fix(mention): skip the bot's own review on someone else's PR

generator/src/tend/templates/mention.yaml.j2

@@ -209,8 +209,10 @@ jobs:
           # comments on purpose: the pull_request_review *submission* kind is
           # deliberately left out, since a review the bot leaves on its own PR
           # is its reviewer role (the prompt is told to action it), not a
-          # self-loop. The only self-review skip is the terminal empty-body
-          # APPROVED gate below.
+          # self-loop. The two self-review skips below are author-keyed, but
+          # each is also narrowed to a case that leaves this run nothing to do:
+          # the empty-body APPROVED gate, and a bot review on a PR the bot did
+          # not author. Neither licenses a blanket self-review skip.
           if { [ "$KIND" = "issue_comment" ] || [ "$KIND" = "pull_request_review_comment" ]; } \
              && [ "$COMMENT_AUTHOR" = "<<cfg.bot_name>>" ]; then
             echo "should_run=false" >> "$GITHUB_OUTPUT"
@@ -304,6 +306,26 @@ jobs:
             echo "reason=participation" >> "$GITHUB_OUTPUT"; exit 0
           fi
 
+          # A review the bot leaves on someone else's PR leaves this run
+          # nothing to do: whatever the review warranted, the tend-review
+          # session that submitted it has already done — left the findings for
+          # a human author to act on (pushing to their branch unbidden is
+          # barred by conduct rules), or, on a dependency-bot PR where no
+          # author will act, pushed the fix itself. The BOT_REVIEWS heuristic
+          # below counts this very review, so without this gate the session
+          # always starts and always exits silently. Keyed on author alone —
+          # the APPROVED + empty-body gate above is the same shape narrowed to
+          # its one terminal leg, and reusing those clauses here would let every
+          # bodied COMMENTED review through. Placement carries the rest of the
+          # design: *after* the PR_AUTHOR short-circuit, which has already
+          # exited when the PR is the bot's own, so the reviewer-to-author
+          # handoff on a bot PR still fires; *after* the body and inline
+          # @-mention scans, so an explicit summons still wins.
+          if [ "$KIND" = "pull_request_review" ] \
+             && [ "$REVIEW_AUTHOR" = "<<cfg.bot_name>>" ]; then
+            echo "should_run=false" >> "$GITHUB_OUTPUT"; exit 0
+          fi
+
           # Captured, not counted — see the note on the issue-comment lookup above.
           BOT_REVIEWS=$(gh api --paginate "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER/reviews" \
             --jq '.[] | select(.user.login == "<<cfg.bot_name>>") | .id')

generator/tests/_regtest_outputs/test_generate.test_sandbox_levers_regtest.out

@@ -195,8 +195,10 @@ jobs:
           # comments on purpose: the pull_request_review *submission* kind is
           # deliberately left out, since a review the bot leaves on its own PR
           # is its reviewer role (the prompt is told to action it), not a
-          # self-loop. The only self-review skip is the terminal empty-body
-          # APPROVED gate below.
+          # self-loop. The two self-review skips below are author-keyed, but
+          # each is also narrowed to a case that leaves this run nothing to do:
+          # the empty-body APPROVED gate, and a bot review on a PR the bot did
+          # not author. Neither licenses a blanket self-review skip.
           if { [ "$KIND" = "issue_comment" ] || [ "$KIND" = "pull_request_review_comment" ]; } \
              && [ "$COMMENT_AUTHOR" = "test-bot" ]; then
             echo "should_run=false" >> "$GITHUB_OUTPUT"
@@ -290,6 +292,26 @@ jobs:
             echo "reason=participation" >> "$GITHUB_OUTPUT"; exit 0
           fi
 
