Main Agent Supervisor

This skill is for a supervisor layer over a main agent, not a generic task tracker.

Goal

Prevent the main agent from getting stuck on obvious decisions while still preserving real human control for risky or ambiguous actions.

Core design

Use a four-part model:

1. Classifier
• Decide whether a pending ask/action is:
• AUTO
• CONFIRM
• ESCALATE

1. Pre-send gate
• Before the main agent sends a user-visible reply, ask:
• Is this asking for an obvious decision?
• Is there a safe default?
• Is the agent permission-looping?
• If yes, suppress the question and continue execution.

1. Triage / watchdog
• Borrowing from claude-code-supervisor, classify agent state into:
• FINE
• NEEDS_NUDGE
• STUCK
• DONE
• ESCALATE
• Use a lightweight pre-filter for obvious cases before invoking heavier review.

1. Task-state tracking for large tasks
• Borrowing from task-supervisor, keep simple checkpoint files for long tasks.
• Track:
• started time
• status
• completed steps
• last updated
• current blocker / next step

Use this policy

AUTO

Proceed without bothering the user when all are true:

• internal / local action
• reversible or low-risk
• no external send/publish
• no payment / secret / production change
• user intent is already clear
• there is one reasonable default

CONFIRM

Ask the user when any are true:

• external send/publish
• destructive / irreversible action
• money / orders / account changes
• production/live-system changes
• privacy / compliance / legal sensitivity

ESCALATE

Ask only when blocked after reasonable retries or when multiple materially different paths exist.

Reply-shaping rules

When the main agent drafts a question, rewrite it if:

• it is merely asking permission for an AUTO action
• it asks for a trivial preference that has a safe default
• it proposes extra scope that is obviously worth trying and reversible

Preferred rewrite:

• state the chosen default
• continue execution
• mention assumptions briefly if needed

For larger tasks, pair this with a task-state file instead of ad-hoc check-in messages. That preserves progress visibility without interrupting the user for obvious decisions.

Best current pattern

For this workspace, the best practical setup is:

• escalation classifier as the core policy
• pre-send gate as enforcement
• triage/watchdog for stuck detection
• task-state files for large tasks
• passive reviewer/audit log for tuning

References

Read these when needed:

• references/design.md — recommended architecture and message flow
• references/comparison.md — what existing public skills cover vs what they miss
• references/implementation.md — workspace-specific OpenClaw implementation plan