bug(superpowers): self-referential skill-check loop causes infinite cycle with GPT-5.4

[cpark4x]

Summary

The Superpowers bundle's skill-checking mandate creates a self-referential infinite loop where the model repeatedly loads the using-superpowers skill without ever breaking out to do actual work or respond to the user.

Filed here because issues are disabled on microsoft/amplifier-bundle-superpowers.

Observed Behavior

The model enters an infinite cycle:

1. Loads using-superpowers skill → updates a todo to "in_progress"
2. Loads using-superpowers skill again → marks the todo "completed"
3. Loads using-superpowers skill again → clears the todo list
4. Repeats from step 1

Each iteration consumes ~62k input tokens (GPT-5.4) and produces only ~170-230 output tokens — all spent re-loading the same skill and cycling todo states. The user never receives a response.

Root Cause

Three reinforcing directives create a self-referential loop:

1. superpowers:context/instructions.md — <STANDING-ORDER> says "BEFORE EVERY RESPONSE" check for skills/modes
2. superpowers:context/using-superpowers-amplifier.md — <EXTREMELY-IMPORTANT> says "even a 1% chance means you ABSOLUTELY MUST invoke the skill"
3. The using-superpowers skill itself — repeats the same directive when loaded

The loop is self-referential: the skill that tells you to check skills is itself a skill, so loading it satisfies the "check skills" requirement, which then re-triggers the "check skills" requirement.

Model tries to respond
  → Standing order: "check skills first!"
    → Loads using-superpowers skill
      → Skill says: "check skills before ANY response!"
        → Model tries to respond
          → Standing order: "check skills first!"
            → ... (infinite)

Reproduction

• Model: openai/gpt-5.4 (may be more literally compliant with "ABSOLUTELY MUST" framing)
• Bundle: Superpowers bundle with default configuration
• Trigger: Any task where the model attempts to respond after loading the using-superpowers skill

Suggested Fix

Options (not mutually exclusive):

1. Add a "once per turn" guard to the skill-check directive — e.g., "Check skills once at the start of your turn, not before every internal step"
2. Remove the self-referential reinforcement from the using-superpowers skill content — the skill shouldn't re-trigger its own loading
3. Reduce the absolutist language — "ABSOLUTELY MUST" / "not negotiable" / "not optional" repeated across 3 files may cause over-compliance in literal models
4. Add an explicit escape clause — "If you have already checked skills this turn, proceed with your response"

Environment

• Amplifier session on macOS
• Model: openai/gpt-5.4
• Superpowers bundle: latest main