opencode run silently saves an empty assistant stop message (tokens counted, zero parts) and exits 0

Summary

Headless opencode run --auto against the ollama provider (OpenAI-compatible /v1/chat/completions)
sometimes ends a session by saving an assistant message with finish:"stop", 118 output tokens counted,
and no parts array at all (no text, no tool call), then exits 0 with no error. The task is left incomplete
with no diagnostic. The streamed response on the wire is complete (ollama HTTP 200, truncated=0) and
contains the content/tool-call that opencode then fails to materialize as parts.

Repro signature (verified 2026-08-14, opencode 1.18.15, ollama on node20)

• opencode run --model ollama-node20/gpt-oss:20b --auto --dir <fixture> "<task>".
• Config for the model: tool_call: true, reasoning: true, options.reasoningEffort: "high",
  no interleaved, limit: { context: 32768, output: 8192 }.
• Fresh session, step 1, ~7K context (not a long-session / context-window-exhaustion effect).

Message sequence in ~/.local/share/opencode/opencode.db:

1. user — the task.
2. assistant — finish: "tool-calls", 301 output tokens. (Glob **/* → 5 matches.)
3. Glob executes.
4. assistant — finish: "stop", 118 output tokens, no parts key. parentID equals message 1's id
   (a sibling of message 2, not a child of it). Run exits 0. Loop never advances past step 1.

data of the dropped message:

{"parentID":"...","role":"assistant","mode":"build","agent":"build",
 "cost":0,"tokens":{"total":7235,"input":7117,"output":118,"reasoning":0},
 "modelID":"gpt-oss:20b","providerID":"ollama-node20",
 "finish":"stop"}

Ollama-side (journalctl -u ollama) shows both requests completed cleanly:

18:25:28 [GIN] 200 POST /v1/chat/completions  (301 tok, truncated=0)
18:25:31 [GIN] 200 POST /v1/chat/completions  (118 tok, truncated=0, 2.9s)

The second response contained the model's next tool call (a read); the CLI rendered its raw JSON —
{"filePath":"/private/tmp/opencode-local-test/mock-svc.sh","limit":200,"offset":1} — as a bare line,
then the run stopped. The tool call never became a part and never executed.

Wider pattern

Same signature (empty stop message, tokens counted, zero parts, exit 0, no error) observed repeatedly:

• 2026-08-13, gpt-oss:20b with interleaved:"reasoning" at 32K — dropped its final assistant message
  after a complete 200 (753 tokens); removal of interleaved previously seemed to fix it, but today's drop
  reproduces without interleaved and at low context, so it is not the interleaved setting.
• 2026-08-14, qwen3:30b-a3b — died silently on its 2nd LLM turn (complete 200, 308 tokens, empty stop, exit 0).
• 2026-08-14, qwen3.5:9b — three near-complete runs each ended at this signature before a final report.

Expected

If a stream produces output tokens and a normal finish_reason, the assistant message must materialize its
parts (text and/or tool calls) and the agent loop must continue. An empty finish:"stop" assistant message
that silently ends the run (exit 0) is data loss.

Environment

• opencode 1.18.15 (current); drop observed under the same config that produced a clean PASS earlier, so a
  version change between 2026-08-13 and 2026-08-14 is suspected as the trigger.
• Provider: ollama via api: "openai", baseURL http://<host>:11434/v1, streaming enabled.
• Models: gpt-oss:20b (reasoning model), qwen3 family (reasoning models with interleaved:"reasoning_content").

Suggested investigation

• Message write path when an assistant turn returns finish_reason:"tool_calls" (or "stop") — why are
  parts dropped while tokens are counted, and why does the dropped message share the user message as parent.
• Whether the drop correlates with reasoning:true (all observed cases are reasoning models).
• Whether stream chunk flushing has a race that drops the final chunk when the tool call ends the stream.