Assistant text preceding a tool call is reclassified as reasoning and summarized into "Thought for Ns" (never shown to the user)

Describe the bug

When the model emits a substantial reasoning block, followed by a multi-paragraph user-facing text block, followed by a tool call in the same turn, the CLI does not display the text block. Instead, the text is folded into the collapsed "Thought for Ns" region, where it appears only as a summarized paraphrase appended to the reasoning summary. The verbatim message never appears anywhere in the transcript.

This is not specific to any one tool. In one session it happened before a bash (Shell) call and before an ask_user call. The ask_user case is where it's most damaging: the model writes its explanation, then asks a question. The user receives a bare form with no explanation, which reads as the assistant "thinking privately and then asking a terse question."

Evidence that the text is being merged into the reasoning summary, not merely hidden: expanding the "Thought for 19s" block shows the reasoning summary ending at …before proceeding. immediately followed with no space by Reading surfaced real issues grep missed: inline SVG octicons…, which is a compressed version of the user-facing message the model had written (a numbered findings list plus a 3-bullet proposal). The missing space is the seam where two separately-summarized chunks were concatenated.

The bug is intermittent, which has made it hard to pin down; over months it presented as "the assistant seems to think I have context I was never given." Within one session, six turns with text-before-tool-call were observed:

Only the two turns with both a long reasoning block and a multi-paragraph text block lost the text. Either alone rendered fine.

Also observed, secondary: the ask_user message parameter does appear in the transcript afterward ("Asked user …"), but while the form is live, only the field title and option labels are visible. So if the explanation was dropped, the user sees just e.g. "How to proceed" with three options.

Affected version

GitHub Copilot CLI 1.0.84-1

Steps to reproduce the behavior

Reproduction is probabilistic; the trigger appears to be the combination below.

1. Start an interactive copilot session on a non-trivial task (one that makes the model reason at length, e.g. analysing a codebase and proposing a plan).
2. Ask a question whose natural answer is: think hard → write a multi-paragraph explanation → then call a tool (e.g. "investigate X and then ask me how to proceed", or anything that ends with a shell command after an explanation).
3. Observe the transcript: where the multi-paragraph explanation should be, there is only Thought for Ns, followed directly by the tool call (or the ask_user form).
4. Expand the Thought for Ns block. Its final sentences are a paraphrase of the missing explanation, sometimes concatenated to the reasoning summary without a separating space.

To confirm the text existed on the model side: ask the assistant in the next turn "what exactly did you emit last turn, in order?". Tt will report a text block before the tool call that was never displayed.

Sanity check that rules out "the model just didn't write anything": in the same session, a deliberately long (~120 words, multi-paragraph, bulleted) preamble before an ask_user call did render, in a turn where the model did very little reasoning first. Same text shape, different reasoning length, different outcome.

Expected behavior

• Every user-facing text block the model emits is displayed verbatim, regardless of what precedes or follows it in the turn.
• Reasoning and user-facing text are never merged. The "Thought for Ns" region should only ever contain (a summary of) reasoning content.
• When ask_user is invoked, any assistant text emitted earlier in the same turn should be visible above the form so the question has context.

Additional context

• OS: CachyOS Linux (Arch-based), kernel: Linux
• CPU architecture: x86_64
• Terminal emulator: Ghostty (TERM=xterm-ghostty)
• Shell: zsh
• Model in the session: Claude Fable 5.1 (claude-fable-5.1); reasoning display shows summarized reasoning, not raw. claude-fable-5 also previously ran into this problem multiple times

Screenshots:

• Transcript view. Note Thought for 34s → Shell "Baseline check…" with no assistant text between them, and Thought for 19s → Asked user … with no assistant text between them, while shorter preambles in neighbouring turns ("Content moved a lot…", "You're right, and the baseline already proves it…") render normally.
• Same region with the Thought for 19s block expanded. The second paragraph ends …before proceeding.Reading surfaced real issues grep missed: …, the portion after the missing space is the summarized user-facing message.

Happy to provide --log-level debug --log-file output from a reproduction if that would help; not captured for the session above.

Inizia riproducendo una sessione interattiva di copilot con un ragionamento lungo, un blocco di testo composto da più paragrafi e una chiamata successiva a bash o ask_user. Traccia il rendering della trascrizione intorno alla regione compressa "Thought for Ns" e la gestione delle chiamate agli strumenti. Il lavoro è completo quando ogni blocco visibile all’utente appare alla lettera prima della chiamata allo strumento, il ragionamento rimane separato e ask_user mostra la spiegazione precedente.