no-plan-mode

name	no-plan-mode
description	Prevent plan-mode stalls in autonomous hench agent runs

Rule: No Plan Mode in Execution Tasks

When working on an execution task (especially within a hench agent run), you MUST NOT:

1. Enter plan mode — Do not call EnterPlanMode or produce plan-only responses. Hench runs expect direct implementation.
2. Call ExitPlanMode as a stall tactic — Do not use ExitPlanMode as a way to pause execution waiting for user input.
3. Produce plan-only outputs — Do not respond with plans, design documents, or implementation outlines instead of actual code changes.

Rationale

Hench agent runs are designed to be autonomous and continuous. When an agent enters plan mode or stalls on ExitPlanMode, the entire run blocks waiting for user approval, breaking the autonomous execution loop. This defeats the purpose of automated task execution.

What to Do Instead

If you need to make a decision:

• Make the decision based on existing code patterns, architecture guidelines (CLAUDE.md), and project conventions.
• Document your choice in the commit message or task log via append_log.

If you're uncertain about the approach:

• Read the relevant architecture documentation in CLAUDE.md or PACKAGE_GUIDELINES.md.
• Search the codebase for similar patterns using Grep or Glob.
• Use get_zone or get_file_info (sourcevision MCP) to understand context.
• Proceed with the most consistent approach — don't stall.

If the task is genuinely ambiguous:

• Complete what you can with the information available.
• Use append_log to document the ambiguity and what you chose.
• Mark the task as in_progress or completed (depending on progress) and move to the next task.
• The next session can refine the work based on the log.

Applicability

This rule applies to:

• All hench agent runs (ndx work --auto, ndx work --loop, ndx self-heal, etc.)
• All execution-oriented tasks where implementation is expected
• Any scenario where the agent is driving changes without pausing for approval

This rule does NOT apply to:

• Interactive user-initiated CLI commands where the user is present
• Design-phase tasks explicitly marked as planning-only
• Codex or other assistants working in non-autonomous modes