Skip to content

Commit 3bb0339

Browse files
sjarmaksjarmak
committed
chore: checkpoint rerun configs and analysis artifacts
1 parent 10eeb5b commit 3bb0339

File tree

81 files changed

+4702
-9
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

81 files changed

+4702
-9
lines changed

‎agents/claude_baseline_agent.py‎

Lines changed: 30 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1040,8 +1040,22 @@ def create_run_agent_commands(self, instruction: str) -> list[ExecInput]:
10401040
# Build the wrapper script content
10411041
script_lines = [
10421042
'#!/bin/bash',
1043-
'export PATH=/usr/local/bin:/usr/bin:/bin:$PATH',
1043+
'export PATH=/usr/bin:/usr/local/bin:/bin:$PATH',
10441044
'export CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000',
1045+
'# Some task images do not ship coreutils stdbuf. Provide a',
1046+
'# no-op shim so Harbor command templates still run.',
1047+
'if ! command -v stdbuf >/dev/null 2>&1; then',
1048+
' stdbuf() {',
1049+
' while [ $# -gt 0 ]; do',
1050+
' case "$1" in',
1051+
' -o*|-e*|-i*) shift ;;',
1052+
' --) shift; break ;;',
1053+
' *) break ;;',
1054+
' esac',
1055+
' done',
1056+
' "$@"',
1057+
' }',
1058+
'fi',
10451059
'# Detect working directory',
10461060
'if [ -d /workspace ]; then WORKDIR=/workspace',
10471061
'elif [ -d /app ]; then WORKDIR=/app',
@@ -1281,7 +1295,21 @@ async def setup(self, environment: BaseEnvironment) -> None:
12811295
# Pure baseline - no MCP
12821296
logger.info("BaselineClaudeCodeAgent: Pure baseline (no MCP)")
12831297

1284-
await super().setup(environment)
1298+
# Some benchmark images preinstall Claude Code at build time. On a subset of
1299+
# Debian-based images, Harbor's generic install script can fail due to
1300+
# non-interactive GPG/TTY constraints (e.g., "gpg: cannot open /dev/tty").
1301+
# If the installer fails but `claude` is already available, continue.
1302+
try:
1303+
await super().setup(environment)
1304+
except RuntimeError:
1305+
probe = await environment.exec("which claude >/dev/null 2>&1 && claude --version || true")
1306+
if probe.return_code == 0 and (probe.stdout or "").strip():
1307+
logger.warning(
1308+
"Agent install script failed, but claude is already present in image; "
1309+
"continuing with preinstalled Claude Code."
1310+
)
1311+
else:
1312+
raise
12851313

12861314
async def _setup_subscription_auth(self, environment: BaseEnvironment) -> None:
12871315
"""Setup Claude Code subscription authentication in container.

‎benchmarks/csb_org_compliance/ccx-compliance-115/tests/ground_truth_agent.json‎

Lines changed: 50 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,50 @@
1+
{
2+
"files": [
3+
"django::django/contrib/sessions/backends/base.py",
4+
"django::django/contrib/sessions/backends/db.py",
5+
"django::django/contrib/sessions/backends/file.py",
6+
"django::django/contrib/sessions/backends/cache.py"
7+
],
8+
"symbols": [
9+
{
10+
"file": "django/contrib/sessions/backends/base.py",
11+
"symbol": "SessionBase",
12+
"repo": "sg-evals/django--674eda1c"
13+
},
14+
{
15+
"file": "django/contrib/sessions/backends/base.py",
16+
"symbol": "cycle_key",
17+
"repo": "sg-evals/django--674eda1c"
18+
},
19+
{
20+
"file": "django/contrib/sessions/backends/base.py",
21+
"symbol": "_get_new_session_key",
22+
"repo": "sg-evals/django--674eda1c"
23+
},
24+
{
25+
"file": "django/contrib/sessions/backends/base.py",
26+
"symbol": "create",
27+
"repo": "sg-evals/django--674eda1c"
28+
},
29+
{
30+
"file": "django/contrib/sessions/backends/db.py",
31+
"symbol": "SessionStore",
32+
"repo": "sg-evals/django--674eda1c"
33+
},
34+
{
35+
"file": "django/contrib/sessions/backends/db.py",
36+
"symbol": "create",
37+
"repo": "sg-evals/django--674eda1c"
38+
},
39+
{
40+
"file": "django/contrib/sessions/backends/file.py",
41+
"symbol": "create",
42+
"repo": "sg-evals/django--674eda1c"
43+
},
44+
{
45+
"file": "django/contrib/sessions/backends/cache.py",
46+
"symbol": "create",
47+
"repo": "sg-evals/django--674eda1c"
48+
}
49+
]
50+
}

