# The instance's assistants (Halkyon Learn's ADR 0021, points 4 and 8, D37, D38, D63):
# copy to deploy/assistants.env, restrict it (chmod 600), and fill in. It is git-ignored,
# goes to Halkyon Learn's server alone (compose.learn.yaml, `env_file`), never into an image,
# the database or its backups; keep it with deploy/.env's secrets (in Proton Pass).
# Without the file, the instance offers no assistant, as before. A change takes a
# restart: docker compose up -d server. Anything the server can't use stops it as it
# starts, naming the variable.
#
# Nobody has an assistant until an operator grants it to them in the app, with a budget
# per day and month for all their assistants together, in money and in tokens (D58,
# D59); each person agrees once to what goes where. The instance keeps what each person
# used (requests, tokens, cost) per day for 3 months, never what was asked.

# The assistants' IDs, in the order people see them, by commas or spaces: a letter,
# then letters, digits or _. Each names its variables below (claude: ASSISTANT_CLAUDE_…)
# and its grants; renaming one leaves its grants unused.
ASSISTANTS=claude

# The key the providers' opaque IDs of people are made with (D30): each request carries
# one, so a provider can act against one person's abuse without blocking the key, and
# can't tell who it is. At least 32 characters; a new one gives everyone a new ID.
# Generate: openssl rand -hex 32
ASSISTANT_ID_KEY=

# The name people see; the ID if unset.
ASSISTANT_CLAUDE_NAME=Claude Sonnet
# anthropic (Anthropic's Messages API), openai (OpenAI's chat completions, and every
# service compatible with them: Mistral, Gemini, OpenRouter, Ollama, vLLM, llama.cpp),
# or openai-responses (OpenAI's Responses API, which OpenAI's newest models need for
# tools; Ollama, vLLM and OpenRouter speak it too).
ASSISTANT_CLAUDE_PROTOCOL=anthropic
# The provider's address; unset: its own (https://api.anthropic.com, or
# https://api.openai.com/v1 for both of OpenAI's). With /v1 for the OpenAI protocols
# (https://api.mistral.ai/v1, https://openrouter.ai/api/v1,
# https://generativelanguage.googleapis.com/v1beta/openai), without it for Anthropic's.
# https only, but for a model on this machine or network (http://ollama:11434/v1).
#ASSISTANT_CLAUDE_BASE_URL=
ASSISTANT_CLAUDE_MODEL=claude-sonnet-5-5
# The operator's key at the provider; may be unset only with a BASE_URL of a server of
# one's own that takes none.
ASSISTANT_CLAUDE_KEY=
# on: people may add images (a card's picture, a PDF page) for a model that reads them.
#ASSISTANT_CLAUDE_IMAGES=on
# What the model costs per million tokens, as the provider bills it, in
# ASSISTANT_CURRENCY, with a point for decimals: both or none. With prices, it counts
# against a budget's money and tokens; without, against its tokens alone. A change
# counts from then on.
#ASSISTANT_CLAUDE_PRICE_INPUT=2
#ASSISTANT_CLAUDE_PRICE_OUTPUT=10
# How much the model thinks before it answers: low, medium, high, xhigh or max
# (Anthropic's names, sent in each provider's own field; Gemini takes low to high,
# Mistral none). EFFORT is what a question takes unless the person picks one;
# MAX_EFFORT the highest they may pick (unset: EFFORT). Both unset: the model's own,
# and nothing to pick. More effort spends more tokens, at the same price per token, and
# leaves more room for the answer (8,192 tokens at low to 65,536 at max).
#ASSISTANT_CLAUDE_EFFORT=medium
#ASSISTANT_CLAUDE_MAX_EFFORT=high

# A second one, through OpenAI's Responses API:
#ASSISTANTS=claude gpt
#ASSISTANT_GPT_NAME=GPT
#ASSISTANT_GPT_PROTOCOL=openai-responses
#ASSISTANT_GPT_MODEL=
#ASSISTANT_GPT_KEY=

# The currency of the prices (ISO 4217); unset: USD, as the providers bill.
#ASSISTANT_CURRENCY=USD

# on: people may add their own keys in the app, which then calls their provider
# directly; the instance sees neither key nor question and counts nothing (D5, D6).
#ASSISTANT_OWN_KEYS=on