Land PR #2 on main, remove dead code, rewrite README for customers

.gitignore

@@ -8,61 +8,24 @@ e2b.toml
 # Python environments and caches
 .venv/
 venv/
-env/
-ENV/
 __pycache__/
 *.py[cod]
-*$py.class
-.cache/
 .pytest_cache/
 .mypy_cache/
 .ruff_cache/
-.tox/
-.nox/
 
-# Packaging, builds, and coverage
+# Packaging and builds
 build/
 dist/
 *.egg-info/
-.eggs/
-pip-wheel-metadata/
-.coverage
-.coverage.*
-htmlcov/
 
-# Generated benchmark output and runtime state
+# Generated by cybergym-e2b: template manifest and ledger, image lock, results
 artifacts/
-campaigns/
-reports/
-docs/validation/
-experiments/
-results/
-outputs/
-trajectories/
-logs/
-*.log
-*.pid
-*.pid.lock
-template-manifest*.json
-build-ledger*.json
-credentials.json
-credentials.toml
-credentials.yaml
-credentials.yml
-*.key
-*.p12
-*.pfx
-/*-tasks.txt
-/scripts/analyze_profiles.py
 
-# Downloaded, gated, vendored, and local-only assets
-downloads/
+# Pinned upstream checkout created by `cybergym-e2b sync-upstream`
 vendor/
-benchmark-data/
-local-assets/
 
 # Editors and operating systems
 .idea/
 .vscode/
 .DS_Store
-Thumbs.db

FOLLOW-UPS.md

@@ -0,0 +1,29 @@
+# Follow-ups
+
+Issues are disabled on this repository, so this file is the tracker. Delete entries as they land.
+None of these affects the correctness of a completed run.
+
+## Compatibility patch (`assets/patches/openai-compatible.patch`)
+
+1. The `poc_file is None or ...` and `patch_file is None or ...` guards turn upstream
+   `run_agent()`'s exception path into a graded `no_poc` / `no_patch` failure instead of
+   upstream's retryable `error` status. Codex results are rescued by the trajectory check in
+   `_benchmark_result`; OpenHands results are not.
+2. `LLM_MAX_INPUT_TOKENS=131072` equals DeepSeek V3.2's full context window, leaving no headroom
+   for the 8192-token completion on that model via OpenHands.
+
+## Coverage
+
+3. Add a plain `openai` provider (host `api.openai.com`) so customers without Fireworks or Bedrock
+   can run Codex. The runtime already speaks the OpenAI Responses API; this needs a
+   `--provider` choice, a host entry in both packaged policies, and a key name.
+4. Upstream's default agent is `claude-code` with the Anthropic API. Supporting it needs
+   `ANTHROPIC_API_KEY` delivery through the egress proxy (upstream passes it as a container
+   environment variable) and `api.anthropic.com` in the packaged policies; upstream installs the
+   CLI with `npm install -g @anthropic-ai/claude-code` inside the agent container. Without it the
+   paper's reference configuration cannot be reproduced on E2B.
+
+## Notes
+
+5. The template tag-to-build check is cached per process on purpose; rebuilding or re-verifying
+   per run is not wanted.

pyproject.toml

@@ -1,13 +1,13 @@
 [project]
 name = "cybergym-e2e-e2b"
-version = "0.2.0"
+version = "0.3.0"
 description = "Run CyberGym-E2E with one Docker-in-Docker E2B sandbox per task"
 readme = "README.md"
 license = "Apache-2.0"
 license-files = ["LICENSE"]
 requires-python = ">=3.12"
 dependencies = [
-  "e2b>=2.37.0,<2.38",
+  "e2b>=2.37,<3",
   "httpx>=0.28,<0.29",
   "tomli-w>=1.2,<2",
 ]

scripts/check_public_tree.py

@@ -14,19 +14,8 @@
     ".ruff_cache",
     "__pycache__",
     "artifacts",
-    "benchmark-data",
     "build",
-    "campaigns",
     "dist",
-    "downloads",
-    "experiments",
-    "local-assets",
-    "logs",
-    "outputs",
-    "results",
-    "trajectories",
-    "reports",
-    "validation",
     "vendor",
 }
 
