Repo skeleton, pinned environment, free smoke test

Public from commit one, so no key ever enters this history.

compose.yaml is ImpossibleBench's, plus working_dir: /workspace. The image has no
WORKDIR, so inspect resolves it to "/" and the task files land at the filesystem root
among twenty-odd entries. This experiment turns on whether an agent notices a scratch
directory, so that is a bad place to put one.

ImpossibleBench installs with --no-deps to keep the swebench tree out; datasets is
declared here instead because hf_dataset genuinely needs it.

Verified: docker run prints "/", impossiblebench imports, and the real task against
mockllm/model completes with a real score and tracebacks rooted at /workspace.

Claude-Session: https://claude.ai/code/session_01Cq98H7sNoSJdL3W98f18bu
This commit is contained in:
pj committed 2026-08-31 21:14:58 +05:30
commit bf3bb0d144
12 files changed
+2495

No files matched your search

+22
View File
@@ -0,0 +1,22 @@
# Copied from impossiblebench/src/impossiblebench/compose.yaml (commit 061dc3d),
# with working_dir added.
#
# The image has no WORKDIR, so inspect resolves the container working directory to
# "/" and the task's func.py and test.py land at the filesystem root, mixed in with
# the twenty-odd entries a bare `ls` returns there. This experiment turns on whether
# an agent notices and uses a scratch directory, so burying it in / is a bad idea.
# /workspace holds exactly func.py, test.py and scratch/. Docker creates the
# directory on start; it does not exist in the image.
#
# This is a deviation from the published benchmark's default environment and belongs
# in the paper's limitations. It applies identically to both conditions, so it cannot
# produce the effect.
services:
default:
image: "aisiuk/inspect-tool-support"
command: "tail -f /dev/null"
working_dir: /workspace
init: true
network_mode: none
stop_grace_period: 1s
mem_limit: 1gb