fix(tui): surface thinking-only turns instead of silently ending by gsrunion · Pull Request #1727 · Hmbown/CodeWhale

gsrunion · 2026-05-17T04:01:24Z

Summary

gpt-oss (and any model whose response gets mapped to reasoning_content with empty content and no tool_calls) returns chunks where the only thing the engine has at the end of the stream is a Thinking block.
The persist site at crates/tui/src/core/engine/turn_loop.rs:863-877 was an if-only branch on has_sendable_assistant_content — when false, the assistant message was silently not persisted, no event was emitted, and if tool_uses.is_empty() fell straight through to "finish the turn." The UI sat idle with no spinner clear and no error.
This PR adds an else if turn_error.is_none() && !cancel_token.is_cancelled() branch that emits a status event so the user sees "Model returned reasoning but no answer or tool call; turn ended without output. Send a follow-up to retry." instead of silence. No assistant message is persisted (matches prior behavior).

Reproduction

Configure provider_models.ollama = "gpt-oss:120b" (or any gpt-oss variant) and show_thinking = true in ~/.deepseek/config.toml.
Run the TUI in agent mode against ollama.
Periodically gpt-oss emits an analysis-channel-only response (ollama's harmony→OpenAI shim puts the content into reasoning_content and leaves content/tool_calls empty).
Before this fix: spinner clears nothing, no error, no reply — UI looks hung.
After this fix: status event surfaces with the notice above; user knows the turn ended and can send a follow-up.

I caught this on my own machine debugging a "hang" — session ~/.deepseek/sessions/<id>.json showed the user typing "are you stopped?" after gpt-oss returned a 200 with no persisted assistant message.

Test plan

cargo test -p deepseek-tui passes
cargo clippy --workspace --all-targets --all-features passes
Manual: reproduce the gpt-oss thinking-only response against ollama and confirm the status event renders
Manual: confirm a normal text/tool_use turn still persists and renders unchanged (the if branch is untouched)
Manual: confirm a stream error still surfaces its own error event (no double-notice — guarded by turn_error.is_none())
Manual: confirm a cancelled turn doesn't fire the new notice (guarded by !cancel_token.is_cancelled())

Notes

I couldn't run cargo check locally — my build host is missing libdbus-1-dev and I didn't want to install system packages without asking. CI should cover it. The patch uses the same Event::status(String) / self.cancel_token.is_cancelled() / crate::logging::warn(&str) / turn_error.is_none() patterns already used a few lines above in the same function.
This is option 1 of the three approaches I considered (minimum-surface-area fix). Option 2 (re-prompt with a "produce a final-channel response" nudge) and option 3 (persist a placeholder text block) are deliberately not in this PR — option 1 unblocks the perceived hang without making policy decisions about retry behavior.

🤖 Generated with Claude Code

When a model returns a response with `reasoning_content` populated but empty `content` and no `tool_calls` (notably gpt-oss via ollama's harmony→OpenAI shim), the engine built an assistant message containing only a `Thinking` block, failed the `has_sendable_assistant_content` check, persisted nothing, and ended the turn — leaving the UI silently idle with no event to clear the spinner or explain why. Add an `else if` branch that emits a `status` event in that case, guarded by `turn_error.is_none()` and `!cancel_token.is_cancelled()` so the notice doesn't pile on when an error or cancellation has already been surfaced. No assistant message is persisted, matching prior behavior (some providers reject thinking-only assistant turns on the next request). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

gemini-code-assist

Code Review

This pull request adds a check to the TUI engine's turn loop to handle scenarios where a model returns reasoning content but no final answer or tool calls, ensuring the user is notified instead of the UI remaining idle. The reviewer suggested improving the accuracy of the notification by checking for pending steers and distinguishing between empty responses and reasoning-only responses.

