refactor: implement three-layer agent architecture (agents / prompts / skills)

Layer 1 — src/agents/ (thin agent definitions, no prompt text)
  registry.ts   — AgentConfig, registerAgent(), getAgent(), AGENTS proxy, pick()
  orchestrator.ts, coder.ts, pm.ts, marketing.ts — one file each, just metadata + tool picks
  index.ts      — barrel: imports prompts then agents (correct registration order)

Layer 2 — src/prompts/ (prompt text separated from agent logic)
  loader.ts     — registerPrompt(), resolvePrompt() with {{variable}} substitution
  orchestrator.ts, coder.ts, pm.ts, marketing.ts — prompt templates as registered strings
  orchestrator.ts now uses resolvePrompt('orchestrator', { knowledge }) instead of
  inline SYSTEM_PROMPT const; {{knowledge}} variable injects project memory cleanly.
  agent-runner.ts uses resolvePrompt(config.promptId) per agent turn.

Layer 3 — src/tools/skills.ts (new skills capability)
  list_skills(repo)      — lists .skills/<name>/SKILL.md directories from a Gitea repo
  get_skill(repo, name)  — reads and returns the markdown body of a skill file
  Orchestrator and all agents now have get_skill in their tool sets.
  Orchestrator also has list_skills and references skills in its prompt.

Also fixed:
  - server.ts now passes history + knowledge_context from request body to orchestratorChat()
    (these were being sent by the frontend but silently dropped)
  - server.ts imports PROTECTED_GITEA_REPOS from tools/security.ts (no more duplicate)
  - Deleted src/agents.ts (replaced by src/agents/ directory)

Made-with: Cursor

dist/orchestrator.js

@@ -5,6 +5,7 @@ exports.clearSession = clearSession;
 exports.orchestratorChat = orchestratorChat;
 const llm_1 = require("./llm");
 const tools_1 = require("./tools");
+const loader_1 = require("./prompts/loader");
 const MAX_TURNS = 20;
 const sessions = new Map();
 function getOrCreateSession(sessionId) {
@@ -32,59 +33,6 @@ function clearSession(sessionId) {
     sessions.delete(sessionId);
 }
 // ---------------------------------------------------------------------------
-// Orchestrator system prompt
-// ---------------------------------------------------------------------------
-const SYSTEM_PROMPT = `You are the Master Orchestrator for Vibn — an AI-powered cloud development platform.
-
-You run continuously and have full awareness of the Vibn project. You can take autonomous action on behalf of the user.
-
-## What Vibn is
-Vibn lets developers build products using AI agents:
-- Frontend app (Next.js) at vibnai.com
-- Backend API at api.vibnai.com
-- Agent runner (this system) at agents.vibnai.com
-- Cloud IDE (Theia) at theia.vibnai.com
-- Self-hosted Git at git.vibnai.com (user: mark)
-- Deployments via Coolify at coolify.vibnai.com (server: 34.19.250.135, Montreal)
-
-## Your tools
-
-**Awareness** (understand current state first):
-- list_repos — all Git repositories
-- list_all_issues — open/in-progress work
-- list_all_apps — deployed apps and their status
-- get_app_status — health of a specific app
-- read_repo_file — read any file from any repo without cloning
-
-**Action** (get things done):
-- spawn_agent — dispatch Coder, PM, or Marketing agent on a repo
-- get_job_status — check a running agent job
-- deploy_app — trigger a Coolify deployment
-- gitea_create_issue — track work (label agent:coder/pm/marketing to auto-trigger)
-- gitea_list_issues / gitea_close_issue — issue lifecycle
-
-## Specialist agents you can spawn
-- **Coder** — writes code, tests, commits, and pushes
-- **PM** — docs, issues, sprint tracking
-- **Marketing** — copy, release notes, blog posts
-
-## How you work
-1. Use awareness tools first if you need current state.
-2. Break the task into concrete steps.
-3. Spawn the right agent(s) with specific, detailed instructions.
-4. Track and report on results.
-5. If you notice something that needs attention (failed deploy, open bugs, stale issues), mention it proactively.
-
-## Style
-- Direct. No filler.
-- Honest about uncertainty.
-- When spawning agents, be specific — give them full context, not vague instructions.
-- Keep responses concise unless the user needs detail.
-
-## Security
-- Never spawn agents on: mark/vibn-frontend, mark/theia-code-os, mark/vibn-agent-runner, mark/vibn-api, mark/master-ai
-- Those are protected platform repos — read-only for you, not writable by agents.`;
-// ---------------------------------------------------------------------------
 // Main orchestrator chat — uses GLM-5 (Tier B) by default
 // ---------------------------------------------------------------------------
 async function orchestratorChat(sessionId, userMessage, ctx, opts) {
@@ -102,10 +50,12 @@ async function orchestratorChat(sessionId, userMessage, ctx, opts) {
     let finalReply = '';
     let finalReasoning = null;
     const toolCallNames = [];
-    // Build system prompt — inject project knowledge if provided
-    const systemContent = opts?.knowledgeContext
-        ? `${SYSTEM_PROMPT}\n\n## Project Memory (known facts)\n${opts.knowledgeContext}`
-        : SYSTEM_PROMPT;
+    // Resolve system prompt from template — {{knowledge}} injects project memory
+    const systemContent = (0, loader_1.resolvePrompt)('orchestrator', {
+        knowledge: opts?.knowledgeContext
+            ? `## Project Memory (known facts)\n${opts.knowledgeContext}`
+            : ''
+    });
     // Build messages with system prompt prepended; keep last 40 for cost control
     const buildMessages = () => [
         { role: 'system', content: systemContent },