MCP tools/call cancellation does not terminate Neo4j transaction #829

Summary

When an MCP client cancels a tools/call in flight, the underlying Neo4j transaction keeps running until it completes or is manually terminated at the Neo4j layer. There is no cancel/reset surface exposed on the CGC side, which means any hung query becomes operator-visible CPU load until someone opens cypher-shell and runs TERMINATE TRANSACTION.

Evidence

Observed behaviour

• An MCP client (Claude Code) issued get_repository_stats, then cancelled the call after it had not returned in a reasonable time.
• The Neo4j transaction dispatched by that call continued running for 21+ minutes after cancellation, pegging a CPU core at ~102% on the host machine.
• The only way to stop it was docker exec neo4j-cgc cypher-shell ... "TERMINATE TRANSACTION '<id>'" after locating the id via SHOW TRANSACTIONS.

Likely root cause

src/codegraphcontext/server.py dispatches blocking Neo4j calls like this (paraphrased from line ~361):

result = await asyncio.to_thread(lambda: session.run(query, **params).data())

asyncio.to_thread makes the call awaitable, but once the blocking work is handed to the worker thread it is uncancellable from the asyncio side. Cancelling the surrounding task simply stops the coroutine from receiving the result; the Neo4j session.run continues to completion.

Impact

• Any runaway query hangs the MCP connection indefinitely from the user's point of view.
• The host CPU stays pegged on the Neo4j side until an operator manually terminates the transaction.
• There is no CGC-side surface (tool, admin endpoint, config toggle) that can stop an in-flight query. The escape hatch is always "open cypher-shell, TERMINATE TRANSACTION".

Suggested direction

Transactions started in response to tools/call should be associated with a cancellation token derived from the MCP request. On cancellation, CGC should call the Neo4j dbms.killTransaction / TERMINATE TRANSACTION equivalent for that transaction id. At minimum, per-tool query timeouts (configurable, with a sane default) would prevent unbounded runaway calls.

Environment

• CGC version: 0.2.3 (pip)
• Neo4j: running in Docker container (neo4j-cgc)
• Host: Windows 11, Python 3.14
• MCP client: Claude Code