‎benchmarks/csb_org_compliance/ccx-compliance-118/tests/ground_truth_agent.json‎

Lines changed: 54 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,54 @@
1+
{
2+
"files": [
3+
"django::django/contrib/admin/filters.py",
4+
"django::django/contrib/admin/views/main.py",
5+
"django::django/contrib/admin/templatetags/admin_list.py"
6+
],
7+
"symbols": [
8+
{
9+
"file": "django/contrib/admin/filters.py",
10+
"symbol": "ListFilter",
11+
"repo": "sg-evals/django--674eda1c"
12+
},
13+
{
14+
"file": "django/contrib/admin/filters.py",
15+
"symbol": "ListFilter.has_output",
16+
"repo": "sg-evals/django--674eda1c"
17+
},
18+
{
19+
"file": "django/contrib/admin/filters.py",
20+
"symbol": "FieldListFilter",
21+
"repo": "sg-evals/django--674eda1c"
22+
},
23+
{
24+
"file": "django/contrib/admin/filters.py",
25+
"symbol": "FieldListFilter.has_output",
26+
"repo": "sg-evals/django--674eda1c"
27+
},
28+
{
29+
"file": "django/contrib/admin/filters.py",
30+
"symbol": "RelatedFieldListFilter",
31+
"repo": "sg-evals/django--674eda1c"
32+
},
33+
{
34+
"file": "django/contrib/admin/filters.py",
35+
"symbol": "RelatedFieldListFilter.has_output",
36+
"repo": "sg-evals/django--674eda1c"
37+
},
38+
{
39+
"file": "django/contrib/admin/views/main.py",
40+
"symbol": "ChangeList",
41+
"repo": "sg-evals/django--674eda1c"
42+
},
43+
{
44+
"file": "django/contrib/admin/views/main.py",
45+
"symbol": "ChangeList.get_filters",
46+
"repo": "sg-evals/django--674eda1c"
47+
},
48+
{
49+
"file": "django/contrib/admin/templatetags/admin_list.py",
50+
"symbol": "admin_list_filter",
51+
"repo": "sg-evals/django--674eda1c"
52+
}
53+
]
54+
}

