runtime/microduck-runtime.patch

diff --git a/.github/workflows/_build-release.yml b/.github/workflows/_build-release.yml
index 8647771..e0b84ab 100644
--- a/.github/workflows/_build-release.yml
+++ b/.github/workflows/_build-release.yml
@@ -196,6 +196,7 @@ jobs:
             --include "policies/alpha_stand.onnx=policies/alpha_stand.onnx" \
             --include "policies/alpha_sitstand.onnx=policies/alpha_sitstand.onnx" \
             --include "policies/alpha_ground_pick.onnx=policies/alpha_ground_pick.onnx" \
+            --include "policies/beak_throw.onnx=policies/beak_throw.onnx" \
             --include "policies/ball_kick_left.onnx=policies/ball_kick_left.onnx" \
             --include "policies/ball_kick_right.onnx=policies/ball_kick_right.onnx" \
             --include "policies/roller.onnx=policies/roller.onnx" \
diff --git a/.github/workflows/dev.yml b/.github/workflows/dev.yml
index 8c724ef..e76308d 100644
--- a/.github/workflows/dev.yml
+++ b/.github/workflows/dev.yml
@@ -182,6 +182,7 @@ jobs:
             --include "policies/alpha_stand.onnx=policies/alpha_stand.onnx" \
             --include "policies/alpha_sitstand.onnx=policies/alpha_sitstand.onnx" \
             --include "policies/alpha_ground_pick.onnx=policies/alpha_ground_pick.onnx" \
+            --include "policies/beak_throw.onnx=policies/beak_throw.onnx" \
             --include "policies/ball_kick_left.onnx=policies/ball_kick_left.onnx" \
             --include "policies/ball_kick_right.onnx=policies/ball_kick_right.onnx" \
             --include "policies/roller.onnx=policies/roller.onnx" \
diff --git a/deploy/robotd.toml b/deploy/robotd.toml
index 1a279f8..727c434 100644
--- a/deploy/robotd.toml
+++ b/deploy/robotd.toml
@@ -111,6 +111,7 @@ mode = "walk"
 # stand       = ".../current/policies/alpha_stand.onnx"       standing + body pose
 # sitstand    = ".../current/policies/alpha_sitstand.onnx"    sit <-> stand, posture flag
 # ground_pick = ".../current/policies/alpha_ground_pick.onnx" A-button pick
+# beak_throw = ".../beak_throw.onnx"                       experimental robot.do beak-throw
 # kick_left   = ".../current/policies/ball_kick_left.onnx"
 # kick_right  = ".../current/policies/ball_kick_right.onnx"
 # roulade     = ".../current/policies/roulade.onnx"          X-button forward roll
@@ -144,6 +145,15 @@ mode = "walk"
 # ground_pick_action_scale = 1.0
 # ground_pick_gain_ratio = 1.0
 
+# Experimental beak throw. The policy sees `[cos(2π phase), sin(2π phase), 0]` over this
+# complete cycle. robotd holds the mouth at +20 degrees, slews it to +30 over phase
+# 0.30..0.34, and bypasses the gait target filters for this skill because it was trained
+# unfiltered. Use the trained 1.0 scale: smaller scales were explicitly evaluated and made
+# the closed-loop throw fall. Begin with an empty beak and the operator ready to disable.
+# beak_throw_period = 2.4
+# beak_throw_action_scale = 1.0
+# beak_throw_gain_ratio = 1.0
+
 # How long a kick window stays on the kick network, seconds.
 # kick_duration = 0.5
 
diff --git a/duck-control/src/policy.rs b/duck-control/src/policy.rs
index db5b7db..2878dbb 100644
--- a/duck-control/src/policy.rs
+++ b/duck-control/src/policy.rs
@@ -182,6 +182,8 @@ pub enum Net {
     SitStand,
     /// Phase-scripted ground pick; the twist slots carry `[cos φ, sin φ, 0]`.
     GroundPick,
+    /// Phase-scripted beak throw; the twist slots carry `[cos φ, sin φ, 0]`.
+    BeakThrow,
     KickLeft,
     KickRight,
     /// Episodic forward roll; trained with every command slot at zero, and it starts
@@ -197,6 +199,7 @@ pub struct PolicyPaths {
     pub stand: Option<PathBuf>,
     pub sitstand: Option<PathBuf>,
     pub ground_pick: Option<PathBuf>,
+    pub beak_throw: Option<PathBuf>,
     pub kick_left: Option<PathBuf>,
     pub kick_right: Option<PathBuf>,
     pub roulade: Option<PathBuf>,
@@ -212,6 +215,7 @@ pub struct Policy {
     stand: Option<Session>,
     sitstand: Option<Session>,
     ground_pick: Option<Session>,
+    beak_throw: Option<Session>,
     kick_left: Option<Session>,
     kick_right: Option<Session>,
     roulade: Option<Session>,
@@ -255,6 +259,7 @@ impl Policy {
                 stand: open_opt(&paths.stand, &zero)?,
                 sitstand: open_opt(&paths.sitstand, &zero)?,
                 ground_pick: open_opt(&paths.ground_pick, &zero)?,
+                beak_throw: open_opt(&paths.beak_throw, &zero)?,
                 kick_left: open_opt(&paths.kick_left, &zero)?,
                 kick_right: open_opt(&paths.kick_right, &zero)?,
                 roulade: open_opt(&paths.roulade, &zero)?,
@@ -292,6 +297,10 @@ impl Policy {
         self.ground_pick.is_some()
     }
 
