fix(v2): handle consecutive None-content messages in merge_consecutive_messages

[chuenchen309]

Describe your changes

merge_consecutive_messages (instructor/v2/core/messages.py) crashes with AttributeError: 'NoneType' object has no attribute 'append' when two consecutive same-role messages both have content=None — a normal shape for back-to-back tool-call-only assistant turns (e.g. {"role": "assistant", "content": None, "tool_calls": [...]}) replayed as conversation history.

Root cause: new_content = message.get("content", "") — dict.get(key, default) only substitutes the default when the key is absent, not when its value is explicitly None. So new_content stays None, fails the isinstance(new_content, str) check, and on the second consecutive message of the same role hits new_messages[-1]["content"].append(new_content) where new_messages[-1]["content"] is itself None.

This runs before any network call — it's invoked directly on user-supplied kwargs["messages"] from the OpenAI (Mode.JSON/Mode.MD_JSON), Mistral, and Writer handlers (instructor/v2/providers/{openai,mistral,writer}/handlers.py).

Fix: normalize None content to "" right after extraction, matching the function's existing convention of treating "" as the safe default for missing/empty content.

new_content = message.get("content", "")
if new_content is None:
    new_content = ""

Issue ticket number and link

None found — searched open PRs/issues touching messages.py / merge_consecutive_messages, no overlap.

Checklist before requesting a review

• I have performed a self-review of my code
• If it is a core feature, I have added thorough tests.
• If it is a core feature, I have added documentation. (bug fix, not a new feature — no doc changes needed)

Testing

• Added test_consecutive_none_content to tests/processing/test_message_processing.py; confirmed (via git stash) it fails with the exact AttributeError against the pre-fix code, passes after.
• Ran tests/processing/test_message_processing.py, tests/processing/test_utils.py, tests/test_utils.py: 77 passed (incl. the new test), same 2 pre-existing failures before/after (TestUpdateGeminiKwargs::test_*_safety_settings, need the optional google-genai package which isn't installed in my dev environment — unrelated to this change, confirmed identical on main).
• ruff check, ruff format --check, and ty check all clean on the changed files.
• Added a CHANGELOG.md entry under [Unreleased] / Fixed per CLAUDE.md's PR guidelines.

AI disclosure

This fix was developed with AI assistance (Claude Code): it located the crash via a targeted search for reachable-but-untested code paths, traced the root cause through dict.get's None-vs-missing-key semantics, and drafted the patch/test/changelog entry. I reviewed the diff, independently re-derived the root cause by reading the function and tracing its callers, and ran the test suite and linters myself before opening this PR.

[ErenAta16]

Confirmed the crash: two consecutive same-role messages with content=None (e.g. back-to-back tool-call-only assistant turns) previously hit the isinstance(new_content, str) branch with None, which isn't a str, so it fell through without normalizing to the [{"type": "text", ...}] list form the merge logic expects, and the subsequent merge step assumed that shape. Normalizing None to "" before the str check brings it in line with how empty/missing content is already handled elsewhere in the same function. Test covers the exact reported shape (two tool-call-only assistant messages back to back). Looks correct.

[jxnl]

Consolidated and shipped in #2495. Closing this focused patch as superseded; thank you for the contribution.