Skip to content

[Agent Mode] Final answer renders inside the thinking block instead of visible text #3293

Description

@brevilabs-agent-bot

Context

On Copilot 4.0.8 in Agent Mode, a turn completes and prints "Thought for n seconds" with no visible reply in the chat. The answer is not lost: expanding the thinking dropdown shows the full, finished response. The turn does not hang, and it is not a refusal.

Appending a phrase like "respond directly with no internal reasoning" to the prompt routes the answer to visible output, which points at where the final assistant text is classified rather than at the model declining to answer.

Reported in https://discord.com/channels/1182847556032659546/1548431547777024050, with a second user confirming it independently in the same thread.

Reproduction details from the reporter

  • Copilot 4.0.8
  • macOS Tahoe 26.6.2
  • Agent Mode, sidebar, Copilot Agent icon selected
  • The turn finishes rather than staying in flight
  • Screenshots in the thread show the completed turn with no visible answer and the same text present in the expanded thinking block

Intermittent rather than constant. The reporter saw it repeatedly in one long-running conversation in which AGENTS.md had been edited heavily during the session, and it "barely" recurred after starting a fresh chat. Whether conversation length or a large edited AGENTS.md is a trigger is unconfirmed.

Logs were not collected. The reporter declined on the grounds that they were too large to review for private content, and the team agreed they were not needed given the screenshots and version details.

User experience

Today a user asks a question, watches the turn finish, and sees nothing. The work was done and paid for, and the only way to reach it is to know that answers sometimes hide inside a dropdown labelled as the model's private reasoning. A user who does not know that reads it as the agent ignoring them.

Afterwards a completed turn always renders its final answer as visible text, and the thinking block holds only reasoning.

Proposed behavior

  • A completed Agent Mode turn renders its final assistant message as visible chat text.
  • Content classified as reasoning is not the only place a finished answer appears.
  • Where a backend can emit a turn with reasoning and no final message, that case is handled explicitly rather than rendering an empty turn.

Open questions / Risks

  • Not yet known whether this is stream parsing assigning the final message to the reasoning channel, or the backend genuinely returning reasoning with no final message.
  • Two reporters on different setups, so it is unlikely to be one corrupted conversation.
  • Intermittent, so a fix needs a reproduction before it can be verified.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions