@qadanm/opencode-smart-router 0.0.0-stage → 0.6.1

package/src/core/models/seed.ts

@@ -0,0 +1,781 @@
+import type { CostTier, ModelRecord, PeakPricing } from "../types.js"
+import { OPENDESIGN_MODELS } from "../providers/opendesign-models.js"
+import { OPENDESIGN_DESIGN_PLAN_MODELS, modelMonthlyAllowanceUsd, resolvePlan } from "../plans/catalog.js"
+
+export interface SeedModel {
+  key: string
+  providerID: string
+  modelID: string
+  name: string
+  context: number
+  maxOutput: number
+  input: string[]
+  tools?: boolean
+  cost: CostTier[]
+  peakPricing?: PeakPricing
+  intelligence: number
+  design: number
+  designEvidence?: ModelRecord["designEvidence"]
+  /** Credit multiplier applied to usage value. One for plain USD billing. */
+  creditMultiplier?: number
+  /** True when the model is funded from a pool reserved for escalation. */
+  escalationOnly?: boolean
+  premium?: boolean
+  /** Per-kind capability evidence, passed through with provenance intact. */
+  capabilities?: ModelRecord["capabilities"]
+  training?: boolean
+  free?: boolean
+  tier: 1 | 2 | 3
+  approved?: boolean
+  monthlyAllowanceUsd: number
+  protocol?: "anthropic" | "openai-compat"
+  notes?: string
+}
+
+const cost = (input: number, output: number, read: number, write = 0): CostTier[] => [
+  { input, output, cache: { read, write } },
+]
+
+const tieredCost = (
+  base: CostTier,
+  size: number,
+  high: { input: number; output: number; read: number; write: number },
+): CostTier[] => [
+  base,
+  {
+    input: high.input,
+    output: high.output,
+    cache: { read: high.read, write: high.write },
+    tier: { type: "context", size },
+  },
+]
+
+/**
+ * Command Code documents these DeepSeek models at off-peak rates and charges
+ * 2x from 01:00-04:00 and 06:00-10:00 UTC, Monday-Friday.
+ *
+ * OpenCode Go publishes the identical peak schedule for DeepSeek V4.1 Flash
+ * and V4 Pro, so the same shape is reused rather than duplicated.
+ */
+const DEEPSEEK_PEAK_PRICING: PeakPricing = {
+  multiplier: 2,
+  windowsUtc: [
+    { startHour: 1, endHour: 4 },
+    { startHour: 6, endHour: 10 },
+  ],
+  weekdaysOnly: true,
+}
+
+const COMMAND_CODE_DEEPSEEK_PEAK = DEEPSEEK_PEAK_PRICING
+
+/**
+ * Curated seed. Intelligence values come from Command Code's published model
+ * cards where available. Models that are new or unscored are deliberately not
+ * approved for automatic substantive work.
+ *
+ * Design scores are curated quality estimates, not benchmark results. They are
+ * separate from intelligence and are documented in docs/MODELS.md.
+ */
+const BASE_SEED_MODELS: ReadonlyArray<SeedModel> = [
+  // ---------------------------------------------------------------------------
+  // OpenCode Go: economy
+  // ---------------------------------------------------------------------------
+  {
+    key: "opencode-go/gpt-6-luna",
+    providerID: "opencode-go",
+    modelID: "gpt-6-luna",
+    name: "GPT-6 Luna",
+    context: 1_050_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(0.1, 0.5, 0.01, 0.125)[0]!, 272_000, { input: 0.2, output: 0.75, read: 0.02, write: 0.25 }),
+    intelligence: 38.1,
+    design: 55,
+    designEvidence: "curated",
+    tier: 1,
+    monthlyAllowanceUsd: 60,
+    notes: "Economy tier. Reliable for narrow mechanical edits, below the standard intelligence floor.",
+  },
+  {
+    key: "opencode-go/mimo-v2.6-flash",
+    providerID: "opencode-go",
+    modelID: "mimo-v2.6-flash",
+    name: "MiMo-V2.6-Flash",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "audio", "video"],
+    cost: cost(0.14, 0.28, 0.0028),
+    intelligence: 37.9,
+    design: 50,
+    designEvidence: "curated",
+    tier: 1,
+    monthlyAllowanceUsd: 120,
+    notes: "Cheap multimodal model. Economy tier only.",
+  },
+  {
+    key: "opencode-go/hy3",
+    providerID: "opencode-go",
+    modelID: "hy3",
+    name: "Hy3",
+    context: 256_000,
+    maxOutput: 128_000,
+    input: ["text"],
+    cost: cost(0.14, 0.58, 0.035),
