GPT-5.6 with superpowers enters a planning loop - keeps re-designing/re-planning instead of executing work

[gzjggg]

Description

When superpowers (obra/superpowers) is installed in Codex and the active model is GPT-5.6 (observed with gpt-5.6-terra; likely affects gpt-5.6-sol and gpt-5.6-luna as well), the model enters a self-reinforcing planning loop: instead of transitioning from design/plan to implement/verify, it repeatedly re-outputs status summaries, re-plans, re-designs, or checks project status -- effectively acting as a project manager that never ships code.

The pattern is:

1. User describes a task (e.g., "add feature X" or "fix bug Y").
2. GPT-5.6 enters the superpowers design phase -- appropriate.
3. Design is written. Plan is written.
4. Instead of executing the plan, the model re-invokes the superpowers brainstorming/planning skills again, or outputs a status message like "Let me review the current state..." / "I should check what phase we are in..." / re-reads the project files and outputs an updated design.
5. This repeats 3-5+ times across user turns, consuming significant token budget before the user manually intervenes.

Expected behavior

After the design and plan phases complete, GPT-5.6 should transition to execution -- running commands, writing code, modifying files -- and only return to planning when the plan fails or new design decisions are needed. A single planning phase (or a single replan on task failure) is expected; indefinite re-planning is not.

Steps to reproduce

1. Install obra/superpowers in a Codex project (via the standard .codex/INSTALL.md method).
2. Set model to gpt-5.6-terra (or gpt-5.6-sol/gpt-5.6-luna).
3. Give a concrete task, e.g. "Add a settings page to my React app" or "Fix the failing test in src/tests/api.test.ts".
4. Observe the model output: it will produce a design doc and plan, then stall -- re-reading project files, re-iterating the status, or re-running the superpowers planning skill -- rather than executing the first task in the plan.
5. Compare with gpt-5.5-codex on the same task with the same superpowers installation: the older model transitions to execution correctly.

Root cause hypothesis

The gpt-5.6-* models appear to have stronger adherence to meta-protocols and structural "rules" than their predecessors. When superpowers establishes a design -> plan -> implement -> verify workflow, GPT-5.6 may interpret every user turn (or even its own output) as a signal to re-validate the workflow phase rather than advance along it. Specifically:

• The superpowers skills define clear phases (brainstorm, design, plan, implement, verify, review, ship).
• Older models (gpt-5.4, gpt-5.5) transition linearly through these phases.
• GPT-5.6 seems to re-enter the phase-selection or status-reporting phase after completing the plan, effectively resetting progress.

Another possibility: the model's context/reasoning chain treats "plan" as an output artifact that must be continuously maintained rather than a stepping stone to execution -- a kind of goal-conditioning collision between superpowers' workflow rules and GPT-5.6's latent behavior.

Environment

• Model: gpt-5.6-terra (also seen on gpt-5.6-sol)
• Superpowers version: latest from obra/superpowers (main branch)
• Platform: Windows / Linux
• Codex CLI version: 0.144.x and later

Related issues

• /goal + superpowers:brainstorming: assistant self-loops, generating ~150 duplicate clarifying questions without user input #21786 -- /goal + superpowers:brainstorming turn-yield loop (different root cause: turn control, not phase transition)
• Codex entered a documentation loop and became my project manager #23274 -- Codex entered a documentation loop instead of executing code (related symptom, different trigger)
• After enabling Plan Mode, clicking “Implement Plan” does not execute the plan and remains stuck in Plan Mode #23341 -- Plan Mode "Implement Plan" button re-outputs plan instead of executing (related symptom in GUI mode)
• gpt-5.6-sol/terra/luna hardcode use_responses_lite / multi_agent_version, causing 400s on Azure (and likely any non-ChatGPT-backend) model_provider #31882 -- gpt-5.6 models hardcode backend-specific flags (may contribute to degraded planning-vs-execution reasoning on non-default providers)

Suggested fix direction

• Add a maximum planning depth or phase-transition guarantee in the superpowers integration: once a plan is submitted/accepted, the next turn should always execute a concrete action (file edit, command, tool call) -- no re-planning until an execution step completes or fails.
• Alternatively, consider a prompt-level mitigation in the Codex system prompt when superpowers is active: explicitly instruct GPT-5.6 to advance through workflow phases monotonically and treat a completed plan as a commitment to execute, not a checkpoint for re-validation.

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• Model refuses to exit plan mode. #31546

Powered by Codex Action