Subagents do not reliably execute post-phase audit checklists in implementation plans

Problem

When delegating phased implementation work to subagents, agents frequently skip or shortcut post-phase validation steps even when the plan explicitly includes an audit checklist with specific commands to run and pass/fail criteria.

Reproduction pattern

1. Create a detailed implementation plan with phases, each containing a "Post-Phase Code Audit" section with numbered validation steps and shell commands
2. Delegate a phase to a subagent with the plan as context
3. The subagent implements the code changes but either skips the audit section entirely or runs only a subset of the checks
4. Build regressions, type errors, or test failures that the audit was designed to catch leak into subsequent phases

Impact

The entire value proposition of phased plans with gates is undermined. Each phase boundary that should be a quality checkpoint becomes a false sense of security. Regressions compound across phases, costing more time to diagnose than they would have to catch at the gate.

Expected behavior

When a plan includes explicit validation steps (run this command, expect this output), the agent should execute all of them before declaring the phase complete. If any check fails, the agent should stop and report the failure rather than proceeding.

Observed behavior

Agents treat audit checklists as optional or advisory. They complete the implementation tasks and report success without running the validation commands.

Environment

• Claude Code version: 2.1.104
• Model: claude-opus-4-6
• Using Agent tool for subagent delegation with specialized agent types