Opus 5: prose between tool calls returned as summarized 'narration' thinking blocks since 2026-09-11 — code blocks dropped, Desktop never shows them, model believes user saw it

Summary

Since 2026-09-11 ~02:00 UTC, on claude-opus-5 (effort xhigh), the prose Claude writes between tool calls (after a tool_result, before the next tool_use) frequently comes back not as a text block but as a thinking block containing a server-written summary (the block the stream-json schema calls a narration block: "server summaries of the prose between tool calls, not the model's own reasoning"). Three consequences compound into a user-visible bug:

1. The summary rewrites the wording and drops fenced code blocks. A command the model wrote for the user disappears.
2. The Desktop app does not show these blocks at all (the CLI renders them as ordinary text). Two real sessions in the Desktop app: the user asked "when did you ever tell me that?" after the model's instructions existed only in a narration block.
3. The model still has that narration in its context and believes the user saw it, so it finishes the turn with "run the command above" / "as I explained above" — while nothing was ever shown. The turn_updates system-prompt variant that warns the model ("Text you write between tool calls may not be shown to the user … restate it in the final message") is not active for this account.

No client or settings change happened at onset. Same onset time as #93596 ("Opus 5 at xhigh … since ~02:30 UTC 11 Sep, with no client or settings change").

Environment

• Claude Code CLI 2.1.266 → 2.1.278 (macOS, Darwin 27.0.0); Claude Desktop 2.2553.1 (bundles Claude Code 2.1.271 / 2.1.275)
• Model claude-opus-5[1m], effort xhigh, adaptive thinking (defaults), first-party claude.ai login, auto permission mode
• showThinkingSummaries unset, CLAUDE_CODE_THINKING_DISPLAY_UPDATES unset

Evidence (from ~/.claude/projects/**/*.jsonl, 60 days, ~57k assistant messages)

Block shape of a continuation response (the response that follows a tool_result), one message.id:

| | shape |
|---|---|
| healthy | thinking("" + signature) → text → tool_use |
| affected | thinking("" + signature) → thinking(summary prose + signature) → tool_use (no text at all) |

Continuation responses that contain a tool call, by day (first responses of a turn are never affected):

| day | Claude Code | cont. narration as text | as narration thinking | narration % |
|---|---|---|---|---|
| 09-07 | 2.1.260 | 485 | 0 | 0 |
| 09-08 | 2.1.260 | 986 | 36 | 4 |
| 09-09 | 2.1.260 / 2.1.267 | 581 / 10 | 16 / 0 | 3 / 0 |
| 09-11 | 2.1.266 | 37 | 141 | 79 |
| 09-12 | 2.1.266 | 8 | 122 | 94 |
| 09-14 | 2.1.270 | 32 | 216 | 87 |
| 09-16 | 2.1.271 | 408 | 266 | 39 |
| 09-17 | 2.1.271 | 426 | 228 | 35 |
| 09-21 | 2.1.275 / 2.1.278 | 93 / 0 | 57 / 13 | 38 / 100 |

The flip is by date, not client version (2.1.267 on 09-09 is clean; 2.1.266 on 09-11 is 79 %). No CLAUDE.md / hooks / settings changes on 09-10/09-11 (git history of ~/.claude).

Content of the narration blocks is a paraphrase of what the model wrote (e.g. my own text "거친 정규식이어서 잡음이 섞였지만 … 두 건을 상세히 봅니다" was stored as "정규식이 거칠어서 잡음이 섞였지만 … 두 건을 자세히 살펴보겠습니다"), written in first person, sometimes in past tense describing the model's own output ("I explained two ways to copy the secret value…" — while no text block exists).

