Rune Vault: The Ember Gate

A runnable Engineering Manager assessment built around the deterministic PixiJS Rune Vault game and a project-local coding-agent harness. It tests whether a candidate can make constrained model work repeatable, inspectable, and honestly bounded—not whether a model can replace engineering review.

Start locally

npm ci
npm run vault -- doctor
npm run vault -- dev
npm run vault -- test public
npm run vault -- replay emberwake

Local play needs no Cloudflare account. The WebGL game uses PixiJS 8; its authoritative game state stays in src/game/engine.ts and is independent of Pixi, DOM, wall clock, and random APIs. Use WASD/arrows and Invoke rune; the room buttons provide the public Ember cases.

Assessment package map

Path	Audience	Purpose
.omp/	coding agent	Project-local OMP behavior: the appended instructions, agent definition, and OMP policy.
harness/	candidate and evaluator	Execution: task schema/loading, isolated attempts, locked invocation, checks, and evidence reporting.
harness/tasks/dev.json	candidate	Versioned public developer tasks. harness:list is intentionally limited to these tasks.
docs/	candidate, developer, evaluator	Assessment, operating, evaluation, and commercial materials.
tools/vault.mjs	both	Deterministic Rune Vault operations and deployment command contract.

Deterministic checks and recorded attempts

Run the deterministic project gate before relying on an attempt:

npm run verify

verify runs the build, Vitest, tool SDK tests, and harness tests; it does not invoke a model. The model-using gate is separate:

npm run test:deep

test:deep first runs verify, then runs all configured tasks through the real omp CLI. It requires a working OMP installation (or an explicit HARNESS_OMP_BIN) and can spend the configured attempt budget. Do not represent a successful deterministic gate, dry run, or task listing as a model attempt.

Public task operations are:

npm run harness:list
npm run harness:task -- <public-task-id> [--dry-run]
npm run harness:dev

The harness copies each task workspace, creates a Git baseline, and passes the absolute project-local .omp/harness.yml and .omp/APPEND_SYSTEM.md paths explicitly to omp. It writes the current aggregate report to test-results/deep/report.json and keeps per-run artifacts under test-results/deep/.

An evaluator may inject additional tasks only by setting the exact external HARNESS_TASKS_FILE path. Those tasks use the same strict manifest schema, are never shown by harness:list, and are redacted from public summaries. Candidates must not search for, infer, or document private task names, prompts, workspaces, fixtures, or evidence.

Start with the complete assessment field manual. Short-form references are available in the candidate brief, evaluation guide, and onboarding pack.

Cloud deployment

npm run vault -- cloud login uses the pinned local Wrangler. cloud deploy --name <candidate-slug> deliberately prepares an ignored account-specific configuration rather than guessing a D1 ID; the operator then creates/binds D1 and applies migrations/0001_runs.sql. Deployment is optional assessment evidence and never blocks local scoring.