@@ -43,10 +32,6 @@
     re.compile(r"^template-manifest(?:\..+)?\.json$"),
     re.compile(r"^build-ledger(?:\..+)?\.json$"),
 )
-_FORBIDDEN_PATH_PATTERNS = (
-    re.compile(r"^[^/]+-tasks\.txt$"),
-    re.compile(r"^scripts/analyze_profiles\.py$"),
-)
 
 
 def _is_forbidden(path: str) -> bool:
@@ -60,8 +45,6 @@ def _is_forbidden(path: str) -> bool:
         return True
     if any(pattern.fullmatch(filename) for pattern in _GENERATED_FILE_PATTERNS):
         return True
-    if any(pattern.fullmatch(normalized) for pattern in _FORBIDDEN_PATH_PATTERNS):
-        return True
     return any(
         part in _GENERATED_DIRECTORIES
         or part in _LOCAL_ENVIRONMENT_DIRECTORIES

src/cybergym_e2b/assets/patches/openai-compatible.patch

@@ -1,41 +1,100 @@
+diff --git a/scripts/install_validate_deps.sh b/scripts/install_validate_deps.sh
+index 6834b1c..11da429 100644
+--- a/scripts/install_validate_deps.sh
++++ b/scripts/install_validate_deps.sh
+@@ -1,7 +1,9 @@
+ #!/bin/bash
+
+-apt-get update
+-apt-get install -y curl
++if ! command -v curl >/dev/null; then
++    apt-get update
++    apt-get install -y curl
++fi
+
+ curl -LsSf https://astral.sh/uv/install.sh | sh
+ source $HOME/.local/bin/env
 diff --git a/scripts/run_agent.py b/scripts/run_agent.py
-index bb67950..aae4d4f 100644
+index bb67950..2bd8b25 100644
 --- a/scripts/run_agent.py
 +++ b/scripts/run_agent.py
-@@ -631,0 +632 @@ model_provider = "openai_http"
+@@ -629,11 +629,13 @@ EOF''',
+ openai_base_url = "{openai_base_url}"
+ web_search = "disabled"
+ model_provider = "openai_http"
 +model_reasoning_effort = "{args.reasoning_effort}"
-@@ -636,0 +638 @@ supports_websockets = false
+
+ [model_providers.openai_http]
+ name = "OpenAI HTTP"
+ base_url = "{openai_base_url}"
+ supports_websockets = false
 +wire_api = "responses"
-@@ -788,0 +791,6 @@ def run_agent(args, config, script_path, data_path, prompt, attempt, work_dir, t
+ EOF''',
+             "Adding OPENAI_BASE_URL to config.toml"
+         )
+@@ -786,6 +788,12 @@ def run_agent(args, config, script_path, data_path, prompt, attempt, work_dir, t
+
+         print(f"  Container: {container_id[:12]}")
+
 +        e2b_ca_bundle = os.environ.get("E2B_CA_BUNDLE")
 +        if e2b_ca_bundle:
 +            copy_to_container(
 +                container_id, Path(e2b_ca_bundle), "/etc/ssl/certs/ca-certificates.crt"
 +            )
 +
-@@ -881 +890 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+         # Setup workspace (don't copy ground truth PoC - agent shouldn't see it)
+         setup_workspace(container_id, data_path, script_path, args.mode, copy_gt_poc=False, scripts_dir=scripts_dir)
+
+@@ -878,7 +886,7 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+             print(f"  Agent: {agent_time:.1f}s ({agent_time/60:.1f}m), exec: {agent_exec_time:.1f}s ({agent_exec_time/60:.1f}m), exit={exit_code}")
+
+             # Check for generated files
 -            if args.mode == "e2e" and not poc_file.exists():
 +            if args.mode == "e2e" and (poc_file is None or not poc_file.exists()):
-@@ -896 +905 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+                 print("  No PoC generated!")
+                 all_attempts.append({
+                     "attempt": attempt,
+@@ -893,7 +901,7 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+                     feedback = "\n=== Previous Attempt Failed ===\nNo poc.bin was generated."
+                 continue
+
 -            if not patch_file.exists():
 +            if patch_file is None or not patch_file.exists():
-@@ -972 +981,5 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+                 print("  No patch generated!")
+                 all_attempts.append({