+    intelligence: 25.3,
+    design: 45,
+    designEvidence: "curated",
+    tier: 1,
+    monthlyAllowanceUsd: 240,
+    notes: "Economy tier. Text only. Narrow mechanical work.",
+  },
+  {
+    key: "opencode-go/longcat-2.0",
+    providerID: "opencode-go",
+    modelID: "longcat-2.0",
+    name: "LongCat-2.0",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text"],
+    cost: cost(0.3, 1.2, 0.006),
+    intelligence: 19.1,
+    design: 40,
+    designEvidence: "curated",
+    tier: 1,
+    monthlyAllowanceUsd: 240,
+    notes: "Economy tier. Very low intelligence; use only with deterministic validation.",
+  },
+
+  // ---------------------------------------------------------------------------
+  // OpenCode Go: standard
+  // ---------------------------------------------------------------------------
+  {
+    key: "opencode-go/deepseek-v4.1-flash",
+    providerID: "opencode-go",
+    modelID: "deepseek-v4.1-flash",
+    name: "DeepSeek V4.1 Flash",
+    context: 1_000_000,
+    maxOutput: 384_000,
+    input: ["text", "image"],
+    cost: cost(0.15, 0.6, 0.003),
+    peakPricing: DEEPSEEK_PEAK_PRICING,
+    intelligence: 39.5,
+    design: 65,
+    designEvidence: "observed",
+    tier: 2,
+    monthlyAllowanceUsd: 80,
+    notes:
+      "Approved standard floor. Strong general coding value and native image input. Off-peak list price; 2x during 01-04 and 06-10 UTC Mon-Fri.",
+  },
+  {
+    key: "opencode-go/glm-5.3-flash",
+    providerID: "opencode-go",
+    modelID: "glm-5.3-flash",
+    name: "GLM-5.3-Flash",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text", "image", "video"],
+    cost: cost(0.15, 0.5, 0.03),
+    intelligence: 41.8,
+    design: 65,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 60,
+  },
+  {
+    key: "opencode-go/minimax-m3.1-flash",
+    providerID: "opencode-go",
+    modelID: "minimax-m3.1-flash",
+    name: "MiniMax M3.1 Flash",
+    context: 1_000_000,
+    maxOutput: 512_000,
+    input: ["text", "image", "video"],
+    cost: cost(0.15, 0.6, 0.015),
+    intelligence: 42.7,
+    design: 60,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 20,
+  },
+  {
+    key: "opencode-go/claude-haiku-5-5",
+    providerID: "opencode-go",
+    modelID: "claude-haiku-5-5",
+    name: "Claude Haiku 5.5",
+    context: 1_000_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(0.1, 0.5, 0.01, 0.125)[0]!, 100_000, { input: 0.5, output: 2.5, read: 0.05, write: 0.625 }),
+    intelligence: 43.4,
+    design: 70,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 15,
+  },
+  {
+    key: "opencode-go/mimo-v2.6-pro",
+    providerID: "opencode-go",
+    modelID: "mimo-v2.6-pro",
+    name: "MiMo-V2.6-Pro",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "audio", "video"],
+    cost: cost(0.435, 0.87, 0.003625),
+    intelligence: 46.3,
+    design: 62,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 15,
+    notes: "High intelligence for its price. Narrow monthly allowance.",
+  },
+
+  // ---------------------------------------------------------------------------
+  // OpenCode Go: premium
+  // ---------------------------------------------------------------------------
+  {
+    key: "opencode-go/glm-5.3",
+    providerID: "opencode-go",
+    modelID: "glm-5.3",
+    name: "GLM-5.3",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text"],
+    cost: cost(1.4, 4.4, 0.26),
+    intelligence: 44.8,
+    design: 80,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 15,
+    notes: "Premium. Text only, so it can never receive screenshots.",
+  },
+  {
+    key: "opencode-go/kimi-k3",
+    providerID: "opencode-go",
+    modelID: "kimi-k3",
+    name: "Kimi K3",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "video"],
+    cost: cost(3, 15, 0.3, 3),
+    intelligence: 43.6,
+    design: 72,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 15,
+  },
+  {
+    key: "opencode-go/grok-4.7",
+    providerID: "opencode-go",
+    modelID: "grok-4.7",
