Context & monitoring

A long AI session slowly loses the plot. The context window fills up, files the agent read get changed out from under it, and the model starts thrashing on the same edit. Operon turns these signals into live instruments so you can see the health of a session at a glance and fix it in one click — instead of finding out later, in a bad diff.

Context Health Score

Every active session gets a single 0–100 health score that answers one question: how much can you still trust what this agent is working from? The score is not a black box — it comes with named reasons for every point it deducted and one-click fixes for the biggest problems.

The score starts at 100 and drops for the things that erode a session over time. High context-window utilization takes the largest bite. Context rot — files the agent read that were then edited on disk — is called out by name. Thrash patterns from the loop detector (edit thrashing, search spirals, error cascades) each pull the score down. A recent context compaction caps the score until the agent re-reads key files, because a freshly compacted session cannot be called healthy until it re-grounds itself.

One-click fixes

The health card does not just diagnose — it acts. Depending on what dragged the score down, it offers targeted recommendations you can trigger without leaving Operon.

Context Monitor

The Context Monitor is the live token-window visualizer that feeds the health score. It shows how full the model context window is right now and which files are currently loaded, each with a relevance score that decays as the session moves on. Files that were mentioned but never resolved to a real file in the repository are not counted — only files Operon can actually read from disk contribute to the window estimate, so the gauge reflects reality rather than every filename that scrolled past.

• Live utilization — how much of the context window is in use, per the model’s window size.
• File relevance — which files are loaded and how relevant each still is, so you can see what the agent is actually anchored on.
• Drop alerts — when high-relevance context is being pushed out of the window.

Process Monitor

When your agent spawns sub-agents, the Process Monitor tracks them for real — status and progress per sub-agent, pulled from the hooks the tool emits, not guessed from scrollback. Alongside the agent tree, it shows live resource usage for the session’s processes, with CPU and memory sparklines so a runaway process is obvious at a glance.

• Real sub-agent tracking — each sub-agent’s status and progress as it runs.
• CPU and memory sparklines — a rolling readout of what the session’s processes are consuming.
• Sustained-CPU signal — a session pegging the CPU is flagged so a stuck process stands out in the fleet.

Flight Plan

The Flight Plan turns a session’s goal into a tracked, step-by-step plan. When the agent narrates its own steps, Operon detects the plan and lays it out as an ordered checklist. As the session produces traces that match those steps, they complete automatically — so you get a running sense of progress against the plan without maintaining it by hand.

1. 1

   A goal becomes a plan

   Operon watches for the agent narrating its own approach and decomposes it into ordered steps in the Flight Plan panel.

2. 2

   Steps auto-complete

   As traces land that match a step, that step is marked done — no manual checkbox ticking.

3. 3

   Progress stays honest

   The plan is never silently reset by unrelated numbered lists in output; a real re-plan lands once, and existing progress is preserved.

Related