diff --git a/.devcontainer/.env.example b/.devcontainer/.env.example
index 09ae4c9..253f56a 100644
--- a/.devcontainer/.env.example
+++ b/.devcontainer/.env.example
@@ -25,6 +25,13 @@
 # Uncomment to raise the log level of the setup scripts.
 # MUSHER_LOG_LEVEL=debug
 
-# Skip the AI CLI installs on a slow connection.
+# AI harnesses. Claude Code is the default and installs when nothing here is
+# set; Codex is opt-in, because it is an npm package carrying a platform binary
+# and it dominates the time a cold container spends in postCreateCommand.
+#
+# Set CLAUDE=0 to skip the install on a slow connection, CODEX=1 to add Codex.
+# These belong here and nowhere else: devcontainer.json deliberately does not
+# put them in containerEnv, because `docker run -e` outranks --env-file and
+# would override whatever you write below.
 # MUSHER_INSTALL_CLAUDE=0
-# MUSHER_INSTALL_CODEX=0
+# MUSHER_INSTALL_CODEX=1
diff --git a/.devcontainer/devcontainer.json b/.devcontainer/devcontainer.json
index b1be113..fe4d1d4 100644
--- a/.devcontainer/devcontainer.json
+++ b/.devcontainer/devcontainer.json
@@ -59,10 +59,14 @@
 
     "XDG_CACHE_HOME": "/home/vscode/.cache",
     "NPM_CONFIG_CACHE": "/home/vscode/.cache/npm",
-    "BUN_INSTALL_CACHE_DIR": "/home/vscode/.cache/bun",
-
-    "MUSHER_INSTALL_CLAUDE": "1",
-    "MUSHER_INSTALL_CODEX": "1"
+    "BUN_INSTALL_CACHE_DIR": "/home/vscode/.cache/bun"
+
+    // MUSHER_INSTALL_CLAUDE and MUSHER_INSTALL_CODEX are deliberately NOT set
+    // here. containerEnv becomes `docker run -e`, and an explicit -e beats
+    // --env-file for the same name whichever order they appear in — so pinning
+    // them to "1" here silently overrode the `.devcontainer/.env` opt-out that
+    // .env.example tells developers to use. The scripts default to on when the
+    // variable is unset, which leaves .env as the single place that decides.
   },
 
   "remoteEnv": {
@@ -72,17 +76,42 @@
   // Host-side: creates .devcontainer/.env from the template and strips CRLF so
   // --env-file above has a valid target on a fresh clone. Run through `bash -c`
   // so CRLF in the script files cannot break the strip-and-run bootstrap.
+  //
+  // `perl -i`, not `sed -i`. This one runs on the host, and BSD sed — every
+  // macOS host — reads the next argument as the backup suffix, so the strip
+  // exited non-zero and never happened; `2>/dev/null` then hid that it hadn't.
   "initializeCommand": [
     "bash",
     "-c",
-    "find .devcontainer/scripts -name '*.sh' -exec sed -i 's/\\r$//' {} + 2>/dev/null; bash .devcontainer/scripts/initialize.sh"
+    "find .devcontainer/scripts -name '*.sh' -exec perl -i -pe 's/\\r$//' {} + 2>/dev/null; bash .devcontainer/scripts/initialize.sh"
   ],
 
