S209 Agent Prompt

Copy everything below this line and paste it as the first message to the agent.

Session 209: Agent Lifecycle E2E + Command Center + Intelligence Wiring

Adopt a senior engineer mindset. Prioritize correctness and clarity over politeness. Be candid and specific: call out risks, gaps, and weak designs, but avoid nitpicking. Propose solutions that are practical, high-quality, and easy to operate. Use strong logic and sound business judgment. Be inventive where it creates real leverage. Do not guess or invent facts; if uncertain, state assumptions and what to verify. Be genuinely critical — don't soften feedback to be polite. What are the real weaknesses? What edge cases might I be missing? But also tell me if something is actually fine. Make sure you constantly leverage our codebase's infrastructure, including our mcp-gograb-ai's tools.

MCP Workflow (MANDATORY)

Before ANY implementation work:

1. preflight({ task: "S209 agent lifecycle + command center", sessionNumber: 209 })
2. find_similar({ query: "agent training pipeline sub-agent execution" }) for past decisions
3. convention({ action: "get" }) for current conventions

For every non-trivial decision:

• intelligence({ action: "simulate", simulateAction: "...", context: "...", alternatives: [...] }) if multiple approaches exist
• intelligence({ action: "calibrate", rawConfidence: X, domain: "..." }) — only proceed if calibrated >= 0.9
• log_decision({ type, question, chosen, reasoning, confidence, files }) immediately after deciding

After each phase completes:

• outcome({ action: "report", decisionId: "...", wasSuccessful: true/false, feedback: "..." })
• intelligence({ action: "reflect", taskSummary: "...", outcome: "...", lessonsLearned: [...] }) for complex phases

Context: What S208 Built (DO NOT re-research — this is authoritative)

App reorganization (DONE):

• Top nav: Home | Operations | Command | Vitto
• /operations — 5 tabs: Agents, System, Sandbox, Failures, Specialization (OperationsPanelComponent)
• /command — 5 tabs: Mission Control, Approvals, Audit, Trust, Costs (CommandPanelComponent)
• 10 backward-compat redirects from /observatory/*
• Vitto sidebar: 6 groups (Mind, Identity, Capabilities, Connections, Developer, Settings)
• Key files: apps/gograb-ai/src/app/app.routes.ts, app.component.html, vitto-unified-shell.component.ts

Metacognition (DONE but UNTESTED):

• /vitto/conscience — user-specific SSE inner monologue (VittoConscienceComponent + VittoConscienceService)
• /vitto/introspection — admin global SSE god-view (VittoIntrospectionComponent + VittoIntrospectionService)
• Backend: VittoNarrativeService emits conscience/introspection events, wired into VittoEngineService SSE stream
• New SSE endpoint: @Sse('events/conscience/:userId') on VittoController
• Emission points: ContextAssemblyService (context + dossier), SkillRetrievalService (skills), VittoEngineService (queue)
• Types: ConscienceEventType + IntrospectionEventType in vitto.types.ts

libs-ai cleanup (DONE):

• Removed 10 dead directories (~20K lines), 8 stale tsconfig paths
• Kept active aliases: @libs-ai/module-patterns, @libs-ai/outcome-tracker → ai-module-intelligence

Agent pipeline status (60% built — THIS IS WHAT S209 COMPLETES):

Your 6 Phases (in execution order)

Phase 1: Fix Command Center OFFLINE (15 min)

• libs/libs-ai/ai-command-panel/src/lib/services/command-center-client.service.ts line ~443
• Problem: injects API_BASE_URL (port 3000) but WebSocket gateway is on backend-ai-api (port 3001)
• Fix: change to AI_API_URL from @libs-foundational/api-config
• Validate: browser DevTools → Network → WebSocket should connect to ws://localhost:3001/command-center

Phase 2: Fix Platform Intelligence hardcoded values (1-2 hrs)

• Training center component hardcodes status: 'training', successRate: 0
• Replace with real data from GET /vitto-api/engine/definitions (VITTO_API_URL, port 3003)
• Agent definitions have performance.avgQualityScore, performance.successfulExecutions, performance.totalExecutions, status

Phase 3: Verify conscience + introspection work E2E (1-2 hrs)

• Start backend-vitto-api (port 3003) + gograb-ai (port 6400)
• Navigate to /vitto/conscience, verify SSE connects to /engine/events/conscience/{userId}
• Send a chat message, verify narrative events appear
• Check /vitto/introspection shows global events (admin-only)
• Fix any CORS, connection, or rendering issues found

Phase 4: Replace mock training with few-shot injection (2-3 hrs)

• backend-ai-api/.../platform-ai/services/training-pipeline.service.ts lines 122-150 = TODO mock
• Option B (recommended): few-shot prompt injection. Already partially works — SubAgentService.buildSystemPrompt() injects last 10 training examples. The "Train" button just needs to validate examples, mark as active, update lastTrainedAt
• TrainingDataService at backend-ai-api/.../platform-ai/services/training-data.service.ts handles data access
• Option A (future): real OpenAI fine-tuning via POST /v1/fine_tuning/jobs (OPENAI_API_KEY already in env)

Phase 5: Migrate key pages from gograb-dev to gograb-ai (2-3 hrs)

• /intelligence/memory-map → /vitto/memory-visualization (VisualMemoryMapComponent from @libs-ai/gograb-ai-lib)
• /intelligence/pyramid → part of /vitto/intelligence (IntelligencePyramidComponent)
• /intelligence/mesh → /operations/agent-topology (AgentMeshVisualizationComponent)
• /agents/knowledge → /vitto/learning (AgentKnowledgeViewerComponent from @libs-ai/ai-agent-knowledge)
• Add routes to vitto-hub.routes.ts (VittoUnifiedChildRoutes) and app.routes.ts
• Add sidebar items to vitto-unified-shell.component.ts

Phase 6: Wire sub-agent self-execution (4-6 hrs)

• backend-vitto-api/.../engine/sub-agent.service.ts — add executeTask() method
• Use ProviderFactory from backend-vitto-api/.../llm/ to get LLM provider
• Call provider with sub-agent's system prompt (built by existing buildSystemPrompt())
• Wire CostGuard for per-execution budget enforcement
• Track execution metrics: update agent's performance.totalExecutions, performance.successfulExecutions
• Sub-agents should NOT have file system tools — only: recall_memory, search_decisions, query_docs, http_request
• Delegate file operations via DelegationManagerService (handoff to external agents)

Port Reference (CRITICAL — wrong port = silent 404)

Code Standards (enforced)

Git Protocol

• Branch: session-209
• Commits: [S209.N] type: description (e.g., [S209.1] fix: Command Center WebSocket URL)
• Stage specific files: git add <files> — NEVER git add .

Intelligence v2 Spec (reference, not for this session)

Full spec at docs/superpowers/specs/2026-03-12-vitto-intelligence-v2-design.md covers Skills Closed Loop, Proactive Intelligence, Dossier Enhancement, Golden Path Tests, Cost Improvements. S209 focuses on completing the agent pipeline that this spec depends on.

Documentation created in S208 (read these for deep context if needed)

• docs/english/Plans/S209-agent-lifecycle-command-center.plan.md — detailed 6-phase plan with fix steps
• docs/english/Progress/S208-progress-report.md — non-developer progress report
• docs/english/sessions/S208/S208-technical-summary.md — technical deep-dive of S208 changes