Seven States a Coding-Agent Monitor Must Distinguish to Stay Reliable

A technical analysis argues that developer tools monitoring AI coding agents often conflate raw signals — such as process activity or file writes — with actual session states, leading to inaccurate status reporting. The author proposes a minimal but precise set of seven session states: inactive, running, waiting_for_user, completed, blocked, stalled, and unknown, each requiring specific evidence before being assigned. Key distinctions include that 'stalled' can be inferred from silence, while 'completed' requires a definitive terminal event for the current turn, and 'unknown' signals missing or contradictory evidence rather than idleness. The model recommends attaching modifiers like reason, freshness, and confidence to states rather than creating new top-level states for every error variant. Robust monitors must also ensure a new turn does not inherit a prior completion, silence never fabricates success, and one finished session does not mask another that is still running.