mirror of
https://github.com/BerriAI/litellm.git
synced 2026-10-10 03:28:53 +00:00
fcd285facd
A sync fans out one request per skill folder; on a repo with dozens of skills, a single request occasionally hitting a transient connection blip was expected, not exceptional. Previously any one failure aborted the entire asyncio.gather with an unhandled exception, so admin retries on a big repo were coin flips even after the concurrency fix. Two changes: - _http_get retries transient httpx errors (not real HTTP status codes) with a short backoff before giving up. - A single skill that still fails after retries is skipped rather than fatal, and counted via a new persisted skipped_count column so an admin can actually see whether every skill loaded, instead of an opaque all-or-nothing success/error. Verified live: 10/10 registrations succeeded for both a manifest-based repo (alirezarezvani/claude-skills, 83 skills) and the dir-scan fallback path that was actually flaky before (garrytan/gstack, 53 skills). |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| test_claude_code_marketplace.py | ||
| test_claude_code_marketplace_sources_endpoints.py | ||
| test_claude_code_marketplace_sync.py | ||
| test_claude_code_skill_authz.py | ||
| test_endpoints.py | ||