Your agent's context window is a black box. Open it.

A long agent session fills its context window quietly. There is no gauge, no breakdown, and no warning. One accidental read of a lockfile can take a third of the window, and nothing tells you it happened. You are steering a process whose most important state variable is invisible.

When the window fills, the session compacts. Older material is summarized away and dropped. Work you assumed was still in context is gone, and the only signal is the agent starting to behave as if it forgot something. By then you have already paid for it: a repeated explanation, a wrong answer you trusted, a debugging cycle chasing a problem that was never in the code.

ContextClues turns that black box into an instrument panel. It runs beside your session and answers the questions you currently have to guess at: how full is the window, what is taking up the room, what did compaction actually drop, and how long can I keep going.

Claude Code keeps a registry of running sessions and an append only transcript of every event on local disk. ContextClues watches those files, tail parses the transcript incrementally, indexes it into a local database, and streams updates to the dashboard over server sent events. Claude's files are opened read only and are never modified.

It deliberately does not install an agent hook. A hook would give push style events, but installing one edits your configuration, and this tool does not change your setup without being asked.

A context dashboard is only useful if you can tell measurement from guesswork. Some of these figures are genuinely knowable and others are not, so nothing is presented as plain fact. Every value carries one of four confidence labels, visible in the interface, so you always know which one you are reading.

The same care goes into the projection. Growth per request is strongly right skewed: most requests add a little, and an occasional large file read adds a lot. Runway is a question about cumulative growth, so the projection is driven by the mean rather than the median, because the mean counts large requests at the rate they actually occur. A median would describe a typical request while systematically under predicting how fast the window fills. The panel reports both, and when the average sits far above the typical request it says the session is spike dominated and widens the runway into a range.

ContextClues never claims access to the model's hidden reasoning, and it does not pretend to know exactly which messages survive between compactions. That is inference, and it is labeled as inference.