sim-one-alpha

Model System

Flue agents reference models with string specifiers:

provider-id/model-id

Built-in providers such as OpenAI and Anthropic can be referenced directly when their provider credentials are available. Non-built-in or OpenAI-compatible providers, such as Ollama and local model gateways, must be registered in project code before an agent references them.

This project keeps model cards under their owning provider directories in src/core/models/providers/<provider>/cards so the orchestrator does not hardcode model IDs, context limits, or provider-specific capabilities.

Why Cards Exist

Cards are the boundary between model selection and runtime behavior. They make model metadata explicit before the agent, session manager, RAG router, or compaction layer makes token-budget decisions.

The same provider can expose models with very different limits. MiniMax M3, DeepSeek V4 Pro, and Qwen 3.5 all run through ollama-cloud, but they do not have the same context window, output ceiling, modalities, or long-context reliability notes. Keeping those facts in card files lets downstream systems ask the selected card for limits instead of spreading hardcoded numbers across the codebase.

Cards also preserve conflicting-but-useful metadata. For example, MiniMax advertises M3 as a 1M-context model with a guaranteed 512K minimum, while Ollama Cloud currently reports 524288. The card stores each value separately so session budgeting can choose the conservative operational budget while still retaining the advertised/native capability.

Runtime Use

Runtime setup has two layers:

  1. Provider modules in src/core/models/providers/<provider> register Flue provider IDs from src/app.ts.
  2. Model cards in each provider’s cards/ directory provide the model IDs and per-model metadata used by those provider registrations and by the orchestrator model registry.
  3. The shipped gorombo.config.json runtime file selects active project model card keys for the running deployment.
  4. src/core/models/catalog.ts aggregates provider-owned cards for lookup by Flue model specifier.

The orchestrator should not register providers itself. Runtime config selects a primary project model card and an optional backup card. The durable orchestrator agent uses the active card’s specifier as the Flue model string. Session-budget and compaction code uses the selected card before estimating, reserving, or compacting context.

Ollama Cloud, Ollama Local, And Codex Brain

Ollama has two different API paths:

The provider IDs intentionally reflect that split:

ollama-cloud  -> direct Ollama Cloud API
ollama-local  -> local Ollama endpoint
codex-brain   -> DT1 Codex Brain endpoint

Cloud model cards use the direct Ollama Cloud model IDs, such as minimax-m3 and deepseek-v4-pro. The :cloud suffix is for the local Ollama CLI/daemon path.

Current Cards

src/core/models/catalog.ts aggregates eight enabled cards:

Card key Flue specifier Role
minimax-m3-cloud ollama-cloud/minimax-m3 Agentic chat
deepseek-v4-pro-cloud ollama-cloud/deepseek-v4-pro Agentic chat
qwen3-5-cloud ollama-cloud/qwen3.5:397b Agentic chat
kimi-k2.7-code-cloud ollama-cloud/kimi-k2.7-code:cloud Agentic chat
codex-brain codex-brain/gpt-5.5 Agentic chat
nomic-embed-text-cloud ollama-cloud/nomic-embed-text Embedding
nomic-embed-text-local ollama-local/nomic-embed-text Embedding
all-minilm-l6-v2-onnx onnx-local/all-minilm-l6-v2 Embedding

minimax-m3-cloud is the default agentic-chat card. The runtime chat registry uses the seven provider-backed chat and embedding cards from createModelCards(). The embedding chain reads the complete catalog so the bundled ONNX card participates in cloud-to-local fallback.

Card key Advertised context Guaranteed or provider-reported context Output limit or embedding dimensions
minimax-m3-cloud 1,000,000 guaranteed 512,000; provider 524,288 131,072
deepseek-v4-pro-cloud 1,048,576 provider 1,048,576 1,048,576
qwen3-5-cloud 262,144 provider 262,144 65,536
kimi-k2.7-code-cloud 1,000,000 guaranteed 256,000; provider 262,144 32,768
codex-brain 128,000 not separately reported 32,000
nomic-embed-text-cloud 8,192 not separately reported 768 dimensions
nomic-embed-text-local 8,192 not separately reported 768 dimensions
all-minilm-l6-v2-onnx 256 not separately reported 384 dimensions

MiniMax M3 intentionally tracks more than one limit because MiniMax advertises 1M context with a guaranteed 512K minimum, while Ollama Cloud currently reports 524288 for both direct /api/show and local minimax-m3:cloud metadata.

Selection Rules

  1. The shipped gorombo.config.json runtime file is the source of truth for model choice.
  2. models.primary selects the primary project-owned model card key.
  3. models.backup optionally selects a different backup model card key.
  4. GOROMBO_MODEL and GOROMBO_MODEL_BACKUP are rejected.
  5. Raw Flue model specifiers are not accepted as configuration.
  6. If a requested card key is missing, startup fails instead of choosing an unreviewed fallback.

Current default runtime config:

{
  "version": 1,
  "models": {
    "primary": "minimax-m3-cloud",
    "backup": "codex-brain"
  }
}

Adding Models

Add a card file under the provider that owns the model, such as src/core/models/providers/ollama-cloud/cards/new-model.ts, including:

Provider transport belongs in src/core/models/providers/<provider>/provider.ts. Register custom providers from src/app.ts through the model runtime bootstrap before any agent uses the card specifier.

Embedding selection is role-based. The current chain tries nomic-embed-text-cloud, then the bundled all-minilm-l6-v2-onnx, then nomic-embed-text-local.

Budget Policy

The session-management layer uses cards in this order:

  1. Resolve the primary and backup model cards from runtime config.
  2. Use the provider-reported context window for safety when available.
  3. Reserve output tokens before calculating usable input budget.
  4. Warn before the prompt approaches the compaction threshold.
  5. Derive current session usage from stored Flue SessionData when available.
  6. Trigger session.compact() before Flue or the provider rejects the prompt.
  7. Give RAG only the remaining context budget after system instructions, protocol context, memory, current user input, and output reserve are accounted for.

The durable orchestrator agent prompts with the primary card. Backup cards remain explicit model metadata and are not a reason to bypass context budgeting.

RAG should come after this budget layer because retrieved context must fit into the selected card’s remaining budget.

The default web retrieval provider is Ollama Search through POST https://ollama.com/api/web_search, authenticated with the existing Ollama key resolved from model-card/provider env bindings. Web search is owned by the researcher subagent. The researcher-facing web_research tool calls the web-research Flue workflow boundary, which plans searches, uses cache, optionally expands top web results through fetch, packs contexts to a token budget, and records non-fatal provider failures. Other providers can be added behind the same RAG provider interface.

The main orchestrator registers a Flue subagent named researcher. The orchestrator must not call web search or web-capable retrieval directly. It delegates current, external, source-backed, or research tasks through Flue task with agent: "researcher". The researcher may use workflows and tools to decide whether one search, many searches, page fetches, or cache hits are enough.

This site is open source. Improve this page.