fix: the test suite installed a LaunchAgent on the tester's own Mac

docs/STATUS.md

@@ -117,9 +117,19 @@ Everything here needs hardware or a session this branch was written without.
 
 ## Open — older, still true
 
-- **Never exercised, still:** `StopFailure` · `Elicitation` ·
-  permission prompts in the terminal CLI · alert stack compression · click-to-focus
-  for anything other than VS Code.
+- **Never exercised, still:** `StopFailure` · `Elicitation` · alert stack
+  compression · click-to-focus for anything other than VS Code.
+- **Permission prompts in the terminal CLI: verified 2026-08-20.** A CLI session
+  on `gl-login4` was made to edit a tracked source file, `Notification` fired with
+  `notification_type=permission_prompt`, the event reached `events.log`, and the
+  Mac drew the 🔑 alert. That is release 2's one distinct capability — the chat
+  panel has no event for this — confirmed end to end for the first time.
+
+  **Two earlier attempts failed and both were the test's fault, not the code's.**
+  `uptime` and `touch /tmp/…` are auto-approved by Claude Code without raising a
+  permission event at all, so nothing fired and the run looked like a tailbell
+  bug. Anyone re-testing this must pick an action Claude Code will genuinely
+  stop on — editing a tracked file works.
 - The prototype this grew out of still sits in `~/.claude/hooks/` (`cc-notify.sh`,
   `cc-hammerspoon.lua`, `gl-listen.py`, …). It is dead code on a live machine —
   easy to edit by mistake. One part of it is **not** dead: `probe.sh` is still

docs/roadmap.md

@@ -14,7 +14,7 @@ would be wrong.
 | --- | --- | --- | --- |
 | 0 | `v0-ssh-vscode-chatbox` | Claude Code over SSH in the VS Code chat panel → notification on macOS. Zero third-party software or service. | ✅ done |
 | 1 | `v1-rich-presentation` | Centered, clickable overlay: close button, fade in/out, click to focus the originating editor window. Hammerspoon as an **optional** dependency; release 0's zero-dependency path stays the guarantee. | small |
-| 2 | `v2-ssh-terminal-cli` | Claude Code CLI in an SSH terminal. Mostly already working — the addition is `Notification`, which fires in the CLI and covers permission prompts that the chat panel cannot report. | small |
+| 2 | `v2-ssh-terminal-cli` | Claude Code CLI in an SSH terminal. The addition is `Notification`, which fires in the CLI and covers permission prompts the chat panel cannot report. **Verified end to end 2026-08-20** — a permission prompt on `gl-login4` drew the 🔑 alert on the Mac. | ✅ done |
 | 3 | `v3-ssh-any-frontend` | Any editor, terminal or MCP frontend over SSH. Largely documentation and a test matrix rather than new code, for the same reason: hooks are process-level. | small |
 | 4 | `v4-local-and-remote` | Same behaviour whether Claude Code runs on the cluster or on the Mac. The local path already exists; this release is about making one install cover both without branching config. | medium |
 | 5 | `v5-claude-and-codex` | Codex as well as Claude Code. **First real architectural step:** an adapter layer, because each agent's event model differs. Needs a stable internal event schema first. | large |

tests/fake-desktop

@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+# Stands in for anything that would reach the user's actual screen or speakers:
+# osascript, afplay, say. Records the call and returns success.
+#
+# This exists because the suite drew real alerts on a real Mac. On the cluster
+# where most of it was written, the desktop sink is unreachable and every macOS
+# path is dead code, so nothing was visible; on a Mac the same cases pop an
+# overlay and play a sound at whoever is running the tests. A recorder here is
+# the same seam TAILBELL_CURL already provides for the network.
+#
+# FAKE_DESKTOP_RECORD names the log; without it the call is simply swallowed.
+set -u
+if [ -n "${FAKE_DESKTOP_RECORD:-}" ]; then
+  { printf 'CALL|%s|' "$(basename "$0")"
+    for a in ${@+"$@"}; do printf '%s ' "$a"; done
+    printf '\n'
+  } >> "$FAKE_DESKTOP_RECORD" 2>/dev/null
+fi
+exit 0

