Issues

• Team traffic dashboard: apps × models × providers (Prometheus/Grafana) + a questions explorer

  boettiger-lab/open-llm-proxy#152 · Filter by author cboettig opened on Oct 12, 2026

• Runaway generations: no output cap when the client sends no max_tokens (22 responses ran to 65k–131k tokens since July)

  boettiger-lab/open-llm-proxy#150 · Filter by author cboettig opened on Oct 6, 2026

• headless: runner crashes leave no stderr; one stalled MCP call (600 s) eats a 900 s cell

  boettiger-lab/open-llm-proxy#149 · Filter by author cboettig opened on Oct 5, 2026
  1

• Log OpenRouter's upstream provider on every response

  boettiger-lab/open-llm-proxy#145 · Filter by author cboettig opened on Sep 24, 2026

• Routing ignores the model catalog the proxy already fetches

  boettiger-lab/open-llm-proxy#142 · Filter by author cboettig opened on Aug 25, 2026
  1

• Nothing notices when the consolidation CronJob dies — it failed silently for 9 days

  boettiger-lab/open-llm-proxy#122 · Filter by author cboettig opened on Aug 16, 2026
  1

• Namespaced model ids (provider/model) — make callers name the backend and delete routing config entirely

  boettiger-lab/open-llm-proxy#120 · Filter by author cboettig opened on Aug 16, 2026

• Matrix Job silently produced runs with no ##BENCH-VERSIONS## line (2026-08-01); make a provenance-less run fail loudly

  boettiger-lab/open-llm-proxy#103 · Filter by author cboettig opened on Aug 3, 2026
  2

• Consider deploying open-llm-proxy on cirrus — last NRP dependency for headless runs (decide log consolidation first)

  boettiger-lab/open-llm-proxy#97 · Filter by author cboettig opened on Jul 30, 2026

• Own tool-call dialect recovery in the proxy for self-hosted backends — 91% of fleet tool-call leaks are nimbus; OpenRouter/Anthropic are clean

  boettiger-lab/open-llm-proxy#95 · Filter by author cboettig opened on Jul 29, 2026

• Formal benchmarking: migrate question banks to a dedicated repo (provenance, tiers, difficulty, versioned results, public dashboard)

  boettiger-lab/open-llm-proxy#91 · Filter by author cboettig opened on Jul 17, 2026
  6

• Benchmark: nimbus qwen (Qwen3.6-35B-A3B-NVFP4) full gold-standard — 19 pass / 1 partial / 2 fail

  documentation
  boettiger-lab/open-llm-proxy#68 · Filter by author cboettig opened on Jul 7, 2026
  2

• Open-model benchmark (glm-5.2 / nemotron-3-ultra / minimax-m3 / kimi-2.7-code) — results & follow-ups

  boettiger-lab/open-llm-proxy#54 · Filter by author cboettig opened on Jun 27, 2026
  1

• GLM robustness: 5xx tool-session fallback ladder + dual thinking-key normalization

  boettiger-lab/open-llm-proxy#48 · Filter by author cboettig opened on Jun 27, 2026
  1

• Treat the geo-agent harness (not just guidance text) as the optimization target for open-model training

  boettiger-lab/open-llm-proxy#42 · Filter by author cboettig opened on Jun 26, 2026
  1

• Decide policy for full-capture logging (LOG_CAPTURE_MODE=full) — currently summary, toggle unused

  boettiger-lab/open-llm-proxy#32 · Filter by author cboettig opened on Jun 23, 2026
  1

• Investigate prompt caching for the direct Anthropic provider

  boettiger-lab/open-llm-proxy#20 · Filter by author cboettig opened on May 30, 2026