Skip to content

Commit 248f8b7

Browse files
john-b-yangjohn-b-yang
committed
Minor Huskybench fix
1 parent 740c049 commit 248f8b7

2 files changed

Lines changed: 43 additions & 43 deletions

File tree

‎codeclash/games/huskybench/huskybench.py‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -35,7 +35,7 @@ def _construct_game_script(self, agents: list[Player], run_cmd_round: str, verbo
3535
if verbose:
3636
self.logger.debug(f"Starting game engine with command: {run_cmd_round}")
3737
script = [
38-
"!/bin/bash",
38+
"#!/bin/bash",
3939
"rm -rf /app/output/*", # Remove previous outputs
4040
f"kill -9 $(lsof -ti :{HB_PORT})", # Kill previous game if any
4141
run_cmd_round, # Start engine

‎configs/scripts/main_tracker.json‎

Lines changed: 42 additions & 42 deletions
Original file line numberDiff line numberDiff line change
@@ -1,9 +1,49 @@
11
{
2+
"HuskyBench": {
3+
"r15.s1000.p2": {
4+
"claude-sonnet-4-20250514.claude-sonnet-4-5-20250929": "0 (0 rounds)",
5+
"claude-sonnet-4-20250514.grok-code-fast-1": "0 (0 rounds)",
6+
"claude-sonnet-4-20250514.gemini-2.5-pro": "0 (0 rounds)",
7+
"claude-sonnet-4-20250514.glm-4-5": "0 (0 rounds)",
8+
"claude-sonnet-4-20250514.gpt-5": "0 (0 rounds)",
9+
"claude-sonnet-4-20250514.gpt-5-mini": "0 (0 rounds)",
10+
"claude-sonnet-4-20250514.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
11+
"claude-sonnet-4-20250514.o3": "0 (0 rounds)",
12+
"claude-sonnet-4-5-20250929.grok-code-fast-1": "0 (0 rounds)",
13+
"claude-sonnet-4-5-20250929.gemini-2.5-pro": "0 (0 rounds)",
14+
"claude-sonnet-4-5-20250929.glm-4-5": "0 (0 rounds)",
15+
"claude-sonnet-4-5-20250929.gpt-5": "0 (0 rounds)",
16+
"claude-sonnet-4-5-20250929.gpt-5-mini": "0 (0 rounds)",
17+
"claude-sonnet-4-5-20250929.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
18+
"claude-sonnet-4-5-20250929.o3": "0 (0 rounds)",
19+
"gemini-2.5-pro.grok-code-fast-1": "0 (0 rounds)",
20+
"glm-4-5.grok-code-fast-1": "0 (0 rounds)",
21+
"gpt-5.grok-code-fast-1": "0 (0 rounds)",
22+
"gpt-5-mini.grok-code-fast-1": "0 (0 rounds)",
23+
"grok-code-fast-1.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
24+
"grok-code-fast-1.o3": "0 (0 rounds)",
25+
"gemini-2.5-pro.glm-4-5": "0 (0 rounds)",
26+
"gemini-2.5-pro.gpt-5": "0 (0 rounds)",
27+
"gemini-2.5-pro.gpt-5-mini": "0 (0 rounds)",
28+
"gemini-2.5-pro.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
29+
"gemini-2.5-pro.o3": "0 (0 rounds)",
30+
"glm-4-5.gpt-5": "0 (0 rounds)",
31+
"glm-4-5.gpt-5-mini": "0 (0 rounds)",
32+
"glm-4-5.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
33+
"glm-4-5.o3": "0 (0 rounds)",
34+
"gpt-5.gpt-5-mini": "0 (0 rounds)",
35+
"gpt-5.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
36+
"gpt-5.o3": "0 (0 rounds)",
37+
"gpt-5-mini.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
38+
"gpt-5-mini.o3": "0 (0 rounds)",
39+
"o3.qwen3-coder-plus-2025-09-23": "0 (0 rounds)"
40+
}
41+
},
242
"RobotRumble": {
343
"r15.s1000.p2": {
444
"claude-sonnet-4-20250514.claude-sonnet-4-5-20250929": "0 (0 rounds)",
545
"claude-sonnet-4-20250514.grok-code-fast-1": "0 (0 rounds)",
6-
