End to end against a real Ollama (T4)

The demo, configured as examples/demo/config.ollama.toml (embedded Temporal, a loopback Ollama over Chat Completions), completed a task with neither OPENAI_API_KEY nor TEMPORAL_API_KEY in its environment. Run at 420df50 on 2026-10-10.

Environment

Ollama	0.40.2, Linux release archive in a user-local directory, ollama serve on 127.0.0.1:11434 with OLLAMA_CONTEXT_LENGTH=16384, OLLAMA_FLASH_ATTENTION=1, OLLAMA_KV_CACHE_TYPE=q8_0 (to fit the machine's free memory)
Model	qwen3:1.7b (TINY_HARNESS_OLLAMA_MODEL): the demo config's qwen3:8b does not fit this machine (4 CPUs, 7 GB RAM, no GPU)
Temporal	embedded, from the cached CLI

Configuration the test wrote

[temporal]
mode = "embedded"
task_queue = "tiny-harness-e2e-f1db28a4"
search_attributes = false
[openai]
base_url = "http://127.0.0.1:11434/v1"
api = "chat_completions"
model = "qwen3:1.7b"
context_window_tokens = 16384
timeout = "PT600S"
max_output_tokens = 2000

Run

$ env -u OPENAI_API_KEY -u TEMPORAL_API_KEY TINY_HARNESS_OLLAMA_MODEL=qwen3:1.7b \
    uv run pytest tests/e2e -k ollama -s -rs
collected 5 items / 4 deselected / 1 selected

tests/e2e/test_demo_ollama.py [   6.2s] task           TASK_STATE_SUBMITTED     6a473b84-b01a-47d7-8e1a-f69acfaec078
[   6.2s] status_update  TASK_STATE_WORKING
[ 106.5s] status_update  TASK_STATE_COMPLETED
.

================= 1 passed, 4 deselected in 110.06s (0:01:50) ==================

Server log: the endpoint line

{"level": "INFO", "logger": "tiny_harness.runtime", "msg": "model endpoint http://127.0.0.1:11434 api=chat_completions model=qwen3:1.7b", "ts": "2026-10-10T20:18:19.539438+00:00"}

Trace: the chat span

{
  "name": "chat qwen3:1.7b",
  "attributes": {
    "gen_ai.operation.name": "chat",
    "gen_ai.provider.name": "tiny_harness",
    "gen_ai.request.tools": 17,
    "server.address": "127.0.0.1",
    "server.port": 11434,
    "tiny_harness.llm.api": "chat_completions",
    "gen_ai.request.model": "qwen3:1.7b",
    "gen_ai.response.model": "qwen3:1.7b",
    "gen_ai.usage.input_tokens": 2032,
    "gen_ai.usage.output_tokens": 290,
    "gen_ai.usage.cache_read.input_tokens": 0,
    "gen_ai.response.finish_reasons": "stop",
    "gen_ai.response.tool_calls": 0
  }
}

What this run does and does not show

• It shows the real wire path: Ollama accepted the Chat Completions request with the harness's 17 tool definitions, returned usage the adapter mapped, and the task ran to COMPLETED through the embedded worker with no key of either kind.
• The 1.7B model answered the complaint in one turn without calling a tool, so the demo's tool path (get_order, the confirm card) was not exercised against a real local model here. Tool calls over Chat Completions are proved by the recorded-body unit tests and the streaming assembly test (T1); a larger model on capable hardware is what the demo config's qwen3:8b default is for.
• The run had network access; the offline claim is proved separately by T8's loopback-only namespace run (security.md).