@arvorco/relentless 0.3.0 → 0.4.2

package/src/routing/registry.ts

@@ -0,0 +1,501 @@
+/**
+ * Model Registry Module
+ *
+ * Contains model profiles for all supported AI coding agents and harnesses.
+ * Provides data structures and query functions for smart model routing.
+ *
+ * @module src/routing/registry
+ */
+
+import { z } from "zod";
+import { HarnessNameSchema, type HarnessName } from "../config/schema";
+
+/**
+ * Model tier classification for cost-based routing.
+ * - free: Zero-cost models (OpenCode Zen)
+ * - cheap: Low-cost models (Haiku, GPT-5.2 reasoning-effort low, Gemini Flash)
+ * - standard: Balanced cost/performance models (Sonnet, GPT-5.2 reasoning-effort medium)
+ * - premium: High-quality models (Gemini Pro)
+ * - sota: State-of-the-art models (Opus, GPT-5.2 reasoning-effort high)
+ */
+export const ModelTierSchema = z.enum(["free", "cheap", "standard", "premium", "sota"]);
+export type ModelTier = z.infer<typeof ModelTierSchema>;
+
+/**
+ * Complete profile for a model including capabilities, costs, and CLI usage.
+ */
+export const ModelProfileSchema = z.object({
+  /** Unique identifier for the model (e.g., "opus-4.5") */
+  id: z.string(),
+  /** Human-readable display name (e.g., "Claude Opus 4.5") */
+  displayName: z.string(),
+  /** Harness that provides this model */
+  harness: HarnessNameSchema,
+  /** Cost tier for routing decisions */
+  tier: ModelTierSchema,
+  /** Input cost per million tokens (USD) */
+  inputCost: z.number(),
+  /** Output cost per million tokens (USD) */
+  outputCost: z.number(),
+  /** SWE-bench score (if available) */
+  sweBenchScore: z.number().optional(),
+  /** Maximum context window in tokens */
+  contextWindow: z.number(),
+  /** Tokens per second generation rate (if known) */
+  tokensPerSecond: z.number().optional(),
+  /** List of task types this model excels at */
+  strengths: z.array(z.string()),
+  /** List of known limitations */
+  limitations: z.array(z.string()),
+  /** CLI flag used to select this model */
+  cliFlag: z.string(),
+  /** CLI value to pass with the flag */
+  cliValue: z.string(),
+  /** Additional CLI args required for model selection */
+  cliArgs: z.array(z.string()).optional(),
+});
+export type ModelProfile = z.infer<typeof ModelProfileSchema>;
+
+/**
+ * Profile for a harness (AI coding agent) including its available models.
+ */
+export const HarnessProfileSchema = z.object({
+  /** Harness identifier */
+  name: HarnessNameSchema,
+  /** Human-readable display name */
+  displayName: z.string(),
+  /** Available models for this harness */
+  models: z.array(ModelProfileSchema),
+  /** Default model to use when none specified */
+  defaultModel: z.string(),
+  /** Whether the harness supports model selection */
+  supportsModelSelection: z.boolean(),
+  /** Method used to select models (flag, env, config) */
+  modelSelectionMethod: z.enum(["flag", "env", "config"]),
+});
+export type HarnessProfile = z.infer<typeof HarnessProfileSchema>;
+
+/**
+ * Complete registry of all available models across all harnesses.
+ * Ordered by harness, then by tier (SOTA first within each harness).
