Choreography

Choreography

Choreography is BMO’s state-machine layer for session behavior. Use it when you need runtime state, tool gates, prompt shaping, model overrides, or hooks to change together as a session moves between planning, review, execution, or custom phases.

Quick start

The examples below use JSON because the nested DSL is easier to scan that way, but bmo.toml is the canonical config format for BMO.

{
  "choreography": "plan-execute"
}

The agent starts in the plan state (read-only, mutation tools disabled) and advances to execute once you approve the plan.

Visual: state, triggers, and applied behavior

Choreography is easiest to read as a state machine. Each state can apply prompt text, tool gates, model overrides, hooks, and turn limits. Transitions are triggered by approval, slash commands, or other configured events.

stateDiagram-v2 [*] --> Plan: initial Plan: prompt guard Plan: mutation tools denied Plan --> Execute: approval Execute: implementation prompt Execute: write/edit/bash allowed by policy Execute --> Review: command:mode(review) Review: evidence-first output Review --> Plan: command:plan Execute --> Debug: failure hook Debug: retry and inspect failure Debug --> Execute: retry approved Execute --> [*]: task complete

This diagram is conceptual: your exact graph comes from the selected preset and any overrides in merged config.

---

Config shapes

Choose one of three config shapes:

1. Preset name (string)

{
  "choreography": "modal"
}

2. Preset with overrides

Starts from a preset and merges your changes on top:

{
  "choreography": {
    "preset": "plan-execute",
    "override": {
      "states": {
        "execute": {
          "max_turns": 20
        }
      }
    }
  }
}

3. Full inline spec

Define the complete state machine yourself:

{
  "choreography": {
    "initial": "research",
    "directives": {
      "read_only": {
        "prompt": "Read and analyse only. Do not modify files.",
        "tools": { "deny": ["write", "edit", "multiedit", "bash_mutation"] }
      }
    },
    "states": {
      "research": {
        "use": ["read_only"],
        "transitions": [
          { "to": "implement", "trigger": "approval", "label": "approve plan" }
        ]
      },
      "implement": {
        "transitions": [
          { "to": "research", "trigger": "command:plan", "label": "return to planning" }
        ]
      }
    }
  }
}

---

Built-in presets

Preset	Initial state	Description
none	default	Passthrough — no gating or prompt injection
plan-execute	plan	Read-only planning → full execution on approval
plan-auto	plan	Like plan-execute but auto-advances after the first inference turn
modal	coordinator	Seven modes: code, chat, coordinator, debug, review, plan, sre; switch with /mode <name>
plan-modal	plan	Three modes: plan, review, execute; approval gates the execute transition

---

Directives

A Directive is the bundle of behavior that applies when the agent is in a given state.

Field	Type	Description
prompt	string	Text prepended to the system prompt while in this state
tools	ToolGate	Controls which tools are available (see below)
model	string	Override the default model for this state
hooks	Hook[]	Post-execution hooks (reflection, retry, command)
max_turns	int	Maximum inference turns allowed in this state (0 = unlimited)

Directives can be defined once in the top-level directives map and referenced from any state via use:

"directives": {
  "strict_read_only": {
    "prompt": "Do not modify any files.",
    "tools": { "deny": ["write", "edit", "multiedit", "bash_mutation"] }
  }
},
"states": {
  "plan": { "use": ["strict_read_only"] }
}

A state can reference multiple directives — they are merged left-to-right. Inline fields on the state are merged last and take highest priority.

---

Tool gate

The tools field on a directive controls tool availability:

"tools": {
  "allow": ["view", "glob", "grep"],
  "deny": ["write", "edit", "multiedit", "bash_mutation"],
  "require_approval": ["bash"]
}

Field	Behaviour
allow	Only listed tools are available (allowlist mode)
deny	Listed tools are blocked entirely
require_approval	Listed tools show a permission prompt before running

If both allow and deny are specified, deny takes precedence within the allow set. Selectors may be exact tool names or category selectors such as category:write and category:command. When a category selector triggers approval, the permission prompt shows the matched selector in its Gate row.

Blocked tools are real boundaries. If a capability is denied in the current state, the recovery path is to transition to the smallest suitable state and retry, not to route around the gate with a broader tool. For the agent, that recovery path is the session_mode tool; /mode is the operator-facing TUI alias.

