diff --git a/apps/skills/claude/plannotator-annotate/SKILL.md b/apps/skills/claude/plannotator-annotate/SKILL.md
new file mode 100644
index 00000000..a91967ec
--- /dev/null
+++ b/apps/skills/claude/plannotator-annotate/SKILL.md
@@ -0,0 +1,20 @@
+---
+name: plannotator-annotate
+description: Open Plannotator's annotation UI for a markdown file, converted HTML file, URL, or folder and then respond to the returned annotations.
+allowed-tools: Bash(plannotator:*)
+disable-model-invocation: true
+---
+
+# Plannotator Annotate
+
+## Markdown annotations
+
+!`plannotator annotate $ARGUMENTS`
+
+## Your task
+
+The output above will be one of:
+
+1. The exact text `The user approved.`, OR a JSON object with `"decision": "approved"`. The user approved the markdown file(s). Acknowledge with a single sentence ("Approved.") and stop. Do not begin any work.
+2. Empty, OR a JSON object with `"decision": "dismissed"`. The user closed the session without requesting changes. Acknowledge with a single sentence ("Annotation session closed.") and stop. Do not begin any work.
+3. Plaintext annotation feedback, OR a JSON object with `"decision": "annotated"` and a `"feedback"` field. Address the feedback. The user has reviewed the markdown file(s) and provided specific annotations and comments.
diff --git a/apps/skills/claude/plannotator-archive/SKILL.md b/apps/skills/claude/plannotator-archive/SKILL.md
new file mode 100644
index 00000000..d09ebe25
--- /dev/null
+++ b/apps/skills/claude/plannotator-archive/SKILL.md
@@ -0,0 +1,16 @@
+---
+name: plannotator-archive
+description: Browse saved plan decisions in Plannotator's read-only archive UI in the browser.
+allowed-tools: Bash(plannotator:*)
+disable-model-invocation: true
+---
+
+# Plannotator Archive
+
+## Plan archive
+
+!`plannotator archive`
+
+## Your task
+
+The user has finished browsing the plan archive. Acknowledge and continue.
diff --git a/apps/skills/claude/plannotator-last/SKILL.md b/apps/skills/claude/plannotator-last/SKILL.md
new file mode 100644
index 00000000..ddd4170d
--- /dev/null
+++ b/apps/skills/claude/plannotator-last/SKILL.md
@@ -0,0 +1,20 @@
+---
+name: plannotator-last
+description: Open Plannotator on the latest rendered assistant message and use the returned annotations to revise that message or continue.
+allowed-tools: Bash(plannotator:*)
+disable-model-invocation: true
+---
+
+# Plannotator Last
+
+## Message annotations
+
+!`plannotator annotate-last $ARGUMENTS`
+
+## Your task
+
+The output above will be one of:
+
+1. The exact text `The user approved.`, OR a JSON object with `"decision": "approved"`. The user approved your last message. Acknowledge with a single sentence ("Approved.") and stop. Do not begin any work.
+2. Empty, OR a JSON object with `"decision": "dismissed"`. The user closed the session without requesting changes. Acknowledge with a single sentence ("Annotation session closed.") and stop. Do not begin any work.
+3. Plaintext annotation feedback, OR a JSON object with `"decision": "annotated"` and a `"feedback"` field. Address the feedback. The user has reviewed your last message and provided specific annotations and comments.
diff --git a/apps/skills/claude/plannotator-review/SKILL.md b/apps/skills/claude/plannotator-review/SKILL.md
new file mode 100644
index 00000000..d5865538
--- /dev/null
+++ b/apps/skills/claude/plannotator-review/SKILL.md
@@ -0,0 +1,16 @@
+---
+name: plannotator-review
+description: Open Plannotator's browser-based code review UI for the current worktree or a pull request URL, then act on the feedback that comes back.
