diff --git a/AGENTS.md b/AGENTS.md
index ce1e4dc..9181129 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -10,8 +10,8 @@
 ## Key Files
 - `/shared/utils/communicationsEngine.ts` — Core state machine composable (used by `/pm` live ATC). Drives local cursor and TTS; Python backend owns the authoritative state.
 - `/app/composables/useRadioBackend.ts` — Typed wrapper around the Python backend REST API (`createSession`, `transmit`, `deleteSession`, `fetchFlows`)
-- `/server/utils/openai.ts` — Legacy LLM decision router (`routeDecision()`). No longer called by `/pm`; may still be used by other routes.
-- `/server/services/decisionFlowService.ts` — Builds runtime decision trees from MongoDB (used by Nuxt `/api/decision-flows/runtime`; `/pm` now fetches directly from the Python backend)
+- `/server/utils/openai.ts` — `getOpenAIClient()` only (TTS via `/api/atc/say`, STT via `/api/atc/ptt`). The old `routeDecision()` LLM router has been removed; routing lives in the Python backend.
+- `/server/services/decisionFlowService.ts` — Builds/edits MongoDB-backed decision trees for the flow editor (`/server/api/editor/flows*`). There is no Nuxt `/api/decision-flows/runtime` route — `/pm` fetches the runtime flow directly from the Python backend.
 - `/app/pages/pm.vue` — Live ATC page (speech-to-text, PTT, text input)
 - `/app/pages/classroom.vue` — Classroom learning mode (separate system, does NOT use communicationsEngine)
 
diff --git a/CLAUDE.md b/CLAUDE.md
index 233beab..e02e9c4 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -5,26 +5,59 @@
 - **H3 server** handlers in `/server`
 - **Shared types/utils** in `/shared`
 - MongoDB models in `/server/models`
+- **Python backend** — a separate repo, `OpenSquawk-LiveATC-api`, owns the authoritative
+  Live ATC (`/pm`) session state and routing decisions. It runs as its own service
+  (default `http://127.0.0.1:8000`). This Nuxt app owns STT, TTS, client audio, auth,
+  the flow editor, and account/admin features.
+
+> The LLM decision routing that used to live in this repo (`routeDecision()`,
+> `/api/llm/decide`, a Nuxt `/api/decision-flows/runtime` route) has been removed.
+> Decision routing now happens in the Python backend. Don't reintroduce it here.
 
 ## Key Files
-- `/shared/utils/communicationsEngine.ts` — Core state machine composable (used by `/pm` live ATC)
-- `/server/utils/openai.ts` — LLM decision router (`routeDecision()`)
-- `/server/services/decisionFlowService.ts` — Builds runtime decision trees from MongoDB
-- `/app/pages/pm.vue` — Live ATC page (speech-to-text, PTT, text input)
-- `/app/pages/classroom.vue` — Classroom learning mode (separate system, does NOT use communicationsEngine)
+- `/app/composables/useRadioBackend.ts` — Typed client for the Python backend REST API
+  (`createSession`, `transmit`, `deleteSession`, `fetchFlows`). This is how `/pm` talks
+  to the decision engine.
+- `/shared/utils/communicationsEngine.ts` — Local state-machine composable used by `/pm`.
+  It mirrors the flow for cursor tracking and TTS scheduling; the **Python backend owns the
+  authoritative state**. `fetchRuntimeTree()` loads the flow definition from the backend's
+  `/api/decision-flows/runtime`.
+- `/server/services/decisionFlowService.ts` — Builds/edits MongoDB-backed decision trees.
+  Used **only by the flow editor** (`/server/api/editor/flows*`), not by `/pm`.
+- `/server/utils/openai.ts` — `getOpenAIClient()` only. Used for TTS (`/api/atc/say`) and
+  STT (`/api/atc/ptt`). It is **not** a decision router anymore.
+- `/app/pages/pm.vue` — Live ATC page (speech-to-text, PTT, text input).
+- `/app/pages/classroom.vue` — Classroom learning mode (separate system, does NOT use
+  communicationsEngine).
 
