harbor-docker-manual

name: harbor-docker-manual description: Manually reproduce Harbor task containers in local Docker. Use when the user wants to inspect or run a Harbor benchmark image by hand, trace a Harbor job or script such as instance/robinCliScripts/v2.sh to find the task Docker image, install a Cline/clite tarball inside the container, debug Harbor agent setup commands, or run Cline manually from inside a Docker shell.

Harbor Docker Manual

Use this skill to help the user recreate Harbor's container setup manually, usually so they can inspect the task environment and run cline/clite by hand.

Workflow

1. Start from the launcher script or job path the user names.
2. Read the launcher script enough to identify:
   • ROOT_DIR
   • DATASET
   • AGENT
   • ENV_TYPE
   • TARBALL_URL
   • agent kwargs passed with --ak
3. Locate the task config:
   • Existing job: read jobs/<job>/<trial>/config.json.
   • Cached Harbor task: search ~/.cache/harbor/tasks for the task name.
   • New run script: infer the selected task from INCLUDE_TASK_NAME, or tell the user to pick one task first.
4. Find the Docker image:
   • Prefer task.toml field [environment].docker_image.
   • For existing jobs, read .task.path from trial config.json, then find the matching cached task.
   • If no docker_image exists, use the task's environment/Dockerfile and explain that Harbor would build an image rather than pull one.
5. Start a manual Docker shell matching Harbor's runtime architecture.
6. Have the user run install/setup mostly from inside the shell.
7. Run Cline non-interactively with explicit provider/key/model flags.

Useful Discovery Commands

Find the selected settings in a run script:

sed -n '1,180p' /path/to/instance/robinCliScripts/v2.sh

Inspect an existing job and trial:

jq '.environment, .task, .agent' /path/to/jobs/<job>/<trial>/config.json

Find a cached task by name:

find ~/.cache/harbor/tasks -maxdepth 3 -type d -name '<task-name>'

Read the task image:

sed -n '1,120p' ~/.cache/harbor/tasks/<cache-id>/<task-name>/task.toml

List local Harbor-ish Docker images:

docker images --format '{{.Repository}}:{{.Tag}} {{.ID}} {{.Size}}' | rg 'hb__|alexgshaw|terminal|swebench'

Start Container

Use linux/amd64 for Linux x64 tarballs, especially on Apple Silicon Macs. Do not rely on host architecture.

IMAGE="<task-image>"
NAME="harbor-manual"

docker rm -f "$NAME" 2>/dev/null || true

docker run -it \
  --name "$NAME" \
  --platform linux/amd64 \
  -w /app \
  -e API_KEY="$OPENROUTER_API_KEY" \
  -e MODELID="anthropic/claude-sonnet-4.6" \
  "$IMAGE" \
  bash

If the task workdir is not /app, inspect the task Dockerfile or run pwd, ls, and find / -maxdepth 2 -name test.sh 2>/dev/null inside the container.

Install Cline Tarball Inside Container

Run this inside the container. This mirrors Harbor's cline-v2 install path closely enough for manual debugging.

TARBALL_URL="<tarball-url>"

apt-get update
apt-get install -y curl ca-certificates git

curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.2/install.sh | bash

export NVM_DIR="$HOME/.nvm"
. "$NVM_DIR/nvm.sh"

nvm install 22
nvm use 22
nvm alias default 22

npm install -g --ignore-scripts -- "$TARBALL_URL"

which cline
which clite
cline --version
clite --version

If API_KEY or MODELID is empty inside the shell, set them manually:

export API_KEY="sk-or-v1-..."
export MODELID="anthropic/claude-sonnet-4.6"

Run Cline Manually

Avoid plain cline in Docker. It may trigger browser auth or interactive TUI behavior where 127.0.0.1 points at the container, not the host.

Use explicit provider/key/model flags and a prompt:

cline -P openrouter \
  -k "$API_KEY" \
  -m "$MODELID" \
  --yolo \
  --reasoning-effort medium \
  --json \
  "Hello"

For closest Harbor behavior, close stdin and capture all output:

set -o pipefail
cline -P openrouter \
  -k "$API_KEY" \
  -m "$MODELID" \
  --yolo \
  --reasoning-effort medium \
  -- "Task prompt here" \
  < /dev/null 2>&1 | tee /tmp/cline.txt

echo "exit=${PIPESTATUS[0]}"

Common Debug Checks

Environment variables:

echo "API_KEY set? ${API_KEY:+yes}"
echo "MODELID=$MODELID"

Node/Cline path:

export NVM_DIR="$HOME/.nvm"
. "$NVM_DIR/nvm.sh"
which node
which npm
which cline
cline --version

Cline state/log files:

find ~/.cline -maxdepth 5 -type f | sort | tail -100

Terminal escape junk such as 997;1n10;rgb... after running bare cline is usually leaked TTY response text. Run reset, then use non-interactive commands with a prompt instead of the bare TUI.

Cleanup

From the host:

docker rm -f harbor-manual

Use the actual NAME value if different.