swe_agentic: auto-release pulled images at process exit

Per-instance images are 1-4GB each; a --limit 10 run pulls up to 10 of
them and previously stranded the whole footprint on Ctrl+C/exit. Every
image THIS process pulls is now registered and released by an atexit
hook (containers first, then rmi) -- images that already existed
locally are never touched. EVALHARNESS_KEEP_SWE_IMAGES=1 opts out for
prefetch-style runs that want to keep them.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/agent/envs/swe_agentic.py

@@ -13,12 +13,56 @@ import asyncio
 import json
 import re
 import subprocess
+import sys
 import uuid
 from typing import Any, Dict
 
 from ...data.sample import ChatMessage, Sample
 from ..loop import Environment, register_env
 
+# ---- image lifecycle tracking -------------------------------------------
+# per-instance swe images are BIG (1-4GB each, 500 total ~2TB): every
+# image THIS process pulled is registered here and released at exit
+# (normal end, Ctrl+C kill, crash) so a run never strands its footprint.
+import atexit as _atexit
+
+_IMAGES_PULLED: list = []
+
+
+def _release_pulled_images() -> None:
+    import os as _os
+
+    if not _IMAGES_PULLED or _os.environ.get('EVALHARNESS_KEEP_SWE_IMAGES'):
+        return
+    # in-use containers first (docker refuses rmi otherwise)
+    try:
+        _docker(['ps', '-aq', '--filter', 'name=eh-swe-'], timeout=30)
+    except Exception:
+        pass
+    try:
+        _docker(['ps', '-aq', '--filter', 'ancestor=none'], timeout=30)
+    except Exception:
+        pass
+    import subprocess as _sp
+
+    try:
+        ids = _sp.run(['docker', 'ps', '-aq', '--filter', 'name=eh-swe'],
+                      capture_output=True, text=True, timeout=30).stdout.split()
+        if ids:
+            _docker(['rm', '-f'] + ids, timeout=120)
+    except Exception:
+        pass
+    for img in _IMAGES_PULLED:
+        try:
+            _docker(['rmi', '-f', img], timeout=120)
+            print(f'· released swe image {img}', file=sys.stderr, flush=True)
+        except Exception:
+            pass
+
+
+_atexit.register(_release_pulled_images)
+# ------------------------------------------------------------------------
+
 SENTINEL = 'COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT'
 
 # mini-swe-agent swebench.yaml contract (verbatim sections that matter)
@@ -123,10 +167,13 @@ class SWEAgenticEnvironment(Environment):
         # output and its 120s timeout kills runs on slow mirrors
         from ...sandbox.docker import ensure_image
 
+        had_it = _docker(['image', 'inspect', image]).returncode == 0
         try:
             ensure_image(image)
         except RuntimeError as e:
             raise RuntimeError(f'image {image} unavailable: {str(e)[:150]}')
+        if not had_it and image not in _IMAGES_PULLED:
+            _IMAGES_PULLED.append(image)   # we pulled it -> we release it
         name = f'eh-swe-{uuid.uuid4().hex[:10]}'
         r = _docker(['run', '-d', '--name', name,
                      '-w', '/testbed',