mirror of
https://github.com/rennf93/roboco.git
synced 2026-08-03 07:23:24 +02:00
Renn F
676a87985f
- llm_catalog: OLLAMA_DEFAULT_MODEL minimax-m3:cloud → kimi-k2.7-code:cloud;
role defaults kimi-k2.6→kimi-k2.7-code, developer minimax→kimi, product_owner/
ceo kimi→glm-5.2, documenter glm→kimi; GLM 5.1→5.2 comment fix.
- config + .env.example + docker-compose{.yml,.yaml,.registry.yml} + docs +
memory_distiller + optimal_brain: glm-5:cloud → glm-5.2:cloud.
- panel ai-routing-card: typed SelfHostedModel/boolean annotations; drop the
stale "Minimax M3 default" string (default is now catalog-driven).
- tests: glm-5:cloud → glm-5.2:cloud in pricing + rate-limit-retry fixtures.
133 lines
5.4 KiB
Python
133 lines
5.4 KiB
Python
"""
|
|
Model Catalog
|
|
|
|
Preset list of (model_name, provider_type, display_name) that the Settings
|
|
UI renders. Users never type a model name — they pick from this list, and
|
|
the router maps back to the correct pre-seeded provider row.
|
|
|
|
**Anthropic entries derive from `runtime.MODEL_MAP`** — that's the single
|
|
source of truth for which Claude versions are supported. Bumping a model
|
|
version there (e.g. `claude-opus-4-6` → `claude-opus-4-7`) updates the
|
|
catalog automatically without a second edit here.
|
|
|
|
Ollama Cloud entries are hand-maintained because Ollama's cloud tags
|
|
don't live in the rest of the codebase.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from dataclasses import dataclass
|
|
|
|
from roboco.models.base import ModelProvider
|
|
from roboco.models.runtime import MODEL_MAP
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class CatalogEntry:
|
|
"""One selectable model in the Settings dropdown."""
|
|
|
|
model_name: str
|
|
provider_type: ModelProvider
|
|
display_name: str
|
|
|
|
|
|
# Display labels for the Anthropic short names. Order of this tuple is the
|
|
# render order in the UI dropdown; entries missing from MODEL_MAP are
|
|
# silently skipped so we don't expose a model we can't route to.
|
|
_ANTHROPIC_DISPLAY: tuple[tuple[str, str], ...] = (
|
|
("opus", "Claude Opus"),
|
|
("sonnet", "Claude Sonnet"),
|
|
("haiku", "Claude Haiku"),
|
|
)
|
|
|
|
|
|
def _build_anthropic_entries() -> tuple[CatalogEntry, ...]:
|
|
"""Expand MODEL_MAP into catalog entries with full version in the label.
|
|
|
|
Keeps the UI honest: a user picking "Claude Opus" sees exactly which
|
|
underlying Claude Code model id will be used at spawn
|
|
(e.g. "Claude Opus · claude-opus-4-6"), so version bumps in
|
|
`runtime.MODEL_MAP` are immediately visible.
|
|
"""
|
|
entries: list[CatalogEntry] = []
|
|
for short_name, label in _ANTHROPIC_DISPLAY:
|
|
full_id = MODEL_MAP.get(short_name)
|
|
if not full_id:
|
|
continue
|
|
entries.append(
|
|
CatalogEntry(
|
|
model_name=short_name,
|
|
provider_type=ModelProvider.ANTHROPIC,
|
|
display_name=f"{label} · {full_id}",
|
|
)
|
|
)
|
|
return tuple(entries)
