[Tryanks]

Summary

With thinking on, Claude Code 2.1.312 + Opus 5.5 do not deliver the prose the model writes to the user between tool calls as text. The server replaces it with its own summary, sent as a thinking block. Only the frame's narration_block_indexes tells this block apart from reasoning. Claude Code's terminal UI shows it as a message marked "· summarized". Tcode mapped it to Reasoning, so it was folded into the work log. A reply to a mid-turn steer therefore looked as if the steer had been ignored. In one local database, 335 of 2555 first replies after a steer (13%) ended up folded this way.

 on_assistant(frame)
   thinking block, non-empty
-    → Reasoning
+    listed in frame.narration_block_indexes → AssistantMessage { summarized: true }
+    otherwise                               → Reasoning

The streamed placeholder (Reasoning, same msg:index id) is replaced in place by the completed message, so the text renders once, as a message.

 ItemContent::AssistantMessage {
     text,
+    summarized: bool,   // serde default false, omitted when false
 }
 <AssistantMessage>
   markdown
+  "Summarized" / "已概括" caption (muted, xs) when summarized

summarized is a field on AssistantMessage, not a new variant. Narration is still the assistant talking to the user, so every existing reader of assistant text keeps it: relay, search, export, titles, the latest-message action row. Before the server began summarizing, that same interim prose arrived as plain text and those readers already included it.

Not covered: Claude transcript import and subagent transcript tails. Their JSONL has no narration_block_indexes. Telling narration apart there would mean decoding the signature envelope, which the CLI does internally and does not expose. Event logs written before this change keep their Reasoning entries.

Evidence

• Before: the status update the model sent mid-turn was persisted as {"kind":"reasoning","text":"…"}. The CLI's own transcript has the same block as type: "thinking", and its signature tags it block_kind narration. A live CLI run shows the frame field:
  assistant [('thinking', "f.txt contains two lines…")] {'narration_block_indexes': [0], 'thinking_display': 'updates'}
• After: narration_completes_its_streamed_block_as_a_summarized_message replays those recorded frames. The completion is AssistantMessage { summarized: true } under the streamed id msg_n:1. The test fails with the narration branch disabled and passes with it.
• assistant_messages_are_summarized_only_when_recorded checks the wire literals. A legacy record with no field decodes to false, and "summarized":true is written only when set.
• Gallery section 7 has a new "SUMMARIZED NARRATION" card, checked in light and dark at 1100 and 420 pt.
• cargo fmt --check, cargo clippy --workspace --all-targets -D warnings and cargo nextest run --workspace (1359 passed) were run on macOS. The iOS, Android and Web checks are left to CI.

Merge Danger

Door: two-way

The new field is additive and optional on the wire. Older logs and peers decode it as false; an older client ignores it and shows the narration as a plain message. Noted as Unreleased above PROTOCOL_VERSION.

Blast Radius: Claude timelines

Mid-turn narration moves out of the collapsed work log into the timeline as messages, so Claude turns with thinking on show more prose between activity runs.