diff --git a/.github/workflows/persona-mention-reusable.yml b/.github/workflows/persona-mention-reusable.yml
index 177a8d76b..eceb54adc 100644
--- a/.github/workflows/persona-mention-reusable.yml
+++ b/.github/workflows/persona-mention-reusable.yml
@@ -386,7 +386,13 @@ jobs:
     # triggered. Same two recursion axes and conservative §4 default floor as the
     # comment path's pre-filter — a PR event is no laxer:
     #   1. bot actor — a PR opened by an agent identity never routes
-    #   2. marker    — a PR body carrying '<!-- persona:' never routes
+    #   2. marker    — a PR body AUTHORED as a persona (it starts with
+    #                  '<!-- persona:', pr_ensure_marker's convention) never routes.
+    #                  A marker quoted later in the body does not count (#1286):
+    #                  no persona writes mid-body, so that is a quotation. This
+    #                  startsWith is the cheap pre-filter; pm_pr_should_route
+    #                  (pm_is_agent_authored_body) stays authoritative and also
+    #                  skips leading blank lines and whitespace.
     # There is deliberately NO '@petry-projects/' check: the PR path derives the
     # firing personas from their manifests, it does not read a handle from the body.
     #
@@ -412,7 +418,7 @@ jobs:
       github.event.pull_request.user.login != 'github-actions[bot]' &&
       github.actor != 'donpetry-bot' &&
       github.actor != 'github-actions[bot]' &&
-      !contains(github.event.pull_request.body, '<!-- persona:') &&
+      !startsWith(github.event.pull_request.body, '<!-- persona:') &&
       contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.pull_request.author_association)
 
     steps:
diff --git a/scripts/lib/persona-mention.sh b/scripts/lib/persona-mention.sh
index 91804a05e..3f5fe6c9e 100644
--- a/scripts/lib/persona-mention.sh
+++ b/scripts/lib/persona-mention.sh
@@ -78,6 +78,25 @@ pm_is_agent_comment() {
   esac
 }
 
+# pm_is_agent_authored_body <body> — 0 if this body was AUTHORED as a persona:
+# its first non-blank line begins with the agent marker (leading blank lines and
+# leading whitespace ignored). Axis 2 of the PR path's recursion guard. Unlike a
+# comment, a PR body is never written by a persona mid-text — the runner's
+# pr_ensure_marker prepends the marker as the first line — so a marker anywhere
+# else is a quotation, and substring matching would skip that PR silently (#1286).
+pm_is_agent_authored_body() {
+  local line
+  while IFS= read -r line || [ -n "$line" ]; do
+    line="${line#"${line%%[![:space:]]*}"}"
+    [ -n "$line" ] || continue
+    case "$line" in
+      "$PERSONA_AGENT_MARKER"*) return 0 ;;
+      *) return 1 ;;
+    esac
+  done <<< "$1"
+  return 1
+}
+
 # pm_extract_slugs <body> — emit each DISTINCT persona slug addressed in body.
 #
 # Matches '@<org>/<slug>' where slug is kebab-case, mirroring the schema's
@@ -533,7 +552,7 @@ pm_pr_should_route() {
 
   pm_is_bot_actor "$author" && return 1       # axis 1a: bot PR author
   pm_is_bot_actor "$actor" && return 1        # axis 1b: bot pusher/reopener (github.actor)
-  pm_is_agent_comment "$body" && return 1     # axis 2: agent marker
+  pm_is_agent_authored_body "$body" && return 1  # axis 2: body authored with the marker
   pm_trust_ok "$assoc" OWNER MEMBER COLLABORATOR || return 1
   return 0
 }
diff --git a/tests/persona_mention.bats b/tests/persona_mention.bats
index 9cc6c18d5..207896358 100644
--- a/tests/persona_mention.bats
+++ b/tests/persona_mention.bats
@@ -88,6 +88,41 @@ MENTION_ON='    - surface: mention
   [ "$status" -ne 0 ]
 }
 