+    pub fn has_beak_throw(&self) -> bool {
+        self.beak_throw.is_some()
+    }
+
     pub fn has_roulade(&self) -> bool {
         self.roulade.is_some()
     }
@@ -317,6 +326,7 @@ impl Policy {
             Net::Stand => self.stand.as_mut(),
             Net::SitStand => self.sitstand.as_mut(),
             Net::GroundPick => self.ground_pick.as_mut(),
+            Net::BeakThrow => self.beak_throw.as_mut(),
             Net::KickLeft => self.kick_left.as_mut(),
             Net::KickRight => self.kick_right.as_mut(),
             Net::Roulade => self.roulade.as_mut(),
diff --git a/duck-ipc-proto/src/lib.rs b/duck-ipc-proto/src/lib.rs
index 33224ab..489cc82 100644
--- a/duck-ipc-proto/src/lib.rs
+++ b/duck-ipc-proto/src/lib.rs
@@ -161,7 +161,7 @@ pub const JSONRPC_VERSION: &str = "2.0";
 /// results are not `deny_unknown_fields`. An older `updaterd` answers `update.show` with
 /// [`code::METHOD_NOT_FOUND`] naming it, which is the designed skew behaviour rather than a
 /// handshake refusal.
-pub const API_VERSION: u32 = 16;
+pub const API_VERSION: u32 = 17;
 
 /// The longest an update may legitimately go quiet, in seconds — the pre-install hook's ceiling.
 ///
@@ -1675,6 +1675,9 @@ pub struct ThereminState {
 pub enum Skill {
     /// Phase-scripted pick from the ground. One shot, ~3 s.
     GroundPick,
+    /// Synchronized phase-scripted body throw and beak release. Experimental hardware
+    /// bring-up skill; one request is one complete 2.4 s cycle.
+    BeakThrow,
     /// Left-leg kick. One shot, half a second, blind to any ball.
     KickLeft,
     /// Right-leg kick.
@@ -1796,6 +1799,8 @@ pub struct SubscribeResult {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub ground_pick: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
+    pub beak_throw: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
     pub kick_left: Option<String>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub kick_right: Option<String>,
diff --git a/policies/README.md b/policies/README.md
index a1760e6..6fccd86 100644
--- a/policies/README.md
+++ b/policies/README.md
@@ -33,6 +33,7 @@ names to specific training runs:
 | `alpha_stand.onnx` | `BEST_alpha_stand_body_control.onnx` | standing + body-pose |
 | `alpha_sitstand.onnx` | `BEST_alpha_sitstand.onnx` | sit ↔ stand (posture flag) |
 | `alpha_ground_pick.onnx` | `alpha_ground_pick.onnx` | ground pick (phase command) |
+| `beak_throw.onnx` | `q2p/beak-hardware-straight-v3-20260901-a100`, `model_2150.pt` | experimental collision-trained beak throw |
 | `ball_kick_left.onnx` | `ball_kick_left.onnx` | left-leg kick |
 | `ball_kick_right.onnx` | `ball_kick_right.onnx` | right-leg kick |
 | `roller.onnx` | `BEST_roller.onnx` | roller-mode locomotion |
@@ -72,3 +73,27 @@ stand = "/home/radxa/my_stand.onnx"
 Then `sudo systemctl restart robotd`. A policy that fails to load is reported through
 `robot.health` as `policy unavailable: <reason>` while the loop keeps ticking and holding its
 pose, so a bad file is visible without putting the robot on the floor.
+
+The beak throw is deliberately not a default capability. Opt in explicitly and invoke it
+from a cleared range. Use the policy's trained scale of 1.0: evaluation found that reducing
+the scale changes the closed-loop trajectory enough to make the duck fall; safety comes from
+the throw-specific anatomical target clamp and an empty-beak first trial, not rescaling.
+
+The selected ONNX has SHA-256
+`8638bbd29b672d84bcd426b02fa49eb6224ab57b29f5556b4003f3052bf5cecd`. It was selected
+over the final training checkpoint because a randomized throw-to-stand screen found better
+lateral accuracy and no near-flip. Its raw outputs still exceed an anatomical target limit,
+so the throw-specific runtime clamp is mandatory; this remains an experimental supervised
+test policy rather than an accepted production policy.
+
+```toml
+[policy]
+beak_throw = "/opt/robot/current/policies/beak_throw.onnx"
+beak_throw_action_scale = 1.0
+```
+
+```bash
+robotctl robot enable
+robotctl robot do beak-throw
+robotctl robot disable
+```
diff --git a/robotctl/src/configure.rs b/robotctl/src/configure.rs
index e8b6db5..e40ff61 100644
--- a/robotctl/src/configure.rs
+++ b/robotctl/src/configure.rs
@@ -194,6 +194,7 @@ impl Model {
             "policy.stand" => path(policy.stand),
             "policy.sitstand" => path(policy.sitstand),
             "policy.ground_pick" => path(policy.ground_pick),
+            "policy.beak_throw" => path(policy.beak_throw),
             "policy.kick_left" => path(policy.kick_left),
             "policy.kick_right" => path(policy.kick_right),
             "policy.roulade" => path(policy.roulade),
@@ -202,6 +203,8 @@ impl Model {
             "policy.legs_lowpass" => policy.legs_lowpass.and_then(float),
             "policy.ground_pick_period" => float(policy.ground_pick_period),
             "policy.ground_pick_action_scale" => float(policy.ground_pick_action_scale),
+            "policy.beak_throw_period" => float(policy.beak_throw_period),
+            "policy.beak_throw_action_scale" => float(policy.beak_throw_action_scale),
             "media.bitrate" => Some(params.media.bitrate_resolved().to_string()),
             "audio.pet_detect" => Some(
                 params
diff --git a/robotctl/src/main.rs b/robotctl/src/main.rs
index c7f0fc5..801822e 100644
--- a/robotctl/src/main.rs
+++ b/robotctl/src/main.rs
@@ -350,6 +350,18 @@ enum SystemCommand {
 