 ## Live ATC Flow (/pm)
-1. User inputs (PTT or text) → `handlePilotTransmission()`
-2. `processPilotTransmission()` logs the pilot message
-3. `buildLLMContext()` builds candidates from `nextCandidates`
-4. POST `/api/llm/decide` → `routeDecision()` selects next state
-5. `applyLLMDecision()` moves to next state, updates vars/flags
-6. `collectAtcStatesUntilPilotTurn()` advances through ATC/system states
-7. Each ATC `say_tpl` is spoken via TTS (`scheduleControllerSpeech`)
+1. `startMonitoring()` → `fetchRuntimeTree('icao_atc_decision_tree', radioBackendUrl)`
+   loads the flow definition from the Python backend into the local engine (cursor / TTS).
+2. `startMonitoring()` → `radioBackend.createSession(...)` creates the authoritative
+   server-side session and stores `backendSessionId`.
+3. User input (PTT or text) → `handlePilotTransmission()`.
+   PTT audio is transcribed first via Nuxt `POST /api/atc/ptt` (Whisper).
+4. `radioBackend.transmit(backendSessionId, transcript)` → the Python backend runs regex
+   routing, readback evaluation, side effects, and flow chaining; it returns
+   `next_state_id`, `controller_say_template`, `auto_advanced_states`, `flags`,
+   and `session_complete`.
+5. `moveToSilent(stateId)` is called for each auto-advanced state and the final state —
+   advances the local cursor without re-triggering auto-transitions.
+6. `scheduleControllerSpeech(...)` speaks the ATC reply via TTS (`/api/atc/say`).
 
 ## Decision Tree States
-States have `role: 'pilot' | 'atc' | 'system'`. ATC states have `say_tpl` (what controller says). Pilot states have `utterance_tpl` (expected pilot response). Transitions: `next`, `ok_next`, `bad_next`, `timer_next`.
+States have `role: 'pilot' | 'atc' | 'system'`. The local engine handles two template
+schemas transparently (via `stateSayTpl()` / `stateUtteranceTpl()`):
+- Legacy (MongoDB editor): `say_tpl`, `utterance_tpl`, vars as `{var}`
+- Python backend YAML: `say_template`, `expected_pilot_template`, vars as `{{var}}`
+
+Both `{{var}}` and `{var}` are rendered by `renderTpl()`. Transition fields seen across
+schemas: `next`, `ok_next`, `bad_next`, `timer_next`, `auto_transitions`.
+
+## Environment Variables
+- `NUXT_PUBLIC_RADIO_BACKEND_URL` — URL of the Python backend (default `http://127.0.0.1:8000`)
 
 ## Commands
-- `bun run dev` — dev server
-- Decision trees are stored in MongoDB and fetched via `/api/decision-flows/runtime`
+- `yarn dev` — dev server (Nuxt)
+- `yarn test` — runs the `tsx --test` suite (`tests/`, `server/`, `shared/`)
+- Python backend: see `OpenSquawk-LiveATC-api/README.md`
diff --git a/shared/types/llm.ts b/shared/types/llm.ts
index 43f940c..e3f9b92 100644
--- a/shared/types/llm.ts
+++ b/shared/types/llm.ts
@@ -1,15 +1,5 @@
 import type { DecisionNodeCondition, DecisionNodeTrigger } from './decision'
 
-export interface LLMDecisionInput {
-    state_id: string
-    state: any
-    candidates: Array<{ id: string; state: any; flow?: string }>
-    variables: Record<string, any>
-    flags: Record<string, any>
-    pilot_utterance: string
-    flow_slug?: string
-}
-
 export type FlowActivationMode = 'main' | 'parallel' | 'linear'
 
 export interface FlowActivationInstruction {
@@ -17,14 +7,6 @@ export interface FlowActivationInstruction {
     mode?: FlowActivationMode
 }
 
-export interface ActiveNodeSummary {
-    flow: string
-    state: string
-    role?: string
-    say_tpl?: string
-    controller_say_tpl?: string
-}
-
 export interface CandidateTraceEntry {
     id: string
     flow: string
@@ -81,17 +63,6 @@ export interface DecisionCandidateTimeline {
     autoSelected?: CandidateTraceEntry | null
 }
 
-export interface LLMDecision {
-    next_state: string
-    updates?: Record<string, any>
-    flags?: Record<string, any>
-    controller_say_tpl?: string
-    off_schema?: boolean
-    radio_check?: boolean
-    activate_flow?: string | FlowActivationInstruction
-    resume_previous?: boolean
-}
-
 export interface LLMDecisionTraceCall {
     stage: 'readback-check' | 'decision'
     request: Record<string, any>
@@ -116,10 +87,3 @@ export interface LLMDecisionTrace {
         reason?: string
     }
 }
-
-export interface LLMDecisionResult {
-    decision: LLMDecision
-    trace?: LLMDecisionTrace
-    active_nodes?: ActiveNodeSummary[]
-    pilot_intent?: string | null
-}