Mirrored verbatim from utopia-models/knowledge _rules/every-tree-holds-another-agents-work.md by the allowlisted publish pipeline. Edit it there, not here.

name: every-tree-holds-another-agents-work
cases: [2026-07-01-agent-always-add-all, 2026-07-01-agent-never-reset-hard-dirty-tree, 2026-07-01-agent-sync-with-ff-only-never-reset-hard, 2026-07-01-dmv-verify-branch-in-commit-command, 2026-08-01-agent-feature-branch-at-start, 2026-08-01-agent-no-second-shell-assumption, 2026-08-01-agent-spec-creates-branch, 2026-08-01-dmv-check-shared-clone-dirty-before-writing, 2026-08-01-dmv-check-sibling-branches-before-fixing, 2026-08-09-a-silent-failure-path-is-worse-than-a-missing-one, 2026-08-09-i-am-not-a-master-agent-and-tyler-is-not-waiting, 2026-08-10-tyler-runs-four-shells-not-one, 2026-08-28-an-edit-from-a-stale-base-silently-reverts-the-fix, 2026-10-04-stash-and-rebase-in-the-shared-knowledge-clone, 2026-10-16-check-the-asset-against-tylers-source-and-cd-into-the-worktree-in-the-same-chain]
status: active
mechanism: pretooluse: git reset --hard | git stash in a clone with >1 known writer
audience: all
last-recurrence: 2026-10-16

Every working tree holds another agent's uncommitted work. Branch first, never reset --hard.

Fires when

About to git reset --hard, git stash, sync a branch, stage by filename, or write into a shared clone — and any time work starts on main/dev instead of a feature branch. Also when explaining an unfamiliar diff as "left over from a previous session". 🔴 The same applies to a LINE. Editing a shared file from the version in my head reverted a peer's fix (18f8c91 restored a jq filter dd090a6 had removed); a sed on a stale pattern and on the current line look identical, and fixtures for THIS change are blind to what it also did. Re-read the exact region immediately before rewriting, git diff the hunk asking what else did I change, read the file's recent commit MESSAGES. A live mechanism armed from pre-fix text keeps the bug.

Trigger detail (full text — kept out of the always-loaded bundle)

About to git reset --hard, git stash, sync a branch, stage by filename, or write into a shared clone — and any time work is starting on main/dev instead of a feature branch. Also any time I am about to explain an unfamiliar diff to myself as "left over from a previous session".

🔴 AND THE SAME RULE APPLIES TO A LINE, NOT ONLY TO A TREE (2026-08-28). Editing a shared file from the version in my head silently reverted a peer's fix: 18f8c91 rewrote a jq expression to add a glyph and carried back a select(.author.login != "kxagent") filter that dd090a6 had deliberately removed, because my memory of the line predated theirs. A sed that matches a stale pattern and one that matches the current line look identical in their output, and fixtures written to prove THIS change are structurally blind to what the change also did. Re-read the exact region immediately before rewriting it, git diff the hunk to ask what else did I change, and read the recent commit MESSAGES on that file first — dd090a6's said exactly what the filter breaks. And a live mechanism armed from pre-fix text keeps the pre-fix bug: the running copy does not update itself.

The rule

🔴 Tyler runs 2–4 concurrent Architect sessions as his normal working style (his words, three times in five days: "I have 4 shells open and I am planning and executing with all"). Assume N live sessions in every shared clone. A file I did not write is another live shell's work, not residue — the opposite default, from a narrow 2026-05-22 correction, was wrong for three months and pointed at a /second-shell skill that has never existed.

Create and check out the feature branch myself at the start of work, and again when I spec something new. Sync with merge --ff-only; reset --hard silently discards a concurrent master's WIP and untracked files. Never git stash in a shared clone — the stash stack is global, so one shell pops another's work into its tree. git add -A so Tyler's parallel edits ride along: that is his standing correction (2026-05-31, case 2026-07-01-agent-always-add-all), because he wants his in-flight changes committed rather than stranded dirty, and .gitignore keeps secrets out. Including another shell's work is the safe direction; discarding it is not. Before writing into a clone with no per-run worktree, read git status and git branch -vv — a sibling dispatch may already be mid-fix. A checkout in one Bash call may not still be HEAD in the next, so verify the branch in the same command that commits.

Session identity is $CLAUDE_PID, exported into every child process. When something must be attributed to this shell, key it on that — never on recency, proximity, or "the newest thing".

Cost when it fired

The author of the never-reset-hard-dirty-tree rule broke it on 2026-07-07: git reset --hard origin/main on the knowledge clone wiped an untracked FEEDBACK/ directory that had never pushed. Nine cases across agent, dmv and cloud.