AI: personal assistants without the dead API, and keep Ollama warm

Two things the field report asked for.

1) PERSONAL ASSISTANTS (replacing the sunset OpenAI Assistants API)

The old implementation gave three capabilities. Two are reimplemented here,
the third was confirmed unused and is deliberately not replaced:

 * per-user persona - it already lived in system_gpt_settings.json; it was
   only ever being shipped to OpenAI. It is now the system prompt.
 * per-user conversation thread - OpenAI held this server-side. It now lives
   in assistant_memory.json, keyed by discord user id, trimmed to the most
   recent turns (CONJURER_ASSISTANT_MEMORY_TURNS) and written atomically so a
   torn write cannot lose someone's history. Deliberately a plain trim, not
   the AI summarisation used for the bar's shared memory: these are private
   DMs and must not end up in a public "legend".
 * file_search - not replaced. Confirmed not in use.

The conversation goes through handle_response with request_type="NONE" and an
explicit message list, which keeps it out of the bar's shared memory. The big
win: create_chat_assistant hardcoded model="gpt-4o", so assistants were locked
to OpenAI. They now run on whatever $gadaj_teraz selects - Claude and Ollama
included.

create_chat_assistant / chat_with_assistant are gone, and with them the last
call to beta.threads in the startup path - so the cog cannot be killed by that
API again. (add_files_to_vector_store / delete_files_from_vector_store still
reference beta.assistants but are dead code - nothing calls them - so they
cannot crash anything; left alone rather than widening this change.)

2) KEEPING A SELF-HOSTED MODEL WARM

Loading is the slow part - the GPU is shared with other users - so we preload
via Ollama's documented mechanism: /api/generate with a model, a keep_alive
and NO prompt. It loads the model and generates nothing.

 * on switching to ollama, $gadaj_teraz fires a preload in the BACKGROUND
   (not awaited: loading can take minutes and the command must answer at
   once), so the wait lands on the operator rather than the first user;
 * a warm loop re-asserts keep_alive every CONJURER_OLLAMA_WARM_MINUTES.

Both are hard-guarded on the ACTIVE provider being ollama. Warming a metered
API would burn tokens and money for nothing, so that guard is pinned by a test
asserting the preload is never called for gpt/claude, and another asserting the
preload body carries no prompt (a prompt would make every warm-up generate).

Tests: 82 unit + 71 integration green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

constants.py

@@ -245,6 +245,16 @@ DELIVERED_DIR = os.getenv(
 )
 DELIVERED_MAX = int(os.getenv("CONJURER_DELIVERED_MAX", "10000"))
 
+# Personal DM assistants. Replaces the OpenAI Assistants API (threads/runs),
+# which was sunset and answers 404: the persona now rides handle_response, so it
+# works on EVERY backend, and the conversation lives here instead of on OpenAI's
+# server. Kept per user so private DMs never bleed into the bar's shared memory,
+# and trimmed to the most recent turns so it cannot grow without bound.
+ASSISTANT_MEMORY_FILE = os.getenv(
+    "CONJURER_ASSISTANT_MEMORY", os.path.join(_STATE_ROOT, "assistant_memory.json")
+)
+ASSISTANT_MEMORY_TURNS = int(os.getenv("CONJURER_ASSISTANT_MEMORY_TURNS", "40"))
+
 FILE_SERVICE_ADDRESS = os.getenv("CONJURER_FILE_SERVICE", "http://192.168.1.15:5000")
 RADIO_HARBOR_ADDRESS = os.getenv("CONJURER_RADIO_HARBOR", "http://192.168.1.15:54321")
 # Betoniarka (radio-operator service colocated with Liquidsoap). Falls back to
@@ -411,6 +421,19 @@ else:
 AI_TIMEOUT_SECONDS = int(os.getenv("CONJURER_AI_TIMEOUT_SECONDS", "120"))
 
 OLLAMA_URL = os.getenv("CONJURER_OLLAMA_URL", "").rstrip("/")
+# Keeping a self-hosted model resident. Loading it is the slow part (it is
+# offloaded to a GPU shared with other users), so we preload it - Ollama's
+# /api/generate with a model and NO prompt loads it and generates nothing, which
+# costs no tokens and no money. KEEP_ALIVE is how long Ollama should then hold
+# it; the warm loop re-asserts that well inside the window.
+# STRICTLY Ollama-only: doing this against a paid API would burn tokens for
+# nothing, so every caller checks the active provider first.
+OLLAMA_KEEP_ALIVE = os.getenv("CONJURER_OLLAMA_KEEP_ALIVE", "30m")
+OLLAMA_WARM_MINUTES = float(os.getenv("CONJURER_OLLAMA_WARM_MINUTES", "10"))
+# A preload waits for the model to finish loading, which on a shared GPU is the
+# slow path we are trying to move off the user's first message.
+OLLAMA_PRELOAD_TIMEOUT = int(os.getenv("CONJURER_OLLAMA_PRELOAD_TIMEOUT", "600"))
+
 OLLAMA_LATEST_MODEL = os.getenv("CONJURER_OLLAMA_MODEL", "llama3.1:8b")
 OLLAMA_CHEAP_MODEL = os.getenv("CONJURER_OLLAMA_CHEAP_MODEL", OLLAMA_LATEST_MODEL)
 if openai and OLLAMA_URL: