diff --git a/AGENTS.md b/AGENTS.md
index be82167..d555307 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -48,6 +48,13 @@ Three layers, kept deliberately separate (see `docs/experiment-architecture.md`
 
 ## Tooling
 - We primarily develop in a dev container and use `uv` for project and package management.
+- If your agent session runs *outside* the dev container (e.g. no Remote-Containers support), don't run training/tests on the bare host. Use `scripts/devcontainer.sh` to drive the existing container instead:
+  - `scripts/devcontainer.sh up` — bring the container up (safe to call anytime; no-op if already running).
+  - `scripts/devcontainer.sh exec -- <cmd>` — run a command inside the container with correct user, cwd, and `remoteEnv` (GPU/CUDA vars). Prefer this over raw `docker exec`, which silently drops `remoteEnv`.
+  - `scripts/devcontainer.sh shell` — open an interactive shell in the container.
+  - `scripts/devcontainer.sh status` / `down` — check or stop the container.
+  - `scripts/devcontainer.sh gpu-check` — verify the GPU is *actually* usable (real CUDA context, not just `nvidia-smi`, which can report healthy while training still fails). `gpu-recover` auto-heals (cheap `docker restart` first, full recreate as fallback).
+  - The script auto-resolves the main worktree checkout the container was built against, so it works correctly even when invoked from a linked `git worktree`.
 
 ## Project Management
 - GitHub Project board: https://github.com/orgs/AmbiqAI/projects/36
diff --git a/scripts/devcontainer.sh b/scripts/devcontainer.sh
new file mode 100755
index 0000000..7693737
--- /dev/null
+++ b/scripts/devcontainer.sh
@@ -0,0 +1,204 @@
+#!/usr/bin/env bash
+# Helper for driving the project's dev container from an agent/session that is
+# running *outside* the container (e.g. a VS Code Remote-Containers window
+# already has it open, or you're on a plain host shell / git worktree).
+#
+# Why this exists:
+#   - `devcontainer` CLI looks up a running container by matching the
+#     *workspace folder path* baked into its labels at creation time. If you
+#     run it from a git worktree (a different path than the one the
+#     container was originally created for), it fails with
+#     "Dev container not found" even though the container is running.
+#   - Plain `docker exec` works from anywhere, but silently drops the
+#     `remoteEnv` values from devcontainer.json (e.g. TF_FORCE_GPU_ALLOW_GROWTH,
+#     CUDA LD_LIBRARY_PATH additions) since it doesn't know about them.
+#
+# This script always resolves the *main* worktree checkout (the one the dev
+# container was actually built against) and passes `--container-id` so
+# commands work correctly regardless of which worktree/directory you're in.
+#
+# Usage:
+#   scripts/devcontainer.sh up                  # bring the container up (idempotent)
+#   scripts/devcontainer.sh id                   # print the running container id
+#   scripts/devcontainer.sh exec -- <cmd> [args] # run a command in the container
+#   scripts/devcontainer.sh shell                # open an interactive shell
+#   scripts/devcontainer.sh down                 # stop the container
+#   scripts/devcontainer.sh status               # show container state
+#   scripts/devcontainer.sh gpu-check            # verify the GPU is actually usable (not just nvidia-smi)
+#   scripts/devcontainer.sh gpu-recover          # auto-heal a broken GPU: restart, else full recreate
+#
+# Examples:
+#   scripts/devcontainer.sh exec -- uv run pytest tests/test_two_stage.py -q
+#   scripts/devcontainer.sh exec -- uv run python scripts/train_rvq_prior.py --help
+#
+# Why gpu-check/gpu-recover instead of just running `nvidia-smi`:
+#   `nvidia-smi` inside the container only proves the NVML library can talk to
+#   the driver — it does NOT prove a CUDA context can actually be created on
+#   the container's own device file descriptors. After a host-side kernel/
+#   driver churn event (e.g. an unattended-upgrades kernel update rebuilding
+#   the NVIDIA DKMS module without a reboot), `nvidia-smi` can keep working
+#   while real framework calls (TensorFlow/PyTorch CUDA init) fail inside an
+#   already-running container. gpu-check does a real minimal CUDA init via
+#   TensorFlow so it catches that failure mode; gpu-recover tries the cheap
+#   fix first (`docker restart`, which re-runs the NVIDIA container-runtime
+#   hook and re-binds fresh device state) before falling back to removing
+#   and recreating the container entirely via `devcontainer up`.
+
+set -euo pipefail
+
+# Resolve the main worktree path: the dev container is created against the
+# primary checkout, not any linked `git worktree add` copy. `git worktree
+# list` always prints the main worktree first.
+resolve_main_worktree() {
+    if git -C "$(dirname "${BASH_SOURCE[0]}")" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+        # Strip only the `worktree ` prefix and print the rest of the line
+        # verbatim (not `{print $2}`, which splits on whitespace and would
+        # silently truncate a workspace path containing spaces).
+        git -C "$(dirname "${BASH_SOURCE[0]}")" worktree list --porcelain \
+            | awk '/^worktree /{sub(/^worktree /, ""); print; exit}'
+    fi
+}
+
+MAIN_WORKSPACE="${DEVCONTAINER_WORKSPACE:-$(resolve_main_worktree)}"
+if [[ -z "${MAIN_WORKSPACE}" ]]; then
+    echo "error: could not resolve main worktree path; set DEVCONTAINER_WORKSPACE" >&2
+    exit 1
+fi
+
+container_id() {
+    local ids
+    ids="$(docker ps -q --filter "label=devcontainer.local_folder=${MAIN_WORKSPACE}")"
+    if [[ -n "${ids}" ]] && [[ "$(wc -l <<< "${ids}")" -gt 1 ]]; then
+        echo "error: multiple running containers match workspace ${MAIN_WORKSPACE}:" >&2
+        echo "${ids}" >&2
+        echo "Disambiguate manually with --container-id, or stop the extras." >&2
+        exit 1
+    fi
+    echo "${ids}"
+}
+
+cmd_up() {
+    devcontainer up --workspace-folder "${MAIN_WORKSPACE}"
+}
+
+cmd_id() {
+    local id
+    id="$(container_id)"
+    if [[ -z "${id}" ]]; then
+        echo "error: no running dev container for ${MAIN_WORKSPACE} (run 'up' first)" >&2
+        exit 1
+    fi
+    echo "${id}"
+}
+
+cmd_exec() {
+    local id
+    id="$(cmd_id)"
+    devcontainer exec --container-id "${id}" --workspace-folder "${MAIN_WORKSPACE}" -- "$@"
+}
+
+cmd_shell() {
+    local id
+    id="$(cmd_id)"
+    devcontainer exec --container-id "${id}" --workspace-folder "${MAIN_WORKSPACE}" -- bash
+}
+
+cmd_down() {
+    local id
+    id="$(container_id)"
+    if [[ -z "${id}" ]]; then
+        echo "no running dev container for ${MAIN_WORKSPACE}"
+        return 0
+    fi
+    docker stop "${id}"
+}
+
+cmd_status() {
+    local id
+    id="$(container_id)"
+    if [[ -z "${id}" ]]; then
+        echo "stopped: ${MAIN_WORKSPACE}"
+        return 0
+    fi
+    docker ps --filter "id=${id}" --format 'table {{.ID}}\t{{.Image}}\t{{.Status}}\t{{.Names}}'
+}
+
+cmd_gpu_check() {
+    local id
+    id="$(cmd_id)"
+    echo "-- nvidia-smi (NVML-level check) --"
+    if ! devcontainer exec --container-id "${id}" --workspace-folder "${MAIN_WORKSPACE}" -- nvidia-smi --query-gpu=name,memory.used --format=csv,noheader; then
+        echo "gpu-check: FAILED at nvidia-smi (driver/NVML not visible in container)" >&2
+        return 1
+    fi
+    echo "-- CUDA context init via TensorFlow (real usability check) --"
+    if ! devcontainer exec --container-id "${id}" --workspace-folder "${MAIN_WORKSPACE}" -- \
+        uv run python -c "
+import tensorflow as tf
+gpus = tf.config.list_physical_devices('GPU')
+assert gpus, 'no GPU visible to TensorFlow'
+with tf.device('/GPU:0'):
+    x = tf.constant([1.0, 2.0]) + tf.constant([3.0, 4.0])
+    x.numpy()
+print('CUDA context OK:', gpus)
+" 2>&1 | tail -20; then
+        echo "gpu-check: FAILED — nvidia-smi worked but a real CUDA context could not be created." >&2
+        echo "This is the 'stale device binding' failure mode; run '$(basename "$0") gpu-recover'." >&2
+        return 1
+    fi
+    echo "gpu-check: OK"
+}
+
+cmd_gpu_recover() {
+    local id
+    id="$(container_id)"
+    if [[ -z "${id}" ]]; then
+        echo "no running container; bringing one up..."
+        cmd_up
+        # Bringing the container up only proves it was created — it says
+        # nothing about whether the GPU/CUDA issue gpu-recover exists to fix
+        # is actually resolved. Verify with gpu-check, same as the other two
+        # recovery paths below, so a caller checking this function's exit
+        # code gets a real answer instead of a false "success".
+        cmd_gpu_check
+        return
+    fi
+    echo "Attempting cheap recovery: docker restart (re-runs the NVIDIA container-runtime hook)..."
+    docker restart "${id}" >/dev/null
+    sleep 3
+    if cmd_gpu_check; then
+        echo "gpu-recover: fixed via docker restart."
+        return 0
+    fi
+    echo "docker restart did not fix it; falling back to a full recreate (stop + remove + devcontainer up)..."
+    # `docker stop` alone is not enough: `devcontainer up` finds and restarts
+    # an existing (even stopped) container for this workspace rather than
+    # creating a fresh one, so the stale device bindings this fallback exists
+    # to clear would otherwise survive. Remove the container so `up` has no
+    # choice but to create a new one from the image.
+    docker stop "${id}" >/dev/null 2>&1 || true
+    docker rm "${id}" >/dev/null 2>&1 || true
+    cmd_up
+    cmd_gpu_check
+}
+
+subcommand="${1:-}"
+[[ $# -gt 0 ]] && shift || true
+
+case "${subcommand}" in
+    up) cmd_up ;;
+    id) cmd_id ;;
+    exec)
+        [[ "${1:-}" == "--" ]] && shift
+        cmd_exec "$@"
+        ;;
+    shell) cmd_shell ;;
+    down) cmd_down ;;
+    status) cmd_status ;;
+    gpu-check) cmd_gpu_check ;;
+    gpu-recover) cmd_gpu_recover ;;
+    *)
+        echo "usage: $(basename "$0") {up|id|exec -- <cmd>|shell|down|status|gpu-check|gpu-recover}" >&2
+        exit 1
+        ;;
+esac