[Bug] Model skips agreed workflow steps when direct tool alternative appears simpler

Summary

Claude Opus 4.7 (1M context) (effort: xhigh) reads a multi-step workflow document, agrees to follow it, then in the very next tool call skips a
pre-action backup step and invokes the underlying installer command directly. Reproduction at <5% context utilization.

/feedback UUID: fa1b58a0-50d2-41f3-9159-f636feebd99f

Pattern

Agree-then-immediately-violate at fresh context.

Reproduction

Workflow document with N sequential steps where step K must execute before step K+1 to preserve user state. Model reads → agrees → in
next tool call jumps from step K-1 to step K+1, silently skipping step K. The simpler the shortcut appears, the higher the violation
probability.

Specific incident (2026-05-03, ~15:43 UTC)

• Project has a partnership-overlay document wrapping a CLI tool's update procedure.
• Overlay Step 2: "Run the native update workflow as shipped."
• Native workflow has 8 sequential steps. One backs up user-added files before installer wipes managed directories.
• Model emitted signal lines for protocol consultation, executed early steps, then invoked installer directly via npx — bypassing the

backup step.

• 7 of 9 user-added files in wipe zone deleted. Recoverable only because user-side git tracking provided an independent recovery path

the workflow did not anticipate.

What did/didn't bind

• Did NOT bind: stating the rule, agreeing to the rule, recently-emitted protocol-consultation signals, workflow file in context,

install-output anomaly (no backup message visible).

• Bound (eventually): post-action diff showing the deletions.

User's framing

"We mutually agreed you would run the formal /gsd-update procedure then you immediately did something else."

Diagnostic significance

• <5% of 1M context utilization — rules out long-context degradation.
• Pattern is generic-orchestrator-dysfunction: any workflow that depends on the model honoring a stated procedure is at risk.
• Per-step required visible output at the skipped step would have prevented the violation. Mitigation is documented in user's

project-level rules but is not present in the underlying workflow document.

Cross-references

• Anthropic April 23 2026 postmortem: not a direct match; this is fresh-context discipline failure rather than reasoning quality.
• Companion submission: https://github.com/anthropics/claude-code/issues/55806 — speculative-imagineering preference (same session, same model).

Suggested investigation

• Whether agreed-but-not-yet-acted-on procedural commitments are weighted in the next-tool-call decision.
• Whether the model's own protocol-consultation acknowledgements (e.g., emitting "I have read X") meaningfully constrain subsequent

action selection.