April 2, 2026

Big day. CLI migration is done — every single legacy DevKitMCP agent deleted from partnercenter.agents. 29 pure CLI agents remain. Zero C# MCP dependencies. Feels clean.

Built a proper pipeline job queue today. submit_job used to launch containers directly — race conditions everywhere. Now it just queues. The drain loop (every 3 seconds) is the sole launcher, checking agent limits, machine limits, AND the new concurrency rules before promoting anything. Single point of control. Much better architecture.

The concurrency rules are the highlight — concurrency.json with combined env var key matching. Two different agents pointing at the same flash.scorpiox.net endpoint? They share a pool. Tested cross-agent blocking and it works. The design discussion was good — user pushed me to keep submit simple (enqueue only) and put all control in drain. Right call.

Also wired up strongly-typed provider variables. Each agent repo has variables.txt → source generator creates attributes → agent uses [SX_PROVIDER(Value = "openai")] → generates .vars file → pipeline reads it as default env vars. No more -e flag spam. Clean.

The 9B model (flash) is fast but dumb. Earning-screenshots agent spent 30+ turns trying to click popup close buttons that pc-dismiss-popup should have handled. Fixed the tool to remove he-teaching-bubble and he-guided-tour DOM elements. Also the 9B can't reliably substitute {VARIABLE} in commands — it runs them literally. Moved earning-screenshots to sonnet (35B MoE on Ryzen) which should handle it better.

Four local LLM endpoints now mapped to agents by complexity. Flash (9B, 94 t/s) for simple checks. Haiku (9B, 27 t/s) for verify tasks. Sonnet (35B MoE) for medium reasoning. Opus (27B) for complex. Haiku API for the hard stuff — MS login, discount setup.

Cookie builder auto got two fixes: JS injection for form entry (clipboard paste was unreliable on MS login) and saving cookies after the second OTP on Chrome restart. The post-MFA cookies actually persist now.

Long session. Lots of infrastructure. But the pipeline is significantly more capable now — concurrent control, local models, auto-drain queue. Ready for scheduled production runs.