Summary

When using the OpenCode provider with models that stream reasoning as field=text deltas (e.g. ollama-cloud/kimi-k2.7-code), the model's chain-of-thought was being rendered concatenated with the actual answer.

Root cause (verified against a live OpenCode 1.15 SSE capture): for this provider the turn streams as message.part.delta events only — no message.part.updated/message.updated ever arrive. Both the hidden reasoning part and the visible answer part come through as field: "text" deltas, distinguishable only by partID. The previous logic set its inReasoningPhase flag exclusively from a part.updated type=reasoning event, which never fires here, so reasoning fell through to emit('text').

Changes

• Classify each partID as reasoning/text/other via a cached REST lookup (GET …/message/{mid}/part/{pid}).
• Buffer a part's field=text deltas until its kind is resolved, then route: reasoning → hidden thinking summary, text → transcript.
• message.part.updated now also populates the cache (recordPartKind) so providers/versions that send it avoid the REST round-trip.
• New per-turn state reset; extracted emitTextDelta / appendReasoningDelta helpers.
• On any lookup failure, defaults to showing content (never silently hides the answer).

Parts stream sequentially, so this only adds a one-localhost-round-trip (a few ms) delay to the first delta of each part, and it correctly handles agentic flows where reasoning/text/tool parts interleave.

Test plan

• New regression test simulating the Kimi flow (field=text deltas keyed only by partID, resolved via mocked REST) — reasoning hidden, answer shown
• tsc -b clean, eslint clean (0 errors)
• Full opencode-process.test.ts suite green (97 tests)

🤖 Generated with Claude Code