Incident A (Desktop, 2026-09-16)

1. Model runs git worktree remove … via Bash → denied by the auto-mode classifier.
2. Next response: thinking("") + narration thinking: "…함장님이 터미널에서 제시된 명령어를 직접 실행해주셔야 합니다" ("you must run the command I presented in your terminal") + AskUserQuestion whose option description says "run the command above".
3. Final text: "위 명령을 터미널에서 실행하시면 …" ("if you run the command above…"). No text block in the whole turn contains the command; it only ever existed in the denied Bash tool_use input and in the narration block.
4. User: "아니 명령어를 또 안 줬네?" ("you didn't give me the command again").

Incident B (Desktop, 2026-09-21)

Narration block (no text block in that response): "Google 콘솔에서 시크릿 값을 복사하는 두 가지 방법(터미널 명령 또는 콘솔에서 직접 확인)을 안내해 드렸습니다" ("I have explained two ways…") → AskUserQuestion. User's answer: "네가 언제 안내해줬어?" ("when did you explain that?"). Next response again narration-only: "안내가 모달에 가려 안 보였을 수 있어 다시 정리했습니다" ("I restated it") — again no text block. The gcloud … | pbcopy command reached the user only on the third attempt, as a text block after the user typed "터미널 명령어를 출력하라고".

Controlled repro (claude -p … --output-format stream-json --verbose, 2.1.278)

Prompt asks for a sentence + a fenced echo HELLO-MARKER-42 before the second tool call.

• Run 1 (defaults): the pre-tool prose came back as thinking with narration_block_indexes: [0]; HELLO-MARKER-42 appears in no text block — the fenced command was summarized away.
• Run 2 (defaults): same prompt, prose came back as text, marker intact → the conversion is probabilistic (matches the 35–60 % above).
• Run 3 (CLAUDE_CODE_THINKING_DISPLAY_UPDATES=false): text, marker intact, no narration blocks (n=1, so not proof; the code path suggests it disables the display: "updates" request).

What I found in the client

• 2.1.278 requests thinking.display: "updates" with beta thinking-display-updates-2026-08-18 by default (connector_text mode; CLAUDE_CODE_THINKING_DISPLAY_UPDATES ?? true).
• The stream-json schema documents narration_block_indexes: "indexes … of the thinking blocks whose signature the server tagged block_kind 'narration' (server summaries of the prose between tool calls, not the model's own reasoning), so a renderer can label them as summaries".
• The platform docs ("Progress updates between tool calls") say under display: "updates" progress-update blocks come back as "Summary text", and list only Fable 5.1 / Mythos 5.1 / Fable 5 — Opus 5 started producing them on 09-11.
• Two system-prompt variants exist: one tells the model "Your text output is what the user reads between tool calls"; the other (turn_updates) says "Text you write between tool calls may not be shown to the user … restate it in that final message". This account gets the former.

Expected

• Anything the model writes for the user is either shown verbatim or the model is told it will not be shown. Today neither holds in the Desktop app: the prose is summarized (code blocks lost), the summary is not rendered, and the model is told the opposite.

Asks

1. Desktop: render narration blocks (as the CLI does), or do not request display: "updates" for Desktop sessions.
2. Server: when summarizing prose between tool calls, keep fenced code blocks verbatim (or pass the prose through unchanged when it contains a fence).
3. Client: whenever display: "updates" is requested, also send the turn_updates guidance so the model restates deliverables in its final message.
4. Document CLAUDE_CODE_THINKING_DISPLAY_UPDATES as the opt-out if it is one.

Workaround in use

Two local hooks: a PreToolUse(AskUserQuestion) hook that denies questions referring to "the command above" unless the fenced command is inside the question field, and a Stop hook that blocks a final message which points at an earlier command without restating it in a fence. Replayed over 30 days of transcripts: 27/924 and 1/913 hits, both real incidents caught.

import base64

def is_narration(block):
    if block.get("type") != "thinking":
        return False
    sig = block.get("signature") or ""
    try:
        raw = base64.b64decode(sig + "=" * (-len(sig) % 4))
    except Exception:
        return False
    return b"narration" in raw

# at Stop: if any block in the reply is_narration() and len(final_text) < 150,
# return {"decision": "block", "reason": "<quote the summaries, ask for a restatement>"}