Why: liveToMessage emitted all text buckets then all tools, while the
persisted message interleaves per turn (thinking, text, calls), so parts
reordered on reconcile with 2+ tool turns. Commentary stayed in the reply
area until the next turn's text bucket appeared, because the process flag
was positional (i < length - 1) instead of tool-driven.
Evidence: LiveState now holds an arrival-ordered items stream
(thinking/text deltas merge into their last segment; tools land in place),
so event order - which matches the backend's per-turn append order - maps
1:1 onto persisted parts. A text segment is process iff a tool.call lands
before the next thinking/text segment, aligning with the backend's
Process: len(calls) > 0; the collapse predicate moved to testable
showProcessInline. Long replies (>800 chars) now show a 100-char excerpt
and label the count as Markdown source length. vitest 27/27 pass (was 5
live cases, now 12 incl. multi-turn order, immediate process marking,
collapse predicate); tsc -b clean.
Impact: streaming UI no longer reshuffles parts across turns or on final
reconcile; tool-turn commentary folds into the process area as soon as the
tool call arrives. Residual edge (documented in live.ts): a silent
tool-only turn with no text and no thinking attributes its call to the
previous text segment.
Refs: code review of feature/chat-process-split-r1 (2026-09-30)
weichen
754ce83ffd