Suriyakumarvijayanayagam
1a0dc7e5f1
The key had two incompatible meanings running at once. CLAUDE.md §3 defines subagents as "keys of other specs this may DELEGATE to" and §6 as a tool call from the parent's perspective — the subagent runs its own turn, as the same caller, out of the parent's budget, and returns an answer. The backend implements exactly that. This side did something else: agentSkillIds folded one level of subagent skills into the parent's carried set, so a parent silently gained everything its subagents carried, and the UI described it that way — "other agents whose skills this one may also use", "borrowing them cannot reach data this page does not hold". Both are defensible readings. Only one is the specification, and running both meant krow-workforce-agent carried eight delegation tools AND the flattened skills of those same eight agents — able to answer a question directly or to ask an agent that had already lent it the means to answer. Two ways to do one thing, differing in cost and in what the trajectory records. So an agent carries what it declares. Reaching another agent is delegation, which the runtime does with its own budget and its own trajectory. The blast radius was one check, which is the useful part of the answer: only "a subagent cycle terminates" depended on the folding, because that traversal was the only thing that could loop. Nothing walks subagents here now, so the cycle question moved to where it belongs — refused at publish by definition.FindSubagentCycle, bounded at run time by the depth cap. The replacement checks assert the new meaning rather than deleting the old ones, because the previous behaviour reads as perfectly reasonable and will be reinvented otherwise. The `agents` parameter stays on agentSkillIds and is no longer read. Removing it is a wider edit for no behavioural gain, and agentScopedDisabledWith exists precisely to pass it. 924/924 checks pass; production build clean. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01PJvibeSc1JYXjatankqM1g
KROW Demo
AI-powered hiring for event and hospitality staffing — a standalone demo build of KROW Forge.
Runs with no backend, no API keys and no real login. npm install && npm run dev and the app is fully
populated.
npm install
npm run dev # http://localhost:5173 — opens on /admin/login
Stack
Vite 6 · React 19 · React Router 6 · React Query 5 · Tailwind + shadcn/ui · Framer Motion · Recharts + MUI X Charts · react-hot-toast.
Structure
src/
main.jsx entry — mounts <App/>, imports index.css
App.jsx providers + the entire route table
index.css design tokens (HSL CSS variables) + utility layers
api/ the backend seam — see below
base44Client.js the contract the production app talks to
store.js in-memory entity store, mirrored to localStorage
aiEngine.js local, deterministic InvokeLLM
seed.js demo dataset
attendanceSeed.js shift/attendance dataset
lib/ domain logic and data hooks
krowHooks.js React Query hooks over the entity store
krowAi.js AI workflows
krowScore.js score engine
AuthContext.jsx auth provider
admin/ admin session, permissions, position insights
skills/ the Owliver skill system (registry, resolver, flows)
agents/ the agent system (registry, runtime, lifecycle)
skills/ *.md skill definitions — loaded by import.meta.glob
agents/ *.md agent definitions — loaded by import.meta.glob
components/
ui/ 23 shadcn primitives
ds/ 30 design-system components (the app's real vocabulary)
admin/ admin page shell + role glyphs
krow/ domain components (+ talent/, proving/)
charts/ MUI X Charts wrappers and theme
agents/ agent configure / canvas / test panel
forge/ KROW Forge skill + challenge components
skills/ skill authoring surfaces
ai-assistant/ Owliver — panel, routing, capabilities, history
layouts/ AdminLayout — header nav, ⌘K search, assistant panel
pages/ route components (admin/ holds the live product)
hooks/ shared hooks
assets/ brand asset references
Routes
The app opens on /admin/login. Everything else lives under /admin/* behind two guards:
ProtectedRoute (auth) and AdminRoute (admin session), rendered inside AdminLayout.
| Path | Page |
|---|---|
/admin/login |
Admin sign-in — outside the protected boundary |
/admin |
Control Center |
/admin/positions, positions/new, positions/:id |
Positions, Create, Detail |
/admin/candidates, candidates/:id |
Candidates, Candidate Intelligence profile |
/admin/hired · /admin/talent-pool |
Hired History, Talent Pool |
/admin/university, university/:id |
KROW Forge, challenge detail |
/admin/analytics · /admin/activity |
Analytics, Activity |
/admin/profile · /admin/settings |
Profile, Settings |
/admin/workspace |
Workspace hub |
/admin/workspace/agents, agents/new, agents/:id |
Agent list, agent editor |
/admin/workspace/skills, skills/new, skills/:id |
UI skill list and editor |
/admin/workspace/skills/owliver/new, owliver/:id |
Owliver skill editor |
/admin/workspace/skill-development |
Skill Development |
* |
lib/PageNotFound.jsx |
Legacy top-level paths (/overview, /positions, /candidates, /hired, /talent-pool,
/university, /analytics, /activity, /tracking) all Navigate into their /admin/*
equivalent, id preserved. Static route segments are declared before dynamic ones so
workspace/agents/new cannot be read as an agent whose id is "new".
Note.
App.jsxstill imports a set of pre-redesign Employer/Talent pages (Overview,Apply,WorkerProfile,KrowIdentity,EmployeeDashboard,DesignSystem, …) andlayouts/Layout.jsx. None of them are mounted on a route — they are unreachable at runtime and retained pending a decision on whether to re-route or remove them.
