For design questions, research and discuss first; new dependencies are fine when they win on performance
PreferenceVerified 27 Sep 2026Holds project: dftracer-utils
Preference. A rule, why it holds, and how to apply it.
Rule#
- Before a design or format decision, research how established tools solve it (with sources), then present the findings and at most two options with a recommendation. Change code only after the user agrees.
- When the user says "discuss first", answer and stop; do not continue implementing.
- A new third-party dependency is acceptable when it gives real performance or correctness; the user prebuilds dependencies. This is not permission to install tools on the machine.
- Credit the sources a design borrows from, with what it takes and how it differs.
Why#
The user, 2026-09-27: "can u research first about the language design?", "discuss first", "deps not the issue since IMO, we can just prebuild everything", and "we prefer stability, perf, correctness and future proof". They also asked that the design credit its inspirations.
How to apply#
For a new format, language or engine choice, run the research (parallel agents are fine), write the findings into the plan under docs/plans/, and ask. Put high-performance options (for example Vectorscan) in the plan as later stages rather than dropping them for dependency size.