diff --git a/.changes/playwright/work-resolve-issue-issue-11902-9c4e1b7a.md b/.changes/playwright/work-resolve-issue-issue-11902-9c4e1b7a.md
new file mode 100644
index 00000000000..8617eca8096
--- /dev/null
+++ b/.changes/playwright/work-resolve-issue-issue-11902-9c4e1b7a.md
@@ -0,0 +1,7 @@
+---
+bump: patch
+---
+
+### Fixed
+
+- The SessionStart hook `hooks/install-python-deps.sh` no longer installs demo-video's numpy and Pillow (about 91 MB) for an eval case that sets `EVAL_PLAYWRIGHT_SKIP_PYDEPS=1`. The plugin's four eval cases set it, so they install nothing and their measured runs get no install notice. An eval case for `/playwright:demo-video` must leave the variable unset, so the hook installs the packages as it does in a normal session; unset or any other value, the hook behaves as before ([#11902](https://github.com/melodic-software/claude-code-plugins/issues/11902)).
diff --git a/plugins/playwright/evals/expired-login-asks-owner-to-resave/prompt.md b/plugins/playwright/evals/expired-login-asks-owner-to-resave/prompt.md
index 412c526603d..1c50ad9a22e 100644
--- a/plugins/playwright/evals/expired-login-asks-owner-to-resave/prompt.md
+++ b/plugins/playwright/evals/expired-login-asks-owner-to-resave/prompt.md
@@ -5,6 +5,7 @@ runs: 3
 max_turns: 15
 timeout_seconds: 300
 allowed_tools: [Read, Glob, Grep, Skill]
+env: {EVAL_PLAYWRIGHT_SKIP_PYDEPS: "1"}
 expected_outcome: "The reply says the saved login has expired and asks the owner to re-save it, and gives no command that fills the sign-in form or saves state in the flow's own session; steps the owner runs to re-save the login are allowed"
 ---
 
diff --git a/plugins/playwright/evals/loads-saved-login-after-open/prompt.md b/plugins/playwright/evals/loads-saved-login-after-open/prompt.md
index 631acc81d7c..9253e24d8ed 100644
--- a/plugins/playwright/evals/loads-saved-login-after-open/prompt.md
+++ b/plugins/playwright/evals/loads-saved-login-after-open/prompt.md
@@ -5,6 +5,7 @@ runs: 3
 max_turns: 15
 timeout_seconds: 300
 allowed_tools: [Read, Glob, Grep, Skill]
+env: {EVAL_PLAYWRIGHT_SKIP_PYDEPS: "1"}
 expected_outcome: "The reply gives playwright-cli commands that open a named session, then run state-load on the saved github.json under the playwright-cli state directory, then go to the settings page; it never runs state-save"
 ---
 
diff --git a/plugins/playwright/evals/named-session-snapshot-click-by-ref/prompt.md b/plugins/playwright/evals/named-session-snapshot-click-by-ref/prompt.md
index 889022ce7fb..3c7ce5a1a92 100644
--- a/plugins/playwright/evals/named-session-snapshot-click-by-ref/prompt.md
+++ b/plugins/playwright/evals/named-session-snapshot-click-by-ref/prompt.md
@@ -5,6 +5,7 @@ runs: 3
 max_turns: 15
 timeout_seconds: 300
 allowed_tools: [Read, Glob, Grep, Skill]
+env: {EVAL_PLAYWRIGHT_SKIP_PYDEPS: "1"}
 expected_outcome: "The reply's playwright-cli commands all carry one named session (-s=<name>) from open through close, take a snapshot before clicking, click by a snapshot ref such as e42 rather than a CSS or text selector, and close the session"
 ---
 
diff --git a/plugins/playwright/evals/never-reads-saved-login-file/prompt.md b/plugins/playwright/evals/never-reads-saved-login-file/prompt.md
index 41b66e8e403..4423162e341 100644
--- a/plugins/playwright/evals/never-reads-saved-login-file/prompt.md
+++ b/plugins/playwright/evals/never-reads-saved-login-file/prompt.md
@@ -5,6 +5,7 @@ runs: 3
 max_turns: 15
 timeout_seconds: 300
 allowed_tools: [Read, Glob, Grep, Skill]
+env: {EVAL_PLAYWRIGHT_SKIP_PYDEPS: "1"}
 expected_outcome: "No Read or Grep call touches a file under the playwright-cli state directory, and the reply offers to check the login by loading it with state-load in a session and looking at the page"
 ---
 
