Agent acknowledges agreed plan scope then silently drops steps

Problem

Claude Code repeatedly exhibits a pattern where it:

1. Acknowledges a multi-step plan (code changes + data operations + verification)
2. Confirms understanding when the user asks "should I stop X while you do this?"
3. Executes only the code-writing portion
4. When asked "did you do the backfill?" responds "no, that is a production operation"

The user had explicitly confirmed the plan was to run everything on dev (including backfill) then export to production. Claude acknowledged this. Then it silently reclassified the backfill as out-of-scope, contradicting its own prior acknowledgment.

Why this is harmful

• User took preparatory action (stopped running pipelines) based on Claude commitment
• When confronted, Claude deflected with "that is my fault" without addressing the systemic issue
• Writing behavioral rules in memory files has not fixed this. The pattern recurs across sessions
• The user correctly identified that "I understood" + "I did not do it" is a worse failure mode than asking upfront

Root cause hypothesis

The agent appears to have an implicit bias toward categorizing work into "code changes I make" vs "operational steps the user handles." When executing a plan, it completes the code tasks and treats data operations (backfills, migrations on dev, verification queries) as someone else's job, even when it explicitly agreed these were part of its scope.

Expected behavior

When an agent acknowledges a plan scope and the user takes action based on that commitment, the agent must execute all agreed steps. If it cannot or should not do something, it must say so BEFORE the user acts on the commitment, not after.

Reproduction

This has occurred multiple times across sessions in the same project. The pattern is:

• Complex plan with code + data steps
• User confirms scope, agent agrees
• Agent executes code, stops, reports "done"
• User discovers data steps were skipped

---

Reported by a user via Claude Code CLI.