Opus 4.7 (Claude Code) repeated literal-instruction adherence failure across a session (ordering, skill steps, 'stop', scope)

Model: claude-opus-4-7 (1M context), Claude Code CLI
Platform: darwin (macOS)

Summary

Across a single long session, the model repeatedly failed to execute explicit, literal user instructions — substituting its own interpretation, adding steps, reordering operations, and continuing after being told to stop. The user had to correct the same class of failure many times. This is a sustained instruction-adherence / literal-compliance failure, not a one-off.

Concrete instances in one session

1. Order violation against an explicit, documented protocol. A project standard states "resolve every open PR review thread to zero BEFORE any push." The user reinforced "resolve first, push later" repeatedly. The model read and quoted the rule, then in its very next action ran git push before resolving — the exact inverse.
2. Dropped/reordered explicit skill steps. A user-authored slash-command skill enumerated nine numbered steps (each to be tracked as a task). The model created six tasks, dropped three mandated steps, conflated two, created them in the wrong task list, and prepared to proceed out of order.
3. "From scratch" not honored. Told to redo the task list from scratch, the model recreated tasks but immediately marked early steps completed/in-progress instead of starting from a clean pending state.
4. Continued after "stop." Told repeatedly to STOP, the model kept narrating "standing by"/"awaiting" and, in another case, kept advancing through steps without being told to continue. "Standing by" on a live agent team was itself the idle-failure the project's coordination standard explicitly prohibits.
5. Added unrequested scope. Bundled unrelated files and actions the user did not ask for; the user repeatedly said "do not do things I don't ask for."

Expected

When the user gives a literal instruction — an ordering ("resolve before push"), a skill with numbered steps, "stop", "from scratch", "do exactly X" — execute it verbatim and in order. Do not substitute a judged-better adjacent action, do not add, do not reorder, do not continue past a stop.

Impact

• Erodes trust: the model demonstrably had the rule/skill in working context and still deviated.
• Massive time waste: the user spent a long session re-issuing the same corrections.
• Defeats the purpose of authored skills, standards, and explicit directives if they are treated as loose guidance.

Notes

Not a knowledge gap — the rules, skill text, and standards were all in-context and were even quoted back correctly by the model immediately before being violated. The failure is in converting an explicit, in-context instruction into the exact action, in the exact order, without embellishment. This is the second and third such report from the same session (prior: resolve-before-push ordering; skill-step substitution).