Problem

SWT-bench benchmark images (source-minimal target) bundle three heavy dependency groups that benchmarks never use:

These add build time, disk footprint (3+ GiB/image), and push time to every benchmark image — for functionality benchmarks don't exercise.

However, these dependencies are critical for other OpenHands users (ACP for Claude Code/Codex agent support, boto3 for Bedrock model discovery, browser-use for browser automation). We cannot simply remove them.

Proposal

Add build-time flags with safe defaults that preserve current behavior for all existing users, while allowing benchmarks to opt out of unused dependencies:

# New build args (in base-image-minimal stage)
ARG INSTALL_ACP=true
ARG INSTALL_BOTO3=true
ARG INSTALL_BROWSER=true

• Default true = identical to today. No user sees any change.
• Benchmarks pass false = lighter images, faster builds.

Dependency-by-dependency analysis

1. npm ACP packages — trivial

The ACP npm packages are installed unconditionally in base-image-minimal (Dockerfile L88-96):

npm install -g @zed-industries/claude-agent-acp @zed-industries/codex-acp

ACP is architecturally isolated — only loaded when running in ACP server mode. The agent server and benchmark evaluation paths never import it.

Fix: Wrap in a conditional:

ARG INSTALL_ACP=true
RUN set -eux; \
    if ! command -v npm >/dev/null 2>&1; then \
        curl -fsSL https://deb.nodesource.com/setup_22.x | bash - && \
        apt-get install -y --no-install-recommends nodejs && \
        rm -rf /var/lib/apt/lists/*; \
    fi; \
    if [ "$INSTALL_ACP" = "true" ]; then \
        npm install -g @zed-industries/claude-agent-acp @zed-industries/codex-acp; \
    fi

2. boto3/botocore — trivial

boto3 is already an optional extra in openhands-sdk/pyproject.toml L29-30:

[project.optional-dependencies]
boto3 = ["boto3>=1.35.0"]

And the runtime already handles its absence gracefully via lazy import in unverified_models.py:

def _get_boto3():
    try:
        return importlib.import_module("boto3")
    except ModuleNotFoundError:
        return None

If boto3 isn't installed, Bedrock model listing is skipped with a warning. Everything else works fine.

The only reason it's always installed is that the Dockerfile unconditionally passes --extra boto3 (Dockerfile L34):

uv sync --frozen --no-editable --managed-python --extra boto3

Fix: Conditionally include the extra:

ARG INSTALL_BOTO3=true
RUN ... uv sync --frozen --no-editable --managed-python $([ "$INSTALL_BOTO3" = "true" ] && echo "--extra boto3")

3. browser-use — moderate (but well-positioned)

browser-use>=0.8.0 is currently a hard dependency of openhands-tools (pyproject.toml L14):

dependencies = [
    ...
    "browser-use>=0.8.0",
    ...
]

However, the runtime already treats it as optional. Browser tools are conditionally loaded behind an enable_browser flag in preset/default.py:

if enable_browser:
    from openhands.tools.browser_use import BrowserToolSet

CLI mode explicitly disables browser tools (enable_browser=not cli_mode). Benchmarks also don't use them.

Fix (two parts):

1. Move browser-use to an optional extra in openhands-tools/pyproject.toml:

   dependencies = [
       "openhands-sdk",
       "bashlex>=0.18",
       "binaryornot>=0.4.4",
       "cachetools",
       "libtmux>=0.53.0",
       "pydantic>=2.11.7",
       "func-timeout>=4.3.5",
       "tom-swe>=1.0.3",
   ]
   
   [project.optional-dependencies]
   browser = ["browser-use>=0.8.0"]

2. Add a try/except guard in preset/default.py for when the package isn't installed:

   if enable_browser:
       try:
           from openhands.tools.browser_use import BrowserToolSet
           logger.debug(f"Tool: {BrowserToolSet.name} registered.")
       except ImportError:
           logger.warning("browser-use not installed — browser tools unavailable")

3. Add a corresponding Dockerfile build arg and conditionally include --extra browser in uv sync.

Changes required

SDK repo (software-agent-sdk)

Benchmarks repo

Expected impact

Combined with the ARG cache fix from #531 (SDK PR #2522), cold builds for 433 images could drop below 4 hours.

Non-breaking guarantee

• All build args default to true — existing docker build invocations produce identical images
• pip install openhands-tools continues to work (browser-use becomes an extra, but the Dockerfile includes it by default)
• Runtime code already handles missing browser tools and missing boto3 gracefully
• Only benchmark builds explicitly opt out via --build-arg

Related

• SWT-Bench image building slowness: root cause analysis and fix plan #531 — Root cause analysis of SWT-bench build slowness (ARG ordering fix)
• SDK #2522 — ARG ordering fix (restores registry cache)
• SDK #2465 — PR that added npm ACP install to every image
• Set SWT-bench cache-mode default to max #547 — Re-enable cache-mode=max after ARG fix