feat(model-routing): per-tier model access gates and billing multipliers

.github/workflows/ci.yml

@@ -44,6 +44,7 @@ jobs:
           pnpm --filter @maschina/events build
           pnpm --filter @maschina/nats build
           pnpm --filter @maschina/jobs build
+          pnpm --filter @maschina/model build
           pnpm --filter @maschina/telemetry build
           pnpm --filter @maschina/usage build
           pnpm --filter @maschina/billing build
@@ -127,6 +128,7 @@ jobs:
           pnpm --filter @maschina/events build
           pnpm --filter @maschina/nats build
           pnpm --filter @maschina/jobs build
+          pnpm --filter @maschina/model build
           pnpm --filter @maschina/telemetry build
           pnpm --filter @maschina/usage build
           pnpm --filter @maschina/billing build
@@ -231,8 +233,9 @@ jobs:
           uv pip install -e packages/agents --system
           uv pip install -e packages/risk --system
           uv pip install -e "packages/sdk/python[dev]" --system
+          uv pip install -e "services/runtime[dev]" --system
           uv pip install pytest pytest-asyncio pytest-mock --system
-      - run: pytest packages/runtime/tests packages/agents/tests packages/risk/tests packages/sdk/python/tests -v
+      - run: pytest packages/runtime/tests packages/agents/tests packages/risk/tests packages/sdk/python/tests services/runtime/tests -v
 
   # ── Gate ──────────────────────────────────────────────────────────────────
 

.lintstagedrc.mjs

@@ -0,0 +1,23 @@
+import { existsSync } from "node:fs";
+
+// Filter helper — lint-staged 15 can temporarily drop newly-tracked files
+// from disk during its stash/restore cycle. Filter to only existing files
+// before passing to formatters/linters.
+const existing = (files) => files.filter(existsSync);
+
+export default {
+  "*.{ts,tsx,js,jsx,mjs,cjs,json,jsonc}": ["biome check --write --no-errors-on-unmatched"],
+
+  "*.py": (files) => {
+    const ex = existing(files);
+    if (!ex.length) return [];
+    const paths = ex.join(" ");
+    return [`ruff check --fix ${paths}`, `ruff format ${paths}`];
+  },
+
+  "*.rs": (files) => {
+    const ex = existing(files);
+    if (!ex.length) return [];
+    return [`rustfmt --edition 2021 ${ex.join(" ")}`];
+  },
+};

CHANGELOG.md

@@ -8,6 +8,24 @@ Format: [Semantic Versioning](https://semver.org) — `[version] YYYY-MM-DD`
 
 ## [Unreleased]
 
+### Added (2026-03-08 — Model routing)
+- `packages/model/src/catalog.ts` — TypeScript model catalog: 3 Anthropic cloud models + 3 Ollama local models, per-tier access gates, billing multipliers (Haiku 1x, Sonnet 3x, Opus 15x, Ollama 0x)
+- `packages/model/src/index.ts` — Barrel export
+- `packages/model/src/catalog.test.ts` — 20 vitest tests covering multipliers, tier access, validation, resolution
+- `packages/model/tsconfig.json` + build script — TS package alongside existing Python code
+- `packages/validation` — `RunAgentSchema` gains optional `model` field
+- `packages/jobs` — `AgentExecuteJob` gains `model` + `systemPrompt` fields; `dispatchAgentRun` updated
+- `services/api` — Model access validation at run dispatch; resolves system prompt from `agent.config.systemPrompt`; passes model + system prompt through job queue
+- `services/daemon` — `AgentExecuteJob`, `JobToRun` gain `model` + `system_prompt`; `RuntimeRequest` now sends all fields the Python runtime needs (`plan_tier`, `model`, `system_prompt`, `max_tokens`, `timeout_secs`); URL fixed from `/execute` → `/run`
+- `services/daemon` — `RunOutput.payload` renamed to `output_payload` to match Python `RunResponse`
+- `services/runtime` — Full model routing in `runner.py`: routes by model ID prefix (ollama/* vs Anthropic), applies billing multiplier, lazy-imports Anthropic client per request; drops global Ollama flag
+- `services/runtime/tests/test_runner_routing.py` — Unit tests for multiplier + routing helpers (no real LLM calls)
+- CI + pytest scripts updated to include `services/runtime` tests
+
+### Fixed (2026-03-08 — Model routing)
+- Daemon was calling `/execute` endpoint on Python runtime — correct endpoint is `/run`
+- Daemon `RuntimeRequest` was missing `plan_tier`, `model`, `system_prompt`, `timeout_secs` fields that the Python `RunRequest` model requires
+
 ### Fixed (2026-03-07 — Session N+1: backend boot + E2E)
 - All 31 TS packages now build clean (`pnpm turbo build --filter='./packages/*'`)
 - `packages/cache/src/client.ts` — ioredis ESM default import via `(Redis as any)` constructor cast

