The ticket that starts it
Here is the whole problem, sitting in one line of somebody else's code:
def _retrieve_node(state: AgentState):
# TODO: add re-ranking to improve retrieval quality
No strategy. No candidate count. Nobody said what quality means, or which files are fair game. Hand that to an agent and it won't stop to ask — it will invent a strategy, commit to it, and rewrite your retriever with total confidence.
That ambiguity is the fuel. Every failure that follows is downstream of a decision nobody wrote down.
The instinct is to write a better prompt. The session that follows argues for something else: a smaller, sharper task beats a better prompt — and then spends the last third of the deck on where that advice stops being true.