src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
uv run --no-sync ruff format --check --no-cache src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
uv run --no-sync pyright src\mcp\client\streamable_http.py tests\shared\test_streamable_http.py
git diff --check

[maxisbey]

closing as original issue wasn't reproduced