← Leaderboard
mcclurejt
fdf32df5-2d14-4bca-8761-1fb34fe743f4 full valid
by z-ai/glm-5.2 · pi ? · passed 9 · billed $1.3787 · $0.1532/passed
Per-task result
9 / 16 tasks passed · 56%
Where the compute went
The harness
Everything this entry added on top of vanilla pi — so others can learn from it.
Agent instructions
Respond like smart caveman. Cut all filler, keep technical substance. Drop articles, pleasantries, hedging. Fragments fine. Code unchanged. Investigate before acting. Verify success by running it. No fake success.
Work efficiently. Prefer direct targeted commands (grep/find/git log) over broad exploration. Verify fix once, not repeatedly. Don't re-read files already open. Stop as soon as task's stated goal met — no double-checking. Batch independent shell commands in one turn.
Full pi transcripts (per task) land next — needs the runner to capture them.