diff --git a/.config/git/ignore b/.config/git/ignore
index 3f86ce2..06948f5 100644
--- a/.config/git/ignore
+++ b/.config/git/ignore
@@ -11,3 +11,13 @@
 # 9 repos this way and left 2 of them permanently dirty — and a permanently
 # dirty tree is what lets genuinely stranded work hide in the noise.
 .playwright-mcp/
+
+# Per-session agent worktrees. `_claude_autoworktree_enter` creates these in
+# whatever repo a session opens in, so the set of affected repos is "all of
+# them, eventually" — which is the test at the top of this file.
+#
+# It was done the other way first: 21 repos got a .gitignore line, 2 got a
+# .git/info/exclude (machine-local, so invisible to review and to every other
+# checkout), and 2 were missed — leaving solon reported as stranded work for
+# five days over an empty directory. That is the 19th-repo problem, measured.
+**/.claude/worktrees/
diff --git a/.gitignore b/.gitignore
index 6313533..bbb3575 100644
--- a/.gitignore
+++ b/.gitignore
@@ -3,7 +3,13 @@
 __pycache__/
 *.pyc
 .claude/hooks/dangerous-commands.log
-.claude/hooks/fleetcrown-capture.sh
+
+# FleetCrown rewrites its own hooks here on every startup ("Do not edit; Fleet
+# Runner rewrites this file", says the file). Only fleetcrown-capture.sh was
+# listed, so when a second generated hook appeared this tree went permanently
+# dirty — and a permanently dirty tree is where genuinely stranded work hides.
+# Match the family, not the member.
+.claude/hooks/fleetcrown-*.sh
 
 # Per-session agent worktrees — scratch, never part of the repo.
 .claude/worktrees/