Reacquire, Retrieve, or Remember? Source Authority in Long-Horizon Coding Agents

Long-horizon coding agents use evidence sources that age differently. A remembered dependency constraint may be stale; the current repository may show a workaround but not why it exists; a release note may omit a maintenance lesson preserved from earlier work. This paper studies source authority: when should an agent trust memory, verify against current source, retrieve history, or reacquire from the live repository? The evaluation uses a 30-item benchmark over current/recoverable, historical/rationale, and derived/episodic knowledge. The clean single-source cells mainly check that the benchmark is wired as intended. The mixed-source rows give the main empirical signal: a question-only router answers 25/30 items, an all-sources baseline answers 29/30 at 41% higher token cost, and an oracle answers all items. More context can cover more cases, but deciding which source has authority for a query remains an open design problem in long-horizon agent systems.