ai-providers: one shared SSE frame loop (FrameHandler) and response-to-event-stream helper #855

Goal

The streaming loop that reads SSE frames with content timeout, keep-alive handling, transport errors and abrupt end of stream exists once. Providers only implement how a frame is parsed and how the stream is finalised.

Context

• process_frame_stream exists three times: anthropic.rs:1453, openai_compatible.rs:1380, openai_responses.rs:577. Each is a futures::stream::unfold with tokio::time::timeout_at(state.content_deadline, state.frames.next()) (anthropic.rs:1488, openai_compatible.rs:1416, openai_responses.rs:615), the same transport-error, timeout and abrupt-EOF branches (68–74 % identical lines).
• event_stream_from_response exists three times: anthropic.rs:1433, openai_compatible.rs:1358, openai_responses.rs:229 (94 % identical, two differing lines).
• Provider state machines with the parts that differ: AnthropicStreamState::handle_event/finalize (anthropic.rs:995,1196), OpenAiStreamState::handle_chunk/finalize (openai_compatible.rs:1108,1218), handle_event in openai_responses.rs:422.
• Shared building blocks already exist in sse.rs (sse_frame_stream, frame_is_content, CONTENT_TIMEOUT, SSE_INACTIVITY_TIMEOUT).
• Specs: 0065 / ADR 0056 (a tool call from a truncated or incomplete response is never forwarded; only an allowlist of stop reasons counts as complete), issue AI request hangs forever with only "Automatic mode running": content timeout, total deadline and a visible wait state #694 (content timeout and total deadline).

Requirements

• New module crates/ai-providers/src/common/frame_loop.rs with a trait, e.g. FrameHandler { fn on_frame(&mut self, frame: &SseFrame) -> Vec<RawEvent>; fn finalize(&mut self, abrupt: bool) -> Vec<RawEvent>; } (plus whatever the error mapping needs), and one generic process_frame_stream.
• One shared event_stream_from_response; the two differing lines become a parameter or a trait method.
• The three providers implement the trait on their existing state structs; their own loop copies are deleted.
• The timeout semantics stay exactly as today: keep-alive frames do not extend the content deadline, content deltas do, the inactivity timeout yields a network error instead of hanging.
• The withheld-tool-call rule (security invariant 5) stays in the providers' finalize and is not weakened.

Acceptance Criteria

• grep -rn "timeout_at(" crates/ai-providers/src (production code) finds exactly one call site, in common/frame_loop.rs.
• grep -rn "fn process_frame_stream\|fn event_stream_from_response" crates/ai-providers/src finds each once.
• Unit tests for the shared loop with a fake FrameHandler: content timeout, keep-alive does not reset the deadline, transport error, abrupt EOF calls finalize(true), normal end calls finalize(false).
• The existing timeout tests of all three providers pass (test_inactivity_timeout_yields_network_error_instead_of_hanging_forever, test_keep_alive_frames_do_not_restart_content_timeout, test_slow_deltas_over_three_minutes_complete_without_error, test_continuous_deltas_for_eleven_minutes_keep_streaming, test_responses_keep_alive_frames_time_out_but_deltas_do_not).
• The existing truncated/abrupt tool-call tests pass unchanged (e.g. test_truncated_but_parseable_tool_call_is_never_forwarded_and_triggers_retry, test_abrupt_disconnect_mid_tool_call_triggers_retry_not_silent_parse).
• The number of tests in cargo test -p ai-providers -- --list does not decrease.
• Gate green (cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace, cargo build --workspace).

Technical Notes

• Security-relevant (incomplete tool calls must never be executed, security invariant 5 in .claude/skills/smart-ssh-coder/references/security-invariants.md); review priority elevated.
• Pure restructuring: event order and contents emitted to the app stay identical.
• New code comments in English.

Spec impact: none

Open questions

• Static dispatch (generic process_frame_stream<H: FrameHandler>) or Box<dyn FrameHandler>. Recommendation: generic; the handler type is known per provider.

Related Work

• Previous steps: ai-providers: one shared formatter for command results and rejected actions (single fencing call site) #845, ai-providers: one shared HTTP send loop with 429/5xx retry for all providers #851; next: ai-providers: shared max-tokens retry, openai_responses without openai_compatible internals, dead budget field removed #860