package.json

@@ -29,8 +29,9 @@
     "cargo:run:cli": "cargo run -p maschina-cli",
     "cargo:run:code": "cargo run -p maschina-code",
 
-    "pytest": "pytest packages/runtime packages/agents packages/ml packages/model packages/risk packages/sdk/python services/worker",
+    "pytest": "pytest packages/runtime packages/agents packages/ml packages/model packages/risk packages/sdk/python services/worker services/runtime",
     "pytest:runtime": "pytest packages/runtime",
+    "pytest:runtime-service": "pytest services/runtime",
     "pytest:agents": "pytest packages/agents",
     "pytest:ml": "pytest packages/ml",
     "pytest:model": "pytest packages/model",
@@ -186,7 +187,7 @@
     "ci": "pnpm check && pnpm build:packages && pnpm test",
     "ci:ts": "pnpm check && pnpm build:packages && turbo test --filter='!@maschina/daemon' --filter='!@maschina/cli' --filter='!@maschina/code' --filter='!@maschina/rust'",
     "ci:rust": "pnpm check:rust && pnpm build:rust && pnpm test:rust",
-    "ci:python": "pytest tests packages/runtime packages/agents packages/ml packages/risk packages/sdk/python services/worker",
+    "ci:python": "pytest tests packages/runtime packages/agents packages/ml packages/risk packages/sdk/python services/worker services/runtime",
     "ci:e2e": "turbo test --filter=@maschina/tests",
     "ci:integration": "vitest run --root tests/integration"
   },

packages/jobs/src/dispatch.ts

@@ -31,6 +31,8 @@ export async function dispatchAgentRun(opts: {
   agentId: string;
   userId: string;
   tier: string;
+  model: string;
+  systemPrompt: string;
   inputPayload: unknown;
   timeoutSecs: number;
 }): Promise<string> {

packages/jobs/src/types.ts

@@ -15,6 +15,8 @@ export interface AgentExecuteJob {
   agentId: string;
   userId: string;
   tier: string;
+  model: string;
+  systemPrompt: string;
   inputPayload: unknown;
   timeoutSecs: number;
 }

packages/model/package.json

@@ -3,11 +3,27 @@
   "version": "0.0.0",
   "private": true,
   "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "import": "./dist/index.js",
+      "types": "./dist/index.d.ts"
+    }
+  },
   "scripts": {
+    "build": "tsc",
     "train": "python -m maschina_model.train",
     "eval": "python -m maschina_model.eval",
     "infer": "python -m maschina_model.infer",
     "test": "pytest",
     "clean": "rm -rf dist __pycache__ .pytest_cache"
+  },
+  "dependencies": {
+    "@maschina/plans": "workspace:*"
+  },
+  "devDependencies": {
+    "@maschina/tsconfig": "workspace:*",
+    "typescript": "^5"
   }
 }

