Run as a LiteLLM sidecar: reasoning, failover errors, usage, image

[seankoji]

Summary

• Patches the upstream proxy for use behind LiteLLM on the NAS (nas-ai stack in seankoji-com/careynas.net). The full list is in the README's "Fork changes" table.
• Correctness fixes found against the live gateway and the reverse-engineered wire protocol:
  • reasoning is now sent as reasoning_content instead of being mixed into content
  • an upstream error before any output returns 502, so the router can fail over; previously it was a 200 carrying an SSE error
  • late events after an error could write to an ended response and crash Node
  • usage was always 0
  • tool-call input was sent as a JSON string instead of an object
  • image parts had the wrong shape
  • the x-session-id and x-cli-environment headers were missing
  • the CLI version was pinned at 1.54.0 while npm is at 1.66.0; it now tracks npm
• Container: Dockerfile running as the non-root node user with a healthcheck, logs to stdout only, graceful SIGTERM, and the Windows netstat/taskkill code removed.
• CI: unit tests plus an image build that publishes ghcr.io/seankoji-com/proxy-commandcode:sha-<commit> on every branch push. It runs on GitHub-hosted runners because the org's self-hosted runner groups refuse public repos.

Test plan

• npm test: 6/6 pass locally and in CI
• Branch image sha-c922ce4 built, pulled on the NAS (amd64, user node)
• Live on the NAS through the proxy:
  • sync, streaming and tool-call round trip all succeed
  • a bad key returns 401 and a model outside GOAT returns 403
  • 23 of 25 mapped model ids answer

🤖 Generated with Claude Code