+# pm_is_agent_authored_body is the PR-path axis 2: a PR body is AUTHORED as a
+# persona only when the marker leads it (pr_ensure_marker's convention). A marker
+# anywhere else is a quotation and must not suppress routing (#1286).
+
+@test "pm_is_agent_authored_body matches a body whose first line is the marker" {
+  run pm_is_agent_authored_body $'<!-- persona:qa-lead -->\n## Summary\nautomated PR'
+  [ "$status" -eq 0 ]
+}
+
+@test "pm_is_agent_authored_body ignores leading blank lines and indentation" {
+  run pm_is_agent_authored_body $'\n\n   \n<!-- persona:qa-lead -->\nbody'
+  [ "$status" -eq 0 ]
+  run pm_is_agent_authored_body $'  \t<!-- persona:dev-lead -->\nbody'
+  [ "$status" -eq 0 ]
+  run pm_is_agent_authored_body $'\r\n<!-- persona:dev-lead -->\r\nbody'
+  [ "$status" -eq 0 ]
+}
+
+@test "pm_is_agent_authored_body does not match a marker quoted mid-body in prose" {
+  run pm_is_agent_authored_body $'## Summary\n\nThe `<!-- persona:solution-architect -->` marker is unchanged.'
+  [ "$status" -eq 1 ]
+}
+
+@test "pm_is_agent_authored_body does not match a marker inside a later fenced code block" {
+  run pm_is_agent_authored_body $'## Summary\n\n```\n<!-- persona:qa-lead -->\n```'
+  [ "$status" -eq 1 ]
+}
+
+@test "pm_is_agent_authored_body does not match an empty or blank body" {
+  run pm_is_agent_authored_body ''
+  [ "$status" -eq 1 ]
+  run pm_is_agent_authored_body $'\n  \n'
+  [ "$status" -eq 1 ]
+}
+
 @test "pm_is_bot_actor recognises the agent identities" {
   run pm_is_bot_actor donpetry-bot
   [ "$status" -eq 0 ]
@@ -693,6 +728,11 @@ runtime:
 # Same two recursion axes and trust floor as pm_should_route, but WITHOUT the
 # @-mention requirement: a PR is derived, not addressed. A PR event must not
 # become an ungated write surface, so the guards apply on this path too (AC #4).
+#
+# The workflow's pull_request pre-filter uses !startsWith(body, '<!-- persona:')
+# for axis 2 (not substring contains). The tests below verify this prefix-matching
+# semantics: a marker quoted mid-body does NOT block routing, but a marker at the
+# start (or after leading blank lines, per pm_is_agent_authored_body) DOES block.
 
 @test "pm_pr_should_route allows a trusted human's PR with no @-mention in the body" {
   run pm_pr_should_route don-petry don-petry OWNER "Ordinary PR description, no handles."
@@ -716,11 +756,40 @@ runtime:
   [ "$status" -eq 1 ]
 }
 
+@test "pm_pr_should_route allows a trusted human's PR that QUOTES the marker on a later line (#1286)" {
+  run pm_pr_should_route don-petry don-petry OWNER $'## Summary\n\nThe `<!-- persona:solution-architect -->` marker is unchanged.'
+  [ "$status" -eq 0 ]
+}
+
+@test "pm_pr_should_route blocks a PR whose body leads with the marker after blank lines (axis 2)" {
+  run pm_pr_should_route don-petry don-petry OWNER $'\n\n<!-- persona:qa-lead -->\nautomated PR'
+  [ "$status" -eq 1 ]
+}
+
 @test "pm_pr_should_route blocks a PR from an author below the default floor" {
   run pm_pr_should_route drive-by drive-by CONTRIBUTOR "PR body"
   [ "$status" -eq 1 ]
 }
 
+# The workflow's route-pull-request job uses a cheap if: condition that must match
+# the authoritative pm_pr_should_route function. The condition uses startsWith for
+# the marker axis (not contains), matching pm_is_agent_authored_body's semantics:
+# a marker quoted mid-body is not considered agent-authored. These verifications
+# ensure a regression from startsWith back to contains would be caught by CI.
+@test "pull_request workflow pre-filter uses prefix matching (startsWith, not contains)" {
+  local workflow="$SCRIPT_DIR/.github/workflows/persona-mention-reusable.yml"
+  # Inspect each job's own `if:` so a match elsewhere in the file cannot mask a
+  # lost guard.
+  local pr_if comment_if
+  pr_if="$(yq '.jobs.route-pull-request.if' "$workflow")"
+  comment_if="$(yq '.jobs.route.if' "$workflow")"
+  [[ "$pr_if" == *"!startsWith(github.event.pull_request.body, '<!-- persona:')"* ]]
+  [[ "$pr_if" != *"contains(github.event.pull_request.body, '<!-- persona:')"* ]]
+  # The comment path keeps substring matching: any persona marker in a comment
+  # means the comment is agent-authored.
+  [[ "$comment_if" == *"!contains(github.event.comment.body, '<!-- persona:')"* ]]
+}
+
 # --- pm_pr_route_verdict (the composed PR gauntlet — AC #2/#4 proof) --------
 # Given a manifest that DECLARES the pull_request surface enabled and its
 # interaction contract, decide the verdict for one PR: stop-marker → write-gate