+allowed-tools: Bash(plannotator:*)
+disable-model-invocation: true
+---
+
+# Plannotator Review
+
+## Code review feedback
+
+!`plannotator review $ARGUMENTS`
+
+## Your task
+
+If the review above contains feedback or annotations, address them in the same conversation. If no changes were requested (an approval/LGTM-style result), acknowledge that review passed and continue.
diff --git a/scripts/install.cmd b/scripts/install.cmd
index 804962f8..0c655953 100644
--- a/scripts/install.cmd
+++ b/scripts/install.cmd
@@ -716,20 +716,30 @@ if !ERRORLEVEL! equ 0 (
     pushd "!SKILLS_TMP!\repo"
     git sparse-checkout set apps/skills apps/kiro-cli apps/opencode-plugin/commands apps/gemini/commands >nul 2>&1
 
-    REM Core skills -> Claude + shared agent scope (all 4, copy-if-present).
-    if exist "apps\skills\core" (
+    REM Claude Code reads apps\skills\claude\* (injection `!`plannotator … $ARGUMENTS``
+    REM + allowed-tools, so /plannotator-* run with no permission prompt); Codex
+    REM reads apps\skills\core\* (prose). The `!`…`` injection is Claude-Code-only,
+    REM so the two are sourced separately. Replace rather than merge on each run.
+    if exist "apps\skills\claude" (
         if not exist "!CLAUDE_SKILLS_DIR!" mkdir "!CLAUDE_SKILLS_DIR!"
+        for %%S in (plannotator-review plannotator-annotate plannotator-last plannotator-archive) do (
+            if exist "apps\skills\claude\%%S" (
+                if exist "!CLAUDE_SKILLS_DIR!\%%S" rmdir /s /q "!CLAUDE_SKILLS_DIR!\%%S" >nul 2>&1
+                xcopy /s /i /y /q "apps\skills\claude\%%S" "!CLAUDE_SKILLS_DIR!\%%S\" >nul 2>&1
+            )
+        )
+        echo Installed Claude Code skills to !CLAUDE_SKILLS_DIR!\
+    )
+    if exist "apps\skills\core" (
         if not exist "!AGENTS_SKILLS_DIR!" mkdir "!AGENTS_SKILLS_DIR!"
         for %%S in (plannotator-review plannotator-annotate plannotator-last plannotator-archive) do (
             if exist "apps\skills\core\%%S" (
                 REM Replace rather than merge so files removed upstream don't linger.
-                if exist "!CLAUDE_SKILLS_DIR!\%%S" rmdir /s /q "!CLAUDE_SKILLS_DIR!\%%S" >nul 2>&1
                 if exist "!AGENTS_SKILLS_DIR!\%%S" rmdir /s /q "!AGENTS_SKILLS_DIR!\%%S" >nul 2>&1
-                xcopy /s /i /y /q "apps\skills\core\%%S" "!CLAUDE_SKILLS_DIR!\%%S\" >nul 2>&1
                 xcopy /s /i /y /q "apps\skills\core\%%S" "!AGENTS_SKILLS_DIR!\%%S\" >nul 2>&1
             )
         )
-        echo Installed core skills to !CLAUDE_SKILLS_DIR!\ and !AGENTS_SKILLS_DIR!\
+        echo Installed shared agent skills to !AGENTS_SKILLS_DIR!\
     ) else (
         echo Tag !TAG! predates the core/extra skill layout — skipping core skill install
     )
diff --git a/scripts/install.ps1 b/scripts/install.ps1
index 977b3e9b..f17a60e1 100644
--- a/scripts/install.ps1
+++ b/scripts/install.ps1
@@ -644,20 +644,32 @@ try {
         try {
             git sparse-checkout set apps/skills apps/kiro-cli apps/opencode-plugin/commands apps/gemini/commands 2>$null
 
-            # Core skills -> ~/.claude/skills and ~/.agents/skills (all 4).
+            # Claude Code and Codex consume different skill bodies. Claude Code
+            # reads apps/skills/claude/* (dynamic-context injection
+            # `!`plannotator … $ARGUMENTS`` + allowed-tools, so /plannotator-*
+            # run with no permission prompt — like the old slash commands).
