Hook-enforced ownership for AI coding session handoffs.
Most "handoff" tools solve amnesia: capture state to a markdown file, restore it after compaction or a new session. That problem is well covered. This one solves the problem nobody enforces: concurrent clobber. When two sessions work the same repo, or you resume on a second machine, or a background agent runs alongside an interactive one, they overwrite each other's handoff notes and you do not find out until the context you needed is gone.
The fix here is not a better template. It is a PreToolUse hook that makes a cross-session overwrite structurally blocked, not merely discouraged.
Every handoff file's first line is an ownership marker:
<!-- claude-session: 9e0d3802-... -->
There is no sidecar .lock
file. Ownership travels with the artifact through git, across devices,
through a mv
. A PreToolUse hook reads the calling session's id and compares it to the marker in the content being written and the marker already on disk. Mismatch blocks the write.
To write a handoff that says who wrote it, the session needs to know its own id. It does not. The
model has no native access to its session_id
.
So the first write of a fresh handoff is designed to fail. The block reason carries the id:
Handoff write missing or wrong ownership marker.
Your session_id: `9e0d3802-4f...`.
Prepend exactly this as line 1:
<!-- claude-session: 9e0d3802-4f... -->
Then retry.
The model copies the id from the failure and retries. One block per fresh handoff, and the file is now self-identifying for every future session. The missing capability becomes a one-time handshake.
A model that is blocked on Write
will route around you. So the guard covers every way a file can be mutated:
Write validates the marker in the new content.Edit validates the marker on disk (and blocks edits to legacy marker-less files until you take ownership with a Write).Bash matches shell redirects to handoff paths (>
,>>
,tee
,sed -i
) and blocks unowned writes that try to sneak past the file tools.
It also accepts both the Claude Code tool schema (Write
/Edit
/Bash
) and the Gemini CLI schema
(write_file
/replace
/run_shell_command
) in one hook, because gating on one silently disables the guard for the other client.
hooks/handoff-write-guard.mjs PreToolUse: the ownership guard
hooks/handoff-session-start.mjs SessionStart: surface existing handoffs + slug overlaps
hooks/handoff-stop-gate.mjs Stop: once-per-session "you have no handoff" nudge
hooks/pre-commit-staged-marker-check.mjs git pre-commit: block commits mixing two sessions' handoffs
hooks/test/handoff-write-guard.test.mjs node --test suite (8 cases)
scripts/handoff-migrate-archive.mjs archive stale, marker-less legacy handoffs
scripts/install-git-hooks.sh per-device installer for the pre-commit hook
skills/handoff/SKILL.md the /handoff slash command
rules/session-handoff.md the convention the hooks enforce
settings.example.json hook wiring to merge into ~/.claude/settings.json
Handoffs are expected under the standard Claude Code memory layout:
~/.claude/projects/<encoded-cwd>/memory/handoff-<branch>-<topic>.md
, where <encoded-cwd>
is the
absolute working directory with /
, \
, and .
replaced by -
.
cp hooks/*.mjs ~/.claude/hooks/
cp -r hooks/test ~/.claude/hooks/
cp scripts/* ~/.claude/scripts/
cp -r skills/handoff ~/.claude/skills/
cp rules/* ~/.claude/rules/
bash ~/.claude/scripts/install-git-hooks.sh
node --test ~/.claude/hooks/test/*.test.mjs
Every hook wraps its body in try/catch and exits 0 on any internal error. A bug in the guard
degrades to convention; it never bricks a session. The Stop nudge is non-blocking and fires at most
once per session. The escape hatches (touch /tmp/handoff-guard-bypass-<file>
, or
HANDOFF_GUARD_BYPASS=1
) exist precisely because a structural guard you cannot override becomes a structural guard you rip out. Bypass use is logged so silent disabling is auditable.
This makes the unaware clobber impossible. It does not make the chosen one impossible.
- A session correctly shown a foreign marker can still archive the file or set the bypass env. Human review of the visible block message is the backstop for that class.
- The
<branch>
token in a filename is not checked against the real branch (intentional: cross-device resume deliberately inherits a foreign branch's topic). - A TOCTOU race exists if two sessions create the same newfilename in the gap between the hook's read and the tool's write. Vanishingly rare for solo dev; deliberately not locked.
MIT.