Skip to content

Oversight System ​

Flightdeck's oversight system controls how much autonomy agents have during a session. It replaces the older Intent Rules feature with a simpler, more intuitive three-tier model.

Oversight Tiers ​

TierAutonomyBest For
SupervisedLow — agents ask before most actionsLearning the system, critical codebases, high-risk changes
BalancedMedium — routine work flows, structural changes need approvalDay-to-day development, most teams
AutonomousHigh — agents work independently, you monitor resultsTrusted workflows, experienced users, batch operations

What Each Tier Controls ​

Each tier injects behavioral instructions into agent system prompts at spawn time. These instructions tell agents how to behave — not just which actions are gated:

Supervised:

  • Agents explain their reasoning before acting
  • File modifications require explicit confirmation
  • Architecture decisions are surfaced for review
  • Agents prefer asking over assuming

Balanced (default):

  • Routine code changes proceed automatically
  • New file creation and dependency changes trigger confirmation
  • Architecture and security decisions require review
  • Agents use good judgment on when to check in

Autonomous:

  • Agents work independently with minimal interruption
  • Only critical errors and security issues trigger alerts
  • Agents make and execute their own plans
  • Results are reviewed after completion, not during

Custom Instructions ​

In addition to the tier presets, you can provide natural language custom instructions that are appended to every agent's system prompt. Use this for project-specific guidance:

Always run tests before committing.
Use conventional commits (feat:, fix:, docs:).
Never modify the database schema without discussing first.
Prefer TypeScript strict mode patterns.

Custom instructions work alongside the oversight tier — they add to the tier's behavioral instructions, not replace them.

Scope ​

Oversight settings are per-project with a global default:

  • Global default — Set in Settings → Oversight. Applies to new projects.
  • Per-project override — Each project can override the global default. Set when creating the project or in project settings.

When a session starts, Flightdeck reads the project's oversight tier (or the global default), generates the behavioral instructions, and injects them into each agent's system prompt.

Configuration ​

Via the UI ​

Open Settings → Oversight to:

  • Select the global default tier (Supervised / Balanced / Autonomous)
  • Add custom instructions
  • Preview the instructions that will be injected

Per-project overrides are set in the project creation dialog or project settings panel.

Via Config File ​

yaml
oversight:
  level: balanced  # supervised | balanced | autonomous
  customInstructions: |
    Always run tests before committing.
    Use conventional commits.

Notification Interaction ​

Oversight level affects in-app notifications only. External channel routing (Telegram, Slack) uses separate routing preferences and fires regardless of oversight setting.

The three layers of the notification system:

  1. Events — Generated by agent actions, system alerts, etc.
  2. Notification Routing — Delivers to external channels per user preferences and quiet hours
  3. Oversight Level — Client-side gating of in-app toasts by severity

This means: even in Autonomous mode, critical alerts still reach your Telegram/Slack. The oversight tier only controls how much the in-app UI interrupts you.

Migration from Intent Rules ​

The Intent Rules system (per-category allow/alert/review rules) has been removed. The oversight tiers provide equivalent control with less configuration overhead:

Old Intent RulesNew Oversight Equivalent
Conservative presetSupervised tier
Moderate presetBalanced tier
Autonomous presetAutonomous tier
Per-category rulesCustom instructions (natural language)

If you had custom intent rules, translate them into natural language custom instructions. For example, if you had "Architecture → Require Review", add the custom instruction: "Always ask before making architecture changes."

Documentation generated by AI

Documentation generated by AI