Template Reference

Templates are TOML files that define everything about a worker — its identity, model, capabilities, compute mode, and evaluation criteria. The template is the source of truth.

Full example

template_version = "1.0"

[worker]
name = "my-worker"

[workstation]
model = "claude-sonnet-4-20250514"
task = "Analyze the codebase and generate a dependency report."
instructions = "Focus on circular dependencies and unused imports."

[workstation.settings]
max_tokens = 8192
temperature = 0.7

[workstation.capabilities]
mcps = ["filesystem", "github"]
skills = ["code-analysis"]

[compute]
mode = "sandbox"
timeout = "2h"

[eval]
suite = "analysis-quality"

Sections

`template_version`

Required. Currently "1.0".

template_version = "1.0"

`[worker]`

Worker identity.

Field	Type	Required	Description
`name`	string	Yes	Unique name for this worker type

[worker]
name = "code-reviewer"

`[workstation]`

Model configuration and task definition.

Field	Type	Required	Description
`model`	string	Yes	Model identifier (e.g., `claude-sonnet-4-20250514`)
`task`	string	No	One-shot task description. Worker completes and exits
`instructions`	string	No	Ongoing instructions for daemon workers

Use task for one-shot workers that complete a single job. Use instructions for long-running daemon workers that process multiple episodes.

[workstation]
model = "claude-sonnet-4-20250514"
task = "Review the latest PR for security issues."

`[workstation.settings]`

Model parameters.

Field	Type	Default	Description
`max_tokens`	integer	Model default	Maximum tokens per response
`temperature`	float	Model default	Sampling temperature (0.0 - 1.0)

[workstation.settings]
max_tokens = 4096
temperature = 0.3

`[workstation.capabilities]`

Tools and skills available to the worker.

Field	Type	Default	Description
`mcps`	string[]	`[]`	MCP server names (e.g., `filesystem`, `github`, `slack`)
`skills`	string[]	`[]`	Skill names (e.g., `code-review`, `testing`)

Capabilities drive credential inference. When you list mcps = ["github"], arpi’s registry resolves what credentials are needed (GitHub token), provisions them via Infisical, and injects them into the worker’s sandbox.

[workstation.capabilities]
mcps = ["filesystem", "github"]
skills = ["code-review"]

`[compute]`

Compute environment configuration.

Field	Type	Default	Description
`mode`	string	`"bare"`	`"sandbox"` (isolated container) or `"bare"` (host process)
`timeout`	string	`"1h"`	Maximum worker lifetime (e.g., `"30m"`, `"2h"`, `"24h"`)

Sandbox mode provides kernel-level isolation and is required for credential injection. Bare mode runs without isolation — suitable for trusted, low-risk tasks.

[compute]
mode = "sandbox"
timeout = "2h"

`[eval]`

Evaluation configuration.

Field	Type	Required	Description
`suite`	string	No	Name of the eval suite to run against the worker’s output

Eval suites score worker output using three grader types:

Code graders — Deterministic checks (regex, assertions, test execution)
Model graders — LLM-as-judge scoring (rubric-based evaluation)
Human graders — Manual review queue for subjective quality assessment

[eval]
suite = "code-review-quality"