From bdef8e1766295afc2a04ceb339f4d311111b8d15 Mon Sep 17 00:00:00 2001
From: echobt <154886644+echobt@users.noreply.github.com>
Date: Wed, 29 Jul 2026 13:14:35 +0000
Subject: [PATCH] fix(agent-challenge): allow PyPI index for agent dependency
 install

Host-local evaluation has egress to PyPI; --no-index caused instant
empty-index failures for real agent deps (e.g. litellm>=1.55.0).
Drop offline-only pip flags and use network-sane retries/timeouts.
---
 .../challenges/agent-challenge/Dockerfile     |  6 +--
 .../src/agent_challenge/evaluation/runner.py  | 15 +++-----
 .../agent-challenge/tests/test_evaluation.py  | 37 +++++++++----------
 3 files changed, 27 insertions(+), 31 deletions(-)

diff --git a/packages/challenges/agent-challenge/Dockerfile b/packages/challenges/agent-challenge/Dockerfile
index 5fc15696..5797349e 100644
--- a/packages/challenges/agent-challenge/Dockerfile
+++ b/packages/challenges/agent-challenge/Dockerfile
@@ -134,9 +134,9 @@ COPY golden/dataset-digest.json ./golden/dataset-digest.json
 
 RUN pip install --no-cache-dir .
 
-# Pre-bake the toolchain a submitted agent needs so its install resolves fully
-# offline. Runner jobs run on an egress-free network, so own_runner installs the
-# agent with `--no-index --no-build-isolation` (see runner._own_runner_script).
+# Pre-bake common PEP 517 build backends and agent runtime deps so installs
+# are faster and more reliable. own_runner still resolves remaining agent
+# deps from PyPI (see runner._own_runner_script).
 RUN pip install --no-cache-dir \
         "setuptools>=61" \
         "wheel>=0.40" \
diff --git a/packages/challenges/agent-challenge/src/agent_challenge/evaluation/runner.py b/packages/challenges/agent-challenge/src/agent_challenge/evaluation/runner.py
index 367e15c2..1749e68e 100644
--- a/packages/challenges/agent-challenge/src/agent_challenge/evaluation/runner.py
+++ b/packages/challenges/agent-challenge/src/agent_challenge/evaluation/runner.py
@@ -2000,14 +2000,11 @@ def _own_runner_script(
     # (no inner dockerd). The default socket path needs no DOCKER_HOST, but we
     # set it explicitly so a custom broker socket path still resolves.
     #
-    # Offline agent install: runner jobs are attached to an egress-free network
-    # (a security boundary for untrusted agent code), so the install must resolve
-    # entirely from packages pre-baked into the runner image. --no-build-isolation
-    # reuses those baked PEP 517 build backends instead of fetching them into a
-    # fresh isolated build env (the setuptools>=61 fetch that previously failed
-    # every install), and --no-index keeps a missing/exotic dep failing fast
-    # instead of hanging on unreachable pypi retries. `|| true` preserves the
-    # best-effort behaviour so a partially-satisfiable agent still attempts to run.
+    # Agent dependency install: resolves from PyPI on the host-local evaluation
+    # path (CHALLENGE_NO_PHALA / unattested), which has egress. The egress-free
+    # --internal overlay applies only to the broker/Swarm job path, not here.
+    # Network-sane retries/timeouts; `|| true` keeps best-effort behaviour so a
+    # partially-satisfiable agent still attempts to run.
     return f"""
 set -u
 cd /workspace/agent
@@ -2016,7 +2013,7 @@ def _own_runner_script(
 {replay_env}
 TMO="timeout -k 10 -s KILL 600"
 PIP="python -m pip install --no-input --disable-pip-version-check"
-PIP="$PIP --no-index --no-build-isolation --retries 0 --default-timeout 15"
+PIP="$PIP --retries 3 --default-timeout 30"
 if [ -f requirements.txt ]; then $TMO $PIP -r requirements.txt || true; fi
 if [ -f pyproject.toml ]; then $TMO $PIP -e . || true; fi
 mkdir -p {output_dir}
diff --git a/packages/challenges/agent-challenge/tests/test_evaluation.py b/packages/challenges/agent-challenge/tests/test_evaluation.py
index e617c33d..e2538020 100644
--- a/packages/challenges/agent-challenge/tests/test_evaluation.py
+++ b/packages/challenges/agent-challenge/tests/test_evaluation.py
@@ -2428,18 +2428,15 @@ def test_own_runner_script_cache_wiring_honours_settings_overrides(monkeypatch):
     assert "--digest-manifest /custom/golden/digest.json" in script
 
 
-def test_own_runner_script_pip_installs_are_offline_and_hang_proofed():
-    """Runner jobs run on an egress-free network, so the agent install must
-    resolve entirely from packages pre-baked into the runner image and must fail
-    fast rather than hang on unreachable pypi.
-
-    Regression (score-0.0 bug): every terminal-bench task scored 0.0 because the
-    agent install reached out to pypi and failed/hung. The fix installs with
-    ``--no-build-isolation`` (reuse the pre-baked PEP 517 build backends instead
-    of fetching setuptools>=61 into a fresh isolated build env) and ``--no-index``
-    (a missing/exotic dep fails immediately instead of retrying pypi for ~150s).
-    Both installs stay wrapped in a hard ``timeout -k 10 -s KILL`` safety net and
-    keep ``|| true`` so a partially-satisfiable agent still attempts to run.
+def test_own_runner_script_pip_installs_use_pypi_with_sane_retries():
+    """Agent install resolves from PyPI with network-sane pip flags.
+
+    Host-local evaluation has egress to PyPI; ``--no-index`` was blocking real
+    agent deps (e.g. litellm>=1.55.0) with an instant empty-index failure.
+    The install keeps ``--no-input`` / ``--disable-pip-version-check``, uses
+    retries/timeouts suitable for network installs, stays wrapped in a hard
+    ``timeout -k 10 -s KILL`` safety net, and keeps ``|| true`` so a
+    partially-satisfiable agent still attempts to run.
     """
 
     job = EvaluationJob(job_id="job-hangproof", selected_tasks_json="[]")
@@ -2454,16 +2451,18 @@ def test_own_runner_script_pip_installs_are_offline_and_hang_proofed():
 
     assert 'TMO="timeout -k 10 -s KILL 600"' in script
     assert "python -m pip install --no-input --disable-pip-version-check" in script
-    # Offline-first: no isolated build env, no pypi index, fail fast.
-    assert "--no-build-isolation" in script
-    assert "--no-index" in script
-    assert "--retries 0 --default-timeout 15" in script
-    # Both install paths carry the offline flags and stay best-effort.
+    # PyPI-reachable install: no offline-only flags; network-sane retries.
+    assert "--no-index" not in script
+    assert "--no-build-isolation" not in script
+    assert "--retries 3 --default-timeout 30" in script
+    # Both install paths stay best-effort under the hard timeout.
     assert "$TMO $PIP -r requirements.txt || true" in script
     assert "$TMO $PIP -e . || true" in script
     pip_flag_line = next(line for line in script.splitlines() if line.startswith('PIP="$PIP'))
-    assert "--no-index" in pip_flag_line
-    assert "--no-build-isolation" in pip_flag_line
+    assert "--no-index" not in pip_flag_line
+    assert "--no-build-isolation" not in pip_flag_line
+    assert "--retries 3" in pip_flag_line
+    assert "--default-timeout 30" in pip_flag_line
 
 
 def test_own_runner_command_args_wires_cache_root_and_digest_manifest():