+    name: "Grok 4.7",
+    context: 500_000,
+    maxOutput: 500_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(2, 6, 0.5)[0]!, 200_000, { input: 4, output: 12, read: 1, write: 0 }),
+    intelligence: 46.4,
+    design: 75,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 15,
+  },
+  {
+    key: "opencode-go/qwen3.8-max",
+    providerID: "opencode-go",
+    modelID: "qwen3.8-max",
+    name: "Qwen3.8 Max",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text", "image", "video"],
+    cost: cost(2, 6, 0.25, 2.5),
+    intelligence: 40.2,
+    design: 70,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 15,
+  },
+  {
+    key: "opencode-go/step-5-preview",
+    providerID: "opencode-go",
+    modelID: "step-5-preview",
+    name: "Step 5 Preview",
+    context: 1_000_000,
+    maxOutput: 65_536,
+    input: ["text", "image", "video"],
+    cost: cost(1, 2.7, 0.1),
+    intelligence: 43.7,
+    design: 70,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 15,
+  },
+
+  // ---------------------------------------------------------------------------
+  // Command Code GOAT: standard
+  // ---------------------------------------------------------------------------
+  {
+    key: "commandcode/deepseek/deepseek-v4.1-flash",
+    providerID: "commandcode",
+    modelID: "deepseek/deepseek-v4.1-flash",
+    name: "DeepSeek V4.1 Flash (GOAT)",
+    context: 1_000_000,
+    maxOutput: 384_000,
+    input: ["text", "image"],
+    cost: cost(0.15, 0.6, 0.003),
+    peakPricing: COMMAND_CODE_DEEPSEEK_PEAK,
+    intelligence: 39.5,
+    design: 65,
+    designEvidence: "observed",
+    tier: 2,
+    monthlyAllowanceUsd: 80,
+    notes: "Approved standard floor on Command Code. Off-peak list price; 2x during 01-04 and 06-10 UTC Mon-Fri.",
+  },
+  {
+    key: "commandcode/z-ai/glm-5.3-flash",
+    providerID: "commandcode",
+    modelID: "z-ai/glm-5.3-flash",
+    name: "GLM-5.3-Flash",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text", "image", "video", "pdf"],
+    cost: cost(0.15, 0.5, 0.03),
+    intelligence: 41.8,
+    design: 65,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 80,
+  },
+  {
+    key: "commandcode/claude-haiku-5-5",
+    providerID: "commandcode",
+    modelID: "claude-haiku-5-5",
+    name: "Claude Haiku 5.5",
+    context: 1_000_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(0.1, 0.5, 0.01, 0.125)[0]!, 100_000, { input: 0.5, output: 2.5, read: 0.05, write: 0.625 }),
+    intelligence: 43.4,
+    design: 70,
+    designEvidence: "curated",
+    tier: 2,
+    monthlyAllowanceUsd: 20,
+  },
+  {
+    key: "commandcode/gpt-6-luna",
+    providerID: "commandcode",
+    modelID: "gpt-6-luna",
+    name: "GPT-6 Luna",
+    context: 1_050_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(0.1, 0.5, 0.01, 0.125)[0]!, 272_000, { input: 0.2, output: 0.75, read: 0.02, write: 0.25 }),
+    intelligence: 38.1,
+    design: 55,
+    designEvidence: "curated",
+    tier: 1,
+    monthlyAllowanceUsd: 20,
+  },
+
+  // ---------------------------------------------------------------------------
+  // Command Code GOAT: premium
+  // ---------------------------------------------------------------------------
+  {
+    key: "commandcode/claude-sonnet-5-5",
+    providerID: "commandcode",
+    modelID: "claude-sonnet-5-5",
+    name: "Claude Sonnet 5.5",
+    context: 1_000_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: cost(2, 10, 0.1, 2.5),
+    intelligence: 56,
+    design: 90,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 10,
+    notes: "Premium design and architecture candidate. Very small monthly allowance.",
+  },
+  {
+    key: "commandcode/meta/muse-spark-1.3",
+    providerID: "commandcode",
+    modelID: "meta/muse-spark-1.3",
+    name: "Muse Spark 1.3",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "video", "pdf", "audio"],
+    cost: cost(1.25, 4.25, 0.15),
+    intelligence: 48.1,
+    design: 82,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 20,
