Skip to content

HTTPd M1.5 — inference live (/v1 runtime)

Summary

Completes plan todo m15-inference-live: live build/li-httpd serves OpenAI-compatible /v1 proxy routes with global rate limits (429), required traceparent on inference routes (400 without), upstream cancel when the client disconnects during SSE relay, and gate smoke test-m15-inference-live.sh.

Agent continuation

  1. Run ./scripts/test-m15-inference-live.sh after ./scripts/build-li-httpd.sh.
  2. Full plan gates: ./scripts/httpd-plan-gates.sh.
  3. Example config: packages/li-net-httpd/examples/inference_live.toml.

Changed

Area Paths
Runtime runtime/li_rt_net.c — Li-proxy ingress traceparent gate + inject export
Li proxy packages/li-net-httpd/src/lib.li — OTel inject, cancel upstream on client hangup
Gates scripts/test-m15-inference-live.sh, scripts/httpd-plan-gates.sh
Example packages/li-net-httpd/examples/inference_live.toml
Plan docs/superpowers/plans/2026-05-16-li-httpd-plan.md — m15-inference-live: completed

Not changed

  • Tier-5 bench matrix / live Pages (no new CSV).

Breaking

N/A.

Security

Smoke uses loopback-only listeners; same SSRF/upstream allowlist as other M1 proxy gates.