Agent Workspace
Agent Workspace centralizes interaction and run state in OriginRouter while the background service manages Codex or Claude sessions. After CLI quickstart, daily operation starts by entering the Workspace and submitting objectives; one long-lived Workspace Session can contain multiple independent Runs.
Use the Workspace for objectives that require coordination, task assignment, review, or cross-device follow-up. When the native Claude Code or Codex terminal is sufficient, return to Quickstart and use the corresponding entry. See Devices and security and Data and execution boundaries for cross-device control and data handling.
Open and submit an objective
originrouter
Without an objective, this opens the interactive workspace; with an objective, it creates the next Run:
originrouter "Investigate the performance regression and fix the most likely cause"
Session and Run
- A Session preserves long-lived conversation and team continuity.
- A Run represents one objective with its own plan, budget, approvals, and result.
- Completing one Run does not automatically end the Session.
/newcreates a new Session instead of continuing the current context.
Resume
Use /resume in the interactive workspace to select a recent Session, or /resume <session-id> to choose one directly. A Run ID is for inspection, retry, and audit; it is not a history branch continuation point.
Collaboration modes
| Mode | Use it for |
|---|---|
auto | Select the smallest effective team automatically |
solo | Questions and tightly scoped tasks |
build-review | Independent review after implementation |
plan-build-verify | Large or production-sensitive changes |
parallel-research | Independent investigation across several directions |
review-panel | Compare multiple architecture proposals |
remote-ops | Include another trusted device in execution |
Use --mode at launch, or /mode <name> inside the workspace.
Startup confirmation and action approval
The default policy is required: when a new Session needs a generated Team or task plan, Workspace shows the configuration and plan and waits for you to confirm or request changes before execution begins. auto selects a team mode; it does not skip these confirmations.
--yes is the explicit automated-start option (the internal always_auto policy) and accepts the Team/configuration and plan gates. --review remains a compatible spelling for the default behavior, and cannot be combined with --yes.
This is separate from in-session action approval. /approval, Shift+Tab, and Manual, Guarded, AI Review, Rules, and Full policies determine whether individual Agent actions require approval after a Run begins; --yes never bypasses those approvals.