tests/run-tests.sh

@@ -24,6 +24,15 @@ export TAILBELL_LOG="$TMP/events.log"
 export TAILBELL_DEBUG_LOG="$TMP/debug.log"
 export TAILBELL_STATE_DIR="$TMP/state"
 export TAILBELL_MIN_SECONDS=60
+# Nothing in this suite may reach the screen or the speakers. On the cluster that
+# was free -- every macOS path is dead code there -- so it went unnoticed until
+# the suite was run on a real Mac and popped overlays at the person running it.
+# Same seam as TAILBELL_CURL, for the same reason.
+export TAILBELL_OSASCRIPT="$REPO/tests/fake-desktop"
+export TAILBELL_AFPLAY="$REPO/tests/fake-desktop"
+export TAILBELL_SAY="$REPO/tests/fake-desktop"
+export TAILBELL_RENDERER="osascript"   # never probe for a live Hammerspoon
+export FAKE_DESKTOP_RECORD="$TMP/desktop.log"
 mkdir -p "$TAILBELL_STATE_DIR"
 : > "$TAILBELL_LOG"
 
@@ -392,6 +401,15 @@ export TAILBELL_CHANNELS="file"
 ########################################################################
 echo
 echo "── 声音与语音 ──"
+# The seams must be pointed at the recorder for the whole run, not just inside
+# the cases below. Without this the suite passes identically on the cluster and
+# draws real overlays on a Mac -- which is exactly how it shipped.
+if [ "$TAILBELL_OSASCRIPT" = "$REPO/tests/fake-desktop" ] \
+   && [ "$TAILBELL_AFPLAY" = "$REPO/tests/fake-desktop" ] \
+   && [ "$TAILBELL_SAY" = "$REPO/tests/fake-desktop" ] \
+   && [ -x "$REPO/tests/fake-desktop" ]; then
+  ok "屏幕与扬声器的三个出口全程指向录音器"
+else no "有出口没被拦住 —— 在 Mac 上跑会弹真弹窗"; fi
 #
 # tailbell-show is macOS code, and this suite runs on the cluster. What IS
 # testable here is every decision it makes before touching a Mac API: which
@@ -674,12 +692,30 @@ else no "install.sh 没生成 cluster-id —— 监听器会静默跳过这台
 # point with no coverage at all. Cloned from this working tree at this commit, so
 # it needs no network either.
 OH="$TMP/onelinerhome"; mkdir -p "$OH"
+# ⚠️ The one-liner branches on `uname`, and on Darwin it runs mac/install.sh,
+# which writes a LaunchAgent and `launchctl load`s it. Run unguarded on a Mac
+# this test installed a listener into the tester's own login session, pointed at
+# this throwaway HOME. The temp directory was later cleaned; the loaded job was
+# not, and went on drawing test fixtures onto a real screen for days.
+#
+# So the machine is told it is Linux for this one call. What is under test here
+# is the agent side the one-liner installs -- the workstation installer has its
+# own coverage and must never be executed by the suite.
+mkdir -p "$TMP/unamebin"
+printf '#!/bin/sh\necho Linux\n' > "$TMP/unamebin/uname"
+chmod +x "$TMP/unamebin/uname"
 HOME="$OH" TAILBELL_REPO="$REPO" TAILBELL_DIR="$OH/tailbell" \
   TAILBELL_BRANCH="$(git -C "$REPO" rev-parse HEAD)" \
+  PATH="$TMP/unamebin:$PATH" \
   bash "$REPO/packaging/get-tailbell.sh" >/dev/null 2>&1
 if [ -s "$OH/.tailbell/cluster-id" ] && [ -x "$OH/tailbell/bin/tailbell-notify" ]; then
   ok "一行安装脚本能装出一个完整的 agent 侧"
 else no "一行安装脚本没装成"; fi
+# The bug this guards: a LaunchAgent loaded into the tester's session, pointing
+# at a temp directory, outliving both the test and the directory.
+if [ -z "$(ls -A "$OH/Library/LaunchAgents" 2>/dev/null)" ]; then
+  ok "并且没有往用户会话里装 LaunchAgent"
+else no "测试装了一个 LaunchAgent —— 它会活过这次测试"; fi
 # And that what it installed actually notifies, which is the claim being made.
 s=$(sid o001)
 mkdir -p "$OH/.tailbell/state"; echo $(( $(date +%s) - 400 )) > "$OH/.tailbell/state/$s.start"