 #[derive(Subcommand, Debug)]
 enum RobotCommand {
+    /// Enable the policy and bring a limp robot to the home pose before driving.
+    Enable {
+        #[arg(long)]
+        json: bool,
+    },
+
+    /// Disable the policy and return to the home pose while keeping torque on.
+    Disable {
+        #[arg(long)]
+        json: bool,
+    },
+
     /// Power the joints and ramp to the home pose, over about two seconds.
     ///
     /// **This moves every joint.** Have the robot on its stand, or hold it. Needs no policy — a
@@ -681,6 +693,8 @@ fn bar(fraction: f64) -> String {
 #[derive(clap::ValueEnum, Clone, Copy, Debug)]
 enum SkillArg {
     GroundPick,
+    /// Experimental synchronized beak throw.
+    BeakThrow,
     KickLeft,
     KickRight,
     /// Sit if standing, stand if sitting.
@@ -693,6 +707,7 @@ impl SkillArg {
     fn as_skill(self) -> proto::Skill {
         match self {
             SkillArg::GroundPick => proto::Skill::GroundPick,
+            SkillArg::BeakThrow => proto::Skill::BeakThrow,
             SkillArg::KickLeft => proto::Skill::KickLeft,
             SkillArg::KickRight => proto::Skill::KickRight,
             SkillArg::Sit => proto::Skill::SitToggle,
@@ -2165,6 +2180,20 @@ fn run_robot(socket: &Path, command: RobotCommand) -> Result<(), Failure> {
     client.hello()?;
 
     let (call, json) = match &command {
+        RobotCommand::Enable { json } => (
+            proto::Call::RobotEnable(proto::EnableParams {
+                on: true,
+                toggle: false,
+            }),
+            *json,
+        ),
+        RobotCommand::Disable { json } => (
+            proto::Call::RobotEnable(proto::EnableParams {
+                on: false,
+                toggle: false,
+            }),
+            *json,
+        ),
         RobotCommand::Init { json } => (proto::Call::RobotInit, *json),
         RobotCommand::Relax { json, .. } => (proto::Call::RobotRelax, *json),
         RobotCommand::Do { skill, json } => (
@@ -2232,6 +2261,12 @@ fn run_robot(socket: &Path, command: RobotCommand) -> Result<(), Failure> {
         return Err(Failure::new(exit::REFUSED, reason));
     }
     match command {
+        RobotCommand::Enable { .. } | RobotCommand::Disable { .. } => {
+            println!(
+                "{}",
+                outcome.reason.unwrap_or_else(|| "accepted".to_owned())
+            )
+        }
         RobotCommand::Init { .. } => println!("standing up — about two seconds to the home pose"),
         RobotCommand::Relax { .. } => println!("torque off"),
         RobotCommand::Do { skill, .. } => println!("{skill:?} queued"),
diff --git a/robotctl/src/monitor.rs b/robotctl/src/monitor.rs
index 232767c..5f225dd 100644
--- a/robotctl/src/monitor.rs
+++ b/robotctl/src/monitor.rs
@@ -1615,6 +1615,9 @@ impl View {
         if policy.ground_pick.is_some() {
             skills.push("pick");
         }
+        if policy.beak_throw.is_some() {
+            skills.push("beak-throw");
+        }
         match (policy.kick_left.is_some(), policy.kick_right.is_some()) {
             (true, true) => skills.push("kicks"),
             (true, false) => skills.push("kick-left"),
diff --git a/robotd-params/src/lib.rs b/robotd-params/src/lib.rs
index 5aa8ba0..96178b3 100644
--- a/robotd-params/src/lib.rs
+++ b/robotd-params/src/lib.rs
@@ -521,6 +521,9 @@ pub struct PolicyParams {
     pub sitstand: Option<PathBuf>,
     /// Phase-scripted ground pick. In roller mode this slot holds the crouch.
     pub ground_pick: Option<PathBuf>,
+    /// Phase-scripted beak throw. Disabled by default; it is an experimental hardware
+    /// bring-up capability rather than part of the ordinary release bundle.
+    pub beak_throw: Option<PathBuf>,
     pub kick_left: Option<PathBuf>,
     pub kick_right: Option<PathBuf>,
     /// Episodic forward roll. Ships by default in both modes, as the prototype now does.
@@ -547,6 +550,13 @@ pub struct PolicyParams {
     pub ground_pick_action_scale: Option<f64>,
     /// Gain multiplier while the ground pick runs.
     pub ground_pick_gain_ratio: f64,
+    /// One complete beak-throw cycle, including recovery, seconds.
+    pub beak_throw_period: f64,
+    /// Action scale during the beak throw. The trained value is 1.0; hardware bring-up
+    /// starts lower and promotes only after empty-beak trials.
+    pub beak_throw_action_scale: f64,
+    /// Gain multiplier during the beak throw.
+    pub beak_throw_gain_ratio: f64,
     /// How long a kick window stays on the kick network, seconds.
     pub kick_duration: f64,
     /// One roulade — one forward roll, seconds. Holding the button chains rolls; this is
@@ -581,6 +591,7 @@ pub struct ResolvedPolicy {
     pub stand: Option<PathBuf>,
     pub sitstand: Option<PathBuf>,
     pub ground_pick: Option<PathBuf>,
+    pub beak_throw: Option<PathBuf>,
     pub kick_left: Option<PathBuf>,
     pub kick_right: Option<PathBuf>,
     pub roulade: Option<PathBuf>,
@@ -593,6 +604,9 @@ pub struct ResolvedPolicy {
     pub ground_pick_period: f64,
     pub ground_pick_action_scale: f64,
     pub ground_pick_gain_ratio: f64,
+    pub beak_throw_period: f64,
+    pub beak_throw_action_scale: f64,
+    pub beak_throw_gain_ratio: f64,
     pub kick_duration: f64,
     pub roulade_duration: f64,
     pub roulade_action_scale: f64,
@@ -642,6 +656,7 @@ impl PolicyParams {
             stand: path(&self.stand, stand),
             sitstand: path(&self.sitstand, sitstand),
             ground_pick: path(&self.ground_pick, ground_pick),
+            beak_throw: path(&self.beak_throw, None),
             kick_left: path(&self.kick_left, kick.then_some("ball_kick_left.onnx")),
             kick_right: path(&self.kick_right, kick.then_some("ball_kick_right.onnx")),
             roulade: path(&self.roulade, Some("roulade.onnx")),
@@ -663,6 +678,9 @@ impl PolicyParams {
                 Mode::Roller => 0.8,
             }),
             ground_pick_gain_ratio: self.ground_pick_gain_ratio,
+            beak_throw_period: self.beak_throw_period,
+            beak_throw_action_scale: self.beak_throw_action_scale,
+            beak_throw_gain_ratio: self.beak_throw_gain_ratio,
             kick_duration: self.kick_duration,
             roulade_duration: self.roulade_duration,
             roulade_action_scale: self.roulade_action_scale,
@@ -740,6 +758,7 @@ impl Default for PolicyParams {
             stand: None,
             sitstand: None,
             ground_pick: None,
+            beak_throw: None,
             kick_left: None,
             kick_right: None,
             roulade: None,
@@ -753,6 +772,9 @@ impl Default for PolicyParams {
             ground_pick_period: None,
             ground_pick_action_scale: None,
             ground_pick_gain_ratio: 1.0,
+            beak_throw_period: 2.4,
+            beak_throw_action_scale: 1.0,
+            beak_throw_gain_ratio: 1.0,
             kick_duration: 0.5,
             roulade_duration: 1.0,
             roulade_action_scale: 1.0,
diff --git a/robotd-params/src/registry.rs b/robotd-params/src/registry.rs
index bda0238..b3e48ae 100644
--- a/robotd-params/src/registry.rs
+++ b/robotd-params/src/registry.rs
@@ -138,6 +138,11 @@ pub const REGISTRY: &[Entry] = &[
         Kind::OptionalPath,
         "Ground-pick policy (roller: the crouch)",
     ),
+    entry(
+        "policy.beak_throw",
+        Kind::OptionalPath,
+        "Experimental phase-scripted beak-throw policy",
+    ),
     entry("policy.kick_left", Kind::OptionalPath, "Left-kick policy"),
     entry("policy.kick_right", Kind::OptionalPath, "Right-kick policy"),
     entry("policy.roulade", Kind::OptionalPath, "Forward-roll policy"),
@@ -186,6 +191,21 @@ pub const REGISTRY: &[Entry] = &[
         Kind::Float,
         "Gain multiplier during the ground pick",
     ),
+    entry(
+        "policy.beak_throw_period",
+        Kind::Float,
+        "One beak-throw cycle including recovery, seconds",
+    ),
+    entry(
+        "policy.beak_throw_action_scale",
+        Kind::Float,
+        "Action scale during the beak throw",
+    ),
+    entry(
+        "policy.beak_throw_gain_ratio",
+        Kind::Float,
+        "Gain multiplier during the beak throw",
+    ),
     entry("policy.kick_duration", Kind::Float, "Kick window, seconds"),
     entry(
         "policy.roulade_duration",
diff --git a/robotd/src/control.rs b/robotd/src/control.rs
index bd258dd..ff418a7 100644
--- a/robotd/src/control.rs
+++ b/robotd/src/control.rs
@@ -36,12 +36,47 @@ use duck_control::model::{DEFAULT_POSITION, NUM_JOINTS};
 use duck_control::obs::{ACTION_LEN, Command, Observation};
 use duck_control::policy::{Net, Policy, PolicyError};
 
