End-to-end test against a live provider

This proves the full pipeline — token estimation, classification, routing, forwarding, real token/cost accounting, and SQLite logging — works against a real cloud provider.

Prerequisites

• A provider key in the environment. Either:

  • OPENROUTER_API_KEY (recommended — covers all catalog models), or
  • OPENAI_API_KEY (covers the openai provider models only).

• A config file at config/llmux.yaml. Copy the template:

  cp config/llmux.example.yaml config/llmux.yaml

  The example config has auth.llmux_key: "mp_dev_changeme" — used as the Authorization: Bearer token below.

1. Start the server

export OPENROUTER_API_KEY=sk-or-...      # your key
cargo run --release

Expected startup log:

INFO llmux: Konfiguration geladen und validiert source=config/llmux.yaml
INFO llmux: SQLite-Log geöffnet path=data/llmux.sqlite
INFO llmux: llmux läuft auf http://localhost:3456  (Dashboard unter /)

The read-only dashboard is now served at http://localhost:3456/; the requests below show up in its live feed.

If the config is malformed the server exits before binding, e.g.:

Error: Konfiguration ungültig:
  - Modell 'x' verweist auf unbekannten Provider 'ghost'

2. Send a simple_text request

Routed to the cheapest tier-1 model (a free/cheap model in the example catalog).

curl -s http://localhost:3456/v1/chat/completions \
  -H "Authorization: Bearer mp_dev_changeme" \
  -H "Content-Type: application/json" \
  -H "x-llmux-tool: e2e-test" \
  -d '{"messages":[{"role":"user","content":"Say hello in one short sentence."}]}' | jq .

Expected: a normal OpenAI-shaped response with a choices[0].message.content string and a usage block.

3. Send a code_review request

code_review has min_tier: 3, so the selector picks a higher-tier model (e.g. gpt-4.1-mini).

curl -s http://localhost:3456/v1/chat/completions \
  -H "Authorization: Bearer mp_dev_changeme" \
  -H "Content-Type: application/json" \
  -H "x-llmux-tool: e2e-test" \
  -d '{"messages":[{"role":"user","content":"Review this Rust function for bugs: fn add(a:i32,b:i32)->i32{a-b}"}]}' | jq .

The server log line for each request shows the chosen model and real cost:

INFO llmux::api: request ok task=simple_text provider=openrouter model=... tier=1 prompt_tokens=23 completion_tokens=9 cost_usd=0.0000...
INFO llmux::api: request ok task=code_review provider=openai model=gpt-4.1-mini tier=3 prompt_tokens=41 completion_tokens=120 cost_usd=0.00021...

4. Verify the logged entries (Definition of Done)

The DoD requires a log entry with status=200, real_cost_usd > 0, and the correct model. Query the SQLite log directly:

sqlite3 -header -column data/llmux.sqlite \
  "SELECT task_type, model, provider, tier, status, prompt_tokens,
          completion_tokens, real_cost_usd
   FROM requests
   WHERE tool = 'e2e-test'
   ORDER BY id DESC LIMIT 2;"

Expected (illustrative — exact tokens/cost depend on the model and prompt):

task_type    model          provider    tier  status  prompt_tokens  completion_tokens  real_cost_usd
-----------  -------------  ----------  ----  ------  -------------  -----------------  -------------
code_review  gpt-4.1-mini   openai      3     200     41             120                0.0002104
simple_text  ...            openrouter  1     200     23             9                  0.0000...

Pass criteria:

• status = 200 for both rows.
• real_cost_usd > 0 for the code_review row (a tier-3 paid model). A free tier-1 model legitimately logs real_cost_usd = 0; to assert a non-zero cost on the cheap request too, force a paid model: -H "x-llmux-model: google/gemini-flash-1.5".
• prompt_tokens / completion_tokens come from the provider’s usage block (not the local estimate), and real_cost_usd is derived from them.

Note: this run requires a live provider key and therefore must be executed manually. Everything else in the pipeline is covered by cargo test (101 tests, no network, as of 2026-09-27).