+  },
+  {
+    key: "commandcode/google/gemini-4-argon",
+    providerID: "commandcode",
+    modelID: "google/gemini-4-argon",
+    name: "Gemini 4 Argon",
+    context: 1_000_000,
+    maxOutput: 65_536,
+    input: ["text", "image", "video", "audio", "pdf"],
+    cost: cost(2, 10, 0.1),
+    intelligence: 0,
+    design: 85,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 40,
+    notes: "Unscored in the current model cards. Excluded from automatic routing until qualified.",
+  },
+  {
+    key: "commandcode/xai/grok-4.7",
+    providerID: "commandcode",
+    modelID: "xai/grok-4.7",
+    name: "Grok 4.7",
+    context: 500_000,
+    maxOutput: 500_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(2, 6, 0.5)[0]!, 200_000, { input: 4, output: 12, read: 1, write: 0 }),
+    intelligence: 46.4,
+    design: 75,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 20,
+  },
+  {
+    key: "commandcode/Qwen/Qwen3.8-Max",
+    providerID: "commandcode",
+    modelID: "Qwen/Qwen3.8-Max",
+    name: "Qwen3.8 Max",
+    context: 1_000_000,
+    maxOutput: 131_072,
+    input: ["text", "image", "video", "pdf"],
+    cost: cost(2, 6, 0.25, 2.5),
+    intelligence: 40.2,
+    design: 70,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 20,
+  },
+  {
+    key: "commandcode/stepfun/Step-5-Preview",
+    providerID: "commandcode",
+    modelID: "stepfun/Step-5-Preview",
+    name: "Step 5 Preview",
+    context: 1_000_000,
+    maxOutput: 1_000_000,
+    input: ["text", "image", "video"],
+    cost: cost(1, 2.7, 0.05),
+    intelligence: 43.7,
+    design: 70,
+    designEvidence: "curated",
+    premium: true,
+    tier: 3,
+    monthlyAllowanceUsd: 20,
+  },
+
+  // ---------------------------------------------------------------------------
+  // Known but deliberately unapproved models. They are discovered so the user
+  // can qualify them explicitly, but automatic routing never selects them.
+  // ---------------------------------------------------------------------------
+  {
+    key: "commandcode/claude-opus-5-5",
+    providerID: "commandcode",
+    modelID: "claude-opus-5-5",
+    name: "Claude Opus 5.5",
+    context: 1_000_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: cost(4, 20, 0.2, 5),
+    intelligence: 0,
+    design: 92,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 10,
+    notes: "Unscored in the current model cards. Explicit approval required.",
+  },
+  {
+    key: "commandcode/gpt-6.1-sol",
+    providerID: "commandcode",
+    modelID: "gpt-6.1-sol",
+    name: "GPT-6.1 Sol",
+    context: 1_050_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(2, 10, 0.1, 2.5)[0]!, 272_000, { input: 4, output: 15, read: 0.2, write: 5 }),
+    intelligence: 55,
+    design: 88,
+    designEvidence: "unverified",
+    capabilities: { coding: 56, debugging: 65, agentic: 41, architecture: 86 },
+    premium: true,
+    tier: 3,
+    approved: true,
+    monthlyAllowanceUsd: 20,
+    notes:
+      "Intelligence 55 on Artificial Analysis v5.0.1 (2026-11-06). Capabilities: Terminal-Bench 4.0 56 percent, AutomationBench 65 percent, FrontierCode 2.0 41 percent, same index and configuration. Long-context tier doubles input to 4 and raises output to 15 above 272K. Auto-approved for tier-3 escalation by owner instruction 2026-11-06; the Command Code gateway advertises 1.05M but that window is unverified, so treat it as pending a passed probe.",
+  },
+  {
+    key: "commandcode/gpt-6-astra",
+    providerID: "commandcode",
+    modelID: "gpt-6-astra",
+    name: "GPT-6 Astra",
+    context: 1_050_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: tieredCost(cost(10, 50, 1, 12.5)[0]!, 272_000, { input: 20, output: 75, read: 2, write: 25 }),
+    intelligence: 0,
+    design: 88,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 10,
+    notes: "Frontier-priced model. Never selected automatically without explicit approval.",
+  },
+  {
+    key: "commandcode/claude-fable-5-1",
+    providerID: "commandcode",
+    modelID: "claude-fable-5-1",
+    name: "Claude Fable 5.1",
+    context: 1_000_000,
+    maxOutput: 128_000,
+    input: ["text", "image", "pdf"],
+    cost: cost(10, 50, 0.25, 12.5),
+    intelligence: 0,
+    design: 90,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 10,
+    notes: "Frontier-priced model. Never selected automatically without explicit approval.",
+  },
+  {
+    key: "opencode-go/gemini-4-argon",
+    providerID: "opencode-go",
+    modelID: "gemini-4-argon",
+    name: "Gemini 4 Argon",
+    context: 1_048_576,
+    maxOutput: 65_536,
+    input: ["text", "image", "video", "audio", "pdf"],
+    cost: cost(2, 10, 0.1),
+    intelligence: 0,
+    design: 85,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 120,
+    notes: "Unscored in the current model cards. Explicit approval required.",
+  },
+  {
+    key: "opencode-go/kimi-k3.1-pro",
+    providerID: "opencode-go",
+    modelID: "kimi-k3.1-pro",
+    name: "Kimi K3.1 Pro",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "video"],
+    cost: cost(2, 10, 0.2, 2),
+    intelligence: 0,
+    design: 75,
+    designEvidence: "unverified",
+    premium: true,
+    tier: 3,
+    approved: false,
+    monthlyAllowanceUsd: 60,
+    notes: "Unscored in the current model cards. Explicit approval required.",
+  },
+  {
+    key: "commandcode/meta/muse-spark-1.3-contributor",
+    providerID: "commandcode",
+    modelID: "meta/muse-spark-1.3-contributor",
+    name: "Muse Spark 1.3 (Contributor)",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "video", "pdf", "audio"],
+    cost: cost(0.1, 0.2, 0.002),
+    intelligence: 0,
+    design: 80,
+    designEvidence: "unverified",
+    training: true,
+    tier: 1,
+    approved: false,
+    monthlyAllowanceUsd: 20,
+    notes: "Training-enabled contributor tier. Blocked unless privacy.allowTrainingModels is true.",
+  },
+  {
+    key: "opencode-go/muse-spark-1.3-contributor",
+    providerID: "opencode-go",
+    modelID: "muse-spark-1.3-contributor",
+    name: "Muse Spark 1.3 Contributor",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text", "image", "video", "pdf", "audio"],
+    cost: cost(0.1, 0.2, 0.002),
+    intelligence: 0,
+    design: 80,
+    designEvidence: "unverified",
+    training: true,
+    tier: 1,
+    approved: false,
+    monthlyAllowanceUsd: 120,
+    notes: "Training-enabled contributor tier. Blocked unless privacy.allowTrainingModels is true.",
+  },
+  {
+    key: "opencode-go/iceman-contributor-free",
+    providerID: "opencode-go",
+    modelID: "iceman-contributor-free",
+    name: "Iceman Contributor Free",
+    context: 262_144,
+    maxOutput: 131_072,
+    input: ["text"],
+    cost: cost(0, 0, 0),
+    intelligence: 0,
+    design: 0,
+    designEvidence: "unverified",
+    training: true,
+    free: true,
+    tier: 1,
+    approved: false,
+    monthlyAllowanceUsd: 0,
+    notes: "Free model that trains on data. Blocked by default.",
+  },
+  {
+    key: "opencode-go/beam-free",
+    providerID: "opencode-go",
+    modelID: "beam-free",
+    name: "Beam Free",
+    context: 1_048_576,
+    maxOutput: 131_072,
+    input: ["text"],
+    cost: cost(0, 0, 0),
+    intelligence: 0,
+    design: 0,
+    designEvidence: "unverified",
+    free: true,
+    tier: 1,
+    approved: false,
+    monthlyAllowanceUsd: 0,
+    notes: "Free promotional model. Not approved for automatic work.",
+  },
+  {
+    key: "opencode-go/deepseek-v4-pro",
+    providerID: "opencode-go",
+    modelID: "deepseek-v4-pro",
+    name: "DeepSeek V4 Pro",
+    context: 1_000_000,
+    maxOutput: 384_000,
+    input: ["text"],
+    cost: cost(0.66, 1.98, 0.022),
+    intelligence: 36,
+    design: 55,
+    designEvidence: "curated",
+    tier: 2,
+    approved: false,
+    monthlyAllowanceUsd: 60,
+    notes: "Below the current standard floor and text only. Not approved automatically.",
+  },
+]
+
+/**
+ * OpenDesign: the models served by the OpenDesign gateway.
