diff --git a/.github/workflows/review-swarm.yml b/.github/workflows/review-swarm.yml
new file mode 100644
index 000000000..f8862a721
--- /dev/null
+++ b/.github/workflows/review-swarm.yml
@@ -0,0 +1,95 @@
+name: Review swarm
+
+on:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+permissions:
+  contents: read
+  pull-requests: write
+
+concurrency:
+  group: review-swarm-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  review:
+    # Per RFC-0001 §2 rule 7: "Every PR is met by a review swarm." No author
+    # filter — human PRs get reviewed too. If we ever need a rollout-scoped
+    # narrower filter, express it as a documented rule, not a hardcoded
+    # username list.
+    runs-on: ubuntu-latest
+    # Job cap must exceed the poll deadline (65min) + install overhead.
+    # 75min gives 10min headroom for checkout, npm install, sync, post.
+    timeout-minutes: 75
+    env:
+      GH_TOKEN: ${{ github.token }}
+      PR_NUMBER: ${{ github.event.pull_request.number }}
+      RELAY_WORKSPACE_KEY: ${{ secrets.RELAY_WORKSPACE_KEY }}
+    steps:
+      - name: Check out PR merge commit
+        uses: actions/checkout@v4
+        with:
+          ref: refs/pull/${{ github.event.pull_request.number }}/merge
+          fetch-depth: 0
+
+      - name: Install agent-relay
+        run: npm install --global agent-relay@11.8.7
+
+      - name: Set review target
+        run: echo "$PR_NUMBER" > .review-target
+
+      - name: Launch review swarm
+        id: launch
+        shell: bash
+        run: |
+          set -euo pipefail
+          response=$(agent-relay cloud run workflows/review-swarm.yaml --json)
+          printf '%s\n' "$response"
+          run_id=$(jq -er '.runId // .id // .run.id' <<<"$response")
+          echo "run_id=$run_id" >> "$GITHUB_OUTPUT"
+
+      - name: Wait for review swarm
+        id: wait
+        shell: bash
+        env:
+          RUN_ID: ${{ steps.launch.outputs.run_id }}
+        # Poll deadline (3900s = 65min) must be >= workflows/review-swarm.yaml's
+        # own timeoutMs (3600000 = 60min) plus buffer, or CI abandons swarms
+        # still executing in cloud. The job's timeout-minutes must exceed 65.
+        # Record the terminal status in an output so the next step can post
+        # transcripts EVEN when the swarm rejects — otherwise a rejecting
+        # swarm's evidence never reaches the PR.
+        run: |
+          set -euo pipefail
+          deadline=$((SECONDS + 3900))
+          while (( SECONDS < deadline )); do
+            response=$(agent-relay cloud status "$RUN_ID" --json)
+            printf '%s\n' "$response"
+            status=$(jq -er '.status // .run.status' <<<"$response")
+            case "$status" in
+              completed|failed|errored|cancelled|canceled)
+                echo "swarm_status=$status" >> "$GITHUB_OUTPUT"
+                exit 0 ;;
+            esac
+            sleep 30
+          done
+          echo "swarm_status=timeout" >> "$GITHUB_OUTPUT"
+
+      - name: Sync and post review comments
+        # ALWAYS run, even if wait recorded a non-completed status. A rejecting
+        # swarm's transcripts and marker are the evidence we need on the PR;
+        # skipping this step on failure is exactly the evidence-loss class
+        # ops/DRIVE-LOG.md warns about.
+        if: always() && steps.launch.outputs.run_id != ''
+        env:
+          RUN_ID: ${{ steps.launch.outputs.run_id }}
+        run: bash .github/workflows/scripts/swarm-post.sh "$RUN_ID" "$PR_NUMBER"
+
+      - name: Fail the job when swarm rejected
+        # Post-transcript gate: the marker comment now records the verdict on
+        # the PR, so we can safely fail the CI check to block merge.
+        if: steps.wait.outputs.swarm_status != 'completed'
+        run: |
+          echo "Swarm terminated with status: ${{ steps.wait.outputs.swarm_status }}" >&2
+          exit 1
diff --git a/.github/workflows/scripts/swarm-post.sh b/.github/workflows/scripts/swarm-post.sh
new file mode 100755
index 000000000..b2f70dd8c
--- /dev/null
+++ b/.github/workflows/scripts/swarm-post.sh
@@ -0,0 +1,93 @@
+#!/usr/bin/env bash
+# Sync a completed cloud review-swarm run and post its transcripts + marker
+# back to the PR.
+#
+# Two load-bearing extraction rules (both hard-won repo history):
+#   1. SORT transcripts by FILENAME (starts with YYYYMMDD-HHMM), not by mtime.
+#      Fresh checkouts give transcripts uniform mtimes and mtime-sort picked
+#      stale verdicts — commit b2535aa fixed this class of bug elsewhere.