Tool names match the built-in tool names documented in Tools Reference.

---

States

Each entry in states is a node in the state machine:

"states": {
  "plan": {
    "use": ["no_mutations"],
    "prompt": "Extra note appended after directive prompts.",
    "max_turns": 5,
    "transitions": [
      { "to": "execute", "trigger": "approval", "label": "approve plan" },
      { "to": "review",  "trigger": "command:mode", "label": "review" }
    ]
  }
}

Field	Description
use	Names of shared directives to merge in (in order)
prompt	Additional prompt prefix (merged after use directives)
tools	Inline tool gate (merged after use directives)
model	Inline model override
hooks	Inline hooks
max_turns	Inline turn limit
transitions	Outgoing edges (see below)

---

Transitions

A transition moves the engine from the current state to a target state when its trigger fires.

{ "to": "execute", "trigger": "approval", "label": "approve plan" }

Field	Description
to	Target state name
trigger	Trigger that fires this edge (see trigger reference below)
label	Human-readable description shown in the TUI and /choreography output

Trigger reference

Trigger	Fires when
approval	User approves an agent action in the permission dialog
command:plan	User runs /plan or the TUI “Return to Plan” action
command:mode	User runs /mode <state> where the target label matches a state name
turn:end	An inference turn completes (used for auto-advance)
error	A tool execution returns an unrecoverable error

---

Hooks

Hooks run synchronously after each tool execution in a state.

"hooks": [
  { "type": "reflection",  "on_failure": "warn" },
  { "type": "retry",       "max_retries": 3, "on_failure": "abort" },
  { "type": "command",     "command": "go test ./...", "on_failure": "reinject" }
]

Hook type	Description
reflection	Triggers the reflection runner to evaluate the tool result
retry	Retries the failed operation up to max_retries times
command	Runs a shell command; result is injected back as context

on_failure values

Value	Behaviour
reinject	Feed the failure output back to the agent as context
abort	Stop the current turn and surface the error to the user
warn	Log the failure and continue

---

Examples

Read-only research then implement

{
  "choreography": {
    "initial": "research",
    "directives": {
      "read_only": {
        "prompt": "Analyse the codebase only. Do not make changes.",
        "tools": { "deny": ["write", "edit", "multiedit", "bash_mutation"] }
      }
    },
    "states": {
      "research": {
        "use": ["read_only"],
        "max_turns": 10,
        "transitions": [
          { "to": "implement", "trigger": "approval", "label": "start implementation" }
        ]
      },
      "implement": {
        "transitions": [
          { "to": "research", "trigger": "command:plan", "label": "back to research" }
        ]
      }
    }
  }
}

Plan-execute with a post-tool test hook

{
  "choreography": {
    "preset": "plan-execute",
    "override": {
      "states": {
        "execute": {
          "hooks": [
            { "type": "command", "command": "go test ./...", "on_failure": "reinject" }
          ]
        }
      }
    }
  }
}

Modal with a custom model for review

{
  "choreography": {
    "preset": "modal",
    "override": {
      "states": {
        "review": {
          "model": "anthropic/claude-opus-4"
        }
      }
    }
  }
}

---

Slash commands

When choreography is active, two commands are available in the TUI:

Command	Action
/plan	Fire the command:plan trigger — returns to the plan state if a transition exists
/mode <name>	Fire the command:mode trigger toward <name> — switches to that state if a transition exists

The current state is visible in the TUI header. The mode picker only shows states reachable from command:mode transitions, plus the current state. For new sessions, that list is derived from the choreography’s initial state. Use /choreography (when available) to inspect available transitions.

---

State persistence

Choreography state is persisted to the session SQLite database. If the process restarts mid-session the agent resumes in the same state it was in when it stopped.

---

Relationship to session modes

Session modes (/mode code, /mode chat, etc.) are the user-facing surface of the modal preset. When modal or plan-modal is active, /mode <name> fires the command:mode trigger and transitions to that state. Custom choreographies can expose a narrower mode surface, and the UI follows that graph instead of always advertising the built-in modal states.

Native runs honor choreography prompt, tool-gate, hook, and model overrides. Optional agent_sdk_go and adk_go backends receive the same prepared prompt and session-mode context, but model selection remains backend-owned until those runtimes expose per-run override hooks.