My tool could only ingest traces from one framework, and that narrow door was strangling adoption. The obvious fix screamed at me: the detection engine must be too framework-specific — rewrite it to be agnostic. Days of work, touching the most sensitive code I have (the frozen detector). I was ready to start.
Instead I spent one hour doing nothing but reading. Read-only: no edits, no plan, just trace exactly where the framework-coupling actually lived.
It wasn't in the engine. The detector was already framework-agnostic — waste looks the same whether it came from LangGraph or CrewAI, and the core only ever saw normalized spans. The coupling was a single thing: there was no entry point that read a trace file. The "rewrite" I was about to do would have churned days of risky changes through code that was already correct, to fix a problem located somewhere else entirely.
The actual fix was adding one entrypoint — hours, low risk, the frozen engine untouched. The hour of diagnosis didn't just save time; it saved me from confidently solving the wrong problem, which is the more expensive mistake. A rewrite that "works" but addressed a non-issue still leaves you with the original bottleneck and a pile of new risk.
The discipline I'm taking from this: when something feels broken, the instinct is to open the file you suspect and start changing it. The cheaper move is to first prove where the problem is, read-only, before a single edit. Most of the time the thing you were about to rewrite was never the thing standing in your way.
Code: github.com/JEONSEWON/Clew-by-Custos
Top comments (0)