Snapshot Context
Early stack snapshot capturing the initial tool evaluation phase. Documents OpenCode Zen limits reduction, Kiro CLI changes, and early MCP experiments. Pre-dates the “rate limit reality” documentation and Kimi’s dominance as daily driver.
The Stack
Primary Models
- Opus 4.5 — Deep reasoning and complex architectures
- Kimi k2.5 — Free via OpenCode Zen or near-free via trial paths
- GPT-5.2-codex — General purpose coding
- GLM 4.7 — Fallback when rate-limited
- Gemini 3 Flash — Google’s offering
Tools & Platforms
- VSCode — Primary editor
- Claude Code — AI assistant integration
- Codex — OpenAI coding tool
System Architecture
Agent work runs on a Debian VM with UFW + Tailscale. Isolation-first mindset, disposable VM/VPS when needed. YOLO Safely playbook documents the approach.
Key Changes
| Change | Impact | Notes |
|---|---|---|
| OpenCode Zen limits reduced | High | Generous limits → practically none |
| Kilo Code free Kimi gone | High | Option removed (at least on this side) |
| Kiro CLI removed Opus 4.5 | Medium | Now Sonnet 4.5 only, 500 credits month one, 50/mo after |
| AGENTS.md preference confirmed | Low | Skills adherence low; AGENTS.md gets respected |
Notes & Observations
Rate Limit Landscape (Feb 4)
The free tier generosity was already showing cracks. OpenCode Zen limits reduced dramatically. Kilo Code’s free option appeared gone. The pattern: free tiers tightening as user bases grow.
Model Economics
OpenAI releasing Codex felt like an answer to the Kimi ecosystem — deep research, slide creation, document creation, and strong frontier performance. This puts real pressure on incumbents chasing massive funding targets. The model economics war has begun.
Infrastructure Decision
$5/mo VPS is safer isolation than a $600+ dedicated box. Not buying dedicated hardware — disposable VPS is the better strategy for agent experiments.
AGENTS.md vs Skills
Leaning more toward AGENTS.md vs skills. Skills adherence is low; AGENTS.md gets respected. Considering multiple AGENTS.md per project with progressive disclosure. Practical guide with templates →
Active at This Time
- Demonstrate early “yolo” experiment leading to MCP exploit (see MCP First Contact field notes - draft)
- Document the full “YOLO Safely” deployment from scratch
- Demonstrate Codex’s new app
Paid Plans (Keeping)
- Claude $20: Claude Code + Opus 4.5
- ChatGPT $20: More generous limits than Opus 4.5 with GPT‑5.2‑codex
- Kimi $3.49/mo: K2.5 via haggling/trial paths
Related
- Current stack →
- Next snapshot: 9 Feb 2026 →
- /value/kimi-access/
- /value/free-stack/
- /value/smart-spend/
- /compare/codex-vs-claude-vs-kimi/
- /deploy/openclaw/yolo-safely/
This is a historical snapshot from 2026-02-04. View current stack →