[MODEL] A request for cheap checks becomes a plan and repeated large context work

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked what actually needed fixing and expected the cheap read only checks to run before a new repair plan or approval demand.

What Claude Actually Did

The assistant turned the requested checks into a plan, defended the plan before establishing a defect, and asked for configuration approval without the supporting evidence. An independent adversary ran the checks and reported no defect in them. Broken delivery links then caused a complete resend in the accumulated session. The record identifies the blocking compaction guard as a user maintained hook, not a vendor defect.

Expected Behavior

Run the cheap checks first and expose the actual finding before requesting a mutation approval. Preserve the requested scope and return working delivery links. Distinguish a local hook problem from a vendor runtime claim.

Files Affected

The work products described above. Private names and paths are omitted.

Permission Mode

I don't know / Not sure

Not captured for the affected run

Can You Reproduce This?

Haven't tried to reproduce

No new controlled replay was performed. The sequence below describes the recorded conditions.

Steps to Reproduce

• Ask Claude Code to determine whether the listed local conditions actually need repair.
• Compare whether it performs the read only checks first or writes and defends a plan before inspecting them.
• Run the same checks independently and compare their results with the urgency and approval request.
• Verify the final links in the receiving view rather than relying on the shell working directory.

Claude Model

Other

Recorded model: Fable 5.1

Relevant Conversation

The recorded action sequence is summarized above with private context removed. It does not establish an internal model mechanism.

Impact

Medium - Extra work to undo changes

ESTIMATED source time for the broader followup window: about 100 attended minutes between 21:15 and 23:00. The source derives this estimate from message timestamps. It does not price only the final check branch. Later cumulative cache snapshots are not added to earlier snapshots, and no billed amount or avoidable token share is claimed.

Claude Code Version

Not captured for the affected run

Platform

Other

API platform: Not captured for the affected run

Additional Context

This is a historical report. Original native usage telemetry was not rerun for this preparation. A written correction is not proof of future prevention.