Vibe Coding Night // Portable Harnesses
You swapped the model and everything broke. Not the API call — the API call worked fine. The model just stopped finishing tasks.
This is the part nobody warns you about. Every model shipped in 2026 was post-trained against a specific harness. Qwen 3.6 was RL'd against particular tool schemas. Kimi K2.6 was optimized at specific sampling settings inside a specific agentic loop. Claude expects every tool_call to have a matching tool result. These aren't preferences. They're baked into the weights.
LiteLLM translates the API surface. OpenRouter normalizes the routing. MCP standardizes what tools exist. None of them touch the middle — the context window itself, where a model trained for one shape of input silently degrades when you hand it another.
The bug trackers tell the real story. Entire repos exist just to fix Qwen's Jinja2 chat templates breaking tool calling in vLLM. Kimi on Bedrock emits premature end_turn signals because the harness and the model disagree about what a completed tool call looks like. Mastra had to build a dedicated schema-compat layer because o3-mini silently ignores the format property in JSON schemas.
The portability tax is real and nobody's built the abstraction layer for it yet.
Tuesday we're measuring it directly. One agentic task — file manipulation, web search, code generation. One fixed MCP tool surface. Three models through the same harness. Then three models through harness adapters that reshape context to match each model's post-training expectations. The delta between those two runs is the tax. The adapter code is the beginning of the layer that doesn't exist.
Bring a laptop. Pick a model. Break it against someone else's harness. Then write the shim.
vibecodingnights.com