* fix(llm): keep reasoning out of answers when thinking is off
Small-budget utility calls on reasoning models spend every token on
reasoning, and llm_call then returned reasoning_content as the answer. Web
search extraction sent that text to SearXNG as the query, which found
nothing, and session titles fell back as unusable.
Treat thinking_mode="off" as "only the answer": non-streaming calls now
return empty instead of substituting reasoning, matching the streaming path.
Callers that do not turn thinking off keep the existing fallback.
Turn thinking off for search query extraction, session titles without a
configured effort, scheduled task names, and deep research classification,
and strip inline think blocks from the extracted query.
* fix(routing): require computer task wording for named hosts
"on <word>" and "from <word>" counted as a named computer on their own, so
"Who is Jensen Huang from NVIDIA?" and "What happened on Monday?" routed to
shell tools and the workspace toolset.
A target now only counts next to wording that describes work on a machine,
such as "check disk usage on gpu-box" or "copy the logs from gpu-box". The
intent classifier and the agent loop's local computer check share the pattern.
* fix(chat): resolve follow-up references in web search queries
Query extraction only saw the latest message, so "Does he work with Lisa
Su?" lost who "he" was. Include the last four turns, capped per turn, and
instruct the model to use them only to resolve references. Incognito turns
do not use saved session history.
* test(llm): stop reasoning fallback tests hanging in isolation
The module replaced src.agent_tools and core.database with MagicMock before
importing agent_loop. strip_tool_blocks then returned a mock, and
_strip_think_blocks looped forever because a mock's find() never returns -1.
This hung whenever the file ran alone or first on a worker. Import the real
modules instead.
* test(search): remove extra blank line at EOF
---------
Co-authored-by: Alexandre Teixeira <general@alexandreteixeira.dev>
* refactor(cookbook): move endpoint registration out of the routes closure
_auto_register_llm_endpoint and _auto_register_image_endpoint use no
closure state and model_serve is their only caller, so they can live in a
plain module and be imported and tested directly. Bodies are unchanged
apart from indentation.
Tests now import the helpers instead of reading them from the route's
closure. That lookup also failed if the route stopped referencing a
helper, so add a test that pins which helper model_serve calls.
Part of #4071
* fix(cookbook): bound the OLLAMA_HOST port scan
The pattern rescanned the rest of the token from every OLLAMA_HOST=
occurrence, so a long serve command with repeated assignments took
quadratic time. model_serve runs this on the event loop, and a 192 KB
command stalled the whole app for about 9s.
Cap the host at 253 characters, the longest DNS name, so every valid host
still matches.
Part of #4071
* feat(admin): rename model endpoints inline
* fix(cookbook): keep renamed endpoint names on relaunch
Relaunching a model at an existing endpoint URL reset the endpoint name
to the launched repo's short name, undoing a rename made in Added Models.
Only replace the name when it is still a name the Cookbook generates for
one of the endpoint's pinned models, so endpoints that were never renamed
still follow the newly served model.
---------
Co-authored-by: Alexandre Teixeira <general@alexandreteixeira.dev>
* fix(cookbook): match stale endpoints by launched model, not display name
After registering an LLM endpoint, the Cookbook deleted every other local-*
endpoint with the same display name at a different URL. The display name is
the repo short name, which repos from different orgs share (for example
meta-llama/Llama-3.1-8B-Instruct and unsloth/Llama-3.1-8B-Instruct), so
launching one deleted the endpoint of the other while its server was running.
Match stale endpoints by the full model id they were last launched for,
which is the last entry of their pinned_models. Reusing an endpoint now
moves the launched model to the end of its pins so that entry stays current.
* fix(cookbook): guard endpoint cleanup by serve ownership
Preserve independently launched models instead of sweeping endpoints by display name or mutable model pins.
Bind Cookbook endpoint registrations to their serve sessions and prevent stale watchdogs or scheduled stops from deleting newer registrations.
Reuse the existing endpoint deletion contract to preserve settings, preferences, session cleanup, and cache invalidation.
Cover LLM and image endpoint lifecycle regressions.
---------
Co-authored-by: Alexandre Teixeira <general@alexandreteixeira.dev>
When the window widens after an auto-collapse, the restore check removes
.hidden and measures the chat width to decide whether the sidebar fits.
html.ody-sidebar-mini also forces the sidebar to width 0 and was only
cleared by syncRailSide() after the measurement, so the check always saw
a 0-width sidebar and restored it even when the chat ended up narrower
than MIN_CHAT_WIDTH. Sync the mode classes before measuring.
The global .hidden utility (display: none) also matches #sidebar.hidden,
so the collapsed sidebar was removed from layout before its width/opacity
transition could run, and opening and closing both snapped in one frame.
Keep the collapsed sidebar rendered and take it out of interaction once
closed: .sidebar.hidden stays display: flex, gains pointer-events: none,
and visibility: hidden that is transitioned so it only flips after the
collapse finishes (keeping the closed sidebar out of the tab order).
On desktop the icon rail now tucks under the sidebar and slides out as it
collapses, so the combined width eases instead of jumping by 48px when
the two swap. On mobile the drawer keeps its width while off so it slides
out as well as in, and switching it to the right edge before opening no
longer animates it across the screen. Transitions are disabled under
prefers-reduced-motion, and the auto-collapse restore measures the chat
width with the transition suppressed.
Re-record the computed-style snapshot baseline for the sidebar and icon
rail.
