work-items: add openai unhandled sse observability

work-items/open/20260527-201313-openai-responses-unhandled-sse-observability/item.md

@@ -0,0 +1,21 @@
+---
+id: 20260527-201313-openai-responses-unhandled-sse-observability
+slug: openai-responses-unhandled-sse-observability
+title: OpenAI Responses 未対応 SSE event を破棄せず観測する
+status: open
+kind: feature
+priority: P1
+labels: [llm, openai, observability, trace]
+created_at: 2026-05-27T20:13:13Z
+updated_at: 2026-05-27T20:13:30Z
+assignee: null
+legacy_ticket: null
+---
+
+## Background
+
+Created by tickets.sh.
+
+## Acceptance criteria
+
+- TBD

work-items/open/20260527-201313-openai-responses-unhandled-sse-observability/thread.md

@@ -0,0 +1,66 @@
+<!-- event: create author: tickets.sh at: 2026-05-27T20:13:13Z -->
+
+## Created
+
+Created by tickets.sh create.
+
+---
+
+<!-- event: plan author: orchestrator at: 2026-05-27T20:13:30Z -->
+
+## Plan
+
+## Background
+
+OpenAI Responses parser currently drops unsupported SSE event types by falling through to `_ => Ok(Vec::new())`. That means provider events that do not yet have a match arm are neither visible in stream trace nor preserved as diagnostics. This made it impossible to inspect the "unexpected event" class of logs after the fact.
+
+Recent work preserved diagnostics for known error-like event types (`response.failed`, `response.incomplete`, top-level `error`), but it did not cover event types that are not matched at all. We need observability for those raw/unhandled SSE frames without turning them into conversation history or model-visible content.
+
+## Requirements
+
+- OpenAI Responses SSE event types that are not otherwise handled must be observable.
+  - Do not silently return `Ok(Vec::new())` without any traceable signal.
+  - Include the raw `event_type` and a bounded preview of `data`.
+  - Include full data length so truncation is visible.
+- The signal must be visible in existing stream trace when `[session].record_event_trace = true`.
+- The signal must not become assistant/user history and must not be sent back to the model as normal content.
+- Timeline / collectors must ignore the signal for generation semantics.
+- Known intentionally ignorable events may be classified separately if needed, but they must still be observable enough for debugging.
+- Add tests for at least one unknown OpenAI Responses event type.
+  - Existing `unknown_event_is_ignored` should be replaced or updated.
+  - Verify event type and data preview are retained.
+  - Verify large data is bounded / marked by length.
+
+## Suggested implementation shape
+
+A small normalized event variant is acceptable, for example:
+
+```rust
+Event::UnhandledSse {
+    provider: "openai_responses",
+    event_type: String,
+    data_preview: String,
+    data_len: usize,
+}
+```
+
+or equivalent. If adding a generic variant to `llm_client::event::Event`, make sure Timeline ignores it and trace serialization captures it.
+
+Avoid plumbing raw SSE into session history. This is observability only.
+
+## Acceptance criteria
+
+- Unknown OpenAI Responses SSE event types appear in trace output instead of disappearing.
+- Timeline semantics / assistant output are unchanged for unknown events.
+- Large raw data is capped in the event payload but original byte length is recorded.
+- Focused tests pass for OpenAI Responses parser and Timeline behavior if touched.
+- `cargo fmt --check` and related crate tests pass.
+
+## Out of scope
+
+- Implementing semantics for every OpenAI Responses event type.
+- Retrying or changing behavior based on unknown events.
+- Raw SSE frame permanent audit log separate from trace.
+
+
+---