+            # Codex reads apps/skills/core/* (prose the model follows via its
+            # own shell). The `!`…`` injection is a Claude-Code-only extension,
+            # so the two are sourced separately rather than sharing one body.
             # Route each through Copy-SkillIfPresent (which pre-removes the
-            # existing target dir) so re-runs replace rather than nest,
-            # matching install.sh's per-skill structure.
-            if ((Test-Path "apps\skills\core") -and (Get-ChildItem "apps\skills\core" -ErrorAction SilentlyContinue)) {
+            # existing target dir) so re-runs replace rather than nest.
+            if ((Test-Path "apps\skills\claude") -and (Get-ChildItem "apps\skills\claude" -ErrorAction SilentlyContinue)) {
                 New-Item -ItemType Directory -Force -Path $claudeSkillsDir | Out-Null
+                foreach ($skill in @("plannotator-review", "plannotator-annotate", "plannotator-last", "plannotator-archive")) {
+                    Copy-SkillIfPresent "apps\skills\claude\$skill" $claudeSkillsDir
+                }
+                Write-Host "Installed Claude Code skills to $claudeSkillsDir\"
+            } else {
+                Write-Host "Tag $latestTag predates the per-agent skill layout — skipping Claude Code skill install"
+            }
+            if ((Test-Path "apps\skills\core") -and (Get-ChildItem "apps\skills\core" -ErrorAction SilentlyContinue)) {
                 New-Item -ItemType Directory -Force -Path $agentsSkillsDir | Out-Null
                 foreach ($skill in @("plannotator-review", "plannotator-annotate", "plannotator-last", "plannotator-archive")) {
-                    Copy-SkillIfPresent "apps\skills\core\$skill" $claudeSkillsDir
                     Copy-SkillIfPresent "apps\skills\core\$skill" $agentsSkillsDir
                 }
-                Write-Host "Installed core skills to $claudeSkillsDir\ and $agentsSkillsDir\"
+                Write-Host "Installed shared agent skills to $agentsSkillsDir\"
             } else {
-                Write-Host "Tag $latestTag predates the core/extra skill layout — skipping core skill install"
+                Write-Host "Tag $latestTag predates the core/extra skill layout — skipping shared agent skill install"
             }
 
             # Kiro: hand-maintained skills (origin baked in) + two extras.
diff --git a/scripts/install.sh b/scripts/install.sh
index 6ce1c257..4c3053d7 100644
--- a/scripts/install.sh
+++ b/scripts/install.sh
@@ -730,13 +730,29 @@ if command -v glimpseui >/dev/null 2>&1; then
     glimpse_present=1
 fi
 
-# A wizard needs a real keyboard. Piped installs (curl | bash) still have a
-# terminal at /dev/tty even though stdin is the pipe; CI and scripts do not.
+# A wizard needs a real human at the keyboard. Piped installs (curl | bash)
+# still have a terminal at /dev/tty even though stdin is the pipe; CI and
+# scripts do not. Some automated contexts (docker run -t, devcontainer and
+# provisioner shells) DO expose an openable /dev/tty with nobody behind it —
+# opening /dev/tty succeeds, yet a read would block forever. The per-prompt
+# timeout below (see PROMPT_TIMEOUT / ask_yes_no) handles that: a mis-detected
+# terminal falls through to the safe non-interactive defaults (extras=no,
+# model-invocable=none, glimpse=no) instead of wedging. We deliberately do NOT
+# gate on $CI here — an exported CI var must not silently suppress an explicit
+# --reconfigure or --extras in an otherwise interactive shell.
 can_prompt=0
 if [ "$NON_INTERACTIVE" -eq 0 ] && { : < /dev/tty; } 2>/dev/null; then
     can_prompt=1
 fi
 
+# Bound every interactive read so an unattended-but-open /dev/tty auto-takes
+# the default rather than hanging. Set PLANNOTATOR_PROMPT_TIMEOUT=0 to wait
+# indefinitely (restores the old unbounded behavior); non-numeric falls to 30.
