Summary

The agent sometimes opens an approval question that refers to a table, findings, or a proposed change that it has not published in the regular conversation. The user sees choices such as "Approve all rows" or "Add this fix" without the information needed to make that decision.

This initially looked like a table-rendering problem, but inspecting session events found examples where the supporting content was absent from both the normal assistant message and the ask_user question payload. In another example, the explanation appeared only in the expandable Thinking section.

Actual behavior

Example 1: Approval for a missing triage table

During a PR review workflow on October 6, 2026 (times in CEST):

Recorded model for both approval calls: claude-opus-5.5.

I had also reported the same missing-table symptom during a review workflow on September 30.

Example 2: Explanation appears only under Thinking

On October 20, 2026 at 17:23:33 CEST, the agent asked whether to implement a proposed fix, referring to it as "this fix".

The question offered a choice between implementing the fix and documenting a limitation, but did not explain the fix or its tradeoff. Supporting explanation appeared only under Thinking, not as a regular assistant message.

The matching session event had empty assistant message content. The ask_user arguments contained only the short question and its two choices.

Recorded model for this approval call: claude-opus-5.5.

The relevant UI observation is that explanatory material was displayed under Thinking rather than in the normal conversation before approval was requested. No reasoning transcript, code, project details, or session identifiers are included in this report.

Expected behavior

Before asking for approval, the agent should publish the supporting table, findings, proposed change, and relevant tradeoffs in the regular conversation, or include sufficient context directly in the question.

Approval should not depend on expanding Thinking, and prompts should not refer to "the table above" or "this fix" when that context has not been presented to the user.

Reproduction and investigation

This is intermittent. The observed flow is:

1. Ask the agent to assess review feedback or investigate a problem.
2. The agent prepares findings or a proposal and requests approval through ask_user.
3. The approval prompt appears without the supporting context in the regular conversation.

I tried reproducing a table-rendering failure by putting a small table and then a longer, 14-row table directly inside an ask_user question. Both were fully visible and readable. Table length alone therefore did not reproduce the issue.

The evidence points to supporting content not being published in these examples, rather than content being truncated by the question renderer. It does not establish whether the underlying cause is model output, message handling, or another part of the app.

Impact and workaround

The user cannot make an informed approval decision and must ask the agent to show the missing context, sometimes more than once.

Workaround: have the agent publish supporting information as a standalone chat message before requesting approval.