TUI: streamed markdown list items truncated when items contain emphasis spans (full text persisted in storage)

Environment

• OpenCode: 1.18.29
• Terminal: iTerm2 on macOS
• Model where observed: GLM 5.3 flash (ollama-cloud provider); not observed with other models used in the same setup

Bug

Numbered-list items in streamed assistant responses get truncated in the TUI display when the item text contains inline emphasis (**bold**). The truncation is display-only — the full text is persisted correctly in storage.

Evidence that it is display-only

The assistant message is stored complete in ~/.local/share/opencode/opencode.db (part table, type='text') — all list items fully written — while the TUI shows only the first few words of each item. A response with 4 numbered items displayed as:

What was done this session:

1. Four hardening fixes,

2. All four

3. The

4. Path forward locked in the handover, in order: merge

…while the stored text contains the complete items (each several sentences long). Verified by querying the SQLite part table directly.

Pattern

Across multiple occurrences in one session, the cut points align with inline emphasis spans near the start of list items. The affected model habitually begins list items with **bold** lead-ins; models that use emphasis sparingly do not trigger it. Plain-text responses without emphasis render fully, including the same content re-requested as plain text.

Suspected mechanism (inference — not verified against source)

The streaming markdown renderer appears to leave an unterminated emphasis span when a streaming chunk boundary lands mid-span; the parser state then swallows the remainder of the list item (and sometimes subsequent items) from the display. Storage is unaffected, so the loss is purely in the incremental renderer.

Reproduction

Ask a model to produce a streamed numbered list where each item begins with a bold lead-in (e.g. **Point one.** Explanation…). Observe items truncated in the TUI at or near the first emphasis span.

Workaround

Requesting plain text without markdown emphasis renders fully.

Happy to provide a stored-vs-displayed diff or additional examples if useful.