Why a flight recorder for AI?
Networks have packet captures. Software has debuggers, logs and traces. When an AI system acts in an unexpected way, looking only at its input and final output leaves the steps between them hidden.
We want CyCast to become a recorder and analyzer for those steps: what entered the run, how generation unfolded, which tool was chosen, what action followed and what changed. The goal is to reconstruct events from evidence, especially when a run fails or behaves differently from a known baseline.
Attention maps and activations do not tell us what a model is “thinking.” They are measurements to inspect alongside outputs, timing and actions, with their limits made clear.
A proposed recording
One run, from context to consequence
- Input and context
- Model and runtime events
- Tokens and available internal signals
- Tool choice and call
- External action and result
The exact evidence will depend on the model, runtime, integrations and capture settings. A trace should identify what it contains, what could not be captured and the overhead of recording it.
Capture. Compare. Reconstruct.
Our long-term direction is to inspect an individual execution, compare a normal run with a failing run, and locate the first observable divergence. Bringing model activity and agent actions onto one timeline could make failures easier to investigate and explain.
We are exploring a portable .cycap recording format, a cycap command-line
companion, adapters and interoperability with existing tracing systems such as OpenTelemetry.
Their design, availability and open-source licensing are still to be decided.
Where CyCast is today
CyCast for macOS runs open-weight models locally and lets you inspect tokens, layers and attention as an answer takes shape. It is the starting point for this vision. The cross-runtime recorder, agent timeline, replay, run comparison and proposed CLI described here are future work.
Explore the current Mac app → See the planned CyCast Trace direction →