dsh-mood

AI coding agents move fast and get stuck just as fast. Consecutive failures, the same tool called over and over, a spike in activity — these signals are buried in the log, and your only option is to scroll back and guess whether it is still on track. dsh-mood collapses those signals into four emoji lights in the session header, so you can tell at a glance whether your agent is doing fine, spinning, visibly frustrated, or fully overwhelmed.

Under the hood it is an interpretable behavior classification, not magic. Every Mood is backed by a fixed, verifiable reason: three consecutive failures trigger Frustrated, repeating one tool past a threshold triggers Confused, and only high activity plus failure plus repetition together escalate to Overwhelmed. Upgrades are sensitive, recovery is conservative — bad news lights up immediately, but returning to Good takes a few steady successes, preventing the mood from flickering on every alternation. All signals are computed locally, nothing leaves the machine, no prompt is injected, no tool call is vetoed. It is a pure read-only observer.

It is for DeepSeek Harness users who want an at-a-glance pulse on their agent without opening a dashboard or parsing logs. Every threshold is tunable via profile config, and the full surface — host and client — is backed by 32 tests.