-  "waitFor": "postCreateCommand",
-  "postCreateCommand": {
-    "fix-crlf": "find .devcontainer/scripts -type f -name '*.sh' -exec sed -i 's/\\r$//' {} +",
-    "setup": ["bash", ".devcontainer/scripts/post-create.sh"]
-  },
+  // Connect as soon as the container exists, not when provisioning ends.
+  //
+  // postCreateCommand still runs to completion and still reports failure — it
+  // simply no longer gates the VS Code connection. That matters because the
+  // cold path is genuinely minutes long (the Claude Code installer dominates
+  // it), and waiting for it meant a developer stared at a connecting window
+  // with no output for the entire install.
+  //
+  // The trade-off, stated plainly: a terminal opened in the first couple of
+  // minutes will not have task, lefthook or claude on PATH yet. They appear as
+  // mise finishes — `remoteEnv` already points PATH at the shim directory, so
+  // no reload is needed, only a new shell.
+  "waitFor": "onCreateCommand",
+
+  // ONE string, deliberately — not the object form this used to be.
+  //
+  // An object-form lifecycle command runs its entries in PARALLEL and buffers
+  // each entry's output until that entry exits. As two named entries, this pair
+  // therefore did neither of the things it looks like it does: the CRLF strip
+  // raced post-create.sh rather than preceding it (both started in the same
+  // millisecond), and the setup entry printed nothing at all for the several
+  // minutes the cold tool install takes — which is indistinguishable from a
+  // hung container, and is what sent a developer looking for a deadlock that
+  // was not there. `&&` restores the ordering, and a string command streams its
+  // output line by line while it runs.
+  "postCreateCommand": "find .devcontainer/scripts -type f -name '*.sh' -exec sed -i 's/\\r$//' {} + && bash .devcontainer/scripts/post-create.sh",
   "postStartCommand": ["bash", ".devcontainer/scripts/startup.sh"],
 
   "customizations": {
diff --git a/.devcontainer/mise.toml b/.devcontainer/mise.toml
index ba28470..6481a45 100644
--- a/.devcontainer/mise.toml
+++ b/.devcontainer/mise.toml
@@ -19,6 +19,10 @@
 # Docs: https://mise.jdx.dev
 
 [tools]
+# @openai/codex is pinned here but NOT installed by default: it is opt-in via
+# MUSHER_INSTALL_CODEX=1 in .devcontainer/.env, which the setup scripts turn
+# into MISE_DISABLE_TOOLS. The pin stays recorded so enabling it is a one-line
+# change that still resolves to a known version.
 "actionlint"        = "1.7.11"
 "task"              = "3.52.0"
 "npm:@openai/codex" = "0.143.0"
diff --git a/.devcontainer/scripts/initialize.sh b/.devcontainer/scripts/initialize.sh
index 2b4b333..2529df2 100644
--- a/.devcontainer/scripts/initialize.sh
+++ b/.devcontainer/scripts/initialize.sh
@@ -39,11 +39,16 @@ ensure_env_file() {
   fi
 }
 
+# `perl -i`, not `sed -i`: this runs on the HOST, and BSD sed — every macOS
+# host — reads the next argument as the backup suffix, so `sed -i 's/\r$//' f`
+# consumes the expression and then treats f as the script. It exits non-zero,
+# which under `set -e` would take the whole initializeCommand down with it.
+# perl -i means the same thing on both host families.
 strip_crlf() {
   [[ -f "${ENV_FILE}" ]] || return 0
   if grep -q $'\r' "${ENV_FILE}" 2>/dev/null; then
     log "Stripping CRLF from .devcontainer/.env"
-    sed -i 's/\r$//' "${ENV_FILE}"
+    perl -i -pe 's/\r$//' "${ENV_FILE}"
   fi
 }
 
diff --git a/.devcontainer/scripts/lib/base-setup.sh b/.devcontainer/scripts/lib/base-setup.sh
index 4573143..da9d1a3 100644
--- a/.devcontainer/scripts/lib/base-setup.sh
+++ b/.devcontainer/scripts/lib/base-setup.sh
@@ -95,7 +95,8 @@ base_install_mise() {
     return 0
   fi
   log "Installing mise (https://mise.run)..."
-  retry 3 5 bash -c 'curl -fsSL https://mise.run | sh'
+  retry 3 5 bounded 300 bash -c \
+    'curl -fsSL --connect-timeout 10 --max-time 120 https://mise.run | sh'
 }
 
