[apucacao]

LangChain messages and agents map model chunks and the tool loop onto the shared AG-UI subset. Argument text and results follow capture_content. config().stream() stays chunk then done.

Verified with the LangChain handler tests and mypy.

  

[cursor]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

bugbot run

[apucacao]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[jeffdupont]

Reviewed at 462209b against base b7a9f3a (#164). langchain-messages 126 passed, langchain-agents 162 passed, client -k agui 16 passed. Findings 2–4 were reproduced with probes (a real create_react_agent with a fake model for 2; real langchain_core AIMessageChunks for 3 and 4).

Must fix

1. Spec §1.8 says handlers yield chunk; this PR stops doing that (langchain-messages/tests/test_handler.py:1175; monorepo TESTING.md §1.8 "Streaming", §3.15)
test_yields_chunk_events now expects TEXT_MESSAGE_CONTENT, so the spec test fails by design. config().stream() still yields chunks via StableProjector (client.py:255), but direct handler.stream() callers get a different shape. There's no AG-UI spec on any monorepo branch, and no js-ai-sdk counterpart.
Fix: land the AG-UI handler-event contract in §1.8 and open the JS counterpart before merging.

2. langchain-agents leaves tool calls open when a run fails (langchain-agents/.../handler.py:582-587)
The except Exception block closes spans and re-raises but never calls emitter.fail_open(). Reproduced with a raising tool: TOOL_CALL_START/ARGS/END for c1 and d1, then the raise, with no TOOL_CALL_RESULT. on_tool_error does append to finished_calls (spans.py:443-447), but that's drained only inside the astream loop. langchain-messages (:1064), #161 and #165 all call fail_open() here.
Fix: for rich in emitter.fail_open(): yield rich before raise at 587, plus a raising-tool test asserting each started call gets a failed TOOL_CALL_RESULT.

3. Streamed tool events and the tool loop disagree about which calls exist (langchain-messages/.../handler.py:834-884 vs :897-899, :961-966)
Event ids come from tool_call_chunks; the loop still uses accumulated_tool_calls = chunk.tool_calls from the last chunk that had any, and AIMessageChunk.tool_calls is parsed from that one chunk only (langchain_core/messages/ai.py:554-570, 1.6.4). With OpenAI-shaped chunks for two parallel calls, events correctly stream call_1/call_2, then the loop sees [{'name': '', 'id': None, 'args': {}}] and raises No handler registered for tool "", and fail_open reports both as failed though neither ran. With a split {"q + ":"x"}, events show {"q":"x"} but the tool gets {}. The loop bug predates this PR (reproduces on b7a9f3a), but these events can't be right on top of it. The tests hand-build FakeAIMessage with whole tool_calls (tests/test_handler.py:1184-1266), which hides it.
Fix: accumulate the LangChain way, gathered = chunk if gathered is None else gathered + chunk, use gathered.tool_calls, and take ids from gathered.tool_call_chunks by index. That replaces tool_chunk_ids/seen_call_ids (~15 lines) and makes event and loop ids the same by construction. Add one test with real AIMessageChunks.

Should fix

4. Argument "prefix repair" requires byte-identical compact JSON (client/.../agui.py:178-187)
{"q": "x (with a space) streamed and final {"q":"xy"} fails startswith, so the truncated JSON is published silently. Worse, when a later chunk carries whole tool_calls, LangChain synthesises a tool_call_chunk with spaced json.dumps (ai.py:518-528), streamed as a second delta: probe gives {"q": "x + {"q": "xy"}, invalid JSON. (Per-provider spacing unverified.)
Fix: once 3 is fixed this repair has no job; delete _arg_text and keep "emit full args only if nothing streamed".

5. The same TurnEmitter.tool_end edit is in #161 (agui.py:107,163,166-190). Whichever merges second conflicts; move it into #164 once if it survives 4.

6. Agents tool results come from new callback state when the graph already returns them (langchain-agents/.../spans.py:307-308,409-414,420-425,443-447,513-518; handler.py:533-543)
The tools step of agent.astream() already yields ToolMessages with tool_call_id, content and status="error" (langgraph prebuilt/tool_node.py:1011). The callback copy needs the OTel span handler (spans.py:549-550), and its id (tool_call_id or run_id) can differ from the message id (tc.id or tc.name, handler.py:525), which would open a second call.
Fix: handle msg.type == "tool" in the existing loop with emitter.tool_result(msg.tool_call_id, msg.content, failed=getattr(msg, "status", None) == "error"); delete the spans.py changes and the finished_calls loop (~35 lines).

7. No test covers agents tool events (langchain-agents/tests/test_handler.py has no TOOL_CALL). One success test (START/ARGS/END/RESULT per call, distinct parallel ids) and one raising-tool test; the latter would have caught 2.

Nice to have

8. Cross-provider mapping differs: #161 closes text before tool_start (no parentMessageId) while this PR and #165 keep it open. The missing-id fallback is the tool name here (tc.get("id") or tc["name"], messages :962,1006,1016, agents :525), so two parallel same-name id-less calls collapse into one (dedupe at agui.py:142,196); #161/#165 use "". Pick one rule in #164/the spec.
9. The len(seen_call_ids) == 1 fallback (handler.py:867-868) puts an id-less, index-less slice on the only call seen so far even if it belongs to a second call; 3's fix removes it.
10. tests/test_handler.py:1285-1291: replace the messageId conditional with any(e["type"] == "TEXT_MESSAGE_CONTENT" and e["delta"] == "visible answer" for e in events).

Verdict: fix 1, 2 and 3 first. ~−50 lines possible (3, 4, 6). Not checked: real provider streams through LangChain; all chunk shapes were hand-built AIMessageChunks.

[apucacao]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Addressed the review. Tool-call accumulation and the failed-run close are in 23a0548. Tool messages are in 3555490, on top of reader c8d198c.

Must fix:

1. config().stream() is still chunk then done. Direct handler.stream() yields the AG-UI subset. The contract is documented on the reader (agui module docstring in c8d198c). The monorepo TESTING.md and a JS counterpart are not in this repository.
2. A failed agent stream calls fail_open after the root span is failed. The raising-tool test expects a TOOL_CALL_RESULT for each started call.
3. Streamed chunks are merged with LangChain addition, and the loop runs gathered.tool_calls. A real AIMessageChunk test covers two parallel calls.

Should fix:
4 and 5. The argument-suffix rule lives once in #164. A non-matching prefix is left as streamed text, so spaced JSON is not concatenated with compact JSON.
6. Agent tool results come from the graph ToolMessage (tool_call_id, content, status), not a second callback list.
7. The success test checks distinct parallel ids through start, args, end, and result. The raising-tool test covers the failure path.

[apucacao]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

bugbot run

[cursor]

bugbot run

[cursor]

bugbot run

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e30bb95. Configure here.}