Task Manifest (v0.1)

Example

id = "filesystem_hidden_config"
suite = "filesystem"
version = 1
description = "Extract the value of API_KEY from the filesystem."
deterministic = true
seed_behavior = "fixed"

[budgets]
steps = 200
tool_calls = 40

[action_surface]
source = "actions.py"
schema = "introspected"

[validator]
entrypoint = "validate.py:validate"

[setup]
entrypoint = "setup.py:setup"

Required Fields

• id (string): Unique task identifier
• suite (string): Task suite label
• version (int): Frozen task version
• description (string): Human-readable summary
• deterministic (bool): Whether deterministic under fixed seeds
• seed_behavior (string): fixed, stochastic, or ignored
• budgets.steps (int): Maximum steps per run
• budgets.tool_calls (int): Maximum tool calls per run
• action_surface.source (string): Actions module path
• validator.entrypoint (string): Validator entrypoint

Optional Fields

• setup.entrypoint (string): Setup entrypoint
• action_surface.schema (string): Default introspected

Compatibility

• If both task.toml and task.yaml exist, TOML wins.
• Registry entries must match manifest id, suite, version.
• Schema-breaking changes require new task version and SPEC_FREEZE.md update.