+// Absolute position limits from the physical beak MJCF, in runtime joint
+// order. They are scoped to the experimental throw: applying them globally
+// would silently change the deployed walking and standing policies. The
+// hardware continuation task uses these identical limits while training.
+const BEAK_THROW_JOINT_MIN: [f64; NUM_JOINTS] = degrees([
+    -25.0, -22.0, -90.0, -90.0, -90.0, -90.0, -90.0, -170.0, -25.0, -5.0, -30.0, -22.0, -90.0,
+    -90.0, -90.0,
+]);
+const BEAK_THROW_JOINT_MAX: [f64; NUM_JOINTS] = degrees([
+    30.0, 22.0, 90.0, 90.0, 90.0, 60.0, 90.0, 170.0, 25.0, 30.0, 25.0, 22.0, 90.0, 90.0, 90.0,
+]);
+
+const fn degrees(values: [f64; NUM_JOINTS]) -> [f64; NUM_JOINTS] {
+    let mut radians = [0.0; NUM_JOINTS];
+    let mut i = 0;
+    while i < NUM_JOINTS {
+        radians[i] = values[i] * std::f64::consts::PI / 180.0;
+        i += 1;
+    }
+    radians
+}
+
+fn clamp_beak_throw_targets(targets: &mut [f64; NUM_JOINTS]) {
+    for (joint, target) in targets.iter_mut().enumerate() {
+        *target = target.clamp(BEAK_THROW_JOINT_MIN[joint], BEAK_THROW_JOINT_MAX[joint]);
+    }
+}
+
 /// Joint indices the head low-pass covers: neck_pitch, head_pitch, head_yaw, head_roll.
 const HEAD_JOINTS: std::ops::Range<usize> = 5..9;
 
 /// The ground pick hands back at this fraction of its cycle — the prototype's cutoff.
 const GROUND_PICK_END_PHASE: f64 = 0.7;
 
