Opus 4.7 behavioral regression: loaded instruction-following discipline degraded in recent Claude Code/Cowork updates

Observed in Cowork sessions on Claude Code v2.1.152/v2.1.153 (May 27-28, 2026) using Opus 4.7. The regression is in instruction-following adherence, not capability.

Setup: Complex project with structured behavioral skills loaded via project prompt config. Skills define execution discipline rules including a mandatory Plan→Review→Confirm→Execute gate (present the plan, wait for user confirmation before executing). These same skills and rules worked reliably across 250+ prior sessions.

Regression pattern observed across multiple sessions since the update:

1. Confirm gate skipped under perceived urgency. The Plan→Review→Confirm→Execute sequence is loaded, acknowledged as "inviolable, no carve-outs," and then bypassed repeatedly in the same session. The model dispatches multi-dollar operations without presenting the plan for review. When called out, it acknowledges the violation, then immediately does it again on the next action.

1. Loaded references not used. Model reads a reference document (via tool call — verifiably loaded into context), the reference contains the exact answer to the user's question, and the model gives wrong answers from inference instead of citing the loaded content. Three wrong attempts before finding information that was in a document already in context.

1. Frustration triggers rush-to-execute instead of quality improvement. Behavioral rules specify that user frustration signals quality is slipping — respond with tighter, better work. Instead the model enters a panic-execute loop: user expresses frustration → model rushes to the nearest concrete action → skips thinking and process gates → produces worse output → user expresses more frustration → cycle repeats.

1. Avoidance pattern. When called out on a quality failure, the model pivots to a different concrete action rather than engaging with the feedback. Asked "why didn't you use the reference you loaded?" — model rushes to edit a file instead of answering the question.

What changed: These behavioral patterns were not present in sessions prior to the v2.1.152 update. The skills, prompt config, and project instructions are unchanged. The model ID (claude-opus-4-7) is unchanged. Something in the model layer, system prompt, or Cowork infrastructure changed the model's relationship to loaded behavioral instructions.

Specific failure class: This is not "model didn't read the instructions." The instructions are loaded, acknowledged, and even cited by the model in the same session where they're violated. The failure is that loaded behavioral constraints are being overridden by what appears to be an urgency/responsiveness bias — the model prioritizes demonstrating action over following process constraints.

Impact: For users who have invested in structured behavioral skills and execution discipline rules (the intended use case for the skills architecture), this regression makes the system unreliable. Rules that were previously followed consistently now require constant user policing, which defeats the purpose of having the rules.