Concept · Observability · 3 min read

BlackBox and God's Eye

Unattended agents are only trustworthy if you can afterwards see exactly what they did and what it cost. Kaptain ships two surfaces for that: BlackBox records the runs; God's Eye shows the code the runs reason over.

Verification needs receipts

"It worked" is not evidence. A reviewable run needs the prompt, every tool call with its arguments and result, the model's turns, and the exact token bill — kept, not inferred.

BlackBox: the flight recorder

Session turns traces every turn — switchable between this session and all sessions, so a 3am scheduled run is as inspectable as the chat you're looking at. Model usage shows exact provider-reported tokens from Claude Code, Codex, Ollama, llama-server, and cloud runs, with daily and monthly totals. When something behaves oddly, the trace answers which tool call, with which arguments, returned what — the difference between debugging and guessing.

God's Eye: the code map

Inside the Views tab, God's Eye renders the KodeGraph index of your project — the same symbol-and-relationship map the kodegraph_* tools answer from. It follows the project directory live and re-indexes in the background, so what the model sees and what you see stay the same map.

Where they live

BlackBox is its own tab; God's Eye is a section of Views, co-branded KodeGraph. Brain Logs completes the trio with the runtime's own log stream — the same lines kaptain.log holds on disk.

Agent-readable version: index.md