+/// The beak throw's simulator/runtime contract. The body policy sees a 2.4 s phase cycle;
+/// the mouth is held at +20 degrees, then slews to its +30 degree limit over 4% of the
+/// cycle. `mouth_target` maps -5..+30 degrees to 0..1, hence 25/35 for +20 degrees.
+const BEAK_THROW_RELEASE_PHASE: f64 = 0.30;
+const BEAK_THROW_OPEN_DURATION_PHASE: f64 = 0.04;
+const BEAK_THROW_HOLD_MOUTH: f64 = 25.0 / 35.0;
+
 /// How long the sitstand network rises (posture flag 0) before the main policy takes over.
 /// 1 s is enough on the robot — velstand owns the tail of the rise fine.
 const RISE_SECS: f64 = 1.0;
@@ -91,6 +126,13 @@ pub struct SkillTuning {
     pub ground_pick_action_scale: f64,
     /// Gain multiplier while the pick runs.
     pub ground_pick_gain_ratio: f64,
+    /// One complete beak-throw cycle, including recovery, seconds.
+    pub beak_throw_period: f64,
+    /// Action scale while the beak throw runs. Start below 1.0 for empty-beak bring-up;
+    /// 1.0 is the training value.
+    pub beak_throw_action_scale: f64,
+    /// Gain multiplier while the beak throw runs.
+    pub beak_throw_gain_ratio: f64,
     /// How long a kick window stays on the kick network, seconds.
     pub kick_duration: f64,
     /// One roulade — one forward roll, seconds. The prototype's measured single-roll time.
@@ -107,6 +149,9 @@ impl Default for SkillTuning {
             ground_pick_period: 4.0,
             ground_pick_action_scale: 1.0,
             ground_pick_gain_ratio: 1.0,
+            beak_throw_period: 2.4,
+            beak_throw_action_scale: 1.0,
+            beak_throw_gain_ratio: 1.0,
             kick_duration: 0.5,
             roulade_duration: 1.0,
             roulade_action_scale: 1.0,
@@ -127,6 +172,9 @@ pub struct Step {
     /// A scripted move is mid-flight — the robot is moving regardless of the twist, so
     /// restarting the daemon now would put it on the floor.
     pub busy: bool,
+    /// A skill-owned mouth opening fraction. The beak throw owns the mouth while active so
+    /// its release cannot race a client mouth intent, the theremin, or the chorale.
+    pub mouth: Option<f64>,
 }
 
 /// Where the robot is in the sit↔stand cycle.
@@ -155,6 +203,8 @@ pub struct Controller {
     previous: Option<[f64; NUM_JOINTS]>,
     /// Ground-pick phase, 0..[`GROUND_PICK_END_PHASE`]. `None` when inactive.
     ground_pick: Option<f64>,
+    /// Beak-throw phase, 0..1. `None` when inactive.
+    beak_throw: Option<f64>,
     /// An active kick window: which leg, and seconds remaining.
     kick: Option<(bool, f64)>,
     /// An active roulade: seconds remaining in the current roll.
@@ -175,6 +225,7 @@ impl Controller {
             last_action: [0.0; ACTION_LEN],
             previous: None,
             ground_pick: None,
+            beak_throw: None,
             kick: None,
             roulade: None,
             roulade_chain: 0.0,
@@ -190,6 +241,8 @@ impl Controller {
     pub fn reset(&mut self) {
         self.last_action = [0.0; ACTION_LEN];
         self.previous = None;
+        // An interrupted one-shot must never resume halfway through its snap/release.
+        self.beak_throw = None;
     }
 
     pub fn has_sitstand(&self) -> bool {
@@ -204,6 +257,7 @@ impl Controller {
     /// parked, not travelling.
     pub fn busy(&self) -> bool {
         self.ground_pick.is_some()
+            || self.beak_throw.is_some()
             || self.kick.is_some()
             || self.roulade.is_some()
             || matches!(self.sit, Sit::Rising { .. })
@@ -223,6 +277,21 @@ impl Controller {
         Ok(())
     }
 
+    /// Start one synchronized body-policy + mouth-release cycle.
+    pub fn start_beak_throw(&mut self) -> Result<(), &'static str> {
+        if !self.policy.has_beak_throw() {
+            return Err("no beak-throw policy loaded");
+        }
+        if self.busy() {
+            return Err("a scripted move is already running");
+        }
+        if self.sit != Sit::Up {
+            return Err("robot must be standing");
+        }
+        self.beak_throw = Some(0.0);
+        Ok(())
+    }
+
     /// Start a kick window. Blocked while any scripted move runs, as the prototype blocks it.
     pub fn start_kick(&mut self, left: bool) -> Result<(), &'static str> {
         if !self.policy.has_kick(left) {
@@ -333,10 +402,22 @@ impl Controller {
         {
             self.sit = Sit::Up;
         }
+        if let Some(phase) = self.beak_throw
+            && phase >= 1.0
+        {
+            self.beak_throw = None;
+        }
 
         // Re-encode the command for the active skill and pick the network. The priority
         // chain is the prototype's: roulade > kick > ground pick > sit/rise > stand > walk.
-        let (net, effective, label) = if self.roulade.is_some() {
+        let (net, effective, label) = if let Some(phase) = self.beak_throw {
+            let angle = std::f64::consts::TAU * phase;
+            let c = Command {
+                twist: [angle.cos(), angle.sin(), 0.0],
+                ..Command::default()
+            };
+            (Net::BeakThrow, c, "beak_throw")
+        } else if self.roulade.is_some() {
             // Trained with every command slot at zero; it rolls as soon as it is switched
             // in, so being selected IS the trigger.
             (Net::Roulade, Command::default(), "roulade")
@@ -403,6 +484,10 @@ impl Controller {
             || (matches!(net, Net::KickLeft | Net::KickRight | Net::SitStand)
                 && self.policy.will_stand(effective.twist_magnitude()));
         let (scale, gain) = match net {
+            Net::BeakThrow => (
+                self.skills.beak_throw_action_scale,
+                (self.tuning.gain as f64 * self.skills.beak_throw_gain_ratio).round() as u16,
+            ),
             Net::Roulade => (
                 self.skills.roulade_action_scale,
                 (self.tuning.gain as f64 * self.skills.roulade_gain_ratio).round() as u16,
@@ -435,7 +520,11 @@ impl Controller {
             targets[joint] = DEFAULT_POSITION[joint] + scale * offsets[joint];
         }
 