+                     "attempt": attempt,
+@@ -969,7 +977,11 @@ def run_agent_loop(args, config, script_path, data_path, run_dir):
+                 final_status = "success"
+                 break
+             else:
 -                trajectory_summary = summarize_trajectory(log_file, attempt, args)
 +                trajectory_summary = (
 +                    summarize_trajectory(log_file, attempt, args)
 +                    if attempt < args.max_attempts
 +                    else None
 +                )
-@@ -1049 +1057 @@ Examples:
+                 feedback = format_feedback(
+                     results, attempt, args.mode,
+                     poc_file=attempt_poc,
+@@ -1045,8 +1057,9 @@ Examples:
+     parser.add_argument("--default-build-image",
+                         default="gcr.io/oss-fuzz-base/base-builder@sha256:8eda74a11e800aead5a041ee479a65b33dab3150d6e89e5694e2b6eb27be98fc")
+
++    parser.add_argument("--reasoning-effort", choices=["low", "medium", "high", "xhigh"], default="high")
+     # LLM configuration
 -    parser.add_argument("--model-provider", choices=["litellm", "bedrock", "anthropic"], default="anthropic",
 +    parser.add_argument("--model-provider", choices=["litellm", "bedrock", "anthropic", "openai-compatible"], default="anthropic",
-@@ -1051,0 +1060 @@ Examples:
-+    parser.add_argument("--reasoning-effort", choices=["low", "medium", "high", "xhigh"], default="high")
+                         help="LLM provider (default: anthropic)")
+     parser.add_argument("--litellm-model-id", default="openai/gpt-5.2-codex")
+     parser.add_argument("--bedrock-model-id", default="us.anthropic.claude-sonnet-4-5-20250929-v1:0")
 diff --git a/scripts/utils.py b/scripts/utils.py
-index f8b8b82..54ae678 100644
+index f8b8b82..2b20e1d 100644
 --- a/scripts/utils.py
 +++ b/scripts/utils.py
-@@ -176,0 +177,12 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+@@ -174,15 +174,25 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+     )
+
+     # Install validation dependencies (sudo for build scripts, tomli for validate.py)
 +    if not scripts_dir:
 +        raise ValueError("scripts_dir required for openhands")
 +    # Install both package-manager entry points before the first update. Task scripts
@@ -48,13 +107,23 @@ index f8b8b82..54ae678 100644
 +        )
 +    exec_run(container_id, "chmod 0755 /usr/local/bin/apt /usr/local/bin/apt-get", workdir="/", verbose=False, check=True)
 +
-@@ -179 +191 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+     exec_run(
+         container_id,
 -        "apt-get update -qq && apt-get install -y -qq sudo git >/dev/null 2>&1",
 +        "if ! command -v sudo >/dev/null; then printf '#!/bin/sh\\nif [ \"$1\" = \"-E\" ]; then shift; fi\\nexec \"$@\"\\n' > /usr/local/bin/sudo && chmod 0755 /usr/local/bin/sudo; fi; if ! command -v git >/dev/null; then grep -rLiE 'archive.ubuntu.com|security.ubuntu.com|ports.ubuntu.com|old-releases.ubuntu.com|deb.debian.org|security.debian.org' /etc/apt/sources.list.d/ 2>/dev/null | xargs -r -I{} mv {} /tmp/ 2>/dev/null; sed -ri '/apt.kitware.com|bazel-apt|apt.llvm.org|packages.cloud.google.com|storage.googleapis.com|download.docker.com|dl.yarnpkg.com|deb.nodesource.com|ppa.launchpad|nginx.org|packagecloud.io|cran.r-project|pgdg/d' /etc/apt/sources.list 2>/dev/null; find /etc/apt/sources.list.d -type f -exec sed -ri '/apt.kitware.com|bazel-apt|apt.llvm.org|packages.cloud.google.com|storage.googleapis.com|download.docker.com|dl.yarnpkg.com|nginx.org|packagecloud.io|cran.r-project|pgdg/d' {} + 2>/dev/null; apt-get update -qq && apt-get install -y -qq git >/dev/null 2>&1; fi",
-@@ -184,2 +195,0 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+         workdir="/",