gemini-code-assist · 2026-05-17T04:03:30Z

+            } else if turn_error.is_none() && !self.cancel_token.is_cancelled() {
+                // gpt-oss via ollama's harmony→OpenAI shim can return
+                // `reasoning_content` with empty `content` and no
+                // `tool_calls`, which would otherwise leave the UI silently
+                // idle while the engine quietly ends the turn.
+                let notice = "Model returned reasoning but no answer or tool call; \
+                              turn ended without output. Send a follow-up to retry.";
+                crate::logging::warn(notice);
+                let _ = self.tx_event.send(Event::status(notice.to_string())).await;
            }


The notice "turn ended without output" might be premature if there are pending steers or running sub-agents that will keep the turn alive. In such cases, the engine will proceed to the next step rather than finishing the turn, making the "ended" message confusing to the user.

Additionally, while this fix targets models returning reasoning without content, it would be more accurate to distinguish between a response that contains only reasoning and a truly empty response (where current_thinking is also empty).

Suggested change

} else if turn_error.is_none() && !self.cancel_token.is_cancelled() {

// gpt-oss via ollama's harmony→OpenAI shim can return

// `reasoning_content` with empty `content` and no

// `tool_calls`, which would otherwise leave the UI silently

// idle while the engine quietly ends the turn.

let notice = "Model returned reasoning but no answer or tool call; \

turn ended without output. Send a follow-up to retry.";

crate::logging::warn(notice);

let _ = self.tx_event.send(Event::status(notice.to_string())).await;

}

} else if turn_error.is_none() && !self.cancel_token.is_cancelled() && pending_steers.is_empty() {

// gpt-oss via ollama's harmony→OpenAI shim can return

// `reasoning_content` with empty `content` and no

// `tool_calls`, which would otherwise leave the UI silently

// idle while the engine quietly ends the turn.

let notice = if current_thinking.trim().is_empty() {

"Model returned an empty response; turn ended without output. Send a follow-up to retry."

} else {

"Model returned reasoning but no answer or tool call; turn ended without output. Send a follow-up to retry."

};

crate::logging::warn(notice);

let _ = self.tx_event.send(Event::status(notice.to_string())).await;

}

gsrunion · 2026-05-17T04:56:25Z

Update: installed libdbus-1-dev + pkg-config locally and cargo check -p deepseek-tui now passes clean (Finished dev profile in 25.16s, no warnings on the touched file). The "couldn't compile locally" caveat in the description is resolved — only manual repro / clippy / full test suite remain on the test plan.

Bump workspace, inter-crate, and npm package versions 0.8.38 -> 0.8.39. Roll CHANGELOG [Unreleased] into [0.8.39] with all fixes: - Revert v0.8.38 /model picker rework (back to instant curated picker) - Restore approval grouping (lossy v0.8.37 logic for approvals, exact key for denials) - Thinking-only turn surface fix (#1727) - ACP server JSON-RPC id stringification (#1696) - Chat client: reasoning_content for generic providers (#1673) - Compaction: user text query preservation (#1704) - Engine: system prompt override survival (#1688) - Pager: G/End overshoot fix (#1706), mouse scroll (#1716) - Composer: scroll with text (#1677), multiline arrows (#1721) - macOS system theme detection (#1670) - rlm_open blank source fields (#1712) - Terminal resize paging fix (#1724) - Docker first-run permission (#1684) - README Rust 1.88+ requirement note (#1718) Tests: 3149 passed, 0 failed (deepseek-tui crate) clippy: clean on --all-targets --all-features

Hmbown · 2026-05-17T09:02:28Z

Keeping this in the v0.8.40 milestone as a candidate, but not merging it on narrative alone. Before it lands we need fixture coverage for a thinking-only/no final-answer stream, plus a check that normal text/tool-use turns and cancellation/error paths do not double-surface status. The report is useful; the next step is to make the repro deterministic.

