coral-new-task

name: coral-new-task description: End-to-end recipe for adding a new task under examples/ — the three pieces that have to line up (task.yaml, seed/, and grader/), what to put in each, the TaskGrader API surface, the coral validate → smoke-test loop, and the common mistakes (repo_path pointing at the wrong dir, score direction backwards, hidden answer keys leaking into seed/, grader writing to codebase_path which the daemon force-removes, private-vs-public confusion, missing run() signature). Use whenever the user wants to add a new CORAL task or port an existing benchmark into CORAL.

Creating a new CORAL task

A CORAL task is three things that must line up:

examples/<task>/
├── task.yaml      # config: name, description, grader entrypoint, agent count
├── seed/          # starter code agents see when they begin (the repo_path)
│   └── solution.py
└── grader/        # standalone Python package
    ├── pyproject.toml
    └── src/<task>_grader/
        ├── __init__.py
        └── grader.py     # class Grader(TaskGrader): ...

The packaged form is the only supported form. The package gives the grader its own venv and ships everything the eval needs — grader code, helper modules, and hidden data (see "Hidden data" below).

Reference implementations

Look at these before writing anything new — copy the closest one and edit:

1. The seed

Whatever lives in seed/ is what the agent sees on first checkout — it's the working directory the grader will later score. The contract between seed/ and the grader is the program file: a Python file with a function the grader imports and calls.

The convention across examples is:

• solution.py (or initial_program.py) defining a top-level run() function.
• The grader passes program_file: "solution.py" via grader.args.
• run()'s signature is whatever the grader expects — usually () -> result or (input_path) -> result.

Put a real, runnable baseline here. Agents should be able to coral eval immediately and get a non-zero score, so they have a starting point to improve. A no-op skeleton that crashes is not a good baseline.

If the task needs data files at runtime (training data, fixtures), put them under seed/data/ and reference them by relative path from solution.py. The grader will see them at <codebase_path>/data/....

2. The grader

Packaged grader — the recommended path

grader/
├── pyproject.toml
└── src/<task>_grader/
    ├── __init__.py
    └── grader.py

pyproject.toml is a thin Hatchling package. Crib from examples/erdos/grader/pyproject.toml:

[project]
name = "<task>-grader"
version = "0.1.0"
description = "CORAL grader for the <task> task."
requires-python = ">=3.11"
dependencies = ["coral", "numpy"]   # Whatever the grader actually imports.

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["src/<task>_grader"]

Subclass TaskGrader and implement evaluate():

# grader/src/<task>_grader/grader.py
from coral.grader import TaskGrader
from coral.types import ScoreBundle


class Grader(TaskGrader):
    def evaluate(self) -> float | ScoreBundle:
        program_file = self.args.get("program_file", "solution.py")
        # self.codebase_path  — the agent's commit checked out detached
        # self.private_dir    — .coral/private/ (your hidden answer keys live here)
        # self.args           — dict from task.yaml grader.args
        # self.timeout        — grader.timeout in seconds (or None)
        # self.eval_logs_dir  — write subprocess logs / artifacts the agent should see post-grade

        try:
            result = run_program_and_score(...)
        except TimeoutError:
            return self.fail(f"Evaluation timed out after {self.timeout}s")
        except Exception as e:
            return self.fail(f"Evaluation failed: {e}")

        return self.score(result, explanation=f"score={result:.4f}")

What you have available on self:

Bundling data files with the grader

If the grader needs reference files (model weights, ground-truth answers, scoring fixtures), ship them inside the package and load via importlib.resources:

import importlib.resources
scorer_dir = str(importlib.resources.files("<task>_grader.scorers"))

examples/dna_design/grader/src/dna_design_grader/grader.py is the canonical pattern — note the scorers/ subpackage. Add the directory to [tool.hatch.build.targets.wheel] if it has non-Python files.

Heavy / optional dependencies

If the grader wants torch, grelu, etc., put them in optional-dependencies and have the grader fall back gracefully when missing — see examples/dna_design/grader/pyproject.toml. Then grader.setup becomes ["uv pip install -e ./grader[ml]"] for the full version.

Hidden data: ship it inside the package

Answer keys, test fixtures, and helper modules live inside the grader package — the convention is a taskdata/ subdir next to grader.py, resolved with _TASKDATA = Path(__file__).parent / "taskdata" (works for editable and wheel installs alike; hatchling includes non-Python files under packages automatically). See examples/mnist/grader/src/mnist_grader/taskdata/ for a data file and examples/ADRS/txn_scheduling/ for helper modules put on sys.path.

For files that genuinely can't live in the package (huge datasets, shared across tasks), list them under grader.private in task.yaml — they get copied to .coral/private/<name> and the grader reads them via self.private_dir.

3. The task.yaml

Fields that must be set; everything else has a sensible default.

task:
  name: "My Task"                       # shows up in results/<slug>/
  description: |                        # rendered into CORAL.md, agents read this
    What the agent should do.
    Reference the program file by name (e.g. solution.py and its run() signature).
  tips: |                               # optional, also rendered into CORAL.md
    - Eval timeout is N seconds.
    - Constraints / scoring details / known baselines.

grader:
  entrypoint: "<task>_grader.grader:Grader"   # required
  setup:
    - "uv pip install -e ./grader"            # runs once in .coral/private/grader_venv/
  timeout: 600                          # seconds; 0 disables, default 300
  direction: maximize                   # or minimize — controls leaderboard ordering
  args:                                 # arbitrary dict, read inside grader as self.args
    program_file: "solution.py"
  private: []                           # extra files copied into .coral/private/ (hidden from agents)
  parallel:
    max_workers: 1                      # bump only when the grader is concurrency-safe
  max_pending_per_agent: 1              # cap on in-flight submissions per agent

agents:
  count: 1                              # raise once the task is known to be stable
  runtime: claude_code                  # claude_code | codex | cursor_agent | kiro | opencode
  model: sonnet                         # default depends on runtime; see coral/agent/registry.py

workspace:
  results_dir: "./results"              # where each run lands
  repo_path: "./examples/<task>/seed"   # MUST point at the seed/ dir
  setup:                                # runs in each agent worktree before agents start
    - "uv pip install numpy"            # task-runtime deps go here, NOT in grader/setup

run:
  verbose: false
  ui: false
  session: tmux                         # local | tmux | docker

The examples/README.md documents the full schema with every default. When in doubt, look there before adding fields.

4. Validate before running agents

coral validate examples/<task>

This:

1. Parses task.yaml and reports schema errors.
2. Bootstraps .coral/private/grader_venv/ and runs grader.setup.
3. Copies seed/ into a tempdir and runs the grader against it once.
4. Prints the resulting score and explanation.

If coral validate succeeds, the grader can score the seed. That's the single most important checkpoint — most "agent stuck" issues trace back to a grader that crashes on the seed.

After validation, smoke-test with one agent:

coral start -c examples/<task>/task.yaml agents.count=1 run.session=local
# Wait for one eval, then:
coral stop

5. Wire it into the index

Add a one-line entry to the table in examples/README.md and a short ### <task> section under "Details" with the bullet points the others use (Agents / Timeout / Session). Skip this only for throwaway local tasks.

Common mistakes