-        if let Some(previous) = self.previous {
+        // The beak policy was trained on raw targets. The ordinary alpha gaits were trained
+        // with these filters, but applying them to the throw delays and weakens the snap.
+        if !matches!(net, Net::BeakThrow)
+            && let Some(previous) = self.previous
+        {
             if let Some(alpha) = self.tuning.head_lowpass {
                 for joint in HEAD_JOINTS {
                     targets[joint] = alpha * targets[joint] + (1.0 - alpha) * previous[joint];
@@ -450,6 +539,9 @@ impl Controller {
                 }
             }
         }
+        if matches!(net, Net::BeakThrow) {
+            clamp_beak_throw_targets(&mut targets);
+        }
         self.previous = Some(targets);
 
         // Advance the windows, after the tick that used them — the prototype advances its
@@ -460,6 +552,10 @@ impl Controller {
                 self.ground_pick = None;
             }
         }
+        let mouth = self.beak_throw.map(beak_throw_mouth);
+        if let Some(phase) = self.beak_throw.as_mut() {
+            *phase += dt / self.skills.beak_throw_period;
+        }
         if let Some((_, remaining)) = self.kick.as_mut() {
             *remaining -= dt;
         }
@@ -476,10 +572,17 @@ impl Controller {
             label,
             gain,
             busy: self.busy(),
+            mouth,
         })
     }
 }
 
+fn beak_throw_mouth(phase: f64) -> f64 {
+    let t = ((phase - BEAK_THROW_RELEASE_PHASE) / BEAK_THROW_OPEN_DURATION_PHASE).clamp(0.0, 1.0);
+    let smooth = t * t * (3.0 - 2.0 * t);
+    BEAK_THROW_HOLD_MOUTH + smooth * (1.0 - BEAK_THROW_HOLD_MOUTH)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -509,6 +612,9 @@ mod tests {
         assert_eq!(s.ground_pick_period, 4.0);
         assert_eq!(s.ground_pick_action_scale, 1.0);
         assert_eq!(s.ground_pick_gain_ratio, 1.0);
+        assert_eq!(s.beak_throw_period, 2.4);
+        assert_eq!(s.beak_throw_action_scale, 1.0);
+        assert_eq!(s.beak_throw_gain_ratio, 1.0);
         assert_eq!(s.kick_duration, 0.5);
         assert_eq!(s.roulade_duration, 1.0, "one roll, the measured time");
         assert_eq!(s.roulade_action_scale, 1.0);
@@ -535,4 +641,30 @@ mod tests {
         assert_eq!(GROUND_PICK_END_PHASE, 0.7);
         assert_eq!(RISE_SECS, 1.0);
     }
+
+    #[test]
+    fn beak_release_matches_the_training_phase_and_hardware_angles() {
+        assert_eq!(beak_throw_mouth(0.0), 25.0 / 35.0);
+        assert_eq!(beak_throw_mouth(BEAK_THROW_RELEASE_PHASE), 25.0 / 35.0);
+        let halfway =
+            beak_throw_mouth(BEAK_THROW_RELEASE_PHASE + BEAK_THROW_OPEN_DURATION_PHASE / 2.0);
+        assert!(
+            (halfway - (BEAK_THROW_HOLD_MOUTH + (1.0 - BEAK_THROW_HOLD_MOUTH) / 2.0)).abs() < 1e-12
+        );
+        assert_eq!(
+            beak_throw_mouth(BEAK_THROW_RELEASE_PHASE + BEAK_THROW_OPEN_DURATION_PHASE),
+            1.0
+        );
+    }
+
+    #[test]
+    fn beak_throw_targets_use_the_training_joint_envelope() {
+        let mut targets = [100.0; NUM_JOINTS];
+        clamp_beak_throw_targets(&mut targets);
+        assert_eq!(targets, BEAK_THROW_JOINT_MAX);
+
+        targets.fill(-100.0);
+        clamp_beak_throw_targets(&mut targets);
+        assert_eq!(targets, BEAK_THROW_JOINT_MIN);
+    }
 }
diff --git a/robotd/src/intents.rs b/robotd/src/intents.rs
index a0c2893..81746d2 100644
--- a/robotd/src/intents.rs
+++ b/robotd/src/intents.rs
@@ -74,6 +74,7 @@ impl Default for PoseIntent {
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
 pub struct SkillRequests {
     pub ground_pick: bool,
+    pub beak_throw: bool,
     pub kick_left: bool,
     pub kick_right: bool,
     pub sit_toggle: bool,
@@ -84,7 +85,12 @@ pub struct SkillRequests {
 
 impl SkillRequests {
     pub fn any(&self) -> bool {
-        self.ground_pick || self.kick_left || self.kick_right || self.sit_toggle || self.roulade
+        self.ground_pick
+            || self.beak_throw
+            || self.kick_left
+            || self.kick_right
+            || self.sit_toggle
+            || self.roulade
     }
 }
 
@@ -94,6 +100,7 @@ const SKILL_KICK_LEFT: u32 = 1 << 1;
 const SKILL_KICK_RIGHT: u32 = 1 << 2;
 const SKILL_SIT_TOGGLE: u32 = 1 << 3;
 const SKILL_ROULADE: u32 = 1 << 4;
+const SKILL_BEAK_THROW: u32 = 1 << 5;
 
