Creating a CORAL task

Creating a CORAL task

A CORAL task is three things that must line up. Scaffold them with coral init, then iterate edit → coral validate until the grader scores the seed.

my-task/
├── task.yaml      # config: name, description, grader entrypoint, agent count
├── seed/          # starter code agents see at t=0 (this is workspace.repo_path)
│   └── solution.py
└── grader/        # standalone Python package — gets its own isolated venv
    ├── pyproject.toml
    └── src/my_task_grader/
        ├── __init__.py
        └── grader.py     # class Grader(TaskGrader): ...

The packaged grader is the only supported form — it gives the grader an isolated venv and bundles everything the eval needs (grader code, helpers, hidden answer keys). There is no eval/grader.py auto-discovery anymore.

Optimizing code the user already has? Scaffold inside a .coral_workspace/ at the root of their project (gitignored), and copy the code to optimize into seed/ — keeps CORAL's task/results out of their source tree. The coral-quickstart skill has the end-to-end .coral_workspace/ flow; this skill covers the grader you'll write once the code is in seed/.

"Optimize this" is a build instruction, not a question — never answer it with a process menu. A 1/2/3 like "point me to a task / create one / optimize outside coral" is the failure mode; do not produce it. The absence of a task.yaml is not ambiguity — it just means you build one from the current repo. Concretely: (1) dig for what's already measurable — a research/framework repo almost always ships an eval/benchmark script, a test suite, or a metric in its README/paper; that's your target and metric. (2) If no single number is obvious, construct one by wrapping the repo's existing evaluation — don't conclude "no measurable objective" just because there's no CORAL scaffold. (3) Scaffold the most plausible target and start building (a .coral_workspace/ + draft grader is cheap and reversible); state your assumption in one line and proceed. (4) Only as a last resort, if you've actually read the repo and it exposes nothing scorable, propose 2-3 concrete optimization targets you found (each with its metric), pick the most likely, and scaffold that — still not a process menu.

The loop

coral init my-task        # scaffold all three pieces (a runnable end-to-end example)
cd my-task
# ... edit the three pieces for your problem ...
coral validate .          # bootstraps the grader venv, runs the grader on seed/, prints a score
# repeat edit → validate until the seed scores as you expect

coral validate succeeding is the one checkpoint that matters — it proves the grader can score the seed. Most "agents are stuck, every eval fails" reports trace to a grader that crashes on the seed, which validate would have caught. Always start from coral init rather than hand-writing the layout; the generated files are the canonical minimal example.

The three pieces

1. The seed (seed/) — what the agent checks out at t=0 and what the grader later scores. The contract between seed and grader is the program file: a file (e.g. solution.py) with a function or stdout convention the grader invokes, named in grader.args.program_file. Put a real, runnable baseline here — agents should coral eval immediately and get a non-zero score to beat. A skeleton that crashes is a bad baseline. Runtime data goes under seed/data/ and is read by relative path.

2. The grader (grader/) — subclass TaskGrader, implement evaluate(), return a number (or ScoreBundle). The minimum: