diff --git a/.claude/rules/claim-acquire-before-worktree-work.md b/.claude/rules/claim-acquire-before-worktree-work.md
new file mode 100644
index 000000000..a295bebf7
--- /dev/null
+++ b/.claude/rules/claim-acquire-before-worktree-work.md
@@ -0,0 +1,198 @@
+# Always `claim acquire` before worktree-creating backlog work
+
+Carved sentence:
+
+> When multiple instances of the **same** agent (e.g., Otto-CLI + Otto-Desktop)
+> share git + bus on one machine, **`--from` must differ** (e.g., `otto-cli`
+> vs `otto-desktop`) for the claim-coordinator (`tools/bus/claim.ts`,
+> B-0400 slice 3) to prevent split-brain — identical `--from` values both
+> exit 0 (same-sender idempotent re-acquire). Until the sender-ID schema is
+> extended, use **lane-based convention** as the only real split-brain
+> prevention; branch-prefix is NOT a workaround because `claim acquire`
+> only filters by `from`, not by `branch`. Before starting work on any
+> backlog row, `claim acquire` first. If already claimed by another
+> agent, pick a different row.
+
+## Operational content
+
+The bus claim-coordinator already exists (PR #2939, merged 2026-05-09):
+
+```bash
+# Before any worktree-creating backlog work:
+bun tools/bus/claim.ts acquire --from <agent> --item <B-NNNN> [--branch <ref>]
+# Exit 0 = claim acquired; proceed
+# Exit 1 = already claimed by another agent; pick a different row
+
+# After PR merges or work abandoned:
+bun tools/bus/claim.ts release --from <agent> --item <B-NNNN>
+```
+
+The claim envelope (`topic: "claim", action: "claim"`) lives on the bus with
+24h TTL by default. It is visible to all agents reading
+`/tmp/zeta-bus/` and to all factory surfaces (Otto-CLI, Otto-Desktop,
+Vera-Codex, Riven-Cursor, Lior-Antigravity, Alexa-Kiro).
+
+## When this rule applies
+
+**APPLIES** to:
+
+- Starting work on a `docs/backlog/P*/B-*.md` row (any slice)
+- Creating a feature branch + worktree for that row
+- Opening a PR that closes / advances a backlog row
+
+**DOES NOT APPLY** to:
+
+- Fixing CI failures on an already-claimed PR
+- Resolving review threads on an already-claimed PR
+- Ad-hoc memory-file writes / substrate-honest disclosure responses to
+  Aaron's messages (these are conversation-driven, not backlog-driven)
+- Hot fixes to broken main / rollback PRs (urgency > coordination)
+
+## Composes with other rules
+
+- `.claude/rules/backlog-item-start-gate.md` — already mandates prior-art
+  search + dependency check; this rule adds `claim acquire` as the
+  zero-th step before those checks
+- `.claude/rules/dont-ask-permission.md` — within authority scope, ship;
+  `claim acquire` exit 0 IS the substrate-level permission grant
+- `.claude/rules/never-be-idle.md` — if claim acquire fails, pick a
+  different row from the same priority tier (don't go idle)
+- `.claude/rules/honor-those-that-came-before.md` — claim acquire
+  preserves the work the holder is doing
+
+## Operational examples
+
+### Example 1: Otto-CLI picks B-0444 (which exists)
+
+```bash
+$ bun tools/bus/claim.ts acquire --from otto --item B-0444 --branch otto/b0444-impl-2026-05-13
+$ echo $?
+0
+# Proceeded with worktree creation + impl
+```
+
+### Example 2: Otto-CLI and Otto-Desktop race — KNOWN GAP
+
+```bash
+# Otto-CLI publishes claim first:
+$ bun tools/bus/claim.ts acquire --from otto --item B-0444
+$ echo $?
+0
+
+# Otto-Desktop tries to claim same row with the SAME --from value:
+$ bun tools/bus/claim.ts acquire --from otto --item B-0444
+$ echo $?
+0   # <-- ALSO succeeds! claim.ts treats same-from as idempotent self-re-acquire
+```
+
+**Known architectural gap (caught by Vera 2026-05-13 in review of this
+rule, PR #3032):** `tools/bus/claim.ts` line ~270 filters existing
+claims by `c.from !== sender`, so two callers passing the same
+`--from otto` are indistinguishable to claim.ts. The canonical
+`SENDER_IDS` (`otto`, `alexa`, `riven`, `vera`, `lior`) does NOT
+distinguish multi-surface instances of the same agent.
+
+**Workarounds** (until the schema supports multi-surface sender IDs):
+
+1. **Lane-based convention** (zero-code; the ONLY real split-brain
+   prevention available today): Otto-CLI takes backlog grinding +
+   slice impl; Otto-Desktop takes substrate + cowork. Different
+   scopes, no claim collision possible because the scopes don't
+   overlap.