+ */
+export const MODEL_REGISTRY: ModelProfile[] = [
+  // ============== Claude Models ==============
+  {
+    id: "opus-4.5",
+    displayName: "Claude Opus 4.5",
+    harness: "claude",
+    tier: "sota",
+    inputCost: 5.0,
+    outputCost: 25.0,
+    sweBenchScore: 80.9,
+    contextWindow: 200000,
+    strengths: ["code_review", "architecture", "debugging", "final_review", "complex_reasoning"],
+    limitations: ["expensive", "slower_start"],
+    cliFlag: "--model",
+    cliValue: "claude-opus-4-5-20251101",
+  },
+  {
+    id: "sonnet-4.5",
+    displayName: "Claude Sonnet 4.5",
+    harness: "claude",
+    tier: "standard",
+    inputCost: 3.0,
+    outputCost: 15.0,
+    contextWindow: 200000,
+    strengths: ["frontend", "refactoring", "daily_coding", "balanced"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "claude-sonnet-4-5-20251020",
+  },
+  {
+    id: "haiku-4.5",
+    displayName: "Claude Haiku 4.5",
+    harness: "claude",
+    tier: "cheap",
+    inputCost: 1.0,
+    outputCost: 5.0,
+    sweBenchScore: 73.0,
+    contextWindow: 200000,
+    tokensPerSecond: 200,
+    strengths: ["prototyping", "scaffolding", "simple_tasks", "fast"],
+    limitations: ["less_reasoning"],
+    cliFlag: "--model",
+    cliValue: "claude-haiku-4-5-20251001",
+  },
+
+  // ============== Codex (OpenAI) Models ==============
+  {
+    id: "gpt-5.2-high",
+    displayName: "GPT-5.2 (reasoning-effort high)",
+    harness: "codex",
+    tier: "sota",
+    inputCost: 1.75,
+    outputCost: 14.0,
+    sweBenchScore: 80.0,
+    contextWindow: 128000,
+    strengths: ["reasoning", "control_flow", "overnight_runs", "complex_logic"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "gpt-5.2",
+    cliArgs: ["-c", "reasoning_effort=\"high\""],
+  },
+  {
+    id: "gpt-5.2-xhigh",
+    displayName: "GPT-5.2 (reasoning-effort xhigh)",
+    harness: "codex",
+    tier: "sota",
+    inputCost: 1.75,
+    outputCost: 14.0,
+    sweBenchScore: 80.0,
+    contextWindow: 128000,
+    strengths: ["deep_reasoning", "complex_logic", "long_horizon_tasks"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "gpt-5.2",
+    cliArgs: ["-c", "reasoning_effort=\"xhigh\""],
+  },
+  {
+    id: "gpt-5.2-medium",
+    displayName: "GPT-5.2 (reasoning-effort medium)",
+    harness: "codex",
+    tier: "standard",
+    inputCost: 1.25,
+    outputCost: 10.0,
+    contextWindow: 128000,
+    strengths: ["balanced", "good_review", "general_coding"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "gpt-5.2",
+    cliArgs: ["-c", "reasoning_effort=\"medium\""],
+  },
+  {
+    id: "gpt-5.2-low",
+    displayName: "GPT-5.2 (reasoning-effort low)",
+    harness: "codex",
+    tier: "cheap",
+    inputCost: 0.75,
+    outputCost: 6.0,
+    contextWindow: 128000,
+    strengths: ["fast", "simple_tasks", "cost_effective"],
+    limitations: ["less_accuracy"],
+    cliFlag: "--model",
+    cliValue: "gpt-5.2",
+    cliArgs: ["-c", "reasoning_effort=\"low\""],
+  },
+
+  // ============== Droid Models ==============
+  {
+    id: "claude-opus-4-5-20251101",
+    displayName: "Claude Opus 4.5 (via Droid)",
+    harness: "droid",
+    tier: "sota",
+    inputCost: 5.0,
+    outputCost: 25.0,
+    contextWindow: 200000,
+    strengths: ["architecture", "debugging", "complex_reasoning"],
+    limitations: ["expensive"],
+    cliFlag: "-m",
+    cliValue: "claude-opus-4-5-20251101",
+  },
+  {
+    id: "claude-sonnet-4-5-20250929",
+    displayName: "Claude Sonnet 4.5 (via Droid)",
+    harness: "droid",
+    tier: "standard",
+    inputCost: 3.0,
+    outputCost: 15.0,
+    contextWindow: 200000,
+    strengths: ["balanced", "daily_coding"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "claude-sonnet-4-5-20250929",
+  },
+  {
+    id: "claude-haiku-4-5-20251001",
+    displayName: "Claude Haiku 4.5 (via Droid)",
+    harness: "droid",
+    tier: "cheap",
+    inputCost: 1.0,
+    outputCost: 5.0,
+    contextWindow: 200000,
+    strengths: ["fast", "simple_tasks"],
+    limitations: ["less_reasoning"],
+    cliFlag: "-m",
+    cliValue: "claude-haiku-4-5-20251001",
+  },
+  {
+    id: "gpt-5.2",
+    displayName: "GPT-5.2 (via Droid)",
+    harness: "droid",
+    tier: "standard",