…own#1742) Address gemini-code-assist review on PR Hmbown#1742 (MEDIUM): the status was emitted at the assistant-persist site, but the same turn can still CONTINUE for pending steers or sub-agent completions -- the user would see a spurious 'turn ended without output' notice immediately before the turn resumed. Capture thinking_only_no_sendable at the persist site (no emission there) and decide at the end of the tool_uses.is_empty() path, just before the terminal break -- reachable only when there were no pending steers, no sub-agent completions, and we were not holding for running children. Extend should_emit_thinking_only_status with steers_pending and holding_for_subagents (false if either), recomputed live at the decision point as defense-in-depth. Unit test updated with the two new no-emit cases. Refs Hmbown#1727.

Hmbown · 2026-05-21T01:00:28Z

Thanks for raising the thinking-only turn behavior. The release-safe fix was harvested into v0.8.40 release PR #1823 as commits 4ab7c53 and 351898d, and #1823 is now CI-green. Closing this older PR as superseded by the release branch; thank you for pushing this regression into view.

…own#1727) When a model streams a turn with only a reasoning block (empty content, no tool_calls -- e.g. gpt-oss via ollama's harmony->OpenAI shim mapping to reasoning_content), has_sendable_assistant_content is false: the if-only persist branch was skipped, NO event was emitted, and the turn fell through to break. The UI spinner hung with no reply and no error. Add an else-if at the same persist site that, only on a clean end (tool_uses empty, turn_error.is_none(), not cancelled), warns and emits an Event::status notice telling the user the turn ended and they can retry. No assistant message is persisted (prior behavior preserved); no retry/re-prompt/placeholder policy is added (out of scope). Guard ordering extracted into a pure should_emit_thinking_only_status() helper with a unit test, matching the existing should_hold_turn_for_subagents style. Maintainer audit Hmbown#1736 confirms Hmbown#1727 is not shipped in v0.8.39 and release-blocking. Refs Hmbown#1727, Hmbown#1736.

…own#1742) Address gemini-code-assist review on PR Hmbown#1742 (MEDIUM): the status was emitted at the assistant-persist site, but the same turn can still CONTINUE for pending steers or sub-agent completions -- the user would see a spurious 'turn ended without output' notice immediately before the turn resumed. Capture thinking_only_no_sendable at the persist site (no emission there) and decide at the end of the tool_uses.is_empty() path, just before the terminal break -- reachable only when there were no pending steers, no sub-agent completions, and we were not holding for running children. Extend should_emit_thinking_only_status with steers_pending and holding_for_subagents (false if either), recomputed live at the decision point as defense-in-depth. Unit test updated with the two new no-emit cases. Refs Hmbown#1727.

gemini-code-assist Bot reviewed May 17, 2026

View reviewed changes

Hmbown mentioned this pull request May 17, 2026

v0.8.40 regression bug bash: audit v0.8.38/v0.8.39 fallout #1736

Closed

21 tasks

Hmbown added this to the v0.8.40 milestone May 17, 2026

LeoLin990405 mentioned this pull request May 17, 2026

fix(tui): surface thinking-only turns instead of silently ending (#1727) #1742

Closed

Hmbown mentioned this pull request May 20, 2026

chore(release): prepare v0.8.40 #1823

Merged

11 tasks

Hmbown modified the milestones: v0.8.40, v0.8.41 May 21, 2026

Hmbown closed this May 21, 2026

Hmbown mentioned this pull request May 23, 2026

fix(subagent): keep agent_eval recoverable after a child terminates (#1738) #1928

Merged

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(tui): surface thinking-only turns instead of silently ending#1727

fix(tui): surface thinking-only turns instead of silently ending#1727
gsrunion wants to merge 1 commit into
Hmbown:mainfrom
gsrunion:fix/surface-thinking-only-turn

gsrunion commented May 17, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot May 17, 2026

Uh oh!

gsrunion commented May 17, 2026

Uh oh!

Hmbown commented May 17, 2026

Uh oh!

Hmbown commented May 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

gsrunion commented May 17, 2026

Summary

Reproduction

Test plan

Notes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot May 17, 2026

Choose a reason for hiding this comment

Uh oh!

gsrunion commented May 17, 2026

Uh oh!

Hmbown commented May 17, 2026

Uh oh!

Hmbown commented May 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants