config/models.yaml

# Logical model profiles → small models served on Modal.
#
# offline:
#   null  = live inference required (default). The build refuses to start if no
#           backend is configured (MODAL_WORKSPACE / MODAL_LLM_BASE_URL / HF_TOKEN);
#           there is no silent fallback to the stub.
#   false = live inference (same as null, stated explicitly).
#   true  = deterministic stub for every profile — a test/dev seam (the test suite
#           sets this), NOT a product mode.
#
# Each profile binds to a model by its **catalogue key** (`endpoint:`) — the model
# slug in `modal/catalogue.py`, the single source of truth for what is deployed.
# The loader (`Registry.from_dir`) expands that key into the concrete binding:
#   model    = openai/<served_model_id>
#   base_url = https://${MODAL_WORKSPACE}--<app>-<endpoint>.modal.run/v1
#              (or $MODAL_LLM_BASE_URL if set; a backend is required to run)
#   api_key  = $MODAL_LLM_KEY   (a self-served vLLM endpoint accepts any token)
#
# So adding/retuning a model is a one-line edit in `modal/catalogue.py`, and
# pointing a tier at a different model is a one-line `endpoint:` change here —
# nothing duplicates the served-id or URL. Env vars MODEL_TINY / MODEL_FAST /
# MODEL_BALANCED / MODEL_STRONG still override the model string per profile
# (highest priority). You may also bind a profile explicitly with `model:` +
# `base_url:` instead of `endpoint:` (an escape hatch for non-catalogue endpoints).
#
# `endpoint:` keys may name a model on EITHER inference backend (ADR-0024): a bare
# slug is a Modal-served model (as below); a `hf:<repo>` key is a Hugging Face
# serverless model from `src/models/hf_catalogue.py`, resolved against HF_TOKEN. E.g.
#   balanced: {endpoint: "hf:google/gemma-2-9b-it", temperature: 0.8, max_tokens: 320}
# The Fishbowl Lab picks the backend + per-agent model interactively; this file is the
# headless/default-run binding.
offline: null
profiles:
  tiny:
    endpoint: nemotron-3-nano-4b     # NVIDIA Nemotron 3 Nano 4B (≤4B, Tiny Titan)
    temperature: 0.7
    max_tokens: 192
  fast:
    endpoint: minicpm-4-1-8b         # OpenBMB MiniCPM4.1-8B
    temperature: 0.9
    max_tokens: 320
  # balanced/strong are served WITH a reasoning parser (modal/catalogue.py:
  # reasoning_parser="gemma4") — they THINK before answering, and that thinking
  # counts against max_tokens. Budget too low → the model is truncated mid-thought
  # and emits an EMPTY answer (the "agents stopped working / just …" symptom). Give
  # the reasoning room; the thinking is captured separately as the mind-reader thought.
  balanced:
    endpoint: gemma-4-12b            # Google Gemma 4 12B (reasoning)
    temperature: 0.8
    max_tokens: 768
  # strong:
  #   endpoint: gemma-4-26b            # Google Gemma 4 26B-A4B-it (MoE, ~4B active; reasoning)
  #   temperature: 0.6
  #   max_tokens: 1024