+          # A review the bot leaves on someone else's PR leaves this run
+          # nothing to do: whatever the review warranted, the tend-review
+          # session that submitted it has already done ? left the findings for
+          # a human author to act on (pushing to their branch unbidden is
+          # barred by conduct rules), or, on a dependency-bot PR where no
+          # author will act, pushed the fix itself. The BOT_REVIEWS heuristic
+          # below counts this very review, so without this gate the session
+          # always starts and always exits silently. Keyed on author alone ?
+          # the APPROVED + empty-body gate above is the same shape narrowed to
+          # its one terminal leg, and reusing those clauses here would let every
+          # bodied COMMENTED review through. Placement carries the rest of the
+          # design: *after* the PR_AUTHOR short-circuit, which has already
+          # exited when the PR is the bot's own, so the reviewer-to-author
+          # handoff on a bot PR still fires; *after* the body and inline
+          # @-mention scans, so an explicit summons still wins.
+          if [ "$KIND" = "pull_request_review" ] \
+             && [ "$REVIEW_AUTHOR" = "test-bot" ]; then
+            echo "should_run=false" >> "$GITHUB_OUTPUT"; exit 0
+          fi
+
           # Captured, not counted ? see the note on the issue-comment lookup above.
           BOT_REVIEWS=$(gh api --paginate "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER/reviews" \
             --jq '.[] | select(.user.login == "test-bot") | .id')

generator/tests/_regtest_outputs/test_generate.test_workflow_minimal_codex_regtest[mention].out

@@ -195,8 +195,10 @@ jobs:
           # comments on purpose: the pull_request_review *submission* kind is
           # deliberately left out, since a review the bot leaves on its own PR
           # is its reviewer role (the prompt is told to action it), not a
-          # self-loop. The only self-review skip is the terminal empty-body
-          # APPROVED gate below.
+          # self-loop. The two self-review skips below are author-keyed, but
+          # each is also narrowed to a case that leaves this run nothing to do:
+          # the empty-body APPROVED gate, and a bot review on a PR the bot did
+          # not author. Neither licenses a blanket self-review skip.
           if { [ "$KIND" = "issue_comment" ] || [ "$KIND" = "pull_request_review_comment" ]; } \
              && [ "$COMMENT_AUTHOR" = "test-bot" ]; then
             echo "should_run=false" >> "$GITHUB_OUTPUT"
@@ -290,6 +292,26 @@ jobs:
             echo "reason=participation" >> "$GITHUB_OUTPUT"; exit 0
           fi
 
+          # A review the bot leaves on someone else's PR leaves this run
+          # nothing to do: whatever the review warranted, the tend-review
+          # session that submitted it has already done ? left the findings for
+          # a human author to act on (pushing to their branch unbidden is
+          # barred by conduct rules), or, on a dependency-bot PR where no
+          # author will act, pushed the fix itself. The BOT_REVIEWS heuristic
+          # below counts this very review, so without this gate the session
+          # always starts and always exits silently. Keyed on author alone ?
+          # the APPROVED + empty-body gate above is the same shape narrowed to
+          # its one terminal leg, and reusing those clauses here would let every
+          # bodied COMMENTED review through. Placement carries the rest of the
+          # design: *after* the PR_AUTHOR short-circuit, which has already
+          # exited when the PR is the bot's own, so the reviewer-to-author
+          # handoff on a bot PR still fires; *after* the body and inline
+          # @-mention scans, so an explicit summons still wins.
+          if [ "$KIND" = "pull_request_review" ] \
+             && [ "$REVIEW_AUTHOR" = "test-bot" ]; then
+            echo "should_run=false" >> "$GITHUB_OUTPUT"; exit 0
+          fi
+
           # Captured, not counted ? see the note on the issue-comment lookup above.
           BOT_REVIEWS=$(gh api --paginate "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER/reviews" \
             --jq '.[] | select(.user.login == "test-bot") | .id')

generator/tests/_regtest_outputs/test_generate.test_workflow_minimal_regtest[mention].out

@@ -195,8 +195,10 @@ jobs:
           # comments on purpose: the pull_request_review *submission* kind is
           # deliberately left out, since a review the bot leaves on its own PR
           # is its reviewer role (the prompt is told to action it), not a
-          # self-loop. The only self-review skip is the terminal empty-body
-          # APPROVED gate below.
+          # self-loop. The two self-review skips below are author-keyed, but
+          # each is also narrowed to a case that leaves this run nothing to do:
+          # the empty-body APPROVED gate, and a bot review on a PR the bot did
+          # not author. Neither licenses a blanket self-review skip.
           if { [ "$KIND" = "issue_comment" ] || [ "$KIND" = "pull_request_review_comment" ]; } \
              && [ "$COMMENT_AUTHOR" = "test-bot" ]; then
             echo "should_run=false" >> "$GITHUB_OUTPUT"
@@ -290,6 +292,26 @@ jobs:
             echo "reason=participation" >> "$GITHUB_OUTPUT"; exit 0
           fi
 