 # Installs the CLIs pinned in .devcontainer/mise.toml (tools with no Feature),
@@ -112,9 +113,22 @@ base_install_tools() {
   local mise
   mise="$(command -v mise || echo "${_MISE_BIN}")"
   local config="${MISE_GLOBAL_CONFIG_FILE:-${_LIB_DIR}/../../mise.toml}"
+
+  # Claude Code is this repository's default harness; Codex is opt-IN. Codex is
+  # an npm package carrying a platform binary and it dominates the cold-
+  # container install, so a developer who does not use it should not pay for it
+  # on every rebuild. MISE_DISABLE_TOOLS drops it from the resolved set without
+  # editing the manifest, so the pin stays recorded either way and CI still
+  # sees it — set MUSHER_INSTALL_CODEX=1 in .devcontainer/.env to install it.
+  if ! install_wanted "${MUSHER_INSTALL_CODEX:-0}"; then
+    log "MUSHER_INSTALL_CODEX is off — leaving the Codex CLI out of this install"
+    export MISE_DISABLE_TOOLS="npm:@openai/codex${MISE_DISABLE_TOOLS:+,${MISE_DISABLE_TOOLS}}"
+  fi
+
   log "Installing pinned CLIs from ${config}..."
+  debug "MISE_DISABLE_TOOLS=${MISE_DISABLE_TOOLS:-<unset>}"
   "${mise}" trust "${config}" >/dev/null 2>&1 || true
-  retry 3 5 "${mise}" install
+  retry 3 5 bounded 900 "${mise}" install
   "${mise}" reshim >/dev/null 2>&1 || true
 }
 
@@ -127,12 +141,19 @@ base_install_tools() {
 # Returns:
 #   0 on success, non-zero on failure
 base_install_claude() {
+  if ! install_wanted "${MUSHER_INSTALL_CLAUDE:-1}"; then
+    log "MUSHER_INSTALL_CLAUDE is off — skipping Claude Code"
+    return 0
+  fi
   if has_cmd claude; then
     log "Claude Code already installed, skipping"
     return 0
   fi
   log "Installing Claude Code (native installer)..."
-  retry 3 5 bash -c 'curl -fsSL https://claude.ai/install.sh | bash'
+  # install.sh itself downloads a platform binary of its own, so the ceiling has
+  # to cover the whole pipeline rather than just the fetch of the script.
+  retry 3 5 bounded 900 bash -c \
+    'curl -fsSL --connect-timeout 10 --max-time 120 https://claude.ai/install.sh | bash'
 }
 
 # --- Verify ---
@@ -149,7 +170,13 @@ base_install_claude() {
 # Returns:
 #   0 if all tools found, 1 if any are missing
 base_verify_tools() {
-  verify_tools gh task codex lefthook claude actionlint shellcheck
+  # Verify what this container was actually asked to install. Checking a CLI the
+  # developer deliberately switched off would report a self-inflicted failure.
+  local -a tools=(gh task lefthook actionlint shellcheck)
+  if install_wanted "${MUSHER_INSTALL_CODEX:-0}"; then tools+=(codex); fi
+  if install_wanted "${MUSHER_INSTALL_CLAUDE:-1}"; then tools+=(claude); fi
+  debug "verifying: ${tools[*]}"
+  verify_tools "${tools[@]}"
 }
 
 # --- Orchestrator ---
@@ -167,6 +194,14 @@ base_setup() {
   base_install_mise
   base_install_tools
   base_install_claude
-  base_verify_tools
+
+  # Verification reports; it no longer aborts. `set -e` used to let one missing
+  # CLI end post-create.sh right here, which skipped the repo-specific setup
+  # that follows it — git hooks and `bun install` — and left the container in a
+  # worse state than the single missing tool warranted. The status rides out on
+  # the return code instead, so the failure is still loud.
+  local verified=0
+  base_verify_tools || verified=$?
   log "Base setup complete"
+  return "${verified}"
 }
diff --git a/.devcontainer/scripts/lib/common.sh b/.devcontainer/scripts/lib/common.sh
index e6bb13c..cb77be0 100644
--- a/.devcontainer/scripts/lib/common.sh
+++ b/.devcontainer/scripts/lib/common.sh
@@ -20,6 +20,23 @@ log() {
   echo "[$(date '+%H:%M:%S')] $*" >&2
 }
 
