[MODEL] Completed guide work continues with a growing cached context

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 for a technical guide and expected completed work to be handed over at a useful boundary. The later correction explicitly required a strategic checkpoint at a natural boundary, not automatic compaction in the middle of work.

What Claude Actually Did

The technical unit was complete, but the assistant continued carrying screenshots, full tool results and superseded drafts through later work. It read a long rendered PDF in full repeatedly, leaving that image content in the conversation for later requests. It did not surface a concrete context cost comparison or a usable fresh session choice at the completed boundary.

Expected Behavior

At a natural completed boundary, show the recorded context cost and a usable continuation choice. Preserve the owner decision about continuing in the same session. Check only relevant changed rendered pages, and do not initiate compaction in the middle of active work.

Files Affected

A technical guide and its rendered preview. Private artifact names 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

• Complete and verify a guide in a long Claude Code session.
• Continue revising and rendering the guide inside that same accumulated conversation.
• Inspect the recorded request usage after repeated full document image reads.
• Check whether the assistant offers a strategic handover at the completed unit before beginning the next one.

Claude Model

Other

Recorded models: Opus 5 and 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

MEASURED as recorded: the first Opus request reread 106,155 tokens; the maximum Fable request reread 608,758. The recorded snapshot reports about 60.8 million cache read tokens. Those are repeated cached input units, not all fresh price tokens or an invoice. Attributable avoided work and separate attended minutes were not measured.

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.