performance 是做什麼的？
Make it fast, keep tokens low.
How
Batch things. If you need to read 3 files, read them in the same turn. Same for greps. Don't do git status then git diff in two separate calls.
Grep first, read second. Never read a big unknown file just to see what's in it. Grep, then read the exact spot.
Don't load everything. MEMORY summary first, details on demand. Skills load when triggered, not at boot. That's what memory/tiers is for.
Keep context small. Summarize long logs instead of pasting 500 lines. Keep MEMORY.md under 8k. Prefer surgical edit over rewriting a whole file.
Measure. Run harness bench --quick before and after you try to optimize. Track cold start, tokens, tool calls, hook time.
Targets I try to hold:
- cold start <15s
- tokens per task <50k avg
- tool calls per task <60
- hook p99 <100ms
Check
harness bench --task cold-start --harness opencode
harness bench --compare main