+2. **Schema extension** (substrate-level fix; future work): add
+   `otto-cli` and `otto-desktop` (and analogous `alexa-cli`/
+   `alexa-kiro`, etc.) to `SENDER_IDS` in `tools/bus/types.ts`. Then
+   `--from otto-desktop` becomes a distinct claim from `--from
+   otto-cli`. THIS is the substrate-level mechanization.
+
+**Branch-prefix is NOT a workaround**: `claim acquire` filters
+existing claims by `c.from !== sender` only, NOT by branch. Two
+Ottos with `--from otto` but different `--branch` values BOTH
+acquire (both exit 0) — the branch field is post-hoc disambiguation
+metadata, not a coordination key. Vera caught this 2026-05-13 on
+PR #3032.
+
+### Example 3: Otto-CLI and Otto-Desktop race (with schema fix)
+
+```bash
+# Otto-CLI publishes claim first:
+$ bun tools/bus/claim.ts acquire --from otto-cli --item B-0444
+$ echo $?
+0
+
+# Otto-Desktop tries to claim same row with DIFFERENT --from:
+$ bun tools/bus/claim.ts acquire --from otto-desktop --item B-0444
+$ echo $?
+1   # Otto-Desktop sees otto-cli's claim, exits 1
+# Otto-Desktop picks B-0445 instead.
+```
+
+This is the target behavior. Today (Example 2) it doesn't work
+because `otto-cli` and `otto-desktop` aren't in `SENDER_IDS` yet.
+
+### Example 4: Otto-CLI crashes mid-work
+
+```bash
+# Otto-CLI process dies. Claim stays on bus with 24h TTL.
+# Otto-Desktop checks:
+$ bun tools/bus/claim.ts check --item B-0444
+# Output: claimed by otto (TTL expires in 23h 59m)
+
+# After 24h the claim auto-expires (TTL). No PID-liveness reclaim exists;
+# only TTL expiry or explicit `release` ends a claim.
+```
+
+## Why this rule exists (operational evidence)
+
+Aaron 2026-05-13 set up Otto on Claude Desktop alongside Otto-CLI. The
+two-foreground-surface architecture creates a real split-brain risk:
+both Ottos might pick the same backlog row simultaneously, leading to
+duplicate work, race conditions, or worse — both committing to the
+same branch with conflicts.
+
+The claim-coordinator (B-0400 slice 3, PR #2939) was built for exactly
+this. Without a rule enforcing its use, both Ottos might forget to
+acquire claims and the split-brain happens. The substrate-honest fix
+is mechanizing the discipline via this rule.
+
+Per `.claude/rules/encoding-rules-without-mechanizing.md`: this rule
+auto-loads at cold-boot so future-Otto reads it before any backlog
+work.
+
+## Composes with substrate
+
+- B-0400 (bus protocol — the schema this rule uses)
+- B-0400 slice 3 (PR #2939 — claim-coordinator implementation)
+- B-0400 slice 5 (PR #2959 — `--with-bus-claims` gate integration)
+- PR #3017 (B-0440.4 — bus publish pattern; same protocol)
+- `docs/research/2026-05-12-otto-canonical-bootstream-multi-foreground-surface-orchestrator-ifs-format.md`
+  (the multi-foreground-surface design this rule operationalizes)
+- `docs/launch/2026-05-13-otto-claude-desktop-bootstream-tight.md`
+  (PR #3030 — Claude Desktop variant; second Otto surface)
+- `memory/feedback_aaron_good_failure_mode_git_fetch_before_push_catches_multi_agent_duplicate_work_2026_05_13.md`
+  (sibling fetch-before-push discipline; same coordination pattern at
+  git scope)
+
+## Substrate-honest framing
+
+This rule is not a strict requirement at the code level — the
+claim-coordinator returns a useful exit code, but no automated hook
+enforces calling it. The discipline relies on agent-side compliance
+at cold-boot.
+
+A future slice could add a pre-commit / pre-worktree hook that calls
+`claim check` automatically and fails if no claim is held. That's
+substrate-level mechanization. Today's rule is the discipline-level
+mechanization.
+
+## Full reasoning
+
+PR #2939 (B-0400 slice 3 — claim-coordinator implementation)
+PR #2959 (B-0400 slice 5 — bus-gate integration)
+PR #3017 (B-0440.4 — bus publish pattern)
+PR #3030 (Otto Claude Desktop tight bootstream — second Otto surface)
+
+Aaron 2026-05-13 verbatim: *"probalby want to figure out how not to
+split brain with yourself bot any idea?"* — substrate-honest naming
+of the split-brain risk; this rule is the operationally-honest
+answer.