Repository navigation
replay: a step object reused at two flow positions maps to one node id #6015
Description
Summary
In Replay's graph and events (#6010), a step object that is used at two positions in one AgentFlow is mapped by id(obj). Both positions therefore get the same node_id, and the live highlight always lands on the first position.
Reproduce
from praisonaiagents import AgentFlow
def review(ctx): return "ok"
flow = AgentFlow(name="f", steps=[review, "draft", review]) # same object twice
# to_graph(flow): two step nodes (step:review, step:review#2), but step events for
# both executions carry node_id "step:review"Fix direction
Track the step's position at run time (the executor already knows the index or path) and map it to the positional id (step:review#2) when emitting step_start/step_end, instead of looking up id(step_obj). Add a test with the same object at two positions, sequentially and inside a Loop.
Reported by the FINAL review of #6010 (Greptile P1, item 2); deferred as non-blocking.
Activity
dajiaohuang commentedon Oct 26, 2026
I confirmed the current recorder resolves a step span through graph.ids.get(id(step)), so two positions sharing one callable cannot select distinct graph nodes. A simple per-object occurrence counter is not sufficient: a Loop revisits one position and should keep the same node id across iterations, while two positions using the same object need different ids. The executor should pass the graph position/path into span creation, preserving it through nested pattern execution. This also touches the step emission path being modified by the active #6016 variable-recording work; I am not publishing a competing workflows.py rewrite. A regression should assert both step_start and step_end IDs for two shared-callable positions, and stable IDs for those positions over multiple loop iterations. This is diagnosis/integration context, not a claim that #6016 already fixes this issue.