+PROMPT_TIMEOUT="${PLANNOTATOR_PROMPT_TIMEOUT:-30}"
+case "$PROMPT_TIMEOUT" in
+    ''|*[!0-9]*) PROMPT_TIMEOUT=30 ;;
+esac
+
 run_wizard=0
 if [ "$can_prompt" -eq 1 ]; then
     if [ "$RECONFIGURE" -eq 1 ] || [ ! -f "$PREFS_FILE" ]; then
@@ -746,11 +762,32 @@ fi
 
 # Ask a y/n question on the terminal. $1 prompt, $2 default (yes/no).
 ask_yes_no() {
-    local prompt="$1" default="$2" answer suffix
+    local prompt="$1" default="$2" answer suffix rc
     suffix="[y/N]"
     [ "$default" = "yes" ] && suffix="[Y/n]"
     printf '%s %s ' "$prompt" "$suffix" > /dev/tty
-    IFS= read -r answer < /dev/tty || answer=""
+    # Bounded read so an unattended-but-open /dev/tty (e.g. docker run -t with
+    # no human) can't hang the install. Distinguish a human pressing Enter
+    # (read succeeds with an empty answer -> use the prompt's $default) from a
+    # timeout/EOF with nobody there (read fails -> use the SAFE "no", never the
+    # default). Otherwise a prompt whose default is "yes" (Glimpse) would
+    # silently install software on an unattended terminal.
+    # Keep the read in a tested context (`|| rc=$?`) so the read itself never
+    # trips `set -e` (active at the top of this script), without relying on the
+    # subtle rule that -e is suppressed inside a function called in a tested
+    # context. ask_yes_no still returns non-zero on timeout/EOF to signal "no
+    # human", so every caller consumes it with `|| wizard_timed_out=1`.
+    rc=0
+    if [ "$PROMPT_TIMEOUT" -gt 0 ]; then
+        IFS= read -r -t "$PROMPT_TIMEOUT" answer < /dev/tty || rc=$?
+    else
+        IFS= read -r answer < /dev/tty || rc=$?
+    fi
+    if [ "$rc" -ne 0 ]; then
+        printf '\n' > /dev/tty
+        echo "no"
+        return 1
+    fi
     case "$answer" in
         y|Y|yes|YES|Yes) echo "yes" ;;
         n|N|no|NO|No)    echo "no" ;;
@@ -811,6 +848,9 @@ select_skills_checkbox() {
 extras_choice=""
 invocable_choice=""
 glimpse_choice=""
+# Set if any wizard prompt times out / hits EOF (no human answered). A run whose
+# answers are synthetic timeout fallbacks must not be persisted as install-prefs.
+wizard_timed_out=0
 
 if [ "$run_wizard" -eq 1 ]; then
     {
@@ -827,7 +867,7 @@ if [ "$run_wizard" -eq 1 ]; then
         # Flag already answered this question — don't ask and then ignore.
         extras_choice="$EXTRAS_FLAG"
     else
-        extras_choice=$(ask_yes_no "Install the extra skills (compound planning, setup-goal, visual explainer)?" "${saved_extras:-no}")
+        extras_choice=$(ask_yes_no "Install the extra skills (compound planning, setup-goal, visual explainer)?" "${saved_extras:-no}") || wizard_timed_out=1
     fi
     invocable_list="$CORE_SKILL_NAMES"
     if [ "$extras_choice" = "yes" ]; then
@@ -837,7 +877,7 @@ if [ "$run_wizard" -eq 1 ]; then
         # Flag already answered this question — don't ask and then ignore.
         invocable_choice="$MODEL_INVOCABLE_FLAG"
     else
-        want_invocable=$(ask_yes_no "Make any skills callable by the model (instead of user-invoked only)?" "no")
+        want_invocable=$(ask_yes_no "Make any skills callable by the model (instead of user-invoked only)?" "no") || wizard_timed_out=1
         if [ "$want_invocable" = "yes" ]; then
             invocable_choice=$(select_skills_checkbox "$invocable_list" "$saved_invocable")
