Agent Loop

What It Is

YeePilot runs an agentic loop for complex tasks instead of single-shot command generation.

The loop is stage-driven:

1. discover
2. plan
3. edit and/or execute
4. verify
5. review
6. finalize

A clarify stage is inserted when intent is ambiguous.

How to Trigger It

The default interactive runtime is already agent mode.

Use:

/plan <task>

when you want explicit plan-first behavior for a specific task.

Or keep plan-first globally in a session:

/plan on

Runtime Controls

Autonomy profile

/autonomy status
/autonomy guarded
/autonomy strict_review

Profiles: high, guarded, strict_review, ultra.

Thinking depth

/think on
/think low
/think medium
/think high

Token behavior

/tokenmode normal
/tokenmode saver

Recovery System

When verification fails, YeePilot can run bounded recovery cycles with configurable budgets.

Key recovery areas include:

• patch conflicts
• build breakage
• test regressions
• failed verification steps

Related config:

• ai.recovery.*
• ai.agentic.*

This allows retry/fix behavior while preventing unbounded loops.

Verification Expectations

For mutation tasks, verification should produce concrete evidence (tests, build, health checks, etc.).

YeePilot tracks this evidence and uses it for final readiness.

Best Practice

For important changes:

1. /plan <task>
2. keep autonomy at guarded or strict_review
3. require explicit verification artifacts before finalize
4. export session for traceability (/export)