Task Parameters

Task functions (like aider_polyglot(), terminal_bench_2(), etc.) accept the following parameters:

Parameter	Description	Default	Python Example	CLI Example
`ref`	Harbor ref (digest, revision number, tag, or `latest`)	`"latest"`	`"sha256:01e28d85..."`	`sha256:01e28d85...`
`dataset_task_names`	List of task names to include (supports glob patterns)	`None`	`["aime_60", "aime_61"]`	`'["aime_60"]'`
`dataset_exclude_task_names`	List of task names to exclude (supports glob patterns)	`None`	`["aime_60"]`	`'["aime_60"]'`
`n_tasks`	Maximum number of tasks to run	`None`	`10`	`10`
`overwrite_cache`	Force re-download and overwrite cached tasks	`False`	`True`	`true`
`sandbox_env_name`	Sandbox environment name	`"docker"`	`"modal"`	`"modal"`
`override_cpus`	Override the number of CPUs	`None`	`4`	`4`
`override_memory_mb`	Override the memory (in MB)	`None`	`16384`	`16384`
`override_gpus`	Override the number of GPUs	`None`	`1`	`1`

Resource limits: When task.toml omits a resource field, inspect_harbor imposes no limit on it (no CPU cap, no GPU reservation, and — subject to the memory floor below — no memory cap), matching Harbor ≥0.17’s own behavior. Explicit task.toml values are honored as-is, and override_cpus/override_memory_mb/override_gpus take precedence over both. The one exception is memory: inspect_harbor enforces a 6 GB floor so an explicitly low memory_mb can’t starve agent scaffolds — an explicit value below 6 GB is raised to 6 GB (an omitted value stays unlimited, which trivially satisfies the floor). Use override_memory_mb to set a value below the floor.

Multi-service compose & DinD providers: Resource overrides are applied only to the default service (selected by x-default: true, or a service named “default”/“main”, or the first service). Sidecar services run without explicit resource limits, within the sandbox’s total capacity. For DinD-based sandbox providers (e.g. Daytona) that aggregate per-service resources to size the VM, you can control sandbox-level resources directly via the provider’s compose extension (e.g. x-daytona: { resources: { cpu: 4, memory: 8 } }) in your docker-compose.yaml. See the Daytona sandbox provider docs for details.

Example

Here’s an example showing how to use multiple parameters together:

CLI:

inspect eval inspect_harbor/terminal_bench_2 \
  -T n_tasks=5 \
  -T overwrite_cache=true \
  -T override_memory_mb=8192 \
  --model anthropic/claude-sonnet-4-5

Python API:

from inspect_ai import eval
from inspect_harbor import terminal_bench_2

eval(
    terminal_bench_2(
        n_tasks=5,
        overwrite_cache=True,
        override_memory_mb=8192,
    ),
    model="anthropic/claude-sonnet-4-5"
)

This example:

Limits to 5 tasks using n_tasks.
Forces a fresh download with overwrite_cache.
Allocates 8GB of memory.