Repository navigation
tui: reasoning rendered as plain text for models that emit a closing thinking tag without an opening one #55414
Description
Description
Reasoning from the built-in opencode/big-pickle model is rendered in the transcript as normal assistant text instead of a separate collapsible thinking block. Hiding/toggling thinking has no effect, because no reasoning part is produced.
The model streams its reasoning followed by a closing </think> marker but never an opening <think> marker:
<reasoning text...></think><answer...>
OpenCode already has code to turn this pattern into a reasoning part, but it (a) requires the opening <think> marker and (b) is hardcoded to a single model id, so it never runs for this model:
- the streamed extractor only triggers when accumulated content
startsWith("<think>"), and the batch splitter looks forindexOf("<think>"); with no opening marker, everything remainstext; - both are gated by
function UVo(e){return ["qwen3-4b"].includes(e)}, i.e. tag-based splitting only runs forqwen3-4b.
Counting across all assistant messages in an active session: opening <think> appears 0 times, closing </think> appears 39 times (every message).
Expected: reasoning is separated into a reasoning part and can be hidden/toggled.
Actual: reasoning plus </think> is stored and displayed as ordinary text.
Plugins
Local plugins: nudge, pathfix (~/.config/opencode/plugins/)
OpenCode version
2.0.20 (Homebrew)
Steps to reproduce
- Select the model
opencode/big-picklefor a session. - Send any message.
- The response shows the model's raw reasoning as plain text, terminated by
</think>, with no separate thinking block.
Operating System
macOS (Darwin 25.2.0, arm64)
Terminal
iTerm2
Activity
[-]tui: reasoning rendered as plain text when a model emits only a closing thinking tag[/-][+]tui: reasoning rendered as plain text for models that emit a closing thinking tag without an opening one[/+]opencode-agent commentedon Oct 14, 2026
Thanks for the detailed report, @nick-o, especially the tag counts and the pointer to the existing extraction code.
I think this is a duplicate of #54599:
opencode/big-picklesends its reasoning inline, ending with</think>and no opening<think>, and OpenCode saves it as a normal text part instead of a reasoning block. That issue has already been reproduced, and there is a proposed fix in #54612.Possible duplicates and related issues, ordered by likelihood:
</think>stored astext, no thinking block), on v2</think>inside the reasoning makes the rest of the reasoning arrive as visible text, and a later comment there also captures the missing-opening-tag formYour notes on why the existing tag extraction never runs (it needs an opening
<think>and only applies toqwen3-4b) would be a useful addition on #54599.The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different.
nick-o commentedon Oct 14, 2026
Duplicate of #54599 (same model, same symptom:
opencode/big-picklestreams inline reasoning ending in</think>with no openingthinking, stored as atextpart instead of areasoningpart). Fix tracked in #54612. I've moved the code-level detail (why the current extraction never fires) to #54599. Closing as duplicate.opencode-agent commentedon Oct 14, 2026
Thanks, @nick-o, and thanks for moving the details about the extraction code to #54599. That's the best place to follow this, along with the proposed fix in #54612. The bot will close this issue as a duplicate shortly.
opencode-agent commentedon Oct 15, 2026
Closing as a duplicate of #54599, which you confirmed is the same problem:
opencode/big-picklereasoning that ends in</think>with no opening tag is saved as a normal text part. Please follow #54599 and the proposed fix in #54612. Thanks for adding your findings there, @nick-o.