First manual smoke against personal Worldtree (10.250.50.152:8081)
produced httpx.ReadTimeout mid-stream after the worker_phase
BuildingPrompt event. Root cause: httpx's default 5s read timeout
killed the connection during mimir's thinking phase (LLM streaming
has multi-second idle gaps between SSE events).
Fix at the caller layer (where the AsyncClient is owned):
- cli._amain and tui.on_mount now construct AsyncClient with
timeout=httpx.Timeout(connect=10.0, read=None, write=10.0, pool=10.0).
read=None disables the SSE-killing timeout; connect/write/pool
keep modest timeouts so true network failures still surface
promptly.
Defense in depth in sse_client.stream_turn:
- ERROR_ROUTING now also catches httpx.ReadTimeout (was just
ReadError | RemoteProtocolError) and surfaces it as
SseConnectionDropped, so if a caller misconfigures their client
the failure is at least a named exception the presenters handle.
Contract amendments (in-place):
- Issue #1: new [compatibility] constraint documents the read=None
recommendation; ERROR_ROUTING for stream_turn lists ReadTimeout
alongside ReadError/RemoteProtocolError.
- Issues #3 + #4: AsyncClient construction step now spells out the
timeout shape explicitly.
Smoke after fix: SSE stream consumed cleanly, agent responded,
[done] turn_id=88 model=qwen3.6-35-a3b duration_ms=2351. Stdout-only
(2>/dev/null) returned clean agent text + exit 0 — INV-002
stdout/stderr split holds end-to-end against real wire. Wire-compat
envelope (personal v0.16.2 vs ratatoskr's v0.19.0 pin) confirmed.
164/164 tests GREEN; ruff clean; all three drift checks clean.
Note: TUI mode not smoke-tested from this CC session (needs a TTY;
operator-side check via `source env.sh && uv run ratatoskr --new
--agent mimir`).
vh
61c3941ec3