Qwen3.6 27B local Ollama + term-llm artifact
Generated: 2026-04-24
Artifact: pelican-bike-game-term-llm-ollama-mobile.html
Public URL: https://wasnotwas.com/artifacts/qwen3.6-27b/pelican-bike-game-term-llm-ollama-mobile.html

What this is
------------
A single-file standalone HTML5 canvas game generated by Qwen3.6 27B running locally through Ollama, invoked via term-llm's native Ollama chat provider.

Theme: funny pelican riding a bicycle.

The generated game includes a start screen, HUD, score/high score, keyboard controls, touch controls, gravity/jump arc, obstacles, collision logic, progressive difficulty, game-over/restart flow, and responsive/mobile-friendly canvas behavior.

Important integrity note
------------------------
The generated HTML was not manually repaired or edited.

The model returned the HTML wrapped in a Markdown code fence despite being asked not to. The publishing wrapper stripped only the outer ```html / ``` fence before saving the browser artifact. No HTML, CSS, or JavaScript inside the document was changed.

Exact model/backend
-------------------
Model: qwen3.6:27b
Underlying model family: Qwen3.6 27B
Backend: local Ollama native chat API, /api/chat
Ollama version observed: 0.21.1
term-llm provider: native ollama provider, not OpenAI-compatible emulation
term-llm mode: ask, no tools, JSON event stream reconstruction

Runtime/provider config
-----------------------
A temporary XDG config was used for the run:

```yaml
default_provider: ollama
providers:
  ollama:
    type: ollama
    base_url: http://127.0.0.1:11434
    model: qwen3.6:27b
    think: true
    top_k: 20
    min_p: 0.0
    presence_penalty: 0.0
    num_ctx: 12288
    num_predict: 8500
```

term-llm command
----------------
The actual run used a temporary config directory and captured JSONL events:

```bash
XDG_CONFIG_HOME="$RUN_DIR/xdg" \
term-llm --no-session ask \
  --json \
  --max-output-tokens 8500 \
  --timeout 12m \
  "$(cat "$RUN_DIR/prompt.txt")" \
  > "$RUN_DIR/events.jsonl" \
  2> "$RUN_DIR/stderr.log"
```

Prompt
------
```text
Create one compact standalone HTML5 canvas game: funny pelican riding a bicycle.
Output only the final HTML document, no markdown, no explanation, no code fences.
Keep it concise and complete. Inline CSS and JavaScript only. No external assets, libraries, fonts, or network requests.
Make it mobile friendly: responsive canvas, touch controls on screen, works in portrait and landscape, and still supports keyboard controls on desktop.
Include: start screen, readable HUD, score/high score, keyboard controls, touch controls, gravity, visible jump arc, weighty bike landing, lane/balance or steering, obstacles, clear collision logic, progressive speed/difficulty, game over screen, restart flow.
Early game should teach; later game should pressure. Think briefly, then start the HTML within 200 tokens.
```

Output reconstruction
---------------------
term-llm emitted a JSONL event stream. Text was reconstructed by concatenating `text.delta` event text fields. The only post-processing was removal of the surrounding Markdown code fence if present.

Representative reconstruction script:

```python
import json, pathlib, re
run = pathlib.Path(RUN_DIR)
text = []
for line in (run / "events.jsonl").read_text(errors="replace").splitlines():
    try:
        ev = json.loads(line)
    except Exception:
        continue
    if ev.get("type") == "text.delta":
        text.append(ev.get("text") or "")
content = "".join(text)
m = re.search(r"(?is)```html\s*(.*?)\s*```\s*$", content)
if m:
    content = m.group(1)
pathlib.Path("pelican-bike-game-term-llm-ollama-mobile.html").write_text(content)
```

Observed run stats
------------------
Run directory in Jarvis container: /tmp/pelican-termllm-pristine-mobile-20260424-124921
Generated artifact in Jarvis container: /root/Files/pelican-bike-game-termllm-qwen36-pristine-mobile-v1.html
Duration: 271 seconds
Input tokens: 204
Output tokens: 5743
Final artifact size: 13,579 characters
Browser smoke test: desktop start/play and mobile portrait worked; no console/page errors observed.