+#   2. The verdict is the LAST non-empty line's token, not a whole-file grep.
+#      A whole-file grep of REVIEW_FAILED misclassifies a passing review that
+#      quotes the token in prose — commit f59d9cd fixed this elsewhere too.
+# Same for the aggregate: LAST SWARM_ token in the log, not a substring match.
+#
+# Sticky marker: hidden HTML comment identifies the marker; edit-in-place
+# across re-runs so N pushes don't accumulate N marker comments.
+set -euo pipefail
+
+run_id=${1:-}
+pr_number=${2:-}
+
+[[ "$run_id" =~ ^[[:alnum:]-]+$ ]] || {
+  echo "usage: $0 <runId> <PR-number>" >&2
+  exit 2
+}
+[[ "$pr_number" =~ ^[0-9]+$ ]] || {
+  echo "usage: $0 <runId> <PR-number>" >&2
+  exit 2
+}
+
+agent-relay cloud sync "$run_id"
+run_log=$(agent-relay cloud logs "$run_id")
+printf '%s\n' "$run_log"
+
+# Aggregate: last SWARM_ token in the log (not a substring anywhere).
+overall=FAILED
+last_swarm=$(printf '%s\n' "$run_log" | grep -Eo 'SWARM_(PASSED|FAILED)' | tail -1 || true)
+[[ "$last_swarm" == SWARM_PASSED ]] && overall=PASSED
+
+declare -A verdicts
+for lens in maintainability history structure; do
+  # Sort lexicographically by filename (YYYYMMDD-HHMM prefix), take newest.
+  shopt -s nullglob
+  matches=(ops/reviews/*-pr"$pr_number"-"$lens".md)
+  shopt -u nullglob
+  if ((${#matches[@]} == 0)); then
+    echo "missing $lens transcript for PR #$pr_number" >&2
+    verdicts[$lens]=MISSING
+    continue
+  fi
+  transcript=$(printf '%s\n' "${matches[@]}" | sort | tail -1)
+
+  # Verdict = last non-empty line's token.
+  last_line=$(awk 'NF { last=$0 } END { print last }' "$transcript")
+  case "$last_line" in
+    *REVIEW_PASSED*) verdicts[$lens]=PASSED ;;
+    *REVIEW_FAILED*) verdicts[$lens]=FAILED ;;
+    *) verdicts[$lens]=UNCLEAR ;;
+  esac
+
+  gh pr comment "$pr_number" --body-file "$transcript"
+done
+
+# Compute the aggregate from lens verdicts DIRECTLY, not from cloud logs. The
+# logs-derived `overall` above is a first pass but can disagree with the
+# actual transcripts (log parsing missed a lens, aggregate step raced, etc).
+# The transcripts are the load-bearing evidence — a single FAILED lens means
+# aggregate FAILED, per the review-swarm.yaml aggregate step's own rule
+# ("any single honest refusal blocks the merge").
+#
+# Compute aggregate = ALL lenses PASSED; else FAILED. This is fail-closed:
+# MISSING, UNCLEAR, FAILED all degrade to FAILED. The prior bug only
+# degraded MISSING/UNCLEAR, so a log-derived PASSED could survive even when
+# a lens transcript said FAILED.
+overall=PASSED
+for lens in maintainability history structure; do
+  if [[ "${verdicts[$lens]:-MISSING}" != PASSED ]]; then
+    overall=FAILED
+  fi
+done
+
+marker_id='<!-- review-swarm-marker -->'
+body="$marker_id"$'\n'"🎯 review-swarm: $overall (M:${verdicts[maintainability]:-MISSING} H:${verdicts[history]:-MISSING} S:${verdicts[structure]:-MISSING})"
+
+# Sticky comment: find existing by identity marker, edit in place.
+existing_id=$(gh api "repos/${GITHUB_REPOSITORY:?}/issues/$pr_number/comments" --jq \
+  ".[] | select(.body | contains(\"$marker_id\")) | .id" | head -1)
+
+if [[ -n "$existing_id" ]]; then
+  gh api -X PATCH "repos/${GITHUB_REPOSITORY}/issues/comments/$existing_id" \
+    -f body="$body" > /dev/null
+else
+  gh pr comment "$pr_number" --body "$body"
+fi
diff --git a/README.md b/README.md
index 9584dae11..497d158c0 100644
--- a/README.md
+++ b/README.md
@@ -12,6 +12,8 @@ an entire application.
 The constitution is [`docs/RFC-0001-everything-is-a-relayflow.md`](docs/RFC-0001-everything-is-a-relayflow.md).
 Nothing in this repo may contradict it; changing it is a human decision.
 
+The required `RELAY_WORKSPACE_KEY` repository secret authenticates review-swarm GitHub Actions runs to the canonical Agent Relay Cloud workspace.
+
 ## Layout
 
 ```