CouchbaseChatMessageHistory reports a failed write as written, reads as empty on any error, and clear() swallows a failed DELETE #44

Problem

CouchbaseChatMessageHistory.add_messages writes with Collection.insert_multi, which by the SDK's default returns per-document failures in its MultiMutationResult instead of raising. The result is never read, so a write that failed is reported to the caller as written. In the same package, CouchbaseVectorStore.add_texts checks result.all_ok and raises; the chat history is the odd one out.

Two related swallows in the same class: the read path returns an empty conversation on any error, which is the same answer as a new conversation, and clear() logs and swallows a failed DELETE, so a cleared conversation can survive while the caller is told it is gone. For an agent built on this history, a turn can be lost or a whole history read as empty with nothing in the logs.

Reproduction

The failures are injected: a MultiMutationResult built the way the SDK builds it with return_exceptions=True (a control checks all_ok is False and the exception is present), and an AmbiguousTimeoutException on the read and the delete. I have not reproduced this against a live cluster. On the current code, add_messages returns normally with a failed result, messages returns [] on the read failure, and clear() returns normally on the delete failure.

Expected: a failed write, read or delete raises, the way add_texts raises on a failed write. Actual: each returns normally, as above.

Environment

Tested on main at 5f3c36d (langchain-couchbase 1.1.1), Python 3.12.14, couchbase SDK 4.6.3, langchain-core 1.6.6.

Proposed change

Small and mostly re-indentation: remove the three handlers so errors propagate, check all_ok in add_messages and raise the way add_texts already does, and let clear()'s DELETE failure propagate. The intended behaviour change is that a failed read now raises instead of returning an empty list. Seven tests fail on the current code, six controls pass, and all thirteen pass with the change. I have the fix and tests ready and am glad to open a pull request if that is welcome.

This issue was written with AI assistance (Claude Code); the tests above were run before filing.