Model fails to connect adjacent logical steps (downloads large file without saving, then needs to re-download)

Description

During a multi-step task, Claude downloaded a ~70MB file (piped to tar -t for structure inspection), discarded it, then needed to upload that same file to S3 in the very next step. This required re-downloading the entire file — something any person would avoid by simply saving it to disk during the first download.

This isn't a knowledge gap or missing instruction — it's a failure to connect two adjacent facts:

1. "I am downloading a large file right now"
2. "The next step in my plan requires this exact file"

The logical conclusion (save it) is trivial, yet the model didn't make the connection.

Why this matters

• This class of failure (not connecting adjacent physical consequences) can't be addressed with instructions or memory — you can't write a rule for every instance of "think one step ahead"
• It erodes user trust more than knowledge gaps, because it looks like a fundamental reasoning failure rather than a fixable behavior
• The user explicitly asked what feedback/instruction could prevent this, and neither of us could identify one — the issue is upstream in planning/reasoning

Reproduction

1. Give Claude a multi-step plan where step N produces an artifact needed in step N+1
2. In step N, ask it to inspect the artifact (e.g., list tar contents)
3. Observe that it streams/pipes the download without saving, then attempts to re-download in step N+1

Expected behavior

When downloading a file that will be needed in a subsequent step, save it to disk rather than discarding after inspection.

Environment

• Claude Code CLI
• Model: Opus 4.6