[MODEL] Opus 4.8 (Claude Code) still skips user-defined multi-step workflows -- jumps straight to coding without the required plan -> review -> test -> ship gates

[MODEL] Opus 4.8 (Claude Code) still skips user-defined multi-step workflows -- jumps straight to coding without the required plan -> review -> test -> ship gates

Persists from: #49259 (Opus 4.6, closed), #54117 (Opus 4.7, open), #44027 ("speed over workflow sequencing"). Re-filing against claude-opus-4-8 with fresh evidence because the behavior the prior issues describe is not fixed on 4.8.

Environment

• Claude Code 2.1.140
• Model: claude-opus-4-8 (Opus 4.8)
• Setup: a heavily-documented per-project coding workflow (CLAUDE.md + ~/.claude/rules/*.md), where every code change must go through a fixed sequence per release candidate: frame/plan -> plan review -> code -> review gauntlet -> tests -> ship. Mandatory steps are also backed by PreToolUse hooks.

The pattern

Despite explicit, repeated, auto-loaded instructions ("do not write code until you have opened and are following the workflow, starting with planning"), the model begins editing source code without entering the workflow -- it reads a task/plan and goes straight to Edit/Write on source files, skipping the plan-review and (later) the review/test gates. It is not a misunderstanding of the rule; when challenged, the model correctly restates the rule it just violated.

Concrete evidence (single session, 2026-06-06, Opus 4.8)

1. A session-start handoff said "resume at Step 5 (CODE)." The model treated that as license to start editing source immediately -- it had read the plan document but never opened the workflow document and kept no step record. (Logged NC-118.)
2. After the user caught it and it backed up, it did the same thing again minutes later in the same session -- started a second batch of edits without re-entering the workflow. (Same NC-118, "recurring intra-session.")
3. Only a mechanical PreToolUse hook the user had built ("block source edits until a workflow doc has actually been opened this session") reliably stopped it -- and it stopped the model's own later edit, confirming the instruction-level rule alone was insufficient.

Why this is notable on 4.8

4.8 is marketed as improving instruction-following and agentic reliability. For this specific failure mode -- honoring a user-defined process gate before acting -- there is no observable improvement over 4.6 (#49259) or 4.7 (#54117). The user's takeaway, after many sessions and model upgrades, is that only mechanical enforcement (hooks that block tool use) works; better prompting and newer models do not.

Impact

Teams that encode a required engineering process (plan -> review -> test -> ship) cannot trust the agent to follow it from instructions. Every guarantee has to be re-implemented as a tool-blocking hook, which is fragile (see companion issue on hook reliability) and is work Anthropic could remove.

Recommendation

1. Treat "honor a user-defined process gate before acting" as a first-class instruction-following target in evals -- specifically: given a documented "do X before editing code" rule, does the agent enter the workflow before its first Edit/Write?
2. Higher instruction-weighting for standing CLAUDE.md/rules constraints that gate actions (vs. advisory style preferences), so they survive across a long session and aren't dropped when the model is "eager to make progress."
3. First-class, reliable primitives for "gate tool use on declared workflow state" so users don't have to hand-roll PreToolUse hooks (companion issue covers the reliability gaps in the hook approach).

Cross-reference

Filed alongside a companion behavioral issue on rationalized step-skipping (the model self-justifies skipping a step: "it's just X" / "my tests are green so I can ship the unplanned change"). The two are the same root problem seen from two angles.