The backend seam
The production app talks to a Base44 backend through src/api/base44Client.js. This demo keeps that
module path, export name and full method contract, and swaps only the transport:
| Contract | Demo implementation |
|---|---|
base44.entities.<Name>.list/filter/get/create/update/delete |
api/store.js — in-memory records, mirrored to localStorage |
base44.integrations.Core.InvokeLLM |
api/aiEngine.js — local, deterministic |
base44.integrations.Core.UploadFile |
blob URLs, session-scoped |
base44.auth.* |
seeded employer/admin user, always signed in |
Because the seam did not move, everything above it — lib/krowHooks.js, lib/krowAi.js,
lib/krowScore.js, every page and component — is unchanged from the reference implementation.
The AI engine
Every AI workflow funnels through one call, InvokeLLM({ prompt, response_json_schema }).
api/aiEngine.js reimplements it locally: it recognises each workflow by the phrase its prompt opens
with, reads the structured fields the prompt already carries (Years Experience: 6,
Required Certifications: …, the talent-pool JSON block) and scores them deterministically.
Nine workflows are implemented: candidate screening, job-description generation, résumé building, interview questions, interview evaluation, talent matching, Owliver's conversation, Career DNA building and Proving Ground evaluation. Results respond to the data rather than being canned — the same input always returns the same output.
Skills and agents
Two registries discover their definitions from markdown at build time:
| Registry | Glob | Files |
|---|---|---|
lib/skills/registry.js |
import.meta.glob('/src/skills/**/*.md') |
18 Owliver skills, 5 workforce training skills |
lib/agents/registry.js |
import.meta.glob('/src/agents/**/*.md') |
9 agent definitions |
These globs are absolute paths. Moving or renaming src/skills/ or src/agents/ makes the
registry silently return nothing — no build error, no import failure, just an empty registry. Add
definitions by dropping a new .md file into the right folder; nothing else needs to change.
Demo data
Seeded in src/api/seed.js — positions, applicants, AI-scored candidates, interviews, hires,
talent-pool profiles, Proving Ground courses and an activity log; shift data in
src/api/attendanceSeed.js.
Edits persist to localStorage. To restore the shipped dataset:
// in the browser console
localStorage.removeItem('krow_demo_db'); location.reload();
Bumping STORAGE_VERSION in src/api/store.js (currently 8) invalidates saved snapshots
automatically, so changing the seed does not require a manual clear.
Design system
Tokens live in src/index.css as HSL CSS variables consumed by tailwind.config.js.
| Token | Value |
|---|---|
| Primary — KROW Blue | #0A39DF |
| Accent — KROW Yellow | #F9E547 |
| Pale Yellow | #F8E08E |
| Dark Navy | #333F48 |
| Mint | #D1E0D7 |
| Radius | 0.75rem |
| Headings | Sora |
| Body | Inter |
src/components/ds/ is the vocabulary the app actually writes in — Surface, KpiCard,
DataTable, PageHeader, Modal, toast and the rest, exported through ds/index.js.
src/components/ui/ holds the shadcn primitives those are built on. Surfaces use glass and
glass-card; the page background is the gradient-mesh utility.
Toasts go through ds/toast (a wrapper over react-hot-toast), rendered by <HotToaster> in
App.jsx.
Scripts
| Command | Purpose |
|---|---|
npm run dev |
Dev server (honors PORT, default 5173) |
npm run build |
Production build |
npm run preview |
Serve the build |
npm run lint |
ESLint — currently clean |
npm run lint:fix |
ESLint with --fix |
npm test |
scripts/skill-check.mjs — 835 assertions over the skill and agent systems |
npm run typecheck |
tsc -p ./jsconfig.json with checkJs |
The test harness
npm test boots a real Vite dev server in middleware mode and ssrLoadModules the actual
application modules, so it exercises the real @/ alias and import.meta.glob pipeline rather than
mocks. It also diffs Owliver's behaviour across eleven page contexts against a committed snapshot,
scripts/__baseline__/owliver-baseline.json; regenerate that deliberately with
node scripts/owliver-baseline.mjs --write, never to turn a red check green.
Because it loads modules by absolute path, the harness pins the locations of src/api/, src/lib/,
src/lib/skills/, src/lib/agents/, src/components/ai-assistant/, src/App.jsx,
src/layouts/AdminLayout.jsx and a handful of agent files. Move any of them and update
scripts/owliver-capture.mjs and scripts/skill-check.mjs in the same change.
Known-failing checks
Two checks fail on a clean checkout and are tracked as separate work, not regressions:
npm test— 834/835 pass;the seeded overtime climb is foundfails.npm run typecheck— 59 errors, all pre-existing JSDoc/inference gaps inlib/krowAi.js,lib/skills/*,lib/positionModel.jsand a few components.
Deployment
Dockerfile builds with node:22-alpine (capped at 1 GB heap for small hosts) and serves the
static dist/ from nginx:alpine using nginx.conf — gzip, immutable /assets/ caching and an
SPA fallback to index.html so client-side routes resolve on refresh.