Harbor task — JSON department salary totals

A small Harbor benchmark task for agent evaluation: read JSON employee records, compute total salary per department, and write a correctly formatted output.json.

If you are browsing this repo on GitHub, everything you need to understand and run the task is summarized here. Deeper file-level detail lives in the task folder linked below.

---

About this repository

This project is a single, self-contained Harbor task (not the full Harbor framework). It was built for a new-hire / assignment style workflow: the task passes Oracle (reference solution) and NOP (no pre-baked success) checks, and is suitable to show in a portfolio or attach to a pull request against upstream Harbor.

What “success” means for an agent

• Input file: /app/records.json (array of { employee, department, salary }).
• Output file: /app/output.json — object of department → total salary, keys sorted A–Z, 2-space JSON indent, newline at end of file.

---

Repository layout

harbor_tasks/json-dept-salary-totals/
├── task.toml           # Metadata, timeouts, CPU/RAM/disk
├── instruction.md      # Instructions shown to the agent
├── environment/        # Dockerfile + sample input only
├── solution/           # Reference solution (oracle)
└── tests/              # Verifier script + pytest tests

Item	Value
Task name (task.toml)	harbor/json-dept-salary-totals
Difficulty	easy
Category	data-processing

More detail: harbor_tasks/json-dept-salary-totals/README.md

---

Who this is for

• Reviewers / instructors — quick sanity check of task structure and validation expectations.
• You, running it locally — clone this repo, clone Harbor, run harbor run pointing at harbor_tasks/json-dept-salary-totals/.
• Contributors — the same task may also be proposed via a fork/PR to laude-institute/harbor under examples/tasks/ (naming may differ slightly, e.g. test_state.py).

---

Prerequisites

• Docker (Docker Desktop on Windows is fine; WSL is often used with Harbor.)
• uv
• A clone of laude-institute/harbor and uv sync inside that repo so uv run harbor … works.

---

Quick start (run the task)

In a terminal, from your Harbor clone (after uv sync):

# Replace TASK_PATH with the path to this repo’s task folder on your machine.
export TASK_PATH="/path/to/harbor-task-assignment/harbor_tasks/json-dept-salary-totals"

uv run harbor run --agent oracle --path "$TASK_PATH" --job-name oracle-json-dept
uv run harbor run --agent nop    --path "$TASK_PATH" --job-name nop-json-dept

Expected results

• Oracle — mean reward 1.0 (reference solution + tests pass).
• NOP — mean reward 0.0 (agent does nothing; output must not appear by default).

Lint the pytest file:

uvx ruff check "$TASK_PATH/tests/test_outputs.py"

On Windows PowerShell you can set the path instead:

$TASK_PATH = "C:\Users\YOUR_USER\Desktop\HARBOR\harbor_tasks\json-dept-salary-totals"
cd C:\path\to\harbor-repo
uv run harbor run --agent oracle --path $TASK_PATH --job-name oracle-json-dept

---

References

• laude-institute/harbor — framework and CLI
• Task format and concepts: see Harbor’s docs in that repository (e.g. docs/ and examples/tasks/)

---

License

Align with the license of the project you bundle or contribute this task with (for example, Harbor’s license when opening a PR upstream). This repo only adds task files and documentation.