+# Logs a message only when MUSHER_LOG_LEVEL=debug.
+#
+# .env.example advertises MUSHER_LOG_LEVEL as the way to raise the setup
+# scripts' log level. Nothing read it, so the switch was documentation for a
+# feature that did not exist. This is that feature.
+#
+# Globals:
+#   MUSHER_LOG_LEVEL — read, defaults to "info"
+# Arguments:
+#   $@ — message text
+# Outputs:
+#   Writes to stderr via log() when debug is on, nothing otherwise
+debug() {
+  [[ "${MUSHER_LOG_LEVEL:-info}" == "debug" ]] || return 0
+  log "DEBUG: $*"
+}
+
 # --- Command helpers ---
 
 # Checks whether a command exists on the PATH.
@@ -32,6 +49,48 @@ has_cmd() {
   command -v "$1" &>/dev/null
 }
 
+# Reports whether an install guarded by a MUSHER_INSTALL_* switch is wanted.
+#
+# devcontainer.json declares MUSHER_INSTALL_CLAUDE and MUSHER_INSTALL_CODEX and
+# .env.example documents them as the way to "skip the AI CLI installs on a slow
+# connection". No script read either one, so both CLIs installed unconditionally
+# — and they are most of the several minutes a cold container spends inside
+# postCreateCommand. This is what makes the documented switch real.
+#
+# Arguments:
+#   $1 — the switch's value, e.g. "${MUSHER_INSTALL_CLAUDE:-1}"
+# Returns:
+#   0 when the install is wanted, 1 when it is switched off
+install_wanted() {
+  case "${1:-1}" in
+    0 | false | no | off) return 1 ;;
+    *) return 0 ;;
+  esac
+}
+
+# Runs a command under a wall-clock ceiling, when `timeout` is available.
+#
+# Every network install below used to run unbounded. A stalled connection parks
+# postCreateCommand for as long as the kernel keeps the socket, and a lifecycle
+# command shows no output until it exits, so an unbounded install is
+# indistinguishable from a hung container. A ceiling turns "hangs forever" into
+# "fails, and the container is still usable".
+#
+# Arguments:
+#   $1 — ceiling in seconds
+#   $@ — command and arguments to execute
+# Returns:
+#   The command's status, or 124 if the ceiling was reached
+bounded() {
+  local seconds="${1:?usage: bounded <seconds> <command...>}"
+  shift
+  if has_cmd timeout; then
+    timeout --foreground "${seconds}" "$@"
+  else
+    "$@"
+  fi
+}
+
 # Runs a command with sudo if available, otherwise without.
 #
 # Arguments:
diff --git a/.devcontainer/scripts/post-create.sh b/.devcontainer/scripts/post-create.sh
index 93faf24..d6fdec9 100644
--- a/.devcontainer/scripts/post-create.sh
+++ b/.devcontainer/scripts/post-create.sh
@@ -67,10 +67,18 @@ install_spec_tools() {
 #   Writes progress to stderr via log()
 main() {
   log "Starting post-create setup..."
-  base_setup
+  # base_setup reports a missing tool through its status rather than aborting,
+  # so the repo-specific steps below still run. Carry the status to the exit
+  # code so a half-provisioned container is still reported as one.
+  local status=0
+  base_setup || status=$?
   install_lefthook_hooks
   # --- Repo-specific setup ---
   install_spec_tools
+  if ((status != 0)); then
+    log "Post-create setup completed with MISSING TOOLS (see the ✗ lines above)"
+    return "${status}"
+  fi
   log "Post-create setup completed"
 }