Releases: houdry-genomex/houdry-agent
Release list
v0.19.4
Full Changelog: v0.19.3...v0.19.4
v0.19.3
v0.19.2
v0.19.1
v0.19.0
Houdry Agent 0.18.1 — Ollama context window fix
Supersedes 0.18.0, which is withdrawn — its bundled fabric could not talk to a
stock Ollama daemon.
Fixed: Ollama context window
The fabric never sent options.num_ctx, so every request got whatever the
daemon defaults to — 4096 — no matter what the model could actually handle.
deepseek-r1:1.5b supports 131072 tokens, but a 4344-token prompt was rejected
outright:
request (4344 tokens) exceeds the available context size (4096 tokens)
which reads as "this model is too small" when it is really "we never asked for
the room". Any agent transcript with a system prompt and two tool results clears
4096, so on a machine with a stock daemon the fabric failed within a few turns.
The window is now resolved per model, in this order:
- an explicit per-request value
HOUDRY_OLLAMA_NUM_CTX- the model's own maximum from
/api/show— a baked-inPARAMETER num_ctx
beats GGUF metadata, since an author who wrote one meant it - 8192 if detection finds nothing, which still clears the 4096 wall
The result is clamped to 32768 by default. A context window is memory: a
model that can do 131072 will still fail to allocate it on a laptop GPU, and
trading a 4096 wall for an out-of-memory error is not a fix. If you have the
VRAM, lift the clamp:
setx HOUDRY_OLLAMA_NUM_CTX max # the model's full trained window
setx HOUDRY_OLLAMA_NUM_CTX 65536 # or an exact number
There is no truly unlimited setting — the ceiling is what the model was trained
for and what your GPU can hold.
Bundled fabric: houdry 0.6.1.
Also in 0.18.x
Automatic local model detection. Selecting the local endpoint option knocks on
the five ports the mainstream engines bind by default — Houdry fabric (18080),
Ollama (11434), LM Studio (1234), llama.cpp (8080), vLLM (8000) — and prefills
the URL with the first that answers with a real model catalog. Probing goes
through the same round-trip the runtime uses, so a daemon whose weights live on
another volume is found correctly. Detection prefills; Connect stays an explicit
press.
Chat surface: the composer opens five lines tall on an empty chat and drops back
to one line once a transcript exists; transcript type up one step; sidebar rows
at 32px with 14px titles so the selected row reads as a pill; one fixed composer
placeholder instead of a rotating pool.
Known gap: the scan probes 127.0.0.1 only, so a non-default OLLAMA_HOST is
missed by detection (the runtime still honours it if you enter the URL).
Install
Windows x64, per-user, no admin prompt. The installer is unsigned, so
SmartScreen shows "Unknown publisher" — choose More info → Run anyway.