tumix implements a TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture in Go.

HTTP client tracing is off by default to minimize overhead. Set TUMIX_HTTP_TRACE=1 to enable OpenTelemetry spans for outgoing LLM HTTP calls; clients will keep using the shared pooled transport either way.

• -model (default gemini-2.5-flash)
• -max_rounds (default 3; higher improves quality, raises cost)
• -temperature / -top_p / -top_k / -max_tokens / -seed
• -json (emit final answer as JSON on stdout)
• -session_dir (persist sessions to disk; default in-memory)
• TUMIX_SESSION_SQLITE env to use sqlite-backed store instead of session_dir
• -batch_file with -concurrency (one prompt per line)
• -http_trace (enable HTTP spans)
• -otlp_endpoint (export traces)
• -bench_local to run synthetic local benchmark (no LLM calls)
• -max_prompt_chars to fail fast on oversized prompts
• -max_prompt_tokens tokenizer-backed guard (CountTokens) with heuristic fallback; pricing override via TUMIX_PRICING_FILE
• -metrics_addr serve /healthz, /debug/vars, /metrics (Prometheus text)

Env overrides: GOOGLE_API_KEY, TUMIX_MODEL, TUMIX_MAX_ROUNDS, TUMIX_TEMPERATURE, TUMIX_TOP_P, TUMIX_TOP_K, TUMIX_MAX_TOKENS, TUMIX_SESSION_DIR, TUMIX_HTTP_TRACE, TUMIX_CALL_WARN, TUMIX_CONCURRENCY.

• Low cost: ./tumix -model gemini-2.5-flash -max_rounds 2 -temperature 0.2 "Explain X"
• High quality: ./tumix -model gemini-2.5-pro -max_rounds 3 -top_p 0.95 -max_tokens 512 "Explain Y"
• Batch: ./tumix -batch_file prompts.txt -concurrency 4 -json
• Persist & observe: TUMIX_SESSION_SQLITE=/tmp/tumix.db ./tumix -metrics_addr :9090 -http_trace
• CI smoke: ./tools/bin/gotestsum -f standard-verbose -- -race -count=1 -shuffle=on -cover ./...

• /metrics emits counters: tumix_requests, tumix_input_tokens, tumix_output_tokens, tumix_cost_usd.
• /debug/vars exposes the same data via expvar; /healthz returns ok.