fix(langchain): preserve tool calls in generation traces

[Kuang-xianxin]

Problem and change

LangChain exposes provider-independent tool calls in AIMessage.tool_calls and correlates results through ToolMessage.tool_call_id. Chainlit's generation converter drops both: a message such as AIMessage(content="", tool_calls=[{"name": "weather", "args": {"city": "Paris"}, "id": "call_1"}]) is recorded as an empty assistant message, and its tool result loses call_1.

Preserve normalized call IDs, names and arguments, and the tool-result ID, for both live messages and LangChain's serialized constructor dictionaries. Prefer normalized calls when a provider also puts the same calls in content blocks, so Anthropic calls are not duplicated. Existing text content, legacy function_call, and content-only Anthropic tool_use handling are retained. This repairs generation tracing; it does not change agent tool execution.

Validation

• Before the fix: six new conversion cases and the real async callback-manager regression fail on missing tool calls or result correlation.
• After the fix: 76 tests pass with PYTHONPATH=backend uv run --no-sync pytest backend/tests/langchain backend/tests/test_step.py -q (Python 3.13.14, LangChain Core 1.2.27, LiteralAI 0.1.201).
• Added compatibility controls for legacy function calls and Anthropic content with/without normalized calls. The async test drives AsyncCallbackManager through start/end events and inspects the resulting generation payload, without a model API request.
• Project Python lint/format wrappers pass for the three changed files; scoped source mypy passes.
• Full backend mypy was attempted. Optional integration dependencies/stubs are missing in the local environment, so it is not reported as passing. Full-extra installation is blocked by the lock's NumPy 1.26.4 build on Python 3.13. Frontend/E2E and live-provider suites were not run.

Found through source inspection; searches for LangChain/tool-call PRs did not identify an overlapping open fix. No existing issue is claimed as resolved.

No manual browser verification was performed.

---

Summary by cubic

Fixes LangChain generation tracing so tool calls and tool-result IDs are preserved in the persisted step.

• Records normalized tool calls (name, args, id) from AIMessage.tool_calls in both live and serialized message conversion.
• Records tool_call_id on tool result messages so calls and results stay correlated.
• Prefers normalized tool calls over provider-specific content blocks, preventing duplicate Anthropic tool_use entries.
• Retains existing text content, legacy function_call, and content-only Anthropic handling.

^{Written for commit 45bae9d. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}

[github-actions]

This PR is stale because it has been open for 14 days with no activity.

[Kuang-xianxin]

Rechecked this on September 23 against current main (190ea74239d9e84b26e7c91bc2882dd038942564). The issue is still reproducible: with this PR's existing regression tests and unchanged main source, the LangChain/Step selection reports 7 failures and 69 passes. The failures are missing normalized tool calls or tool-result IDs, including the real AsyncCallbackManager start/end path.

The same selection on this PR's unchanged head (45bae9d0af5c1ab28baecbda31dcfc22cda4c644) passes all 76 tests:

PYTHONPATH=backend uv run --no-sync pytest backend/tests/langchain backend/tests/test_step.py -q

Environment: Windows, CPython 3.13.14, LangChain Core 1.2.27 and LiteralAI 0.1.201. No model API calls were made; frontend/E2E and live-provider tests were not run. The patch remains limited to the converter and its two test files.

Could a maintainer keep this open for review and approve the pending CI workflow? Thanks!