fix(log): #0044 log-hygiene batch — artifact planner, ClientDisconnect, default-service setup probe (0.21.135)

[oeway]

Summary

Three prod log-hygiene findings from kth-k8s:zippy-goat's nightly review (2026-08-13). Prod runs 0.21.107, main ~0.21.134 — sub-findings 1a/1b were already fixed on main; 1c, 2, 3 are genuinely present and fixed here. None is a real fault — each restates a benign/expected condition at a level (INFO/ERROR) that either floods the retained kubectl logs forensic window or poisons the ERROR metric that health thresholds key on. Same log-hygiene family as #8 / #19 / #20 / #21.

1c — hypha/artifact.py::search query-planner narration → DEBUG

6 lines (search: stage parameter value, Adding stage filter condition, Adding condition to return ONLY STAGED/COMMITTED …) fired at INFO on every search() — a hot listing path (~300k lines/24h in prod). Internal planner breadcrumbs, no steady-state operator value → demoted INFO→DEBUG.

2 — hypha/http_rpc.py::_handle_rpc_post catches ClientDisconnect

When a client hangs up before/while its request body arrives, await request.body() raises starlette.requests.ClientDisconnect. The broad except Exception logged it ERROR + full traceback + 500 (618/24h). It's a normal network event → caught specifically (before the broad arm), logged at DEBUG with no traceback, answered 499 (no client left to answer).

3 — hypha/core/workspace.py default-service setup probe (the reusable code trap)

The remote-service proxy (get_remote_service → Munch/ObjectProxy) RAISES AttributeError('setup') when a default service simply defines no setup(). The old if svc.setup: inside a broad except Exception as e: logger.error(f"… {e}") turned that benign, expected case into a per-client ERROR whose {e} interpolated to the useless bare string "setup" (~20/24h, unactionable by construction). Fix: setup = getattr(svc, "setup", None) + if callable(setup): (no-setup → silent no-op), a narrower except around the resolve step, and every caught detail reformatted f"{type(e).__name__}: {e!r}" so a setup() that actually raises is diagnosable — naming the real failure (RemoteException: RemoteException('RemoteError:boom-in-setup…')), never the attribute name.

hypha/workers/__init__.py:38 has the same if svc.setup: shape but is intentionally left untouched — worker services define setup; out of scope.

Tests

tests/test_log_hygiene_0044.py — 5 real tests, Docker-free (FakeRedis), no mocks. All conftest fastapi_server* fixtures are subprocess.Popen (caplog can't reach them), so each path is driven in-process: a real ArtifactController(store, s3_controller=None); real HTTPStreamingRPCServer routes over httpx.ASGITransport with a valid Bearer token + monkeypatched Request.body raising ClientDisconnect; two real event-bus create_rpc peers registering a default service (built-in registry key seeded so the real :default@ block runs).

Reproduce-before-fix verified: stashing the 3 fixes makes the 3 regression guards fail, with the captured log showing the exact prod signature ERROR workspace:workspace.py:2401 Failed to run setup for default service …: setup.

5 passed in 3.06s

Version

Bumped 0.21.134 → 0.21.135 across the 10 packaging files. CLAUDE.md and .svamp/issues/0003.md mentions of 0.21.134 are #43 history and left as-is.

Merging does not publish an image (publish.yml is gated on release:published/workflow_dispatch) — the release-cut and deploy remain Wei/crimson-spider-gated.

🤖 Generated with Claude Code