packages/model/src/catalog.test.ts

@@ -0,0 +1,119 @@
+import { describe, expect, it } from "vitest";
+import {
+  DEFAULT_MODEL,
+  getAllowedModels,
+  getModel,
+  getModelMultiplier,
+  resolveModel,
+  validateModelAccess,
+} from "./catalog.js";
+
+describe("getModel", () => {
+  it("returns the model def for a known ID", () => {
+    const m = getModel("claude-haiku-4-5-20251001");
+    expect(m).toBeDefined();
+    expect(m?.provider).toBe("anthropic");
+    expect(m?.multiplier).toBe(1);
+  });
+
+  it("returns undefined for an unknown ID", () => {
+    expect(getModel("gpt-99")).toBeUndefined();
+  });
+});
+
+describe("getModelMultiplier", () => {
+  it("returns 1 for haiku", () => expect(getModelMultiplier("claude-haiku-4-5-20251001")).toBe(1));
+  it("returns 3 for sonnet", () => expect(getModelMultiplier("claude-sonnet-4-6")).toBe(3));
+  it("returns 15 for opus", () => expect(getModelMultiplier("claude-opus-4-6")).toBe(15));
+  it("returns 0 for ollama models", () => expect(getModelMultiplier("ollama/llama3.2")).toBe(0));
+  it("returns 1 for unknown model (safe default)", () =>
+    expect(getModelMultiplier("unknown")).toBe(1));
+});
+
+describe("getAllowedModels", () => {
+  it("access tier only gets local ollama models", () => {
+    const allowed = getAllowedModels("access");
+    expect(allowed.every((m) => m.isLocal)).toBe(true);
+  });
+
+  it("m1 tier can use haiku and ollama", () => {
+    const ids = getAllowedModels("m1").map((m) => m.id);
+    expect(ids).toContain("claude-haiku-4-5-20251001");
+    expect(ids).toContain("ollama/llama3.2");
+    expect(ids).not.toContain("claude-sonnet-4-6");
+    expect(ids).not.toContain("claude-opus-4-6");
+  });
+
+  it("m5 tier can use haiku and sonnet but not opus", () => {
+    const ids = getAllowedModels("m5").map((m) => m.id);
+    expect(ids).toContain("claude-haiku-4-5-20251001");
+    expect(ids).toContain("claude-sonnet-4-6");
+    expect(ids).not.toContain("claude-opus-4-6");
+  });
+
+  it("m10 tier can use all models", () => {
+    const ids = getAllowedModels("m10").map((m) => m.id);
+    expect(ids).toContain("claude-haiku-4-5-20251001");
+    expect(ids).toContain("claude-sonnet-4-6");
+    expect(ids).toContain("claude-opus-4-6");
+  });
+
+  it("internal tier can use all models", () => {
+    const ids = getAllowedModels("internal").map((m) => m.id);
+    expect(ids).toContain("claude-opus-4-6");
+  });
+});
+
+describe("validateModelAccess", () => {
+  it("allows access tier to use ollama", () => {
+    const result = validateModelAccess("access", "ollama/llama3.2");
+    expect(result.allowed).toBe(true);
+  });
+
+  it("denies access tier from using haiku", () => {
+    const result = validateModelAccess("access", "claude-haiku-4-5-20251001");
+    expect(result.allowed).toBe(false);
+    expect(result.reason).toMatch(/m1/);
+  });
+
+  it("denies m1 tier from using sonnet", () => {
+    const result = validateModelAccess("m1", "claude-sonnet-4-6");
+    expect(result.allowed).toBe(false);
+    expect(result.reason).toMatch(/m5/);
+  });
+
+  it("denies m5 tier from using opus", () => {
+    const result = validateModelAccess("m5", "claude-opus-4-6");
+    expect(result.allowed).toBe(false);
+    expect(result.reason).toMatch(/m10/);
+  });
+
+  it("allows m10 tier to use opus", () => {
+    const result = validateModelAccess("m10", "claude-opus-4-6");
+    expect(result.allowed).toBe(true);
+  });
+
+  it("denies unknown model with clear error", () => {
+    const result = validateModelAccess("enterprise", "gpt-99");
+    expect(result.allowed).toBe(false);
+    expect(result.reason).toMatch(/Unknown model/);
+  });
+});
+
+describe("resolveModel", () => {
+  it("returns the requested model if allowed", () => {
+    expect(resolveModel("m5", "claude-haiku-4-5-20251001")).toBe("claude-haiku-4-5-20251001");
+  });
+
+  it("falls back to tier default if requested model is denied", () => {
+    // m1 requesting opus → should fall back to m1 default
+    expect(resolveModel("m1", "claude-opus-4-6")).toBe(DEFAULT_MODEL.m1);
+  });
+
+  it("returns tier default when no model is requested", () => {
+    expect(resolveModel("access")).toBe("ollama/llama3.2");
+    expect(resolveModel("m1")).toBe("claude-haiku-4-5-20251001");
+    expect(resolveModel("m5")).toBe("claude-sonnet-4-6");
+    expect(resolveModel("m10")).toBe("claude-opus-4-6");
+  });
+});