Legacy Common Lisp proxy: Woo/libev signal watcher lifecycle aborts across repeated transport fixtures #89

New issue

Open

opened 2026-09-20 01:19:23 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-20 01:19:23 +00:00

Owner

Symptom

The packaged Common Lisp forwarding runtime aborts during repeated proxy fixture startup/shutdown:

ev_signal_start: Assertion ("libev: a signal must not be attached to two different loops", ...)
WOO.SIGNAL::START-SIGNAL-WATCHERS

On PR #106 run 35480587993 the ordinary CL method/path, binary body, status, duplicate-header and chunked-stream tests pass first. The next fixture aborts while Woo starts signal watchers for a new loop.

This is the same lifecycle class already observed in the stale #74/#76 scheduler stack. It predates and is independent of current Python HTTP/SSE/WebSocket admission.

Current architecture boundary

Per repository AGENTS/current expert architecture:

Python llm_log/ owns current capture/HTTP/SSE/WebSocket transport.
Common Lisp owns the expert runtime/Tek9/SWI-Prolog boundary.
proxy/ remains a packaged alternate/legacy forwarding runtime but is not the production transport authority for #104.

Required fix

isolate Woo signal watchers from per-fixture loop churn, or run one loop/server lifecycle for the contract suite;
prove repeated start/stop does not retain signal watcher ownership;
retain chunked and large-body transport tests;
do not mask a real CL runtime failure as a Python admission failure.

Until fixed, keep this diagnostic visible but do not use it to block unrelated Python transport changes.

## Symptom The packaged Common Lisp forwarding runtime aborts during repeated proxy fixture startup/shutdown: ``` ev_signal_start: Assertion ("libev: a signal must not be attached to two different loops", ...) WOO.SIGNAL::START-SIGNAL-WATCHERS ``` On PR #106 run 35480587993 the ordinary CL method/path, binary body, status, duplicate-header and chunked-stream tests pass first. The next fixture aborts while Woo starts signal watchers for a new loop. This is the same lifecycle class already observed in the stale #74/#76 scheduler stack. It predates and is independent of current Python HTTP/SSE/WebSocket admission. ## Current architecture boundary Per repository AGENTS/current expert architecture: - Python `llm_log/` owns current capture/HTTP/SSE/WebSocket transport. - Common Lisp owns the expert runtime/Tek9/SWI-Prolog boundary. - `proxy/` remains a packaged alternate/legacy forwarding runtime but is not the production transport authority for #104. ## Required fix - isolate Woo signal watchers from per-fixture loop churn, or run one loop/server lifecycle for the contract suite; - prove repeated start/stop does not retain signal watcher ownership; - retain chunked and large-body transport tests; - do not mask a real CL runtime failure as a Python admission failure. Until fixed, keep this diagnostic visible but do not use it to block unrelated Python transport changes.

nsaspy referenced this issue

2026-09-20 06:13:36 +00:00

Define class-specific transport alert policy #103