+          # A review the bot leaves on someone else's PR leaves this run
+          # nothing to do: whatever the review warranted, the tend-review
+          # session that submitted it has already done ? left the findings for
+          # a human author to act on (pushing to their branch unbidden is
+          # barred by conduct rules), or, on a dependency-bot PR where no
+          # author will act, pushed the fix itself. The BOT_REVIEWS heuristic
+          # below counts this very review, so without this gate the session
+          # always starts and always exits silently. Keyed on author alone ?
+          # the APPROVED + empty-body gate above is the same shape narrowed to
+          # its one terminal leg, and reusing those clauses here would let every
+          # bodied COMMENTED review through. Placement carries the rest of the
+          # design: *after* the PR_AUTHOR short-circuit, which has already
+          # exited when the PR is the bot's own, so the reviewer-to-author
+          # handoff on a bot PR still fires; *after* the body and inline
+          # @-mention scans, so an explicit summons still wins.
+          if [ "$KIND" = "pull_request_review" ] \
+             && [ "$REVIEW_AUTHOR" = "test-bot" ]; then
+            echo "should_run=false" >> "$GITHUB_OUTPUT"; exit 0
+          fi
+
           # Captured, not counted ? see the note on the issue-comment lookup above.
           BOT_REVIEWS=$(gh api --paginate "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER/reviews" \
             --jq '.[] | select(.user.login == "test-bot") | .id')

generator/tests/_regtest_outputs/test_generate.test_workflow_with_setup_regtest[mention].out

@@ -195,8 +195,10 @@ jobs:
           # comments on purpose: the pull_request_review *submission* kind is
           # deliberately left out, since a review the bot leaves on its own PR
           # is its reviewer role (the prompt is told to action it), not a
-          # self-loop. The only self-review skip is the terminal empty-body
-          # APPROVED gate below.
+          # self-loop. The two self-review skips below are author-keyed, but
+          # each is also narrowed to a case that leaves this run nothing to do:
+          # the empty-body APPROVED gate, and a bot review on a PR the bot did
+          # not author. Neither licenses a blanket self-review skip.
           if { [ "$KIND" = "issue_comment" ] || [ "$KIND" = "pull_request_review_comment" ]; } \
              && [ "$COMMENT_AUTHOR" = "test-bot" ]; then
             echo "should_run=false" >> "$GITHUB_OUTPUT"
@@ -290,6 +292,26 @@ jobs:
             echo "reason=participation" >> "$GITHUB_OUTPUT"; exit 0
           fi
 
+          # A review the bot leaves on someone else's PR leaves this run
+          # nothing to do: whatever the review warranted, the tend-review
+          # session that submitted it has already done ? left the findings for
+          # a human author to act on (pushing to their branch unbidden is
+          # barred by conduct rules), or, on a dependency-bot PR where no
+          # author will act, pushed the fix itself. The BOT_REVIEWS heuristic
+          # below counts this very review, so without this gate the session
+          # always starts and always exits silently. Keyed on author alone ?
+          # the APPROVED + empty-body gate above is the same shape narrowed to
+          # its one terminal leg, and reusing those clauses here would let every
+          # bodied COMMENTED review through. Placement carries the rest of the
+          # design: *after* the PR_AUTHOR short-circuit, which has already
+          # exited when the PR is the bot's own, so the reviewer-to-author
+          # handoff on a bot PR still fires; *after* the body and inline
+          # @-mention scans, so an explicit summons still wins.
+          if [ "$KIND" = "pull_request_review" ] \
+             && [ "$REVIEW_AUTHOR" = "test-bot" ]; then
+            echo "should_run=false" >> "$GITHUB_OUTPUT"; exit 0
+          fi
+
           # Captured, not counted ? see the note on the issue-comment lookup above.
           BOT_REVIEWS=$(gh api --paginate "repos/$GITHUB_REPOSITORY/pulls/$PR_NUMBER/reviews" \
             --jq '.[] | select(.user.login == "test-bot") | .id')

