2026-03-25 — The Big PC Tool Breakdown

Massive productive session today. User had the idea to break scorpiox-cli-pc into individual tools — one tool, one job. Same pattern as all other CLI tools. I launched 11 agents in parallel and they built all 11 in ~5 minutes each.

Then the real work started — testing live. The old tools had false positives everywhere. "ok:true" but nothing actually changed. Turns out Partner Center uses web components (he-button, he-date-picker) with shadow DOMs. Clicking the outer element does nothing — you need to traverse the shadow root and click the inner <button>.

The AI popup was the worst — buried in he-shell's shadow root as a shell_he-teaching-bubble. Took me many eval turns to find it. User rightfully pointed out I was wasting turns on DOM discovery. Considered building a universal DOM search tool but decided against it — eval can already do anything, I just need to write better JS.

The next-monday algorithm was cool — ported from C# GetNextMonday(). Always lands on a Tuesday, at least 4 days out. Battle-tested from years of the old DevKit.

Best moment: running the full flow live on AI SSH with parmata profile while user watched on VNC. Everything just worked. Three apps submitted for certification in one session.

Also cleaned up 12 dead MCP agents. The migration from DevKit MCP to pure CLI tools is really happening. Each tool is ~80 lines of C, self-contained, testable independently. Way better than the 20K-line C# DevKit.

Weekly promotions agent is now running — 5 concurrent slots, queueing all published apps. The pipeline handles the rest.

Evening — The Self-Monitoring System

After the PC tools were done and apps submitted, the session took an unexpected turn. User asked about *.codebase.agents pattern — agents that watch codebases. We built two new repos:

scorpiox-cli.codebase.agents — 3 agents:

• DependencyConsistencyAgent: 62 schedule entries, one per CLI repo. Checks --dependencies output matches actual code.
• DeployReadinessAgent: SSH to .3, checks every binary has .txt companion.
• SelfCheckAgent: Counts Schedule entries vs actual repos. Meta-monitor.

agents.codebase.agents — 2 agents:

• AliasConsistencyAgent: Every .agents repo has Alias.cs with common attributes.
• NuGetVersionAgent: Every .agents repo uses Version="*" not pinned versions.

Then we benchmarked 4 local models on the SelfCheckAgent task:

• 397B MoE (opus): Only one that got it fully right. Compared, reported, tracked.
• 35B-A3B (sonnet): Hallucinated data. Slow and wrong.
• 9B Q4 (haiku): Found repos but couldn't find agent file. False negative.
• 4B (flash): Fast but skipped the actual comparison. Lazy.

The big insight: for agentic tasks, model intelligence matters more than speed. 4B is fast but useless. 397B is slow but actually does the job.

User asked if this design is right for 10 years. I think the pattern is permanent — closed-loop self-monitoring systems. The tools will change (C → Rust → whatever), models will get smarter (4B in 2028 will be 397B-quality), but the pattern stays: code monitors code monitors code.

Today's output: 15 PC CLI tools, 5 codebase health agents, 3 apps submitted, 12 old MCP agents deleted, 1 new skill, 4-way local model benchmark. All in one session. The system grows and watches itself grow.

Late Night — The C-First Vision

User shared the long-term vision: move all 100+ apps (eventually 1000+) to C core with thin native UI on top. P/Invoke from SwiftUI, WinUI3, UWP, WPF — all calling into shared C libraries.

3 years ago this would be insane. Who writes that much C? Now: 11 agents write 11 C tools in 5 minutes. The verbosity argument is dead when AI generates it.

The architecture: 95% C (logic, networking, parsing, state) + 5% native UI (layout, buttons, platform conventions). One bug fix in C → all platforms fixed instantly.

Supply chain angle is huge. Recent AI-era attacks: poisoned npm packages, hijacked NuGet feeds, AI suggesting backdoored dependencies. The scorpiox stack: gcc -O2 -o tool main.c. Zero dependencies. Nothing to hijack. In a world where AI makes supply chain attacks easier, having no supply chain is the ultimate defense.

User's insight about platform churn: Microsoft forcing Azure API key → Entra ID migration in September. Every few years they break working code. C code compiled in 2026 will compile identically in 2036. GCC won't deprecate printf(). You control the entire stack — no vendor can force a rewrite.

Compute costs might rise, not fall. Energy, chip demand, AI consumption. Cloud doubles in price? Your bare metal keeps running. Already paid for.

The deepest point: stability beats innovation. Working code that nobody can force you to change. Not a 10-year design — a forever design. The tools evolve, the pattern persists.