better summary message for submit test/benchmark

Author: ngc92

src/discord-cluster-manager/cogs/submit_cog.py

@@ -9,7 +9,7 @@
 from consts import SubmissionMode
 from discord import app_commands
 from discord.ext import commands
-from report import MultiProgressReporter, RunProgressReporter, generate_report, private_run_report
+from report import MultiProgressReporter, RunProgressReporter, generate_report, make_short_report
 from run_eval import FullResult
 from task import LeaderboardTask
 from utils import build_task_config, send_discord_message, setup_logging, with_error_handling
@@ -150,12 +150,11 @@ async def _handle_submission(
         else:
             await reporter.update_title(reporter.title + " ✅ success")
 
-        if mode == SubmissionMode.PRIVATE:
-            await reporter.push(private_run_report(result.runs))
-        else:
-            if mode == SubmissionMode.LEADERBOARD:
-                await reporter.push(private_run_report(result.runs))
-
+        await reporter.push(make_short_report(
+            result.runs,
+            full=mode in [SubmissionMode.PRIVATE, SubmissionMode.LEADERBOARD])
+        )
+        if mode != SubmissionMode.PRIVATE:
             try:
                 await generate_report(thread, result.runs)
                 await reporter.push(f"See results at {thread.jump_url}")

src/discord-cluster-manager/report.py

@@ -112,7 +112,7 @@ async def _generate_test_report(thread: discord.Thread, run: RunResult):
     return
 
 
-def private_run_report(runs: dict[str, EvalResult]) -> list[str]:  # noqa: C901
+def make_short_report(runs: dict[str, EvalResult], full=True) -> list[str]:  # noqa: C901
     """
     Creates a minimalistic report for `runs`,
     returned as a list of status strings
@@ -128,31 +128,42 @@ def private_run_report(runs: dict[str, EvalResult]) -> list[str]:  # noqa: C901
     if any_compile:
         result.append("✅ Compilation successful")
 
-    if "test" not in runs or not runs["test"].run.success:
-        result.append("❌ Running tests failed")
-        return result
-    elif not runs["test"].run.passed:
-        result.append("❌ Testing failed")
-        return result
-    else:
-        result.append("✅ Testing successful")
-
-    if "benchmark" not in runs or not runs["benchmark"].run.success:
-        result.append("❌ Running benchmarks failed")
-        return result
-    elif not runs["benchmark"].run.passed:
-        result.append("❌ Benchmarking failed")
-        return result
-    else:
-        result.append("✅ Benchmarking successful")
+    if "test" in runs:
+        test_run = runs["test"].run
+        if not test_run.success:
+            result.append("❌ Running tests failed")
+            return result
+        elif not test_run.passed:
+            result.append("❌ Testing failed")
+            return result
+        else:
+            result.append("✅ Testing successful")
+    elif full:
+        result.append("❌ Tests missing")
 
-    if "leaderboard" not in runs or not runs["leaderboard"].run.success:
-        result.append("❌ Running leaderboard failed")
-    elif not runs["leaderboard"].run.passed:
-        result.append("❌ Leaderboard run failed")
-    else:
-        result.append("✅ Leaderboard run successful")
+    if "benchmark" in runs:
+        bench_run = runs["benchmark"].run
+        if bench_run.success:
+            result.append("❌ Running benchmarks failed")
+            return result
+        elif not bench_run.passed:
+            result.append("❌ Benchmarking failed")
+            return result
+        else:
+            result.append("✅ Benchmarking successful")
+    elif full:
+        result.append("❌ Benchmarks missing")
 
+    if "leaderboard" in runs:
+        lb_run = runs["leaderboard"].run
+        if not lb_run.success:
+            result.append("❌ Running leaderboard failed")
+        elif not lb_run.passed:
+            result.append("❌ Leaderboard run failed")
+        else:
+            result.append("✅ Leaderboard run successful")
+    elif full:
+        result.append("❌ Leaderboard missing")
     return result