         else
@@ -851,7 +891,7 @@ if [ "$run_wizard" -eq 1 ]; then
         # Flag already answered this question — don't ask and then ignore.
         glimpse_choice="$GLIMPSE_FLAG"
     else
-        glimpse_choice=$(ask_yes_no "Install Glimpse so Plannotator opens in a native window instead of a browser tab? (npm i -g glimpseui)" "${saved_glimpse:-yes}")
+        glimpse_choice=$(ask_yes_no "Install Glimpse so Plannotator opens in a native window instead of a browser tab? (npm i -g glimpseui)" "${saved_glimpse:-yes}") || wizard_timed_out=1
     fi
 fi
 
@@ -863,9 +903,11 @@ fi
 [ -z "$invocable_choice" ] && invocable_choice="${saved_invocable:-none}"
 [ -z "$glimpse_choice" ] && glimpse_choice="${saved_glimpse:-no}"
 
-# Persist only when the wizard ran or a flag set something — silent re-runs
-# must not clobber saved answers with defaults.
-if [ "$run_wizard" -eq 1 ] || [ -n "$EXTRAS_FLAG" ] || [ -n "$MODEL_INVOCABLE_FLAG" ] || [ -n "$GLIMPSE_FLAG" ]; then
+# Persist only when the wizard ran with real answers, or a flag set something.
+# Silent re-runs must not clobber saved answers with defaults, and a wizard that
+# timed out to synthetic fallbacks (unattended /dev/tty) must not become sticky
+# prefs that suppress the wizard on a later genuine interactive install.
+if [ "$wizard_timed_out" -eq 0 ] && { [ "$run_wizard" -eq 1 ] || [ -n "$EXTRAS_FLAG" ] || [ -n "$MODEL_INVOCABLE_FLAG" ] || [ -n "$GLIMPSE_FLAG" ]; }; then
     mkdir -p "$_config_dir"
     {
         echo "extras=$extras_choice"
@@ -969,19 +1011,33 @@ checkout_failed=0
     # via --version may predate the core/extra layout — skip core skills
     # but keep installing the command files below (matches install.ps1 and
     # install.cmd, which guard each block independently).
+    # Claude Code and Codex consume different skill bodies. Claude Code reads
+    # the apps/skills/claude/* copies, which use dynamic-context injection
+    # (`!`plannotator … $ARGUMENTS``) + allowed-tools so /plannotator-* run the
+    # binary directly with no permission prompt — matching the old slash
+    # commands. Codex (the OpenAI shared-agent path) reads apps/skills/core/*,
+    # whose prose bodies the model follows via its own shell; the `!`…``
+    # injection is a Claude-Code-only extension, so the two are sourced
+    # separately rather than sharing one body.
+    if [ -d "apps/skills/claude" ] && [ -n "$(ls -A apps/skills/claude 2>/dev/null)" ]; then
+        mkdir -p "$CLAUDE_SKILLS_DIR"
+        copy_skill_if_present apps/skills/claude/plannotator-review "$CLAUDE_SKILLS_DIR"
+        copy_skill_if_present apps/skills/claude/plannotator-annotate "$CLAUDE_SKILLS_DIR"
+        copy_skill_if_present apps/skills/claude/plannotator-last "$CLAUDE_SKILLS_DIR"
+        copy_skill_if_present apps/skills/claude/plannotator-archive "$CLAUDE_SKILLS_DIR"
+        echo "Installed Claude Code skills to ${CLAUDE_SKILLS_DIR}/"
+    else
+        echo "Tag ${latest_tag} predates the per-agent skill layout — skipping Claude Code skill install"
+    fi
     if [ -d "apps/skills/core" ] && [ -n "$(ls -A apps/skills/core 2>/dev/null)" ]; then
-        mkdir -p "$CLAUDE_SKILLS_DIR" "$AGENTS_SKILLS_DIR"
-        copy_skill_if_present apps/skills/core/plannotator-review "$CLAUDE_SKILLS_DIR"
-        copy_skill_if_present apps/skills/core/plannotator-annotate "$CLAUDE_SKILLS_DIR"
