llm: stray closing think marker persisted as visible assistant text (opencode/big-pickle) #56104

Summary

With the built-in opencode/big-pickle model, reasoning is intermittently persisted as an
ordinary text part instead of a reasoning part, and the literal closing thinking marker
</think> leaks into the visible assistant text. In 2.0.29/2.0.30 the marker appeared at
the start of the reply (</think>Готово…); 2.0.31 reduces that shape roughly 16x, but a
rarer variant remains where the whole thinking trace is stored as text ending in the marker.

Environment

• opencode version: 2.0.31 (the issue also reproduced on 2.0.29 and 2.0.30)
• OS: Windows 10 19045.6456 (10.0.19045)
• Terminal: VS Code integrated terminal (TERM_PROGRAM=vscode, TERM=xterm-256color, COLORTERM=truecolor)
• Shell: cmd.exe (C:\Windows\system32\cmd.exe)
• Install/channel: npm global @opencode/cli, prefix D:\Programs\0.IA\npm-global
• Active plugins: none found in config

Reproduction

1. Select model opencode/big-pickle.
2. Send a prompt that makes the model think and then call a tool, e.g.
   вызови инструмент glob 12 раз последовательно, по одному вызову за шаг.
3. Inspect the stored assistant parts (session_message in opencode.db) or watch the
   rendered reply.

The marker sometimes shows up as a separate leading text part and sometimes as part of the
assistant's final message, so the visible symptom is a stray </think> before the answer.

Expected Behavior

Reasoning should be stored in a reasoning part, and the thinking / </think> markers
should never appear in visible assistant text nor be replayed to the model as assistant
content.

Actual Behavior

A text part starts with, or contains, the literal marker </think>, and because it is
persisted as an ordinary text part it is also replayed into the next request as assistant
content.

Raw stored examples from 2.0.29 / 2.0.30 (marker at the start, immediately before the answer):

"</think>Готово.\n\n- `churnCount`: ..."
" responseГотово, можно проверять.\n\n- ..."
" responseГотово.\n\n- ..."

In that session 32 assistant messages were affected. Across the whole database: 191 of
12059 assistant messages, always with the same model (opencode/big-pickle), never with
other providers configured on the machine.

Additional Context

Measured on one machine with the same model, reading the session DB and counting assistant
messages that contain a tool call:

So 2.0.31 largely removes the highly visible "marker before the answer" shape, but the
underlying "reasoning stored as text" problem is not fully gone. Residual 2.0.31 examples
(raw stored text parts):

"The glob tool isn't in the Code Mode catalog - only opencode tools are. ..."
"Try search({query:...}) or search({pattern:...}).</think>"
"</think>"   (lone text part immediately before a tool call)

The leak is frequency-based, not deterministic: the same model and the same prompt produce
it only in a subset of tool-call turns (~18% before the upgrade, ~1% after).

Related: #34126 (standalone </think> before tool_calls persisted as assistant text) and
its fix PR #34698, which was closed without being merged; #43770 (reasoning inlined in
content leaks into text); #11439.

Note: in the v2.0.31 binary the literal marker strings are not present, while
reasoning_content / reasoningField / reasoningEnd are. Worth verifying that the
reasoning-to-tool boundary suppression actually ships in v2 builds (the v1 fix lived in
packages/llm/src/protocols/openai-chat.ts).

Workarounds tried

Upgrading 2.0.29 -> 2.0.31 reduces the visible marker-before-the-answer shape from ~16% to
~1% of tool-call messages, but does not eliminate the leak entirely. No configuration
workaround found.