Current status of the 3 test models on LabChan -- Gemma, Rachel and JW -- for reading web page text
(fetch-page:) and understanding attached images (vision). Generated from the code; release
history is on the Deploy Dashboard, and the per-agent code path for
each is on Agent Service Flows.
| Model Name | Web Text | Web Image |
|---|---|---|
| Gemma | LIVE-TESTED | NOT SUPPORTED |
| Rachel | LIVE-TESTED | LIVE-TESTED |
| JW | SHIPPED, UNTESTED | NOT SUPPORTED |
gemma3:12b running locally via Ollama and Open WebUI.
Integration: a dedicated RX bot (gemma_rx_listener.py) feeds listener/processor.py,
which calls gemma_adapter.py (GemmaBridgeAdapter.invoke). write, list and
discogs commands are handled deterministically with no model call; otherwise the prompt is built within an
8k-token budget and sent via make_open_webui_call_model to Open WebUI/Ollama. A token-count footer is added to
replies, which go out through tx/processor.py, tx/envelope.py and tx/discord_adapter.py
on Bot Gemma Unified.
Web Text: fetch-page: shipped in v1.7 and live-tested 2026-09-26 -- "Gemma: fetch-page:
example.com" returned the page text quoted back (~2 min, local model).
Web Image: no vision support. An attached image gets an honest text marker, not visual interpretation -- same limitation as JW.
Integration: rachel_rx_listener.py feeds listener/processor.py, which calls
rachel_adapter.py (RachelBridgeAdapter.invoke). chart:, fetch-image:,
generate-image: and discogs: are handled deterministically with no model call; otherwise the
prompt is built from persona, shared context, RachelNotesStore and ResponseIdStore continuity,
images and file text, then sent via make_openai_call_model to the OpenAI Responses API. Replies go out through
tx/processor.py, tx/envelope.py and tx/discord_adapter.py on Bot Rachel.
Web Text: fetch-page: shipped in v1.7 and live-tested 2026-09-26 -- correct summary of a
fetched page.
Web Image: real vision, shipped in v1.2. Attached images are base64-encoded and sent to the model
directly. Live-verified 2026-09-20: a Python logo, three kittens, a Dallas-Fort Worth map, Hulk Hogan (via conversation
continuity, no re-attached image), the dashboard diagram itself, and a byte-identical fetch-image: round trip.
The only test model with both capabilities working end to end.
lumo-tamer running on ai-lab-b.
Integration: a dedicated RX bot (jw_rx_listener.py) feeds listener/processor.py,
which calls jw_adapter.py (JWBridgeAdapter.invoke). The prompt is sent via
make_lumo_call_model to lumo-tamer's local OpenAI-compatible endpoint, backed by Lumo. Memory is
JW's own per-sender notes log (last 10 exchanges) plus the shared lab-context file. Hourly cap 20. Replies go out through
the shared LabChan TX bot.
Web Text: fetch-page: shipped in v1.7 (2026-09-26), same code path as Gemma and Rachel,
but blocked live by a Proton "weekly quota" response (a raw passthrough from their backend, not a string in
lumo-tamer's own code). Root-caused 2026-09-28 with a controlled comparison on the same account, same model
tier: the Lumo web chat, explicitly set to Lumo 2.0 Lite, answered normally with no quota error; a
lumo-tamer API call, also explicitly set to Lite, hit the same quota error as before. Browser works, API
doesn't, same account and tier -- this is a genuine limit on API-style access to Lumo, not a shared account-wide quota
and not a model-tier mismatch on our side. Nothing further to investigate from the code side; not yet verified working
end to end.
Web Image: no vision support. An attached image gets an honest text marker, not visual interpretation -- same limitation as Gemma.
Hand-maintained status page -- see Deploy Dashboard for release history and Agent Service Flows for the full per-agent code path.