-        copy_skill_if_present apps/skills/core/plannotator-last "$CLAUDE_SKILLS_DIR"
-        copy_skill_if_present apps/skills/core/plannotator-archive "$CLAUDE_SKILLS_DIR"
+        mkdir -p "$AGENTS_SKILLS_DIR"
         copy_skill_if_present apps/skills/core/plannotator-review "$AGENTS_SKILLS_DIR"
         copy_skill_if_present apps/skills/core/plannotator-annotate "$AGENTS_SKILLS_DIR"
         copy_skill_if_present apps/skills/core/plannotator-last "$AGENTS_SKILLS_DIR"
         copy_skill_if_present apps/skills/core/plannotator-archive "$AGENTS_SKILLS_DIR"
-        echo "Installed core skills to ${CLAUDE_SKILLS_DIR}/ and shared agent skills to ${AGENTS_SKILLS_DIR}/"
+        echo "Installed shared agent skills to ${AGENTS_SKILLS_DIR}/"
     else
-        echo "Tag ${latest_tag} predates the core/extra skill layout — skipping core skill install"
+        echo "Tag ${latest_tag} predates the core/extra skill layout — skipping shared agent skill install"
     fi
 
     # OpenCode slash command stubs (the plugin intercepts execution) —
diff --git a/scripts/install.test.ts b/scripts/install.test.ts
index 9626d400..429a199b 100644
--- a/scripts/install.test.ts
+++ b/scripts/install.test.ts
@@ -78,9 +78,12 @@ describe("install.sh", () => {
     expect(script).toContain("AGENTS_SKILLS_DIR");
     expect(script).toContain("$HOME/.agents/skills");
     expect(script).toContain("copy_skill_if_present");
-    // Core skills (all 4) -> Claude Code and the official OpenAI shared-agent path.
+    // Claude Code reads the injection-form skills from apps/skills/claude;
+    // the OpenAI shared-agent (Codex) path reads the prose skills from
+    // apps/skills/core. Sourced separately because `!`…`` injection is a
+    // Claude-Code-only extension.
     for (const skill of CORE_SKILLS) {
-      expect(script).toContain(`copy_skill_if_present apps/skills/core/${skill} "$CLAUDE_SKILLS_DIR"`);
+      expect(script).toContain(`copy_skill_if_present apps/skills/claude/${skill} "$CLAUDE_SKILLS_DIR"`);
       expect(script).toContain(`copy_skill_if_present apps/skills/core/${skill} "$AGENTS_SKILLS_DIR"`);
     }
     // Codex no longer receives a skills install (core skills live in ~/.agents/skills).
@@ -100,7 +103,7 @@ describe("install.sh", () => {
     // fetch or an old pinned tag never deletes commands without replacement.
     expect(script).toContain('if [ -d "$CLAUDE_SKILLS_DIR/$cmd" ] && [ -f "$CLAUDE_COMMANDS_DIR/$cmd.md" ]');
     const cleanupIndex = script.indexOf('Removed legacy Claude command');
-    const installIndex = script.indexOf('copy_skill_if_present apps/skills/core/plannotator-review');
+    const installIndex = script.indexOf('copy_skill_if_present apps/skills/claude/plannotator-review');
     expect(installIndex).toBeGreaterThan(0);
     expect(cleanupIndex).toBeGreaterThan(installIndex);
   });
@@ -343,10 +346,11 @@ describe("install.ps1", () => {
     expect(script).toContain("agentsSkillsDir");
     expect(script).toContain("$env:USERPROFILE\\.agents\\skills");
     expect(script).toContain("Copy-SkillIfPresent");
-    // Core skills (all 4) copied verbatim to both Claude and the shared-agent
-    // scope, per-skill via Copy-SkillIfPresent so re-runs replace rather than
-    // nest (PowerShell's Copy-Item -Recurse into an existing dir nests).
