Skip to content

Claude's finished answer folds under "Continued for" when a steer is read after Claude ends its turn #40259

Closed

Description

@Bishi

What happened

I asked Claude a question, then sent a steer about something else while it was working. Claude's full answer to my question (4,245 characters, with headings and a numbered list) ended up hidden under "Continued for 1m 43s". Only the short closing message about the steer stayed visible. Nothing in the visible text says an answer is folded above it.

Diagnosis

Claude ended its turn on the answer. T3 then ran the steer in the same run and attempt, so the answer is no longer the run's last message and gets folded.

Provider log: the answer's API response started at 07:58:04, before the steer existed. The steer was offered at 07:58:15 and Claude Code queued it. The answer ended with stop_reason: "end_turn", and Claude Code emitted result: success at 07:58:35.584. The steer started 19 ms later and got its own result at 07:59:58. Since #39018 (in v0.1.11), T3 keeps the Claude run open for the unread steer, so both turns land in one run with one attempt (completed, not superseded).

In MessagesTimeline.logic.ts at v0.1.11, deriveTerminalAssistantMessageIds (line 746) keeps only the last assistant message per run. retainedAssistantReplyEntryIds (line 792) has two escape hatches, and neither applies:

  • phase: "final_answer": only CodexAdapterV2 sets it. ClaudeAdapterV2 never does, at v0.1.11 or on main.
  • A completed reply that finished before its attempt was superseded: this is intended, per the test retains phase undefined ending at 4 only when it is a final reply: true. Claude's in-run steering never supersedes the attempt.

So deriveTurnFolds hides the answer. The steer's createdAt is earlier than the answer's, so the answer also sits below the steer, inside its "Continued for" segment.

#34501 fixed this display bug for Codex only.

Steps to reproduce

  1. In a Claude Agent thread, ask a question that needs a few tool calls before Claude answers.
  2. While Claude is working, and ideally while it streams its answer, send an unrelated steer.
  3. Claude finishes the answer (end_turn), then reads the steer and does more work in the same run, ending with a short message.
  4. When the run settles, the answer is under "Continued for …" below the steer. Only the steer's closing message is visible.

Expected: the completed answer stays visible, as it does for Codex since #34501 and for a superseded-attempt steer. Ideally it also sits above the steer, since it replies to the earlier message.

Version

0.1.11 (server version from /.well-known/t3/environment; triage CLI is also 0.1.11)

Environment

Windows 10 Pro x64 (10.0.19045), desktop app with its bundled local server, Node v26.13.1, Claude Agent provider, Claude Code 2.1.329, claude-opus-5-5

Evidence

# What this shows: Claude finished its answer and ended its turn (result #1).
  # Only then did it read the steer and start a second turn (result #2).
  # T3 stored both turns as ONE run with ONE attempt, so the fold keeps only
  # the last message, and the 4,245-char answer is hidden.

  # Provider event log (Claude Code SDK events, trimmed)
  07:58:04.333Z  answer starts streaming          (steer doesn't exist yet)
  07:58:15.599Z  steer sent to Claude Code        prompt.offer
  07:58:15.700Z  Claude Code queues the steer     command_lifecycle state=queued
  07:58:34.795Z  answer ends                      stop_reason=end_turn
  07:58:35.584Z  result #1                        subtype=success, stop_reason=end_turn
  07:58:35.603Z  Claude Code reads the steer      command_lifecycle state=started
  07:58:39.367Z  steer response starts streaming
  07:59:58.449Z  result #2                        subtype=success, stop_reason=end_turn

  # orchestration_v2_projection_run_attempts: one attempt, never superseded
  attempt:1  status=completed

  # orchestration_v2_projection_turn_items for the run (no item has a phase)
  319000021  user_message       (the steer; sorted above the answer)
  319000022  assistant_message  4,245 chars   answer to the original question  <- FOLDED
  319000024  assistant_message                note about the steer             <- folded
  319000029  assistant_message                note about the steer             <- folded
  319000036  assistant_message  1,903 chars   closing message for the steer    <- only one visible

Related issues

#8879 (closed as not planned, "folding is intended") covers the general case: progress notes that come after an answer in one provider turn. This report is different. Claude's turn ended on the answer (result, end_turn), and only T3's steer handling joined the next turn to it. The equivalent Codex bug was fixed in #34501, and the test above shows a pre-steer reply is meant to stay visible. Also related: #39018 (keeps the Claude run open for an unread steer), #36384 (send_to_user, which only helps when the model calls it), #28161 and #35977 (earlier Claude and steer reports, closed as duplicates of #8879 before #39018 landed), #40108 (the same shape for Antigravity background wakeups).

Fix applied or workaround

Nothing changed on the machine. Workaround: expand "Continued for …" after any run you steered, or send a queued follow-up instead of a steer while an answer is coming.

Filed by

Claude Code (claude-opus-5-5) via t3 triage

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

    via-triageFiled through npx t3 triage

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions