⚡️Making DeepSeek v4 outperform Opus 4.7 with Taste — @AhmadAwais , CommandCode.ai

TL;DR

• 1

  DeepSeek's weakness was often the harness, not the model: Awais says DeepSeek V4 Pro repeatedly sends malformed tool calls, then ignores schema errors for 50-plus failures per session, but deterministic repair logic can fix the call, return the result, and teach the model what went wrong.

• 2

  Command Code learned 16,000 repair patterns across huge usage volume: After first spotting the issue in DeepSeek, the team found the same behavior in Kimi and MiniMax, building a library of repair variations from hundreds of billions of tokens and now processing roughly 600 billion tokens.

• 3

  Taste is automatic repository-level memory for coding preferences: Instead of manually maintaining a giant rules file, Taste watches merges, edits, accepts, and rejects, then writes small markdown preferences like “use pnpm, but npm global link for local CLI linking” directly into the repo.

• 4

  The same repair mindset applies to design slop: By giving models a small set of layout patterns, design “smells,” and preferences like OKLCH over HSL, Awais says they can remove most of the generic indigo-gradient dashboard look that designers spot in 1.5 seconds.

• 5

  Open models are becoming the sweet spot for Command Code: Claude is forgiving enough to recover from bad tool interactions on its own, but open models benefit much more from a smart harness, which is why Command Code found product-market fit there and even launched a $1 plan for 600 million DeepSeek tokens.

• 6

  Command Code is heading toward an Apple-style open source model strategy: Awais says the six-year-old codebase will be open sourced soon, made deeply hackable, but intentionally curated around a smaller set of strong open and closed models instead of supporting every model under the sun.