|
|
|
|
|
|
MODEL_CATALOG: tuple[CatalogEntry, ...] = (
|
|
*_build_anthropic_entries(),
|
|
# --- Ollama Cloud (verbatim tags) ---
|
|
# Pro plan active as of 2026-04-22. Drop any entry that stops working —
|
|
# the catalog is the single source of truth the Settings dropdown renders from.
|
|
CatalogEntry("glm-5.2:cloud", ModelProvider.OLLAMA_CLOUD, "GLM 5.2"),
|
|
CatalogEntry("kimi-k2.6:cloud", ModelProvider.OLLAMA_CLOUD, "Kimi K2.6"),
|
|
CatalogEntry("kimi-k2.7-code:cloud", ModelProvider.OLLAMA_CLOUD, "Kimi K2.7 Code"),
|
|
CatalogEntry("minimax-m3:cloud", ModelProvider.OLLAMA_CLOUD, "Minimax M3"),
|
|
CatalogEntry(
|
|
"nemotron-3-ultra:cloud", ModelProvider.OLLAMA_CLOUD, "Nemotron 3 Ultra"
|
|
),
|
|
# --- Grok (xAI, OpenAI protocol) ---
|
|
# Routes to the GROK provider → GrokCliProvider spawn (api.x.ai/v1). The xAI
|
|
# key is set via PUT /api/providers/grok/key.
|
|
CatalogEntry("grok-build-0.1", ModelProvider.GROK, "Grok Build 0.1"),
|
|
)
|
|
|
|
|
|
# Fast lookup by model_name. Enforces uniqueness at import time.
|
|
MODEL_CATALOG_BY_NAME: dict[str, CatalogEntry] = {
|
|
e.model_name: e for e in MODEL_CATALOG
|
|
}
|
|
assert len(MODEL_CATALOG_BY_NAME) == len(MODEL_CATALOG), "duplicate model_name"
|
|
|
|
|
|
def provider_type_for_model(model_name: str) -> ModelProvider | None:
|
|
"""Return the provider type for a catalog entry, or None if unknown."""
|
|
entry = MODEL_CATALOG_BY_NAME.get(model_name)
|
|
return entry.provider_type if entry else None
|
|
|
|
|
|
# Defaults per role when the user flips to "pure Ollama" mode.
|
|
# Assignments reflect the 2026-04 public benchmarks for each cloud tag:
|
|
# Kimi K2.6 — HLE 44.9%, AIME 95.6%, Agent Swarm (100 parallel sub-agents),
|
|
# 200-300 sequential tool calls. Best at reasoning, orchestration, tool use.
|
|
# MiniMax M3 — SWE-Bench 73.8%, SWE-Pro 56.2%, 10B active params (fastest,
|
|
# cheapest). Explicitly "built for Max coding & agentic workflows".
|
|
# GLM 5.2 — SWE-Bench 77.8% (highest of the three, 94.6% of Claude Opus 4.6),
|
|
# self-correcting across hundreds of iterations, strong creative writing.
|
|
OLLAMA_ROLE_DEFAULTS: dict[str, str] = {
|
|
# High-volume agentic coding — M3 is purpose-built for this.
|
|
"developer": "kimi-k2.7-code:cloud",
|
|
# Deep code review — GLM 5.2 has the highest SWE-Bench and iterates thoroughly.
|
|
"qa": "glm-5.2:cloud",
|
|
# Orchestration + tool coordination — Kimi K2.6's Agent Swarm is the exact fit.
|
|
"cell_pm": "kimi-k2.7-code:cloud",
|
|
"main_pm": "kimi-k2.7-code:cloud",
|
|
# Quality reasoning — Kimi K2.6 leads HLE by a wide margin.
|
|
"auditor": "kimi-k2.7-code:cloud",
|
|
# Product reasoning — same profile as PM work.
|
|
"product_owner": "glm-5.2:cloud",
|
|
# Writing with code-context — GLM 5.2's creative writing + SWE-Bench combo.
|
|
"documenter": "kimi-k2.7-code:cloud",
|
|
# Stylistic writing — GLM 5.2's creative-writing strength.
|
|
"head_marketing": "glm-5.2:cloud",
|
|
# CEO is human-in-the-loop; keep an entry in case someone forces
|
|
# a route to it, but the Settings UI intentionally excludes it.
|
|
"ceo": "glm-5.2:cloud",
|
|
}
|
|
|
|
# The Ollama model picked for "pure Ollama" mode's GLOBAL row when the
|
|
# caller doesn't override. Minimax M3 wins as the generalist because it has
|
|
# the strongest reasoning/tool-use profile and can fall back to coding/writing
|
|
# adequately if a role ends up mapped to the global default.
|
|
OLLAMA_DEFAULT_MODEL: str = "kimi-k2.7-code:cloud"
|