 /// How fresh a wheee hold must be to still count as held. `padd` re-notifies every tick
 /// (20 ms) while the trigger is down, so anything much older means the client stopped
@@ -271,6 +278,7 @@ impl Intents {
     pub fn request_skill(&self, skill: duck_ipc_proto::Skill) {
         let bit = match skill {
             duck_ipc_proto::Skill::GroundPick => SKILL_GROUND_PICK,
+            duck_ipc_proto::Skill::BeakThrow => SKILL_BEAK_THROW,
             duck_ipc_proto::Skill::KickLeft => SKILL_KICK_LEFT,
             duck_ipc_proto::Skill::KickRight => SKILL_KICK_RIGHT,
             duck_ipc_proto::Skill::SitToggle => SKILL_SIT_TOGGLE,
@@ -285,6 +293,7 @@ impl Intents {
         let bits = self.skills.swap(0, std::sync::atomic::Ordering::Relaxed);
         SkillRequests {
             ground_pick: bits & SKILL_GROUND_PICK != 0,
+            beak_throw: bits & SKILL_BEAK_THROW != 0,
             kick_left: bits & SKILL_KICK_LEFT != 0,
             kick_right: bits & SKILL_KICK_RIGHT != 0,
             sit_toggle: bits & SKILL_SIT_TOGGLE != 0,
diff --git a/robotd/src/main.rs b/robotd/src/main.rs
index aa7c46b..1e66d61 100644
--- a/robotd/src/main.rs
+++ b/robotd/src/main.rs
@@ -308,6 +308,7 @@ struct PolicyNames {
     stand: Option<String>,
     sitstand: Option<String>,
     ground_pick: Option<String>,
+    beak_throw: Option<String>,
     kick_left: Option<String>,
     kick_right: Option<String>,
     roulade: Option<String>,
@@ -325,6 +326,7 @@ impl PolicyNames {
             stand: name(&policy.stand),
             sitstand: name(&policy.sitstand),
             ground_pick: name(&policy.ground_pick),
+            beak_throw: name(&policy.beak_throw),
             kick_left: name(&policy.kick_left),
             kick_right: name(&policy.kick_right),
             roulade: name(&policy.roulade),
@@ -1079,6 +1081,9 @@ fn build_controller(
             ground_pick_period: policy_cfg.ground_pick_period,
             ground_pick_action_scale: policy_cfg.ground_pick_action_scale,
             ground_pick_gain_ratio: policy_cfg.ground_pick_gain_ratio,
+            beak_throw_period: policy_cfg.beak_throw_period,
+            beak_throw_action_scale: policy_cfg.beak_throw_action_scale,
+            beak_throw_gain_ratio: policy_cfg.beak_throw_gain_ratio,
             kick_duration: policy_cfg.kick_duration,
             roulade_duration: policy_cfg.roulade_duration,
             roulade_action_scale: policy_cfg.roulade_action_scale,
@@ -1089,6 +1094,7 @@ fn build_controller(
             stand: policy_cfg.stand.clone(),
             sitstand: policy_cfg.sitstand.clone(),
             ground_pick: policy_cfg.ground_pick.clone(),
+            beak_throw: policy_cfg.beak_throw.clone(),
             kick_left: policy_cfg.kick_left.clone(),
             kick_right: policy_cfg.kick_right.clone(),
             roulade: policy_cfg.roulade.clone(),
@@ -1107,6 +1113,7 @@ fn build_controller(
                     stand = ?policy_cfg.stand.as_ref().map(|p| p.display().to_string()),
                     sitstand = ?policy_cfg.sitstand.as_ref().map(|p| p.display().to_string()),
                     ground_pick = ?policy_cfg.ground_pick.as_ref().map(|p| p.display().to_string()),
+                    beak_throw = ?policy_cfg.beak_throw.as_ref().map(|p| p.display().to_string()),
                     kicks = policy_cfg.kick_left.is_some() || policy_cfg.kick_right.is_some(),
                     roulade = ?policy_cfg.roulade.as_ref().map(|p| p.display().to_string()),
                     limp_fall,
@@ -1457,6 +1464,9 @@ async fn control_loop<T: RobotIo>(
                     if requests.ground_pick {
                         outcome("ground_pick", controller.start_ground_pick());
                     }
+                    if requests.beak_throw {
+                        outcome("beak_throw", controller.start_beak_throw());
+                    }
                     if requests.kick_left {
                         outcome("kick_left", controller.start_kick(true));
                     }
@@ -1933,67 +1943,72 @@ async fn control_loop<T: RobotIo>(
             1.0
         };
 
