diff --git a/openspec/changes/confirm-sender/design.md b/openspec/changes/confirm-sender/design.md
new file mode 100644
index 000000000000..e5f002b21318
--- /dev/null
+++ b/openspec/changes/confirm-sender/design.md
@@ -0,0 +1,17 @@
+## Decisions
+
+**D1 Callback, not credentials.** No peer-credential primitive exists in the runtime, so identity is
+established by asking the claimed sender, who alone holds the answer. A forger cannot make the real
+session report a message it never sent.
+
+**D2 Fail to unattributed, never to attributed.** No answer (timeout, dead socket, unknown id) means
+the message arrives with the sender unverified. Delivery is never blocked by the check.
+
+**D3 The sender log is in memory and bounded** (recent ids, short TTL). It is not a record to audit
+and holds no message text.
+
+**D4 Claude Code senders are out of scope** because they cannot be asked. For them the lead check keeps
+matching pid and address, and docs must call that a claim.
+
+**D5 Replay.** A confirmation names the receiver and the `msg_id`; a sender answers yes only for ids it
+sent to that receiver, so a confirmed id cannot vouch for a different recipient.
diff --git a/openspec/changes/confirm-sender/proposal.md b/openspec/changes/confirm-sender/proposal.md
new file mode 100644
index 000000000000..b5449d3ae1e7
--- /dev/null
+++ b/openspec/changes/confirm-sender/proposal.md
@@ -0,0 +1,35 @@
+# Confirm the sender of a peer message by calling it back
+
+Parent: `autonomous-swarm`; unblocks `review-on-done` Phase 1b. Found in review of the lead
+grant (2026-10-03).
+
+## Why
+
+An inbound peer message names its sender in an envelope attribute (`from`). The receiving
+session turns that attribute into a session id (`resolveOpencodeSender`) and `verifyLead` compares
+it with the lead grant. The attribute is written by the sender, so any local process that can
+reach the target's socket can claim to be the lead. `codec.ts` already says provenance must never
+come from these fields; the code does not follow it. Bun exposes no peer credentials on a unix
+socket (probed on 1.3.14: `getPeerName`, `getPeerCredentials`, `getPeerCert` are all undefined), so
+the transport cannot authenticate the sender.
+
+## What Changes
+
+1. Before a message is attributed to an opencode session, the receiver asks the claimed sender's
+   own sidecar "did you send `msg_id` X to me?" and attributes the message only on a yes.
+2. A sender keeps a short in-process log of the `msg_id`s it sent and to whom, and answers
+   from that log.
+3. An unconfirmed claim is delivered as an unattributed message: text is kept, the sender is
+   shown as unverified, and no lead authority is granted.
+
+## What this does NOT do
+
+This **bounds** sender identity; it does not make it unforgeable.
+
+- A same-user process that can replace a registry entry or a socket can still answer for it. Against
+  a same-user, shell-capable agent nothing short of OS-level separation is unforgeable.
+- It covers opencode senders only. A Claude Code session cannot answer the callback, so a Claude Code
+  lead remains a claim (matched on pid and address). This must be stated wherever a Claude Code lead
+  is described as verified.
+- The real controls stay what they were: a human-confirmed grant, scopes, the publish never-list, and
+  the rule that a review record is written by a tool from a session id it holds, never from message text.
diff --git a/openspec/changes/confirm-sender/specs/peer-sender-identity/spec.md b/openspec/changes/confirm-sender/specs/peer-sender-identity/spec.md
new file mode 100644
index 000000000000..06abd1e35ab3
--- /dev/null
+++ b/openspec/changes/confirm-sender/specs/peer-sender-identity/spec.md
@@ -0,0 +1,13 @@
+## ADDED Requirements
+
+### Requirement: an opencode sender is attributed only after it confirms the message
+The system SHALL attribute an inbound peer message to an opencode session only when that session's own
+sidecar confirms it sent that message id to this receiver.
+
+#### Scenario: forged sender
+- **WHEN** a message claims to be from another session that never sent it
+- **THEN** it is delivered unattributed and grants no lead authority
+
+#### Scenario: no answer
+- **WHEN** the claimed sender does not answer in time
+- **THEN** the message is delivered with the sender marked unverified, not dropped
diff --git a/openspec/changes/confirm-sender/tasks.md b/openspec/changes/confirm-sender/tasks.md
new file mode 100644
index 000000000000..b844d8fba80e
--- /dev/null
+++ b/openspec/changes/confirm-sender/tasks.md
@@ -0,0 +1,12 @@
+## 1. Tests first
+- [ ] 1.1 A forged `from` naming another opencode session is attributed to nobody. Assert both halves: verifyLead (lead.ts:166) denies it, AND the delivery-attribution caller (lifecycle.ts:122) does not treat the unattributed result as authoritative.
+- [ ] 1.2 A genuine message is confirmed and attributed; the lead grant works end to end.
+- [ ] 1.3 No answer / dead socket / wrong receiver all yield unattributed delivery, not a drop.
+
+## 2. Implementation
+- [ ] 2.1 `confirm` frame and handler in the sidecar; bounded sent-log recorded on send. The sent-log key is (msg_id, receiver_session_id) where receiver_session_id is the target the sender dialed at send time — not derived from the return path. The confirm must match both fields; a confirm on the right socket with the wrong receiver is a miss, not a hit.
+- [ ] 2.2 Receiver calls back before `resolveOpencodeSender` result is trusted.
+- [ ] 2.3 Show unverified senders as such to the receiving agent.
+
+## 3. Docs
+- [ ] 3.1 State the same-user residual and the Claude Code limitation in the lead docs and WoW.