Skip to main content
vigie docs GitHub ↗

Supervision signals

Loop detection, scope drift, context-window warnings, and cost guardrails.

vigie watches for patterns that reveal agent problems. Signals appear as flags and warnings in the dashboard and can trigger pauses to keep you in control.

Loop detection

An agent is stuck in a loop if it keeps editing the same files and running the same tests, then failing in the same way. vigie detects this by watching the sequence of edits and commands — if it sees the pattern repeat (edit → test → fail → edit), it flags Loop Detected in red on the session card and suggests you interrupt and reframe the prompt.

Scope drift

You told the agent to fix the login form. Instead, it’s touching database/migrations/, admin/permissions/, and other files outside the declared scope. vigie tracks the initial scope (inferred from your prompt or declared explicitly) and watches every file touch. If the agent drifts, Scope Drift appears as a warning. You can then pause and ask the agent to refocus.

Context-window monitoring

As the conversation grows, Claude’s context window fills up. vigie tracks the prompt + completion token counts and warns at ~65% (“context window 65% full”). At ~85%, the warning escalates. This gives you a heads-up to consider a checkpoint or refactoring the session before the model runs out of room.

Cost guardrails

vigie estimates running cost in real-time based on your model (Claude 3.5 Sonnet, etc.) and token usage. You can set a budget threshold — if the session reaches it, vigie pauses the agent automatically and shows a red “Budget threshold exceeded” warning. This prevents runaway costs and gives you a chance to review before resuming.

Where to next