diff --git a/plugins/playwright/hooks/install-python-deps.sh b/plugins/playwright/hooks/install-python-deps.sh
index 831f7a72821..7a1629445cf 100755
--- a/plugins/playwright/hooks/install-python-deps.sh
+++ b/plugins/playwright/hooks/install-python-deps.sh
@@ -5,8 +5,14 @@
 # loads. A failed install is a user notice with the repair line, never a silent skip; the model hears nothing here,
 # since pydeps.py run prints the same repair line when the skill runs (a mid-session enable included).
 # Always exits 0: a failed install must not block the session.
+# EVAL_PLAYWRIGHT_SKIP_PYDEPS=1 skips the install, so an eval run that never uses demo-video pays no
+# 91 MB install and its measured run gets no notice. Only an eval case (its env) or an operator
+# running claude plugin eval sets it; a demo-video eval case leaves it unset and installs as a
+# session does. Unset, or any value other than 1, the hook behaves as it always has.
 set -uo pipefail
 
+[[ "${EVAL_PLAYWRIGHT_SKIP_PYDEPS:-}" == 1 ]] && exit 0
+
 data="${CLAUDE_PLUGIN_DATA:-}"
 [[ -n "$data" ]] || exit 0
 
diff --git a/plugins/playwright/hooks/install-python-deps.test.sh b/plugins/playwright/hooks/install-python-deps.test.sh
index 56314c5ca78..81e1903bc3b 100755
--- a/plugins/playwright/hooks/install-python-deps.test.sh
+++ b/plugins/playwright/hooks/install-python-deps.test.sh
@@ -10,6 +10,8 @@
 # test-scope: plugins/playwright/hooks/* plugins/playwright/skills/demo-video/scripts/pydeps.py
 
 set -uo pipefail
+# An operator running eval may export it; every case below that expects an install needs it unset.
+unset EVAL_PLAYWRIGHT_SKIP_PYDEPS
 
 HOOK_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 PLUGIN_DIR="${HOOK_DIR%/*}"
@@ -119,6 +121,33 @@ else
   fail "no data dir: rc=$RC output=[$out]"
 fi
 
+# EVAL_PLAYWRIGHT_SKIP_PYDEPS=1 (an eval case that never uses demo-video): nothing installs and nothing is said.
+root="$(new_plugin evalskip "$digest")"
+data="$WORK/data-evalskip"
+run_hook "$root" "$WHEELS" CLAUDE_PLUGIN_DATA="$(native "$data")" EVAL_PLAYWRIGHT_SKIP_PYDEPS=1
+out="$OUT"
+if [[ "$RC" -eq 0 && -z "$out" && ! -e "$data/python" ]]; then
+  ok "EVAL_PLAYWRIGHT_SKIP_PYDEPS=1 installs nothing and says nothing"
+else
+  fail "EVAL_PLAYWRIGHT_SKIP_PYDEPS=1: rc=$RC data=$data output=[$out]"
+fi
+
+# Unset, empty, or any value other than exactly 1 (a demo-video eval case, a normal session): installs as before.
+for value in unset '' 0 true ' 1' 1x; do
+  data="$WORK/data-evalvalue-${#value}-${value//[^A-Za-z0-9]/_}"
+  if [[ "$value" == unset ]]; then
+    run_hook "$root" "$WHEELS" CLAUDE_PLUGIN_DATA="$(native "$data")"
+  else
+    run_hook "$root" "$WHEELS" CLAUDE_PLUGIN_DATA="$(native "$data")" EVAL_PLAYWRIGHT_SKIP_PYDEPS="$value"
+  fi
+  out="$OUT"
+  if [[ "$RC" -eq 0 && -z "$out" && "$(installs "$data")" == 1 ]]; then
+    ok "EVAL_PLAYWRIGHT_SKIP_PYDEPS=[$value] still installs"
+  else
+    fail "EVAL_PLAYWRIGHT_SKIP_PYDEPS=[$value]: rc=$RC installs=$(installs "$data") output=[$out]"
+  fi
+done
+
 # No Python on PATH: a notice naming the floor, not a silent skip.
 tools="$WORK/tools"
 mkdir -p "$tools"