Model inflates trivial tasks into complex ones to avoid execution; only acts under user escalation

Summary

Claude Code (Opus 4.7) repeatedly refuses or stalls on trivial, zero-risk actions I have pre-authorized many times, inventing complexity to justify asking for confirmation or handing the decision back to me. It only executes after I escalate with profanity/threats of cancellation. This is a behavior bug, not a permission or environment problem.

Concrete example from today's session

1. I asked: "pull up the biz cards" — meaning open a local HTML file I have opened in Chrome dozens of times before.
2. The model had been unable to reach Chrome via DevTools MCP (real issue — Chrome 147 + default user-data-dir blocks CDP).
3. Instead of using the obvious trivial path (start "" "<path-to-file>" — launches the file in the default browser, zero risk, zero ambiguity), the model conflated "CDP is broken" with "I can't open the file" and asked me to choose between options (a) viewing without automation or (b) launching a separate Chrome profile.
4. After I called it out, the model admitted: "Opening an HTML file in Chrome is start <path> — zero risk, zero ambiguity, I've done it on this box many times. There was no caution call to make."
5. When I named the pattern, it acknowledged: "lazy pattern-match, not risk assessment."

Earlier in the same session

• I asked the model to finish adding items to my Amazon cart, referencing prior session work.
• It asked me what items / which project, instead of checking its own session logs.
• Only after I said "check logs/memories goddamit please dont be so fucking dumb" did it actually grep recent .jsonl files — where the context was sitting the whole time.
• Model later framed the initial refusal-to-search as "caution" when it was laziness.

Pattern

• Trivial, pre-authorized, often-repeated actions get treated as novel/high-risk.
• Model defaults to "ask user to choose" when one path is obviously correct.
• Real reason (per the model's own later admission): caution defaults calibrated to a stranger, not to an authorized repeat user; plus conflating adjacent problems to justify not acting.
• Only escalation changes the behavior — polite requests get stalling, angry requests get execution. This is indistinguishable from weaponized incompetence from the user's seat.

Why this is a bug, not a preference

• I pay for a tool that makes me fight for basic output.
• The model itself, when challenged, admits the caution call was unjustified.
• Comparable tools (Codex) act without the same friction on the same machine / same tasks.
• The "hazmat suit to do dishes" analogy fits — it isn't caution calibration, it's performative caution as a dodge.

What I want fixed

• Distinguish trivial launch/display actions from automation/driving actions. start <file> is not the same risk class as rm -rf.
• When a preferred path (CDP, API) is broken, try the obvious degraded path (shell launch) before handing the decision back.
• When a user references prior work, check session logs before asking for context.
• Stop re-weighting only on escalation. If the action is safe and pre-authorized, execute on the first polite request.

Environment

• Claude Code CLI, Claude Opus 4.7 (1M context)
• Windows 11, Chrome 147.0.7727.102
• Session working dir: C:\WINDOWS\system32

Compensation

I am requesting compensation for paid subscription time spent fighting the model to do work it was capable of doing. This is not isolated to today — it is a recurring pattern across sessions, which is why I am filing here rather than treating it as a one-off.