Kelly agent builder

Agent Builder & Governance Console

Overview

Use this skill as a platform team's local, file-backed governance console for a catalog of mock agent configs. It never provisions or calls a real agent — every action reads or writes a local JSON handoff file under app/.data/. This is a generic, brand-free tool: teams define a name, a trigger/intent description, a set of allowed tools (from a fixed catalog), an approval flag, and a monthly call quota, and this console tracks status and usage and gates the risky transition (draft → live) behind required-field validation.

Default interaction mode: App UI. Unless the user explicitly asks for chat-only handling, start/reuse the local app with app/start.sh and give the actual local URL.

App UI Screenshots

<table> <tr> <td width="50%"><img src="assets/screenshots/overview.webp" alt="Agent Builder overview"></td> <td width="50%"><img src="assets/screenshots/catalog.webp" alt="Agent Builder catalog"></td> </tr> <tr> <td><strong>Overview</strong><br>Governance summary: live agent count, aggregate quota usage, and a list of agents that need attention with reasons.</td> <td><strong>Catalog</strong><br>Sortable, searchable table of every agent config with status badges, owning team, and quota usage.</td> </tr> <tr> <td width="50%"><img src="assets/screenshots/agent-detail.webp" alt="Agent Builder agent detail"></td> <td width="50%"></td> </tr> <tr> <td><strong>Agent detail / edit</strong><br>Tool checklist, quota input, approval toggle, owning team field, trigger/intent textarea, status, and lifecycle actions (activate / pause / archive).</td> <td></td> </tr> </table>

Boundary

• This is a mock governance console. It never provisions, deploys, or calls any real agent, model, or external tool. The "allowed tools" checklist is a fixed local catalog (lib/tool-catalog.ts) used only for governance bookkeeping — selecting a tool here does not grant or invoke it anywhere.
• The app reads and writes local files only: app/.data/agents.json, app/.data/onboarding.json, app/.data/agent.lock. It must not call any remote system.
• No brand-specific integration exists or is implied. Configure org_name in config.local.json for cosmetic display only.

First Run

On invocation, start/reuse the local app with:

skills/kelly-agent-builder/app/start.sh

First run installs hono and @hono/node-server; the frontend is zero-build vanilla. If app/.data/agents.json does not exist yet, seed a mock catalog with:

node skills/kelly-agent-builder/scripts/generate_demo_snapshot.ts

This also writes the app/.data/onboarding.json completion marker — this skill has no external accounts or secrets to configure, so seeding the mock catalog for the first time is itself "setup complete."

Demo Mode

• ?demo=1 opens a deterministic, fully offline mock catalog (8 agent configs spanning draft/live/paused/archived, one over-quota, one missing an owning team) for documentation and screenshots.
• lang=en or lang=zh forces UI chrome language for screenshots.
• Demo API responses never read or write app/.data/agents.json.

UI language: support English and Chinese chrome with Auto default.

Governance Rules

Read references/agent-config-schema.md before editing the app, store, or scripts. In short:

• Draft → live (POST /api/agents/:id/activate) is only allowed when name, trigger_description, at least one allowed_tools entry, non-empty owning_team, and monthly_quota > 0 are all present. This is enforced server-side in app/server/store.ts#activateAgent; on failure the API returns 422 with missing_fields and the UI surfaces the exact reason.
• Archive (POST /api/agents/:id/archive) is allowed from any status. Archived agents become read-only (PUT returns 409).
• Pause (POST /api/agents/:id/pause) is only allowed from live.
• Needs attention = a draft with missing required fields, OR an agent (any status) with no owning team, OR a quota-reached live agent (calls_this_month >= monthly_quota — reached, not strictly exceeded), OR approval_required: true with no owning team assigned.
• PUT validation: a PUT that would leave an already-live agent missing any required field (e.g. clearing owning_team or allowed_tools) is rejected with 422 + missing_fields, the same gate activate uses. Draft agents remain freely editable.

All writes persist to app/.data/agents.json; there is no approval workflow beyond this local file and no outbound network call anywhere in this skill's app.

Local App

• app/index.html + app/app.js + app/styles.css + app/i18n/messages.js: zero-build vanilla frontend with hash routing (#/overview, #/catalog, #/agent/:id, #/agent/new, #/settings).
• app/server/hono.ts: platform-neutral Hono routes for state, tool catalog, and agent CRUD/lifecycle actions.
• app/server/store.ts: status derivation, CRUD/lifecycle helpers, and the read/write surface, all delegated to lib/data-provider/ (never node:fs directly).
• lib/config-validation.ts: pure governance rules (missingRequiredFields, isQuotaReached, deriveAgent) shared by the server and scripts/.
• lib/data-provider/: the DataProvider contract (provider-interface.ts) and the default local-file-provider.ts; selected via KELLY_AGENT_BUILDER_DATA_PROVIDER (default local).
• lib/tool-catalog.ts: the fixed tool catalog (web_search, code_exec, file_read, file_write, send_email, calendar, crm_lookup, db_query, slack_post, http_request).

Safety

• Never provision or call a real agent, tool, or external system from this skill's app.
• Do not commit config.local.json, app/.data/, or app/.cache/.
• Keep owning_team values as free text; do not validate against a real directory service.

Execution reports

Re-read the active provider's decisions immediately before any approved execution. Record each concrete operation, target, status, timestamp, and error in the provider-backed execution report; keep app actions local-only.