Modes

What are modes?

A mode is an enum (chat, plan, review, orchestrate). Each one is declared once, in a capability profile, which every enforcement layer projects from:

Tool surface: which tools exist at all in that mode
Required output verb: what the run must call to have delivered anything
Denied effects: which action classes the guard refuses
Sandbox posture: whether bind-mounted worktrees are writable
System prompt: a generated capability block (derived from the actual tool list, so a prompt can never name a tool the run does not have) plus hand-written guidance

Modes are set per-channel and can be changed during a channel.

Chat mode

The default mode. The agent has full access to its tools and can read and write files in its isolated worktree.

Use chat mode for:

General coding tasks
Bug fixes and feature implementation
Asking questions about the codebase
Ad-hoc requests

The system prompt for chat mode is general-purpose: it instructs the agent to be helpful, precise, and to produce working code.

Plan mode

Plan mode restricts the agent to analysis and planning. The agent cannot write files: it reads the codebase, analyzes the problem, and produces a structured plan.

Use plan mode when you want to:

Get a detailed implementation plan before committing to code
Understand the scope of a change
Have the agent break down a complex task into steps

The plan is delivered by a single submit_plan call carrying a typed dependency graph. It renders as a compact row in the conversation (goal, status, step count) and opens Plan studio in a tab beside it for review. Approving it compiles and runs the work deterministically, and takes the conversation out of plan mode.

A plan-mode run that ends without calling submit_plan is recorded as a failed run, not a silent success.

Plan mode is useful for expensive or risky changes where you want to review the approach before the agent writes any code.

Review mode

Review mode is designed for PR review. The agent receives the PR’s diff as context and produces structured findings.

In review mode:

The agent’s system prompt is tailored for code review
It produces review nodes, structured findings with:
- Finding kind (bug, security, performance, style, suggestion)
- Priority (P0 critical → P3 nit)
- File and line anchor
- Original code and suggested fix
- Confidence score
It produces a review verdict, an aggregate assessment:
- Ship: no significant issues
- Hold: issues that should be addressed before merge
- Block: critical issues that must be fixed

Review mode also gates the MCP tool allowlist to review-relevant tools.

Orchestrate mode

Orchestrate mode is for turning one goal into a whole-team plan. The agent (usually an orchestrator such as the CEO) can read and research — like plan mode, it cannot write files — plus a single privileged verb, propose_orchestration, which emits a structured plan (roles, sub-tickets, a synthesis step, and a budget).

Nothing in the plan executes until you approve it. On approval, a deterministic materializer hires the agents, creates the tickets and a project, and runs a generated pipeline. See Orchestration.

Mode resolution

The ModeResolver resolves the mode for a given channel:

If a channel has an explicit mode set, use that
If the channel was started by a pipeline step with a mode, use that
Default to chat

The mode is also serialized into the wake context so the dispatch service knows which prompt template to use.

Changing modes

You can change the mode of a channel at any time. This affects subsequent messages: the agent will use the new mode’s prompt and policies for its next response.

Some contexts set the mode automatically:

Opening a PR review channel sets mode to review
Starting a planning session sets mode to plan
Regular channels default to chat

Agent dispatch lifecycle: how modes affect prompt assembly
Sandbox and security: how modes gate sandbox writes
Use plan mode: practical guide to planning