+ *
+ * These entries are generated from the provider catalog so the published
+ * credit multipliers live in exactly one place. The gateway's `GET /v1/models`
+ * response is authoritative for entitlement and for the usable context window,
+ * and the discovery adapter merges it over these values at runtime.
+ *
+ * Which funding pool a model draws from is decided by the plan catalog. A
+ * Design Plan model carries its multiplier; a plan model does not, because
+ * the multiplier does not follow it into the plan model pool.
+ */
+const OPENDESIGN_SEED_MODELS: ReadonlyArray<SeedModel> = OPENDESIGN_MODELS.map((model) => ({
+  key: `opendesign/${model.modelID}`,
+  providerID: "opendesign",
+  modelID: model.modelID,
+  name: model.name,
+  context: model.context,
+  maxOutput: model.maxOutput,
+  input: [...model.input],
+  tools: true,
+  cost: model.cost.map((tier) => ({ ...tier, cache: { ...tier.cache } })),
+  // The multiplier only applies to the Design Plan pool. A plan model is
+  // billed at plain usage value from the separate plan model pool, so it
+  // must not carry a Design Plan multiplier into cost estimation.
+  ...(OPENDESIGN_DESIGN_PLAN_MODELS.has(model.modelID) ? { creditMultiplier: model.creditMultiplier } : {}),
+  escalationOnly: !OPENDESIGN_DESIGN_PLAN_MODELS.has(model.modelID),
+  intelligence: model.intelligence,
+  design: model.design,
+  designEvidence: model.designEvidence,
+  // Per-kind benchmark evidence passes through with its provenance intact.
+  ...(model.capabilities ? { capabilities: { ...model.capabilities } } : {}),
+  tier: model.tier,
+  approved: model.approved,
+  // Plan models (Sonnet, Sol, Opus, Astra) are premium: they draw from the
+  // plan-model pool, which premium policy and its budget govern. Design Plan
+  // models are ordinary allowance, never premium.
+  premium: !OPENDESIGN_DESIGN_PLAN_MODELS.has(model.modelID),
+  // Every OpenDesign allowance is a shared provider-level pool, so no model
+  // carries its own monthly allowance.
+  monthlyAllowanceUsd: 0,
+  notes: model.notes,
+}))
+
+/**
+ * Local models served by Ollama.
+ *
+ * These are deliberately *not* approved for automatic online routing. A 4B
+ * local model scoring in the high 30s would be competitive with the hosted
+ * standard floor on paper, and the router would happily send real work to it,
+ * which is not what the owner wants when connected. Offline mode reaches them
+ * through the offline allow-list instead, where the choice is explicit.
+ *
+ * Cost is zero and there is no allowance, because inference runs on hardware
+ * the owner already paid for. `free: true` would route them through the
+ * privacy guard for promotional models, which is the wrong concept, so it is
+ * left unset.
+ */
+const LOCAL_SEED_MODELS: ReadonlyArray<SeedModel> = [
+  {
+    key: "ollama/qwen3.5:4b-q4_K_M",
+    providerID: "ollama",
+    modelID: "qwen3.5:4b-q4_K_M",
+    name: "Qwen3.5 4B (local)",
+    // Matches offline.model.contextTokens. A larger declared window would let
+    // the router believe it can take a transcript the server is configured to
+    // truncate.
+    context: 8_192,
+    maxOutput: 2_048,
+    input: ["text", "image"],
+    tools: true,
+    cost: [{ input: 0, output: 0, cache: { read: 0, write: 0 } }],
+    // Scored below the hosted standard floor on purpose: these numbers describe
+    // real capability, and reporting them low keeps the model out of any online
+    // automatic decision even if someone approves it later.
+    intelligence: 30,
+    design: 40,
+    designEvidence: "unverified",
+    tier: 1,
+    approved: false,
+    monthlyAllowanceUsd: 0,
+    notes:
+      "Local Ollama model for offline development. Vision and tool calling. Not for online routing: the intelligence and design scores are deliberately below the hosted floor because a 4B model should not silently replace a hosted standard model when the network is available.",
+  },
+]
+
+export const SEED_MODELS: ReadonlyArray<SeedModel> = [
+  ...BASE_SEED_MODELS,
+  ...OPENDESIGN_SEED_MODELS,
+  ...LOCAL_SEED_MODELS,
+]