+         verbose=False,
+         check=True,
+     )
 -    if not scripts_dir:
 -        raise ValueError("scripts_dir required for openhands")
-@@ -198,0 +209,9 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+     copy_to_container(container_id, scripts_dir / "install_validate_deps.sh", "/install_validate_deps.sh")
+     exec_run(container_id, "bash -eux /install_validate_deps.sh", workdir="/", verbose=False, check=True)
+
+@@ -196,6 +206,15 @@ def setup_workspace(container_id, data_path, script_path, mode="e2e", copy_gt_po
+     copy_to_container(container_id, data_path / "src.tgz", "/src/src.tgz")
+     exec_run(container_id, "tar xf /src/src.tgz -C /src && rm /src/src.tgz", workdir="/", verbose=False, check=True)
+
 +    # The Opus model host offers an rsync mirror but its HTTPS endpoint is
 +    # intermittently unavailable. Hot E2B templates carry the checksum-pinned
 +    # archive so FFmpeg setup remains reproducible under restricted egress.
@@ -64,7 +133,13 @@ index f8b8b82..54ae678 100644
 +    ).returncode == 0:
 +        copy_to_container(container_id, opus_model_cache, f"/src/opus/{Path(opus_model_cache).name}")
 +
-@@ -331,0 +351,24 @@ def call_llm(
+     # Copy crash.log and poc.bin for patch-only mode (to /src for agent to see)
+     if mode == "patch-only":
+         for f in ["crash.log", "poc.bin"]:
+@@ -330,6 +349,30 @@ def call_llm(
+                 messages=[{"role": "user", "content": prompt}],
+             )
+             return response.content[0].text
 +        elif model_provider == "openai-compatible":
 +            import openai
 +
@@ -89,28 +164,27 @@ index f8b8b82..54ae678 100644
 +                )
 +                return response.choices[0].message.content
 +            raise ValueError(f"unsupported OpenAI-compatible wire API: {wire_api}")
-@@ -385,0 +429,5 @@ def get_llm_env(
+         else:
+             # OpenAI
+             import openai
+@@ -383,8 +426,20 @@ def get_llm_env(
+         "LLM_RETRY_MIN_WAIT": "15",
+         "LLM_RETRY_MAX_WAIT": "120",
+         "LLM_RETRY_MULTIPLIER": "2",
 +        # OpenAI-compatible catalog IDs are absent from LiteLLM's static metadata.
 +        # Pin conservative limits so OpenHands does not assume a 4K context window.
 +        "LLM_MAX_INPUT_TOKENS": "131072",
 +        "LLM_MAX_OUTPUT_TOKENS": "8192",
 +        "LLM_NATIVE_TOOL_CALLING": "true",
-@@ -387,0 +436,7 @@ def get_llm_env(
+     }
+
 +    if os.getenv("E2B_CA_BUNDLE"):
 +        base_env.update({
 +            "SSL_CERT_FILE": "/etc/ssl/certs/ca-certificates.crt",
 +            "REQUESTS_CA_BUNDLE": "/etc/ssl/certs/ca-certificates.crt",
 +            "NODE_EXTRA_CA_CERTS": "/etc/ssl/certs/ca-certificates.crt",
 +        })
 +
-diff --git a/scripts/install_validate_deps.sh b/scripts/install_validate_deps.sh
-index 42ca8a3..4a155df 100644
---- a/scripts/install_validate_deps.sh
-+++ b/scripts/install_validate_deps.sh
-@@ -3,2 +3,4 @@
--apt-get update
--apt-get install -y curl
-+if ! command -v curl >/dev/null; then
-+    apt-get update
-+    apt-get install -y curl
-+fi
+     if model_provider == "bedrock":
+         try:
+             session = boto3.Session(profile_name=aws_profile)

src/cybergym_e2b/assets/policies/network-locked.json

@@ -3,7 +3,7 @@
   "deny_out": [],
   "model_hosts": [
     "api.fireworks.ai",
-    "bedrock-mantle.us-west-2.api.aws"
+    "bedrock-mantle.*.api.aws"
   ],
   "artifact_hosts": [
     "huggingface.co",

src/cybergym_e2b/assets/policies/network.json

@@ -11,7 +11,7 @@
   ],
   "model_hosts": [
     "api.fireworks.ai",
-    "bedrock-mantle.us-west-2.api.aws"
+    "bedrock-mantle.*.api.aws"
   ],
   "artifact_hosts": [
     "huggingface.co",