Assistant text not displayed between tool calls — only permission prompts visible

[madorian]

Description

Intermittently, after submitting a prompt, agent-shell shows only the permission prompts (tool call approvals) but none of the assistant's text output between them. The LLM is clearly working — it makes tool calls, asks permission questions — but the actual response text that should appear between those interactions is missing.

The result is a sequence of permission prompts with no context: the user approves tool calls blindly because the reasoning/explanation text is invisible.

Steps

1. Submit a prompt in agent-shell
2. The prompt appears normally
3. Permission prompts appear (Allow/Reject) — the agent is making tool calls
4. No assistant text is rendered between the prompts
5. After answering a permission prompt, another one appears — still no text
6. The cycle continues with zero visible text output from the assistant

Expected behavior

Between tool calls/permission prompts, the assistant's text output (reasoning, explanations, status updates) should be rendered in the buffer.

Environment

• agent-shell: 1e5d175
• shell-maker: 6377cbd
• Claude Code: 2.1.104
• Emacs 30.2 on NixOS
• Terminal: emacs -nw (over SSH)

Notes

Intermittent — have not found a reliable reproduction. Will attach diagnostic data (buffer state, text properties, traffic log) next time it occurs.

[julianrubisch]

I think I am experiencing the same, however I can clearly pin it down to whenever I interrupt/reject a tool call. Then what the OP described happens. When I interrupt/reject again, the buffer becomes readonly

I'll then close the buffer and resume the session in a new one - but the chat is gone and I have to read through the transcript to gather what I was just working on...

Environment

• agent-shell: 20260408.1635
• shell-maker: 20260330.851
• acp: 20260325.115
• Emacs 30.1 on MacOS
• Claude Code: 2.1.110

[madorian]

New evidence: stale permission requests arriving after end_turn

I've captured ACP traffic from two separate sessions showing a related manifestation of this bug. Instead of (or in addition to) missing agent_message_chunk text, session/request_permission requests arrive after end_turn, creating orphaned permission dialogs that the user cannot meaningfully answer.

Case 1: Pi 400 session — single stale permission, 38-second delay

20:30:57  <- response  id:5  stopReason:"end_turn"
                       (prompt inserted, session complete)

20:31:35  <- request   session/request_permission
                       toolCallId: toolu_017cRYdjTRgiCxyuCzz8S4BR
                       kind: "execute"
                       title: ssh root@quasar "systemctl status dnsmasq-pi4 ..."

A background ssh command finished 38 seconds after the turn ended. The permission dialog appears below the already-visible Claude> prompt.

Case 2: Tripletex session — five stale permissions in rapid succession

09:36:53  <- response  stopReason:"end_turn"
                       (prompt inserted, session complete)

09:37:00  <- request   session/request_permission  (doveadm search 1)
09:37:14  <- request   session/request_permission  (doveadm search 2)
09:37:16  <- request   session/request_permission  (doveadm search 3)
09:37:31  <- request   session/request_permission  (doveadm search 4)
09:37:56  <- request   session/request_permission  (doveadm search 5)

Five background doveadm email searches completed 7-63 seconds after end_turn. All five permission dialogs stack below the prompt.

Buffer state (both cases)

busy: nil
active-requests: nil
buffer-read-only: nil

The session is complete, the prompt is visible, but permission dialogs keep appearing below it.

Impact

• Answering any dialog fails (no active request to correlate with)
• Rejecting via C-c C-c may leave the buffer read-only (Buffer becomes read-only after response: shell-maker--busy stuck at t with no active request #511)
• The user sees permission prompts with no context after the turn already ended
• In the tripletex case, the user reported "asking questions without any text in-between" — the permission dialogs look like the assistant is still working

Root cause

active-requests is cleared on end_turn, but background tasks spawned by the agent can complete afterward and trigger session/request_permission. The permission handler in agent-shell does not check active-requests-p before rendering the dialog, so orphaned dialogs appear.

This is also an upstream issue: background task notifications racing with end_turn is tracked in anthropics/claude-code#39027.

Possible agent-shell mitigation

The session/request_permission handler could check active-requests-p and auto-reject permission requests that arrive when no request is active, rather than rendering an unanswerable dialog.

Environment

• agent-shell: 0.50.1
• Claude Code: 2.1.104
• Emacs 30.2 on NixOS

[madorian]

Update: Further investigation shows that at least some instances of missing text are caused by the ACP server (Claude Code) not sending agent_message_chunk at all, rather than agent-shell dropping them.

In a session with 49 session/update notifications, only 2 contained sessionUpdate: "agent_message_chunk". The text was visible in Claude Code's terminal but never delivered via ACP. Filed and confirmed upstream:

• agent-shell Claude Code ACP: agent_message_chunk not sent between rapid tool calls #568 (closed — redirected upstream by @xenodium)
• Result text lost when output_tokens=0 (no stream events emitted) agentclientprotocol/claude-agent-acp#453 — root cause: text in message.result but no stream events emitted when output_tokens=0
• fix: stale PostToolUse notifications and background task output leak agentclientprotocol/claude-agent-acp#353 — open draft PR fixing stale PostToolUse notifications and background task output leak (related to the stale permission dialogs reported above)

So this issue likely has two independent causes:

1. agent-shell side (original report + our patch): the active-requests-p guard dropping delivered text
2. Upstream ACP side: text never sent via ACP in the first place

[xenodium]

agent-shell side (original report + our patch): the active-requests-p guard dropping delivered text

Working as intended. This is guarding against naughty servers that deliver out of turn notifications. https://agentclientprotocol.com/protocol/prompt-turn

I had come across this issue with Claude before and they eventually fixed it. Please update both claude-agent-acp and claude-agent to the latest.

Upstream ACP side: text never sent via ACP in the first place

This is a server issue. Should be reported to corresponding ACP server.