[Marketen]

Why

A request is charged when the provider reports its usage, at the end of the answer. When the client hung up earlier, the gateway cut the provider off, no usage arrived, and Metering released the hold without a charge. The provider still billed us for what it had generated. Anyone could get answers for free by disconnecting just before the end, on any account. Guest chat (#18) would make that anonymous.

Providers have no API to look up what one cancelled request cost, so the only way to charge the exact amount is to let the answer finish.

What changes

• Once credit is reserved, the provider call no longer follows the client's connection (Proxy). The relay keeps reading to the end, and the usage the provider reports is charged, for every account.
• Pressing Stop no longer lowers the bill. The request's max_tokens (or the model maximum) bounds it, and the hold already covers it.
• A stalled provider still ends: the provider client's existing idle timeout (PROVIDER_SILENCE_TIMEOUT_SECONDS, 15 min) applies.
• A client that hung up gets no fallback answer.
• The handler's 5-second post-finish drain is gone; it's covered by the above.

Provider faults are unchanged: what the provider reports is charged (see dappnode-nexus#152), otherwise nothing.

Tests

• TestStreamingUsageRetention, a real HTTP run through the production OpenAI adapter: the client hangs up after the first chunk, the provider finishes, and usage is charged. It previously asserted the opposite.
• New proxy_hangup_test.go: hang-up mid-stream, hang-up during a non-streaming answer, no fallback after a hang-up. Each fails with its part of the fix removed.
• go test -race ./... and go vet ./... pass.

🤖 Generated with Claude Code

[Marketen]

Merged into #22.