"claude-sonnet-4-20250514.gemini-2.5-pro": "1 (10 rounds)",
46+
"claude-sonnet-4-20250514.gemini-2.5-pro": "0 (0 rounds)",
747
"claude-sonnet-4-20250514.glm-4-5": "0 (0 rounds)",
848
"claude-sonnet-4-20250514.gpt-5": "0 (0 rounds)",
949
"claude-sonnet-4-20250514.gpt-5-mini": "0 (0 rounds)",
@@ -132,7 +172,7 @@
132172
"claude-sonnet-4-5-20250929.grok-code-fast-1": "10 (152 rounds)",
133173
"claude-sonnet-4-5-20250929.gemini-2.5-pro": "10 (154 rounds)",
134174
"claude-sonnet-4-5-20250929.glm-4-5": "0 (0 rounds)",
135-
"claude-sonnet-4-5-20250929.gpt-5": "10 (159 rounds)",
175+
"claude-sonnet-4-5-20250929.gpt-5": "10 (160 rounds)",
136176
"claude-sonnet-4-5-20250929.gpt-5-mini": "10 (160 rounds)",
137177
"claude-sonnet-4-5-20250929.qwen3-coder-plus-2025-09-23": "10 (153 rounds)",
138178
"claude-sonnet-4-5-20250929.o3": "10 (160 rounds)",
@@ -198,45 +238,5 @@
198238
"gpt-5-mini.o3": "10 (160 rounds)",
199239
"o3.qwen3-coder-plus-2025-09-23": "10 (160 rounds)"
200240
}
201-
},
202-
"HuskyBench": {
203-
"r15.s1000.p2": {
204-
"claude-sonnet-4-20250514.claude-sonnet-4-5-20250929": "0 (0 rounds)",
205-
"claude-sonnet-4-20250514.grok-code-fast-1": "0 (0 rounds)",
206-
"claude-sonnet-4-20250514.gemini-2.5-pro": "0 (0 rounds)",
207-
"claude-sonnet-4-20250514.glm-4-5": "0 (0 rounds)",
208-
"claude-sonnet-4-20250514.gpt-5": "0 (0 rounds)",
209-
"claude-sonnet-4-20250514.gpt-5-mini": "0 (0 rounds)",
210-
"claude-sonnet-4-20250514.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
211-
"claude-sonnet-4-20250514.o3": "0 (0 rounds)",
212-
"claude-sonnet-4-5-20250929.grok-code-fast-1": "0 (0 rounds)",
213-
"claude-sonnet-4-5-20250929.gemini-2.5-pro": "0 (0 rounds)",
214-
"claude-sonnet-4-5-20250929.glm-4-5": "0 (0 rounds)",
215-
"claude-sonnet-4-5-20250929.gpt-5": "0 (0 rounds)",
216-
"claude-sonnet-4-5-20250929.gpt-5-mini": "0 (0 rounds)",
217-
"claude-sonnet-4-5-20250929.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
218-
"claude-sonnet-4-5-20250929.o3": "0 (0 rounds)",
219-
"gemini-2.5-pro.grok-code-fast-1": "0 (0 rounds)",
220-
"glm-4-5.grok-code-fast-1": "0 (0 rounds)",
221-
"gpt-5.grok-code-fast-1": "0 (0 rounds)",
222-
"gpt-5-mini.grok-code-fast-1": "0 (0 rounds)",
223-
"grok-code-fast-1.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
224-
"grok-code-fast-1.o3": "0 (0 rounds)",
225-
"gemini-2.5-pro.glm-4-5": "0 (0 rounds)",
226-
"gemini-2.5-pro.gpt-5": "0 (0 rounds)",
227-
"gemini-2.5-pro.gpt-5-mini": "0 (0 rounds)",
228-
"gemini-2.5-pro.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
229-
"gemini-2.5-pro.o3": "0 (0 rounds)",
230-
"glm-4-5.gpt-5": "0 (0 rounds)",
231-
"glm-4-5.gpt-5-mini": "0 (0 rounds)",
232-
"glm-4-5.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
233-
"glm-4-5.o3": "0 (0 rounds)",
234-
"gpt-5.gpt-5-mini": "0 (0 rounds)",
235-
"gpt-5.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
236-
"gpt-5.o3": "0 (0 rounds)",
237-
"gpt-5-mini.qwen3-coder-plus-2025-09-23": "0 (0 rounds)",
238-
"gpt-5-mini.o3": "0 (0 rounds)",
239-
"o3.qwen3-coder-plus-2025-09-23": "0 (0 rounds)"
240-
}
241241
}
242242
}

0 commit comments

Comments
 (0)