Opus 5: long agentic turns degrade to tool-calls-only, hours with no assistant text (not the #84153 transcript loss)

Summary

In long single-turn agentic sessions, claude-opus-5 sometimes stops emitting assistant text entirely: hours of tool_use blocks with interleaved thinking and no prose between tool calls. The session keeps working correctly (tools, hooks, task completion all normal), but a human supervising it has nothing to read, and since thinking content is omitted from display and transcript, the session is completely unobservable while it works.

Environment

• Claude Code 2.1.223 and 2.1.224, macOS (Darwin 24.5), interactive TUI under tmux
• Model claude-opus-5 with 1M context, effort high, bypass permissions
• Autonomous workflow: one user prompt starts a single working turn that runs 1-4 hours, tool-heavy (shell, file edits, simulator UI automation with screenshot reads)

Symptom

Two sessions on consecutive days over the same recurring workload:

• Session A (2.1.223): 303 assistant messages, 193 tool_use, 3 text blocks, all in the final minutes of a ~3.5h turn
• Session B (2.1.224): 352 assistant messages, 242 tool_use, 2 short text blocks in the first 22 minutes, then zero text for the remaining ~100 minutes until completion

18 comparable sessions (same version, model, effort, near-identical ~74k-token opening prompt) emitted 10-97 mid-turn text blocks over similar durations. So this is an occasional degeneration into an act-only cadence, not the model's normal behavior.

Ruled out

• The #84153 / #77651 transcript text-loss family: I could not reproduce it on 2.1.224. A -p --output-format stream-json run produced the text -> thinking -> tool_use shape three times and every text block persisted to the session .jsonl; an interactive TUI run producing the same shape both rendered and persisted all narration. The silent sessions' transcripts therefore look trustworthy: the model genuinely emitted no text.
• Context size: a prose-emitting sibling session peaked at 576k tokens; one silent session peaked at 288k.
• Vision content: a sibling session ingested 18 screenshots and wrote 72 text blocks after the first one.
• Prompt/config: identical opening prompt template, same effort setting, no user input mid-turn.

Impact

Supervised-autonomy setups steer by reading mid-turn narration. When it stops, and thinking is invisible at every layer, the operator flies blind for hours and can only evaluate correctness after the turn ends.

Ask

1. Is there guidance or a planned model-side fix for keeping inter-tool narration alive in very long single turns? The system prompt's communication guidance appears to lose to whatever cadence the model settles into early in the turn.
2. An observability fallback would also resolve the pain: a supported way to surface summarized thinking in the TUI and transcript (equivalent of display: "summarized") so silent stretches remain inspectable.

No minimal repro; it needs an hours-long agentic turn on a real workload and reproduces intermittently on the same workload. I can share per-message block-shape statistics on request (not full transcripts).