LangChain: Retry Middleware Gets an Honesty Fix
A cluster of fixes tightens how retry middleware handles exceptions that fall outside configured retry rules, closing a gap where errors were silently swallowed instead of raised. Separately, a new standard reasoning-effort parameter and Responses API support for OpenRouter expand the chat model surface.
Duration: PT2M40S
Episode overview
This episode is a short developer briefing from LangChain.
It explains recent repository work in plain language.
- Show: LangChain
- Published: 2026-07-17T13:01:55Z
- Audio duration: PT2M40S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
This is LangChain, your developer briefing for July 17, 2026.
The headline today is about trust in error handling. Multiple pull requests this cycle converge on one problem: retry middleware in LangChain was too generous about what it caught, and that made failures harder to detect, not easier.
It started with PR 38845 from ccurme, which fixed Tool Retry Middleware so it only normalizes exceptions into tool messages when they're actually listed in the retry rules. Before that fix, any exception got swallowed and converted, even ones you never asked it to handle. That's dangerous, because a masked failure…
That same logic then moved to the model side. PR 38894 from Nate-git05 fixes Model Retry Middleware so it follows the same rule, propagating non-retryable exceptions instead of fabricating a fake response message. Three linked issues, 38893, 38845, and 38884, show this was a deliberate, tracked effort to make retry…
Second theme: test infrastructure reliability. PR 38888 tracks down a leaked VCR patch causing flaky Anthropic integration test failures on live runs, and PR 38885 raises the minimum VCR-py version to patch a known security advisory. PR 38901 separately fixes a stream…
On…