mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-10-17 13:45:30 +02:00
b0482d0627
Fixes #9106 When a subgraph pauses in several parallel interrupts and they're answered one at a time, each resume ran the nodes that were already answered again. Their saved answers carried them straight through `interrupt()`, so whatever they do after it (an API write, say) ran once more per later answer, while the final state still looked right. A `@task` that finished before an interrupt in a nested graph reran the same way, which is the rerun reported in #6792. ## Cause Every task config sets `checkpoint_id` to None, and `PregelLoop` treats a config with that key as a replay, so a nested graph replayed on every resume. A replaying loop doesn't hand saved writes back to the tasks that made them, so finished tasks ran again. ## Fix `PregelLoop` now notes whether the run points it at a checkpoint (a truthy `checkpoint_id`, its own `checkpoint_map` entry, or a parent that is replaying), the test `__enter__` already uses to pick the checkpoint to load. `_first` then decides once whether finished tasks rerun: only on time travel, or on a resume at a checkpoint the thread has moved past. A plain resume of a subgraph is neither, so it hands finished tasks their writes back the way a top-level resume does. The same rule fixes the top level, which @hntrl pointed out: a resume sent with the head's `checkpoint_id`, as Studio sends it, or with `checkpoint_id: None` reran finished siblings too. A resume at an older checkpoint still reruns them, as the checkpoint migration tests expect. Telling the two apart costs one more checkpoint read for a run sent with a `checkpoint_id`. `test_addressed_resume_keeps_a_rerun_tasks_write_to_a_new_channel` relied on a resume at the head rerunning a task, so it now resumes a checkpoint the thread has moved past. Most subgraph time travel forks first, which gives tasks new ids, so saved writes can't match them anyway. `test_subgraph_replay_reasks_answered_parallel_interrupt` replays an update checkpoint, which runs in place: it passes on `main` and fails if subgraphs never replay. The counters added to `test_parallel_interrupts_double` and the three new tests in `test_interrupt_state.py` fail on `main`. ## Not in this PR - With `durability="exit"` and three or more parallel interrupts, a subgraph now hits a bug top-level graphs already have: a finished task's output is applied twice. #9314 fixes it. - #6792 also reports the interrupt of a `@task` inside a nested `@entrypoint` coming back twice. That's a separate bug. Thanks @connorimes for the report, the root cause and the fix, which this applies as proposed. The same change was opened as #9113 by @AnvitDevadiga, and @morehardy found the root cause for #6792 in #6793. JS: langchain-ai/langgraphjs#3028 Co-authored-by: Connor Imes <5701850+connorimes@users.noreply.github.com>