+    inputCost: 1.25,
+    outputCost: 10.0,
+    contextWindow: 128000,
+    strengths: ["balanced", "general_coding"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "gpt-5.2",
+  },
+  {
+    id: "gpt-5.1",
+    displayName: "GPT-5.1 (via Droid)",
+    harness: "droid",
+    tier: "standard",
+    inputCost: 1.0,
+    outputCost: 8.0,
+    contextWindow: 128000,
+    strengths: ["general_tasks"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "gpt-5.1",
+  },
+  {
+    id: "gpt-5.1-codex",
+    displayName: "GPT-5.1 Codex (via Droid)",
+    harness: "droid",
+    tier: "standard",
+    inputCost: 1.0,
+    outputCost: 8.0,
+    contextWindow: 128000,
+    strengths: ["coding", "refactoring"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "gpt-5.1-codex",
+  },
+  {
+    id: "gpt-5.1-codex-max",
+    displayName: "GPT-5.1 Codex Max (via Droid)",
+    harness: "droid",
+    tier: "premium",
+    inputCost: 1.5,
+    outputCost: 12.0,
+    contextWindow: 128000,
+    strengths: ["long_context", "complex_tasks"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "gpt-5.1-codex-max",
+  },
+  {
+    id: "gemini-3-pro-preview",
+    displayName: "Gemini 3 Pro Preview (via Droid)",
+    harness: "droid",
+    tier: "premium",
+    inputCost: 3.0,
+    outputCost: 15.0,
+    contextWindow: 1000000,
+    strengths: ["long_context", "frontend_ui"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "gemini-3-pro-preview",
+  },
+
+  // ============== OpenCode Zen Models (Free Tier) ==============
+  {
+    id: "glm-4.7",
+    displayName: "GLM-4.7",
+    harness: "opencode",
+    tier: "free",
+    inputCost: 0.0,
+    outputCost: 0.0,
+    sweBenchScore: 73.8,
+    contextWindow: 128000,
+    strengths: ["multilingual", "backend", "tool_use", "agentic", "free"],
+    limitations: ["complex_ui"],
+    cliFlag: "--model",
+    cliValue: "opencode/glm-4.7-free",
+  },
+  {
+    id: "grok-code-fast-1",
+    displayName: "Grok Code Fast 1",
+    harness: "opencode",
+    tier: "free",
+    inputCost: 0.0,
+    outputCost: 0.0,
+    contextWindow: 128000,
+    tokensPerSecond: 92,
+    strengths: ["speed", "tool_calling", "agentic", "bug_fixes", "fastest"],
+    limitations: ["tailwind_v3"],
+    cliFlag: "--model",
+    cliValue: "opencode/grok-code",
+  },
+  {
+    id: "minimax-m2.1",
+    displayName: "MiniMax M2.1",
+    harness: "opencode",
+    tier: "free",
+    inputCost: 0.0,
+    outputCost: 0.0,
+    contextWindow: 128000,
+    strengths: ["fullstack", "web_mobile", "reviews", "free"],
+    limitations: ["newer_less_docs"],
+    cliFlag: "--model",
+    cliValue: "opencode/minimax-m2.1-free",
+  },
+
+  // ============== Amp Models ==============
+  {
+    id: "amp-free",
+    displayName: "Amp Free",
+    harness: "amp",
+    tier: "free",
+    inputCost: 0.0,
+    outputCost: 0.0,
+    contextWindow: 128000,
+    strengths: ["interactive", "refactoring", "smart_mode", "free_daily_grant"],
+    limitations: ["context_caps", "no_execute_mode", "ads"],
+    cliFlag: "-m",
+    cliValue: "free",
+  },
+  {
+    id: "amp-smart",
+    displayName: "Amp Smart",
+    harness: "amp",
+    tier: "standard",
+    inputCost: 2.0,
+    outputCost: 10.0,
+    contextWindow: 128000,
+    strengths: ["intelligent_routing", "best_available", "smart_mode"],
+    limitations: [],
+    cliFlag: "-m",
+    cliValue: "smart",
+  },
+
+  // ============== Gemini Models ==============
+  {
+    id: "gemini-3-pro",
+    displayName: "Gemini 3 Pro",
+    harness: "gemini",
+    tier: "premium",
+    inputCost: 3.0,
+    outputCost: 15.0,
+    contextWindow: 1000000,
+    strengths: ["frontend_ui", "webdev_arena_leader", "algorithms", "long_context"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "gemini-3-pro",
+  },
+  {
+    id: "gemini-3-flash",
+    displayName: "Gemini 3 Flash",
+    harness: "gemini",
+    tier: "cheap",
+    inputCost: 0.5,
+    outputCost: 3.0,
+    contextWindow: 1000000,
+    strengths: ["fast", "long_context", "simple_tasks", "affordable"],
+    limitations: [],
+    cliFlag: "--model",
+    cliValue: "gemini-3-flash",
+  },
+];
+
+/**
+ * Profiles for all supported harnesses including their available models.