-        let (mut targets, gain, moving, policy_label) = match (driving, sensors.as_ref()) {
-            // The limp-fall sequence, before anything else — `driving` is false throughout,
-            // so without this it would fall through to the hold branch and the robot would
-            // be commanded its pre-fall pose at walking gain, which is precisely the thing
-            // the mode exists to stop.
-            //
-            // `moving` stays true for the whole sequence: the joints are travelling (down,
-            // then back to the pose), and `safeToRestart` must not say yes in the middle of
-            // a fall.
-            _ if in_limp_fall => match limp_fall {
-                // Command the joints where they already are, at limp gain. Following the
-                // measurement rather than holding a fixed pose is what makes it soft: a
-                // fixed target grows an error as the robot collapses, and an error at any
-                // gain is a motor pushing back against the floor.
-                LimpFall::Limp { .. } => (
-                    coast.known_positions(hold),
-                    params.safety.gain_limp,
-                    true,
-                    "limp_fall",
-                ),
-                LimpFall::Posing { .. } => (
-                    // Past the end of the ramp the target is the pose itself — the state
-                    // machine above clears `Posing` on the same tick, so this is the one
-                    // frame where the two can disagree.
-                    limp_fall
-                        .pose_target(tick_start, limp_fall_pose)
-                        .unwrap_or(DEFAULT_POSITION),
-                    params.safety.limp_fall_pose_gain,
-                    true,
-                    "limp_pose",
-                ),
-                LimpFall::Idle => unreachable!("in_limp_fall excludes Idle"),
-            },
-            (true, Some(sensors)) => {
-                let controller = controller.as_mut().expect("driving implies a controller");
-                match controller.step(sensors, &command, snapshot.pose.active, dt, scale_mult) {
-                    Ok(step) => (
-                        step.targets,
-                        step.gain,
-                        // A scripted move is motion whatever the twist says; so is walking.
-                        step.busy || command.twist_magnitude() > 0.0,
-                        step.label,
+        let (mut targets, gain, moving, policy_label, skill_mouth) =
+            match (driving, sensors.as_ref()) {
+                // The limp-fall sequence, before anything else — `driving` is false throughout,
+                // so without this it would fall through to the hold branch and the robot would
+                // be commanded its pre-fall pose at walking gain, which is precisely the thing
+                // the mode exists to stop.
+                //
+                // `moving` stays true for the whole sequence: the joints are travelling (down,
+                // then back to the pose), and `safeToRestart` must not say yes in the middle of
+                // a fall.
+                _ if in_limp_fall => match limp_fall {
+                    // Command the joints where they already are, at limp gain. Following the
+                    // measurement rather than holding a fixed pose is what makes it soft: a
+                    // fixed target grows an error as the robot collapses, and an error at any
+                    // gain is a motor pushing back against the floor.
+                    LimpFall::Limp { .. } => (
+                        coast.known_positions(hold),
+                        params.safety.gain_limp,
+                        true,
+                        "limp_fall",
+                        None,
+                    ),
+                    LimpFall::Posing { .. } => (
+                        // Past the end of the ramp the target is the pose itself — the state
+                        // machine above clears `Posing` on the same tick, so this is the one
+                        // frame where the two can disagree.
+                        limp_fall
+                            .pose_target(tick_start, limp_fall_pose)
+                            .unwrap_or(DEFAULT_POSITION),
+                        params.safety.limp_fall_pose_gain,
+                        true,
+                        "limp_pose",
+                        None,
                     ),
-                    Err(e) => {
-                        tracing::warn!(error = %e, "inference failed; holding");
-                        (hold, policy_cfg.gain, false, "held")
+                    LimpFall::Idle => unreachable!("in_limp_fall excludes Idle"),
+                },
+                (true, Some(sensors)) => {
+                    let controller = controller.as_mut().expect("driving implies a controller");
+                    match controller.step(sensors, &command, snapshot.pose.active, dt, scale_mult) {
+                        Ok(step) => (
+                            step.targets,
+                            step.gain,
+                            // A scripted move is motion whatever the twist says; so is walking.
+                            step.busy || command.twist_magnitude() > 0.0,
+                            step.label,
+                            step.mouth,
+                        ),
+                        Err(e) => {
+                            tracing::warn!(error = %e, "inference failed; holding");
+                            (hold, policy_cfg.gain, false, "held", None)
+                        }
                     }
                 }
-            }
-            // Ramping to the home pose. `moving` is true, because it is: the joints are travelling,
-            // and `safeToRestart` must not say yes in the middle of it.
-            _ if bringup.homing_target(tick_start).is_some() => (
-                bringup
-                    .homing_target(tick_start)
-                    .expect("just checked it is Some"),
-                policy_cfg.gain,
-                true,
-                "homing",
-            ),
-            _ => (hold, policy_cfg.gain, false, "held"),
-        };
+                // Ramping to the home pose. `moving` is true, because it is: the joints are travelling,
+                // and `safeToRestart` must not say yes in the middle of it.
+                _ if bringup.homing_target(tick_start).is_some() => (
+                    bringup
+                        .homing_target(tick_start)
+                        .expect("just checked it is Some"),
+                    policy_cfg.gain,
+                    true,
+                    "homing",
+                    None,
+                ),
+                _ => (hold, policy_cfg.gain, false, "held", None),
+            };
         state.moving.store(moving, Ordering::Relaxed);
 
         // The theremin: a hand's distance in front of the beak, turned into a note and a
@@ -2172,7 +2187,11 @@ async fn control_loop<T: RobotIo>(
         // The mouth is not part of any policy; the intent is the only thing that moves it.
         // Only while driving — a held or homing robot keeps whatever its hold pose says, so
         // a restart cannot snap a mouth.
-        if driving && theremin_state.is_none() && chorale_state.is_none() {
+        if let Some(open) = skill_mouth {
+            // The body policy and release clock are one skill. Its mouth target wins over
+            // every ambient mouth owner so release timing cannot race a client intent.
+            targets[duck_control::model::MOUTH_INDEX] = duck_control::model::mouth_target(open);
+        } else if driving && theremin_state.is_none() && chorale_state.is_none() {
             targets[duck_control::model::MOUTH_INDEX] =
                 duck_control::model::mouth_target(snapshot.mouth);
         }
@@ -2722,6 +2741,7 @@ fn dispatch(
                 // One load for the whole decision: these are the *current* mode's networks, and
                 // reading them field by field could straddle a mode switch.
                 proto::Skill::GroundPick => policies.ground_pick.is_some(),
+                proto::Skill::BeakThrow => policies.beak_throw.is_some(),
                 proto::Skill::KickLeft => policies.kick_left.is_some(),
                 proto::Skill::KickRight => policies.kick_right.is_some(),
                 proto::Skill::SitToggle => policies.sitstand.is_some(),
@@ -2901,6 +2921,7 @@ fn dispatch(
                     stand: policies.stand.clone(),
                     sitstand: policies.sitstand.clone(),
                     ground_pick: policies.ground_pick.clone(),
+                    beak_throw: policies.beak_throw.clone(),
                     kick_left: policies.kick_left.clone(),
                     kick_right: policies.kick_right.clone(),
                     roulade: policies.roulade.clone(),
diff --git a/scripts/dev-push.sh b/scripts/dev-push.sh
index ee42b6d..65145da 100755
--- a/scripts/dev-push.sh
+++ b/scripts/dev-push.sh
@@ -400,6 +400,7 @@ cargo run -p xtask -- package \
     --include "policies/alpha_stand.onnx=policies/alpha_stand.onnx" \
     --include "policies/alpha_sitstand.onnx=policies/alpha_sitstand.onnx" \
     --include "policies/alpha_ground_pick.onnx=policies/alpha_ground_pick.onnx" \
+    --include "policies/beak_throw.onnx=policies/beak_throw.onnx" \
     --include "policies/ball_kick_left.onnx=policies/ball_kick_left.onnx" \
     --include "policies/ball_kick_right.onnx=policies/ball_kick_right.onnx" \
     --include "policies/roller.onnx=policies/roller.onnx" \