Submitting requests (Portal)

The Portal is the primary way to interact with Scope when you’re exploring, debugging, or designing benchmarks. This guide covers the full submit-and-inspect workflow.

For automation, see Submitting requests (REST API). For terminal workflows, see Submitting requests (CLI).

What you submit

You submit a request. A request bundles:

1. A task prompt — what to ask the agent.
2. A criteria set — how to judge the result.
3. A profile (or inline runtime configuration) — which agent, model, and tools to use.

When Scope picks up your request, it creates a run (one execution attempt). A retry creates a new run on the same request.

Open the submit form

From the Portal navigation, click New Run (the form name in the Portal — it submits a request). The form is split into the three sections above.

1. Write the task prompt

Type the prompt text into the Task field. You can also pick from recently used prompts in the suggestions list — those are entries from the shared Tasks catalog.

You don’t reference a task prompt by ID; you write the text. Behind the scenes Scope looks up (or creates) the matching task prompt record in the catalog so the same text across requests links to the same record. See Managing task prompts.

2. Pick the criteria

Pick a criteria set from the picker. Criteria sets are reusable and referenced by ID — you can use the same set across many task prompts.

If you need a new set, create it on the Criteria page first. See Defining evaluation criteria.

3. Choose how to run it

You can either select a saved profile or configure the request inline.

Use a profile

Pick a profile from the dropdown. The Portal locks the configuration fields (worker, model, agent version, MCP servers, skills, extensions) to the profile’s latest version. To use an older version explicitly, expand the version selector beneath the profile picker.

Selecting a profile is the only way to guarantee that another user re-running this same combination later will get the exact same configuration — see Defining profiles.

Configure inline

If you don’t pick a profile, set the fields manually:

• Worker — GitHub Copilot CLI, Claude Code CLI, or VS Code Copilot. See Choosing a coding agent.
• Model — pick from the available models for that worker.
• Agent version (optional) — pin a specific agent build.
• MCP servers, skills, extensions (optional) — see Using MCP servers, skills & extensions.

After you’ve configured a request inline, you can save the setup as a profile to capture it for re-use.

4. Submit

Click Submit. The Portal navigates to the request detail page. The request is assigned a unique ID and queued; a run is created when a worker picks it up.

Watch a run

The request detail page has these tabs:

• Logs — the agent’s actions stream here in real time as the worker executes the run.
• Configuration — the exact profile version (or inline configuration), task prompt, and criteria used. This is what makes the run reproducible.
• Report (visible after the run completes) — per-criterion pass/fail with the judge’s rationale.

The status badge at the top tracks the run’s lifecycle: pending → running → completed (or failed).

Re-prioritize, pause, or resume

Pending and running requests can be reordered or paused. See Prioritizing & pausing requests.

Re-run

From a request’s detail page, click Re-run to create another run on the same request — same task prompt, same criteria, same profile. This is useful for sampling variability of an agent on the same prompt.

Tips

• Reproducibility — always pick a profile when running for the record. Inline configurations work for exploration but make later comparisons harder.
• Reuse task prompts and criteria. They’re the building blocks the rest of Scope (prompt features, reports, MDP analyses) is keyed on.
• Hidden cost — requests that pull skills from GitHub or call MCP servers may take longer to start while those resources are resolved.

Next steps

• Automate this workflow with the REST API or the CLI.
• Save your favorite setup as a profile.