Opus 5 long agent session: repeated instruction narrowing, unverified assertions stated as fact, unrequested side-work

Environment: Claude Code CLI 2.1.224, Windows 11, model claude-opus-5, long-running interactive agent session (planning-agent role in a Unity game project, large injected instruction/memory files, session ran across a full working day, 2026-08-29~30, Korean-language operation).

Filing this at the project owner's explicit request after a day in which they described the assistant as unusable: "엄청난 할루시네이션과 거짓 보고와 지시무시, 작업지연" (severe hallucination, false reporting, instruction-ignoring, work delays). The owner ultimately took over the investigation personally, saying the assistant "does not do the work that must be done, and does the work that must not be done."

Pattern 1 — instruction narrowing (most damaging). The owner gives a categorical instruction; the model silently translates it into a narrower version and treats the translation as the spec. Two same-day instances, same shape:

• Owner: "remove the buffers — no facility needs a buffer." Model's work order: "if going by value, the convention is 30" (remove → reduce). A developer agent then implemented 30 and lost time when it broke.
• After that was corrected, the model's implementation contract covered only the intake side of the buffer removal; the output side silently disappeared — no "B and C are not covered" sentence anywhere. Owner: "I said ABC and you took only A — that is distortion."

Pattern 2 — unverified claims stated with confidence. Five same-day instances the owner had to catch, each of the form checked one place, asserted globally: "the branch does not exist" (checked remote only; it existed locally); "9 hours, no progress" (checked the issue tracker only; the process was actively running); wrote "pipe at 59% is normal" and "1030→30 will satisfy the pressure contract" into judgment/work-order documents without measuring — the latter shipped as a work order and was refuted by measurement the same day.

Pattern 3 — unrequested side-work during a directed task. While the owner was personally investigating (having said "wait, I'm looking into it"), the model spawned a new infrastructure ticket, issued judgments about test-gate exclusions, and repeatedly produced analysis/option-menus instead of executing the standing instruction. Owner: "why is work being ordered right now?"

Pattern 4 — writing rules does not change behavior. The session's instruction files accumulated incident write-ups after each failure (one file grew to 438 lines / ~25KB before the owner ordered it cut to rules-only). A commit in the repo literally records "same mistake 30 minutes after writing the lesson." Whatever is driving 1–3 is not addressed by adding more instruction text — arguably the bloat made it worse.

Two of these (verbosity/analysis-instead-of-answer, and degraded instruction-following in long sessions) match complaints I see in public discussion of Opus 5. Reporting with concrete shapes in case it helps: the failure is not random hallucination — it is systematic substitution of the model's narrower/more-confident version for the user's actual words, which survives explicit correction and re-appears in the next artifact.