[BUG] Opus 4.7 could not produce a trustworthy bootstrap plan in a brand-new low-context repo without repeated user correction

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Opus 4.7 1M with max effort could not independently produce a trustworthy bootstrap/environment plan even in a brand-new repo with unusually low context and unusually low environment complexity.

I created a fresh repository specifically to test this under favorable conditions:

• brand-new repo
• minimal repo CLAUDE.md containing only startup-level instructions
• preinstalled Superpowers
• first-phase goal limited to understanding Superpowers and preparing Superpowers + MemPalace
• low context usage throughout the session (~24%)

The point was to let Opus 4.7 decide what minimal CLAUDE.md and bootstrap setup it actually needed, based on its own capabilities and the already-installed workflow plugin.

Instead, it repeatedly failed in ways that made the resulting bootstrap plan untrustworthy.

The main failure sequence was:

1. It misread the task.

I explicitly wanted to understand how Superpowers already works and whether it could serve as the main future workflow plugin.
Instead, Claude drifted toward the assumption that I wanted to design a new governance layer from scratch.

1. It claimed repo-level understanding after reading very little.

It declared that it had already "looked through the whole repo", then treated preparatory material for the first Superpowers run as if it were governance baggage that should be removed.
Later, when challenged, it admitted that it had only read a couple of top-level files and had generalized from titles and directory structure.

1. It generated a formal bootstrap plan before stabilizing the facts.

The plan contained multiple speculative and unverified claims.
Only after I explicitly told it to verify the MemPalace installation/setup options did it go back and investigate.

1. Even its verification pass was unreliable.

A separate transcript audit found obvious problems, including:

• invalid verification logic for the MemPalace plugin
• treating plugin.json as the sole source of truth for plugin behavior
• incorrectly concluding there were "zero hooks"
• listing already-tracked files as new in the commit inventory
• mixing local overlay rules into what was presented as upstream baseline wiring

A key point from the transcript audit:
the model had already obtained partially correct evidence earlier, but later failed to preserve and integrate that evidence.
So this was not just "missing information" — it was also failure to hold onto and use information it had already seen.

For my use case, reviewer-driven correction is not an acceptable fallback.
If I must repeatedly act as a strong reviewer just to force basic fact-checking, then the model is not independently usable for bootstrap planning.

What Should Happen?

.

Error Messages/Logs

Steps to Reproduce

1. Create a brand-new repo with a minimal CLAUDE.md
2. Install/enable Superpowers
3. Keep the first-phase task narrow: understand Superpowers, determine whether it is sufficient, and prepare Superpowers + MemPalace
4. Ask Opus 4.7 1M with max effort to prepare the minimal bootstrap environment in an upstream-first way
5. Observe:

• task misframing
• overclaiming repo understanding
• formal planning before verification
• unstable handling of plugin/MCP/hooks facts
• need for repeated user correction before basic facts are checked

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude 1.3109.0 (35cbf6) 2026-04-16T20:32:01.000Z

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Why This Is Worse Than Ordinary Hallucination

This happened under conditions that should have been unusually easy:

• new repo
• low context
• low environment complexity
• narrow first-phase goal
• existing plugin already installed
• minimal CLAUDE.md

If Opus 4.7 cannot prepare a usable basic environment for itself under these conditions, I cannot trust it to do so in a real project.

This reboot failed for a basic reason:
Claude could not reliably set up even its own minimal working environment without repeated user correction and external review pressure.

Related Prior Incident

This also matches an earlier issue comment I posted here:

https://github.com/anthropics/claude-code/issues/50235#issuecomment-4274080429

In that case, Opus 4.7 confabulated a definition for /fork, one of Claude Code’s own built-in commands, instead of recognizing uncertainty and checking documentation.

The pattern looks similar in both cases:

• it does not reliably recognize when it should verify
• it often invents first and verifies later
• explicit "search when uncertain" instructions do not reliably prevent the behavior
• the user pays the token and attention cost for correction

What Users Actually Need

I do not need longer self-reflection about why the model failed.

I need a Claude Code model that:

• recognizes its own built-in command surface, or immediately looks it up
• treats bootstrap/plugin/config planning as verification-heavy work
• preserves already-gathered evidence instead of drifting back into confident invention
• does not require the user to act as a repeated reviewer/trainer for basic facts
• can maintain established decisions across a conversation instead of re-deriving or drifting away from them

If "search when uncertain" is not a stable default behavior even in a brand-new low-context repo, then the model is not ready to independently plan environment bootstrap work.

• Claude Code
• Opus 4.7
• 1M context
• effort: max / xhigh
• brand-new repository
• minimal repo CLAUDE.md
• preinstalled Superpowers
• first-phase target limited to Superpowers and MemPalace environment setup
• low context usage throughout the session (~24%)

Additional Context

This report was prepared jointly by the user and OpenAI Codex, based on direct review of the affected Claude Code transcripts and artifacts from the session.