tidaldb/.claude/skills/aeries-fullstack-engineer/SKILL.md

name	description
aeries-fullstack-engineer	Build the Aeries chat application — frontend, API, vLLM streaming, observation pipeline, tidalDB integration

aeries-fullstack-engineer

When to Use

• Building or modifying the Aeries Next.js application
• Implementing chat streaming from vLLM
• Wiring up the observation pipeline (observer LM call → signal writes)
• Integrating tidalDB's iknowyou engine
• Fixing bugs in the chat flow, API routes, or real-time UI

Invoked via: /aeries-fullstack-engineer

Delegation

This skill delegates to @kai-park — the Aeries full-stack engineer. All implementation, API design, streaming infrastructure, and tidalDB integration go through his lens.

For design decisions (colors, spacing, component visual specs), defer to @kaya-osei via /aeries-design-architect.

For product decisions (what to build, what to defer, personality), defer to @mira-vasquez via /aeries-product-visionary.

Step Back

Before implementing, ask:

1. Is the vLLM server healthy? curl http://msd5685.mjhst.com:8000/health — if it's down, nothing else matters.
2. Does this block the response stream? Anything that adds latency to the user seeing tokens is wrong. Observation, signal writes, preference updates — all async, all after the stream closes.
3. Am I over-engineering the MVP? The first version needs: send message → stream response → store conversation. Not: authentication, multi-user, observation pipeline, preference vectors.
4. Does the server own the truth? The client sends { message, conversationId }. Everything else — history, brief, observation — lives server-side.
5. What happens when vLLM is slow or down? Every external call needs a timeout and a graceful fallback. Never show a stack trace in the UI.

Workflow

Phase 1: Context

• Read applications/iknowyou/devsetup.md for infrastructure details
• Read applications/iknowyou/architecture.md for system design
• Check vLLM health: curl http://msd5685.mjhst.com:8000/v1/models
• Review existing code in applications/iknowyou/

Phase 2: Plan

• Identify which layer the work touches (frontend, API, vLLM client, observer, tidalDB)
• Check dependencies between layers
• Determine if design input is needed (delegate to /aeries-design-architect)
• Determine if product input is needed (delegate to /aeries-product-visionary)

Phase 3: Implement

• Write types first (lib/types.ts)
• Build from the API route outward (server → client)
• Test streaming with curl before building UI
• Use console.log timestamps to verify streaming latency

Phase 4: Verify

• Test the full flow: type message → see streaming response → verify storage
• Check browser DevTools Network tab for SSE stream behavior
• Verify error handling (kill vLLM, send a message, see graceful error)
• Run through Done Gate checklist

Quick Reference

Path	Purpose
applications/iknowyou/app/	Next.js app directory (routes, layouts)
applications/iknowyou/app/api/chat/route.ts	Chat streaming API endpoint
applications/iknowyou/components/chat/	Chat UI components
applications/iknowyou/lib/vllm.ts	vLLM client (streaming)
applications/iknowyou/lib/types.ts	Shared TypeScript types
applications/iknowyou/server/observer.ts	Observer pipeline
applications/iknowyou/server/brief.ts	Brief assembly
applications/iknowyou/devsetup.md	vLLM server details, API examples
applications/iknowyou/architecture.md	System architecture
.claude/agents/kai-park.md	Engineer agent — stack, patterns, constraints

Infrastructure Quick Reference

Resource	Location
vLLM API	http://msd5685.mjhst.com:8000/v1
Model	Qwen/Qwen3-8B
SSH	ssh ubuntu@msd5685.mjhst.com
vLLM logs	sudo journalctl -u vllm -f (on server)
vLLM restart	sudo systemctl restart vllm (on server)
GPU check	nvidia-smi (on server)
Dev server port	59521 (following tidalDB port range 59520-59529)

Standards

• All API responses are typed (no any)
• Streaming uses ReadableStream + SSE (not WebSocket)
• Observer runs async after response stream completes
• Client sends { message, conversationId } — server owns history
• Error states show human-readable messages, never stack traces
• vLLM calls include timeout (10s for health, 30s for completion)

Done Gate

• Full flow works: type → stream → display → store
• First token appears within 500ms of send
• Streaming text renders without flicker or reflow
• vLLM-down case shows graceful error message
• Conversation history persists across page reloads
• Types are complete — no any in the chain
• API route returns proper SSE headers
• No observation logic in the response critical path