bug(URGENT): claude code incompat

[gkgoat1]

What happened

Drops connections with Claude Code when using subscription auth, Claude can't think without hitting 502 errors.

How are you running llmtrim?

• Proxy (llmtrim setup / serve) — intercepting a tool's traffic

• CLI (llmtrim compress / send / batch)

• Tool / client (proxy path): Claude Code

• Provider: anthropic

• Preset / config: auto

Reproduce

Use Claude Code with a Claude subscription

⚠️ Redact secrets first — strip API keys, authorization headers, and private prompt
content. llmtrim never needs them to reproduce a compression bug. (Security vulnerabilities
go through the private advisory link, not a public issue — see SECURITY.md.)

Expected vs actual

• Expected: no 502 errors
• Actual: extremely frequent 502s

Environment

• llmtrim --version: llmtrim 0.8.0
• llmtrim status (daemon + CA + savings state):
• Install method: cargo install
• OS / arch: macOS 26, aarch64

[fkiene]

Thanks for the report. Two things here, and I want to be straight about which is which.

I found and fixed a real bug, but it's probably not the bug. When a request to Anthropic failed before the API answered (a TLS error, a reset, a dropped connection), llmtrim handed you back a 502 with an empty body and logged the cause nowhere. That's unparseable to Claude Code, so the turn died instead of retrying, and it's why you had nothing to attach to this report beyond "frequent 502s". Fixed in 0.10.0: a connection failure now comes back in a form Claude Code can read and retry, and the cause gets printed.

What I still don't know is why the connection was failing in the first place. I have a candidate: llmtrim reusing a pooled connection that Anthropic's edge had already closed. Extended thinking would make that far more likely, because the model sits idle for tens of seconds between turns, which fits the "Claude can't think" shape of what you're seeing. But I haven't reproduced it, so I held that change back rather than ship a guess.

Your logs would settle it. Upgrade to 0.10.0 (whichever way you installed it), use Claude Code until it fails again, then check the proxy log:

grep "upstream request failed" ~/.llmtrim/serve.log

You should get a line like:

llmtrim: upstream request failed (...) — reported to the client as a retryable error

Paste that line here. Whatever's in the parentheses tells me whether it's the pooled-connection theory, a TLS problem, or Anthropic returning 502s upstream on its own, and I can fix the right thing instead of the likely one.

One note on your version: you're on 0.8.0, six minor releases back. Some of what you hit may already be gone. If 0.10.0 clears it up entirely, say so and I'll close this.

Compression isn't involved, by the way. I checked rather than assumed, and added tests pinning it: Claude Code's thinking blocks pass through llmtrim byte for byte, signatures intact.