feat(agent): SSE streaming for LLM completions with per-provider accumulators

[wpfleger96]

Summary

Replace buffered stream: false LLM requests with Server-Sent Events streaming. Each text delta emits an agent_message_chunk session update as it arrives, providing a natural keepalive that resets the ACP idle clock without relying solely on the 30s ticker.

Stack: #935 (Phase 1: base) → this PR (Phase 2: streaming)

Design Decisions

• Two reqwest::Client instances: http (with global timeout for summarize()) and http_stream (no global timeout — first-byte and inter-chunk timeouts enforced via tokio::time::timeout)
• Two-phase timeout: llm_timeout (120s) until first content delta, then stream_chunk_timeout (30s) between subsequent chunks
• Keepalive via content deltas: each text delta emits agent_message_chunk session update, resetting ACP idle clock; 30s ticker preserved as fallback for reasoning models
• MAX_LLM_RESPONSE_BYTES caps accumulated semantic content (text + tool arg strings), not raw SSE wire bytes

Per-Provider Accumulators

• Anthropic: index-keyed HashMap<usize, (String, String)> for parallel tool-call blocks via content_block index
• OpenAI Chat: Vec-indexed accumulation by tool_calls[].index
• OpenAI Responses API: routes on JSON type field inside data: payload — response.output_text.delta for text, response.function_call_arguments.delta for tool args, response.output_item.done for item completion

SSE Parser

Custom inline parser (~70 lines) handling:

• : comment lines (discarded)
• Multi-line data: fields (joined with \n)
• id: and retry: fields (ignored)
• [DONE] sentinel detection

Changes

crates/sprout-agent/src/llm.rs

• Add http_stream client field (no global timeout)
• post_stream() function with SSE line parser
• Per-provider accumulator structs (Anthropic, OpenAI Chat, Responses)
• Two-phase timeout: llm_timeout for pre-content, stream_chunk_timeout for post-content
• MAX_LLM_RESPONSE_BYTES enforcement on semantic content

crates/sprout-agent/src/config.rs

• Add stream_chunk_timeout field
• Parse from SPROUT_AGENT_LLM_STREAM_CHUNK_TIMEOUT_SECS env var (default 30s)

crates/sprout-agent/src/agent.rs

• Pass WireSender + session_id into complete() for delta emission
• Keep ticker (G) as fallback for reasoning models

Tests

• SSE parser edge cases (comments, multi-line data, id/retry fields)
• Per-provider accumulation (Anthropic parallel tool calls, OpenAI indexed)
• Two-phase timeout behavior
• MAX_LLM_RESPONSE_BYTES enforcement during streaming

Acceptance Criteria

• cargo test passes (existing + new streaming tests)
• cargo clippy clean
• Anthropic streaming: text chunks emit as agent_message_chunk, tool calls accumulate via index-keyed HashMap
• OpenAI Chat streaming: same with Vec-indexed accumulation
• OpenAI Responses streaming: same with correct event taxonomy
• summarize() still works via http client (non-streaming path preserved)
• MAX_LLM_RESPONSE_BYTES enforced during streaming (semantic content)
• Keepalive ticker still fires if no delta arrives within timeout window
• Two-phase timeout: llm_timeout pre-content, stream_chunk_timeout post-content

[wpfleger96]

Pushed Thufir's two review fixes (HEAD now edbe0e40):

Fix 1 — connect_timeout restored to 10s. crates/sprout-agent/src/llm.rs:52 — .connect_timeout(std::time::Duration::from_secs(10)), independent of cfg.llm_timeout. A dead endpoint fails fast at the handshake instead of waiting 120s (and being retried). The .timeout() removal from the original PR stays.

Fix 2 — mid-body read errors retry like stalls. The chunk().await Err arm now backs off and resends when attempts remain, hard-failing only on the final attempt — mirroring the stall arm. reqwest's chunk() wraps a reset/truncated body as a decode error (Kind::Decode, not Kind::Body), so is_retryable_transport_error gained is_decode(). JSON-parse failures take the separate serde_json path on the fully-buffered body and are unaffected.

Test: post_retries_on_read_error_mid_body — real TCP, server sends chunked headers + a partial chunk then resets the socket on attempt 1, serves a complete body on retry. Asserts success and >=2 connect attempts.

cargo test -p sprout-agent (82 lib tests, +1) and cargo clippy --all-targets both clean.

Note: branch was force-updated from a dedicated worktree (duncan/944-review-fixes) to sidestep an edit race with a concurrent SSE-task instance in the shared streaming worktree; nothing from that instance was committed.