[Feature]: Add Plan First, auto-review, and delegation controls to the CLI

[bobo-xxx]

Preflight checklist

• I searched existing issues and discussions and this specific proposal has not been submitted. Issue Headless CLI: remaining headless CLI capabilities (attachments, per-run model/backend selection, npm publication) #363 covers attachments and per-run model/backend selection, but not Plan First, auto-review, or delegation controls.

What problem does this solve?

The headless CLI can create or resume a Session and submit a prompt, but it cannot select several execution controls available in the Open Science interface:

• Start a turn using Plan First
• Enable or disable auto-review
• Bind a Specialist when creating a Session
• Control whether the main agent may create subagents

For example, the following CLI run cannot currently request Plan First or auto-review:

open-science run \
  --project "Analysis" \
  --prompt-file task.md \
  --wait

Users must open the interface and configure these options manually. This limits unattended, scripted, and benchmark-driven workflows and can cause CLI-created Sessions to behave differently from Sessions created interactively.

--approval-profile auto does not solve this problem. It controls tool permissions and is unrelated to automatic review.

Proposed solution

Expose stable Session and turn controls through open-science run.

Suggested interface:

open-science run \
  --project "Analysis" \
  --prompt-file task.md \
  --plan-first \
  --auto-review \
  --specialist <id-or-name> \
  --subagents allow \
  --wait \
  --jsonl

Plan First

Add:

--plan-first

The submitted prompt should use the existing Plan First turn intent.

When the generated Plan requires approval:

• The run should enter an explicit awaiting_plan_approval state.
• --json and --jsonl should expose the Session ID, Plan identity/version, and available actions.
• Waiting for approval should not be treated as run failure or completion.
• A parked Plan should not expire merely because the CLI process exits or remains idle.

Provide commands or equivalent API operations for completing the lifecycle, for example:

open-science plan show <session-id>
open-science plan approve <session-id>
open-science plan revise <session-id> --feedback "..."
open-science plan discard <session-id>

The exact command names can follow the project's preferred CLI structure.

Auto-review

Add:

--auto-review
--no-auto-review

Auto-review for CLI runs must be owned by the daemon/runtime and must not depend on an Open Science renderer or browser being open.

For a new Session, the explicit flag should set its initial auto-review state. For an existing Session:

• An explicit flag should update the Session setting.
• Omitting both flags should preserve the Session's saved setting.

The existing configured reviewer model and review policy can remain the default. Reviewer-model selection can be added separately if it requires another stable public contract.

Review progress and results should be included in --jsonl events and reflected in the final run result.

Specialist selection

Add:

--specialist <id-or-name>

This should bind a Specialist when creating a new Session using the existing internal Specialist identity.

Because the Specialist is a Session-level binding, using --specialist with an incompatible existing --session should return a clear validation error rather than silently switching or ignoring it.

Subagent policy

Add an explicit delegation policy, for example:

--subagents allow
--subagents deny

This would control whether the main agent may use Open Science's existing delegated-work capability. It should not require users to predefine individual child tasks; the agent may still decide when and how to delegate.

For automation:

• Delegated-work lifecycle events should be available through --jsonl.
• The run result should distinguish active, completed, failed, and cancelled delegated work.
• Cancelling a run should use the documented Open Science cancellation policy for its child work.
• Permission requests and questions raised by delegated agents should appear as explicit waiting states rather than an unexplained timeout.

General behavior

• CLI help and documentation should distinguish permission profiles, auto-review, Plan First, Specialists, and subagents.
• Unsupported combinations should fail with exit code 2 and a clear usage message.
• Structured outputs should remain stable and machine-readable.
• Behavior should be covered on Windows, macOS, and Linux.

Alternatives considered

Configure everything through the interface first

This prevents a workflow from being fully reproducible from a command or automation script.

Encode the behavior in prompt text

Instructions such as "create a plan first" or "use subagents" are not equivalent to application-owned Plan First, review, Specialist, and delegation controls.

Use --approval-profile auto

Permission approval and auto-review are separate features. Changing the permission profile does not enable review or Plan First.

Expose all interactive settings immediately

This issue is limited to Plan First, auto-review, Specialist selection, and delegation policy. Attachments and per-run model/backend selection are already tracked in #363.

Additional context

Current Open Science internals already contain much of the required runtime behavior:

• Plan First is represented by the existing plan-first turn intent.
• Session creation accepts a Specialist identity internally.
• Open Science supports delegated work through host.delegate.
• Auto-review already exists as a Session setting, although its current trigger path is associated with the interactive workspace.

The main missing piece is a stable CLI and Task API contract that exposes these capabilities without requiring the interface to remain open.