+ */
+export const HARNESS_PROFILES: HarnessProfile[] = [
+  {
+    name: "claude",
+    displayName: "Claude Code",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "claude"),
+    defaultModel: "sonnet-4.5",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+  {
+    name: "codex",
+    displayName: "Codex CLI",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "codex"),
+    defaultModel: "gpt-5.2-medium",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+  {
+    name: "droid",
+    displayName: "Droid",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "droid"),
+    defaultModel: "claude-opus-4-5-20251101",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+  {
+    name: "opencode",
+    displayName: "OpenCode Zen",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "opencode"),
+    defaultModel: "glm-4.7",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+  {
+    name: "amp",
+    displayName: "Amp",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "amp"),
+    defaultModel: "amp-free",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+  {
+    name: "gemini",
+    displayName: "Gemini CLI",
+    models: MODEL_REGISTRY.filter((m) => m.harness === "gemini"),
+    defaultModel: "gemini-3-flash",
+    supportsModelSelection: true,
+    modelSelectionMethod: "flag",
+  },
+];
+
+/**
+ * Get a model profile by its ID.
+ *
+ * @param id - The model identifier (e.g., "opus-4.5")
+ * @returns The model profile or undefined if not found
+ */
+export function getModelById(id: string): ModelProfile | undefined {
+  return MODEL_REGISTRY.find((model) => model.id === id);
+}
+
+/**
+ * Get all models available for a specific harness.
+ * Models are ordered by tier (SOTA first, then premium, standard, cheap, free).
+ *
+ * @param harness - The harness name
+ * @returns Array of model profiles for the harness
+ */
+export function getModelsByHarness(harness: HarnessName): ModelProfile[] {
+  const tierOrder: ModelTier[] = ["sota", "premium", "standard", "cheap", "free"];
+  return MODEL_REGISTRY.filter((model) => model.harness === harness).sort((a, b) => {
+    return tierOrder.indexOf(a.tier) - tierOrder.indexOf(b.tier);
+  });
+}
+
+/**
+ * Get all models of a specific tier across all harnesses.
+ *
+ * @param tier - The model tier
+ * @returns Array of model profiles matching the tier
+ */
+export function getModelsByTier(tier: ModelTier): ModelProfile[] {
+  return MODEL_REGISTRY.filter((model) => model.tier === tier);
+}
+
+/**
+ * Get the default model ID for a harness.
+ *
+ * @param harness - The harness name
+ * @returns The default model ID for the harness
+ */
+export function getDefaultModelForHarness(harness: HarnessName): string {
+  const profile = HARNESS_PROFILES.find((h) => h.name === harness);
+  return profile?.defaultModel ?? "";
+}
+
+/**
+ * Get the harness that provides a specific model.
+ *
+ * @param modelId - The model identifier
+ * @returns The harness name or undefined if model not found
+ */
+export function getHarnessForModel(modelId: string): HarnessName | undefined {
+  const model = MODEL_REGISTRY.find((m) => m.id === modelId);
+  return model?.harness;
+}