-    expect(script).toContain('Copy-SkillIfPresent "apps\\skills\\core\\$skill" $claudeSkillsDir');
+    // Claude Code reads injection-form skills (apps\skills\claude); the
+    // shared-agent (Codex) scope reads the prose skills (apps\skills\core).
+    // Per-skill via Copy-SkillIfPresent so re-runs replace rather than nest
+    // (PowerShell's Copy-Item -Recurse into an existing dir nests).
+    expect(script).toContain('Copy-SkillIfPresent "apps\\skills\\claude\\$skill" $claudeSkillsDir');
     expect(script).toContain('Copy-SkillIfPresent "apps\\skills\\core\\$skill" $agentsSkillsDir');
     expect(script).toContain('"plannotator-review", "plannotator-annotate", "plannotator-last", "plannotator-archive"');
     // Copy-SkillIfPresent pre-removes the destination to avoid nesting on upgrade.
@@ -467,8 +471,9 @@ describe("install.cmd", () => {
     expect(script).toContain("CLAUDE_SKILLS_DIR");
     expect(script).toContain("AGENTS_SKILLS_DIR");
     expect(script).toContain("%USERPROFILE%\\.agents\\skills");
-    // Core skills (all 4) copied to both Claude and shared-agent scope.
-    expect(script).toContain('xcopy /s /i /y /q "apps\\skills\\core\\%%S" "!CLAUDE_SKILLS_DIR!\\%%S\\"');
+    // Claude Code reads injection-form skills (apps\skills\claude); the shared
+    // agent (Codex) scope reads the prose skills (apps\skills\core).
+    expect(script).toContain('xcopy /s /i /y /q "apps\\skills\\claude\\%%S" "!CLAUDE_SKILLS_DIR!\\%%S\\"');
     expect(script).toContain('xcopy /s /i /y /q "apps\\skills\\core\\%%S" "!AGENTS_SKILLS_DIR!\\%%S\\"');
     expect(script).toContain("for %%S in (plannotator-review plannotator-annotate plannotator-last plannotator-archive) do");
     // No Codex skills install — only the cleanup loop references CODEX skills.
@@ -632,8 +637,17 @@ describe("install shared behavior", () => {
     expect(cmdScript).toContain("timeout /t 0");
     expect(cmdScript).toContain('if "!CAN_PROMPT!"=="1"');
     expect(cmdScript).toContain("set /p");
-    // Silent re-runs must not clobber saved answers with defaults.
-    expect(sh).toContain('if [ "$run_wizard" -eq 1 ] || [ -n "$EXTRAS_FLAG" ] || [ -n "$MODEL_INVOCABLE_FLAG" ] || [ -n "$GLIMPSE_FLAG" ]');
+    // Silent re-runs must not clobber saved answers with defaults, and a wizard
+    // that timed out to synthetic fallbacks (unattended /dev/tty) must not be
+    // persisted — ask_yes_no returns non-zero on timeout/EOF, each prompt ORs
+    // that into wizard_timed_out, and the prefs write is gated on it.
+    expect(sh).toContain('if [ "$wizard_timed_out" -eq 0 ] && { [ "$run_wizard" -eq 1 ] || [ -n "$EXTRAS_FLAG" ] || [ -n "$MODEL_INVOCABLE_FLAG" ] || [ -n "$GLIMPSE_FLAG" ]; }');
+    expect(sh).toContain("wizard_timed_out=0");
+    expect(sh).toContain("|| wizard_timed_out=1");
+    expect(sh).toMatch(/echo "no"\s+return 1/);
+    // The bounded read stays in a tested context (`|| rc=$?`) so `set -e` never
+    // aborts ask_yes_no on a timeout/EOF, regardless of how it's called.
+    expect(sh).toContain('< /dev/tty || rc=$?');
     expect(ps).toContain("if ($runWizard -or $Extras -or $NoExtras -or $ModelInvocable -or $Glimpse -or $NoGlimpse)");
     expect(cmdScript).toContain('if "!DO_PERSIST!"=="1"');
     // Glimpse question: detect-on-PATH skip + global npm install in all three.