Harness Engineering

Models hallucinate and make mistakes. The harness is what decides whether that ruins your session.

Most of what gets reported as a model failing is really a model plus a harness failing, with the harness left out of the description. This section is where we write up the harness side: what breaks, what we fixed, and what the numbers looked like before and after.

Deep dives

Tool Call Repairs Four repeating mistakes cover most open-model tool call failures. Repairing them, and telling the model what it should have sent, is what makes them stop.

lesson: skill issue applies to the harness more often than the model.

The Read Tool The highest-volume tool in the harness, rebuilt around cost per successful read. Bounded windows, recovery notes, a read-before-write ledger, and a benchmark of ten harnesses.

lesson: every read is a token budget decision, multiplied fifty million times a month.

The Shell Tool The token efficiency frontier. 23 shell capabilities priced by the tokens each one saves, benchmarked from source across ten harnesses. Wake-ups instead of polling, honest exit codes, cursor reads, and kill by port.

lesson: there are no bad models, only bad harnesses.

In progress

Cache repairs, on keeping prompt caches intact across long sessions.
Cache thrash repairs, on the edit patterns that quietly invalidate a cache mid-session.
/design engineering, on the same audit-then-treat idea applied to visual design.