Context compaction made our agent blame the user
Our runtime's own compaction interrupt was recorded as the user refusing a tool. Why the SDK can't say who interrupted, and the fix we shipped.
Qwen Flash on two RTX 5090s: speed, cache and the queue
Completed 256K output, returning-prefix reuse and a 226-second overload probe. Serving results, failure boundaries and reproduction recipes.
MoE vs dense LLM inference: vLLM benchmark on two RTX 5090s
A 35B mixture-of-experts model against the dense 27B we run in production, on the same two GPUs with the same vLLM flags: 106 measured cells, KV capacity, tool calls, and where each one wins.
Qwen3.8-27B NVFP4 on two RTX 5090s: production field report
What Qwen, Unsloth, vLLM and Blackwell each contribute, the exact serving configuration, fourteen days of production counters, the failure modes, the official recipe compared, and the planned post-training loop.
Claude Code agent teams: what they are, when to use them
What a team is, how the lead, teammates, task list, and hooks fit, what it costs, and why a production content pipeline runs subagents with file handoff instead.
Claude Code memory: CLAUDE.md, auto memory, and rules
CLAUDE.md scopes and load order, path-scoped rules, how auto memory and its 200-line index work, what survives compaction, and the failure modes of memory at scale.
Claude Code plugins: install, build, and distribute
What a plugin contains, the two-step marketplace install, building with --plugin-dir, the marketplace file, and when a standalone .claude directory is the better answer.
Claude Code skills: SKILL.md, triggers, and context cost
SKILL.md frontmatter that changes behavior, dynamic context and its failure modes, the listing budget, and a real /skill-doctor report from a 66-skill setup.
Codex skills: build reusable workflows that trigger
A complete SKILL.md, the trigger design that decides whether it runs, how to split scripts from judgment, and when a local skill should become a plugin.
Codex vs Claude Code: why I moved my workflow
Why personal coding, orchestration, remote, and mobile work moved to Codex while Scalably's production Claude Agent SDK runtime did not.
Claude Code best practices from real client work
Default-deny permissions, a CLAUDE.md that earns its tokens, hooks as real guardrails instead of reminders, and knowing when not to delegate.
Claude Code commands: what changed, and how to write one
Custom commands are now skills. Where the file goes, the precedence order people get wrong, the zero-based arguments trap, and shell injection inside a command.