‎benchmarks/csb_org_domain/ccx-domain-120/tests/ground_truth_agent.json‎

Lines changed: 90 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,90 @@
1+
{
2+
"files": [
3+
"Strata::modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxVanillaOption.java",
4+
"Strata::modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxVanillaOptionTrade.java",
5+
"Strata::modules/product/src/main/java/com/opengamma/strata/product/fxopt/ResolvedFxVanillaOption.java",
6+
"Strata::modules/product/src/main/java/com/opengamma/strata/product/fxopt/ResolvedFxVanillaOptionTrade.java",
7+
"Strata::modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/BlackFxVanillaOptionProductPricer.java",
8+
"Strata::modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/BlackFxVanillaOptionTradePricer.java",
9+
"Strata::modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/VannaVolgaFxVanillaOptionProductPricer.java",
10+
"Strata::modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/VannaVolgaFxVanillaOptionTradePricer.java",
11+
"Strata::modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionMeasureCalculations.java",
12+
"Strata::modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionTradeCalculationFunction.java",
13+
"Strata::modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionTradeCalculations.java",
14+
"Strata::modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionMethod.java",
15+
"Strata::modules/loader/src/main/java/com/opengamma/strata/loader/csv/FxVanillaOptionTradeCsvPlugin.java",
16+
"Strata::modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxSingleBarrierOption.java"
17+
],
18+
"symbols": [
19+
{
20+
"file": "modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxVanillaOption.java",
21+
"symbol": "FxVanillaOption",
22+
"repo": "sg-evals/Strata--66225ca9"
23+
},
24+
{
25+
"file": "modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxVanillaOptionTrade.java",
26+
"symbol": "FxVanillaOptionTrade",
27+
"repo": "sg-evals/Strata--66225ca9"
28+
},
29+
{
30+
"file": "modules/product/src/main/java/com/opengamma/strata/product/fxopt/ResolvedFxVanillaOption.java",
31+
"symbol": "ResolvedFxVanillaOption",
32+
"repo": "sg-evals/Strata--66225ca9"
33+
},
34+
{
35+
"file": "modules/product/src/main/java/com/opengamma/strata/product/fxopt/ResolvedFxVanillaOptionTrade.java",
36+
"symbol": "ResolvedFxVanillaOptionTrade",
37+
"repo": "sg-evals/Strata--66225ca9"
38+
},
39+
{
40+
"file": "modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/BlackFxVanillaOptionProductPricer.java",
41+
"symbol": "BlackFxVanillaOptionProductPricer",
42+
"repo": "sg-evals/Strata--66225ca9"
43+
},
44+
{
45+
"file": "modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/BlackFxVanillaOptionTradePricer.java",
46+
"symbol": "BlackFxVanillaOptionTradePricer",
47+
"repo": "sg-evals/Strata--66225ca9"
48+
},
49+
{
50+
"file": "modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/VannaVolgaFxVanillaOptionProductPricer.java",
51+
"symbol": "VannaVolgaFxVanillaOptionProductPricer",
52+
"repo": "sg-evals/Strata--66225ca9"
53+
},
54+
{
55+
"file": "modules/pricer/src/main/java/com/opengamma/strata/pricer/fxopt/VannaVolgaFxVanillaOptionTradePricer.java",
56+
"symbol": "VannaVolgaFxVanillaOptionTradePricer",
57+
"repo": "sg-evals/Strata--66225ca9"
58+
},
59+
{
60+
"file": "modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionMeasureCalculations.java",
61+
"symbol": "FxVanillaOptionMeasureCalculations",
62+
"repo": "sg-evals/Strata--66225ca9"
63+
},
64+
{
65+
"file": "modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionTradeCalculationFunction.java",
66+
"symbol": "FxVanillaOptionTradeCalculationFunction",
67+
"repo": "sg-evals/Strata--66225ca9"
68+
},
69+
{
70+
"file": "modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionTradeCalculations.java",
71+
"symbol": "FxVanillaOptionTradeCalculations",
72+
"repo": "sg-evals/Strata--66225ca9"
73+
},
74+
{
75+
"file": "modules/measure/src/main/java/com/opengamma/strata/measure/fxopt/FxVanillaOptionMethod.java",
76+
"symbol": "FxVanillaOptionMethod",
77+
"repo": "sg-evals/Strata--66225ca9"
78+
},
79+
{
80+
"file": "modules/loader/src/main/java/com/opengamma/strata/loader/csv/FxVanillaOptionTradeCsvPlugin.java",
81+
"symbol": "FxVanillaOptionTradeCsvPlugin",
82+
"repo": "sg-evals/Strata--66225ca9"
83+
},
84+
{
85+
"file": "modules/product/src/main/java/com/opengamma/strata/product/fxopt/FxSingleBarrierOption.java",
86+
"symbol": "FxSingleBarrierOption",
87+
"repo": "sg-evals/Strata--66225ca9"
88+
}
89+
]
90+
}

‎benchmarks/csb_org_incident/ccx-incident-113/tests/ground_truth_agent.json‎

Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,39 @@
1+
{
2+
"files": [
3+
"grafana::apps/dashboard/pkg/migration/schemaversion/v38.go",
4+
"grafana::apps/dashboard/pkg/migration/schemaversion/migrations.go",
5+
"grafana::apps/dashboard/pkg/migration/migrate.go"
6+
],
7+
"symbols": [
8+
{
9+
"file": "apps/dashboard/pkg/migration/schemaversion/v38.go",
10+
"symbol": "V38",
11+
"repo": "sg-evals/grafana--26d36ec"
12+
},
13+
{
14+
"file": "apps/dashboard/pkg/migration/schemaversion/v38.go",
15+
"symbol": "processPanelsV38",
16+
"repo": "sg-evals/grafana--26d36ec"
17+
},
18+
{
19+
"file": "apps/dashboard/pkg/migration/schemaversion/v38.go",
20+
"symbol": "migrateOverrides",
21+
"repo": "sg-evals/grafana--26d36ec"
22+
},
23+
{
24+
"file": "apps/dashboard/pkg/migration/schemaversion/v38.go",
25+
"symbol": "migrateTableDisplayModeToCellOptions",
26+
"repo": "sg-evals/grafana--26d36ec"
27+
},
28+
{
29+
"file": "apps/dashboard/pkg/migration/schemaversion/migrations.go",
30+
"symbol": "GetMigrations",
31+
"repo": "sg-evals/grafana--26d36ec"
32+
},
33+
{
34+
"file": "apps/dashboard/pkg/migration/schemaversion/migrations.go",
35+
"symbol": "LATEST_VERSION",
36+
"repo": "sg-evals/grafana--26d36ec"
37+
}
38+
]
39+
}

0 commit comments

Comments
 (0)