Skip to content

v2.3.0

Latest

Choose a tag to compare

@github-actions github-actions released this 11 Mar 15:32

2.3.0 (2026-03-11)

Bug Fixes

  • add --break-system-packages for pip installs + pip.conf bypass PEP 668 (14430c4)
  • allow clippy too_many_arguments for run_task_pipeline (6eb69c2)
  • auto-install deps, python3 symlink, detect full commands in fail_to_pass, language-aware test scripts (a38497f)
  • config test race condition with env var mutex (2963325)
  • correct Basilica API types and SSH key support (63d8174)
  • enable apt/sudo in Basilica containers (d83cb8c)
  • expose agent_output and agent_patch in TaskResult and API responses (348c251)
  • extract_agent_only for /evaluate - no tasks/ dir required (2b90ee1)
  • filter out apt-get/system commands from install (Basilica blocks syscalls), keep project-level installs (e5365da)
  • full clone for commit checkout, explicit pip/pytest symlinks (a0c1d6f)
  • handle null test_patch from HuggingFace API (deserialize null as empty string) (492d068)
  • increase clone/install timeout from 180s to 600s (95cecc3)
  • install base tools, runtimes, and filter redundant deps for Basilica (80a3a0c)
  • install corepack/yarn/pnpm globally via npm in Dockerfile (b7183e8)
  • move workspace to /home/agent/sessions, fix node_modules permissions, improve agent code error handling (1ced355)
  • normalize repo URL in parse_task (add github.com prefix) (398a6fd)
  • pip 22 compatibility for base tools and install commands (68bb93f)
  • remove redundant into_iter() for clippy (eaf2a7c)
  • report task status incrementally during batch execution (4440fd8)
  • resolve all clippy warnings for CI (2b3ae9d)
  • revert Dockerfile git-lfs changes, add GIT_LFS_SKIP_SMUDGE to snapshot clone (7130823)
  • run agent from repo_dir CWD, use absolute path to agent.py (cc6bcde)
  • run as root (Basilica blocks sudo), remove sudo prefix logic (477a433)
  • sudo for apt-get in install commands, add golang/corepack/sudo to Dockerfile (1aceb88)
  • upgrade Go to 1.23 and Node to 20 LTS in Dockerfile (67ca713)
  • use :id path params for Axum 0.7 (not {id} which is 0.8) (5dfa0c1)

Features

  • /evaluate endpoint using stored agent + TRUSTED_VALIDATORS whitelist (b6aee7a)
  • add /code-hash endpoint for code integrity verification (0a8e01b)
  • add /upload-agent-json endpoint for JSON-based agent upload (9cfa1da)
  • add Basilica API client for container provisioning (8a0afca)
  • add install field from swe-forge dataset, fix default split to train, add openssh-client (737ab1f)
  • add POST /submit_tasks endpoint + fix HuggingFace dataset compat (d92444c)
  • agent user with sudo for apt-install, run all commands as non-root agent (e3f574a)
  • agent ZIP upload frontend with env vars + SUDO_PASSWORD auth (3aa5184)
  • auto-install language runtimes from install_config version fields (25b2e94)
  • change default max_concurrent_tasks from 8 to 6, support CONCURRENTLY_TASKS env var (eaba581)
  • extract full agent project instead of concatenating files (3ac1023)
  • fat Docker image with all language runtimes (java, rust, pnpm, unzip, etc.) (3855f2d)
  • fetch task definitions from HF repo (workspace.yaml + tests/), remove auto_install hack (7162a39)
  • propagate agent_env to run_agent and pass --instruction arg to Python agents (d922264)
  • replace per-file HF downloads with bulk git clone snapshot (6036b78)
  • run each task in its own Basilica container via SSH (432107b)
  • swe-bench/swe-forge integration - extend WorkspaceConfig with fail_to_pass/pass_to_pass/install_config/difficulty fields - parse swe-forge workspace.yaml native fields as test script fallback - capture git diff (agent patch) after agent execution - add /dataset endpoint to fetch from HuggingFace CortexLM/swe-forge - wire fail_to_pass/pass_to_pass in dataset entry conversion (814259e)