generator/tests/test_generate.py

@@ -1046,9 +1046,11 @@ def test_mention_self_comment_skip_spares_review_submissions(
     role speaking, not a self-loop — the prompt is told to action it. That
     signal arrives as a `pull_request_review` submission event, distinct from
     the `pull_request_review_comment` inline-comment event the skip covers, so
-    it must still reach the actionable path. The only self-review that is
-    skipped is the terminal empty-body APPROVED case (see
-    test_mention_skips_bot_approved_review).
+    it must still reach the actionable path. The self-reviews that are skipped
+    are skipped for being terminal for the bot, not for being self-authored:
+    the empty-body APPROVED case (see test_mention_skips_bot_approved_review)
+    and a review on a PR the bot did not author (see
+    test_mention_skips_bot_review_on_another_authors_pr).
 
     This pins the boundary against a later "skip self-authored reviews too, for
     consistency" edit that would silently break the review -> fix loop: the
@@ -1081,18 +1083,83 @@ def test_mention_self_comment_skip_spares_review_submissions(
         "submission kind — a bot self-review is actionable reviewer signal"
     )
 
-    # The sole author-keyed skip for a review submission is the terminal
-    # empty-body APPROVED gate; a COMMENTED / non-empty-body bot self-review
-    # falls through to the actionable PR_AUTHOR == bot short-circuit.
-    review_skip = run[run.index('[ "$KIND" = "pull_request_review" ]') :]
-    assert '[ "$REVIEW_STATE" = "approved" ]' in review_skip
-    assert '[ -z "$COMMENT_BODY" ]' in review_skip
+    # Slice to the terminal empty-body APPROVED gate alone — it ends at the
+    # inline-comment fetch, ahead of the other-authors-PR gate that
+    # test_mention_skips_bot_review_on_another_authors_pr pins — so this stays
+    # discriminating rather than passing on either. A COMMENTED /
+    # non-empty-body bot self-review that is neither falls through to the
+    # actionable PR_AUTHOR == bot short-circuit.
+    approved_skip = run[
+        run.index('[ "$KIND" = "pull_request_review" ]') : run.index(
+            "reviews/$PAYLOAD_ID/comments"
+        )
+    ]
+    assert '[ "$REVIEW_STATE" = "approved" ]' in approved_skip
+    assert '[ -z "$COMMENT_BODY" ]' in approved_skip
     assert '[ "$PR_AUTHOR" = "test-bot" ]' in run, (
         "a bot self-review that isn't the terminal empty approval must reach "
         "the actionable PR_AUTHOR == bot short-circuit"
     )
 
 
+def test_mention_skips_bot_review_on_another_authors_pr(tmp_path: Path) -> None:
+    """A review the bot leaves on a PR someone else authored leaves the mention
+    run nothing to do: whatever the review warranted, the tend-review session
+    that submitted it has already done — left the findings for a human author
+    to act on (pushing to their branch unbidden is barred by conduct rules),
+    or, on a dependency-bot PR where no author will act, pushed the fix itself.
+    The BOT_REVIEWS heuristic counts this very review, so without a gate every
+    such review starts a session that can only exit silently (#915). #747's
+    gate covers only the empty-body APPROVED leg —
+    the reviews observed here carry 455-1920 char bodies in APPROVED and
+    COMMENTED states alike, so the gate must key on author alone.
+
+    Placement is the whole design, in both directions:
+
+    - *After* the PR_AUTHOR short-circuit, which exits should_run=true when the
+      PR is the bot's own. A review the bot leaves on its own PR is the reviewer
+      role handing work to the author role, and must keep firing (#166, #761) —
+      an author-keyed gate placed any earlier would swallow it.
+    - *After* the review-body and inline-comment @-mention scans, so an explicit
+      summons the bot quotes still wins.
+    - *Before* the BOT_REVIEWS heuristic, which is what misreads the triggering
+      review as prior engagement.
+    """
+    cfg = Config.load(_minimal_config(tmp_path))
+    wf = generate_mention(cfg)
+    data = yaml.safe_load(wf.content)
+    check_step = next(
+        s for s in data["jobs"]["verify"]["steps"] if s.get("id") == "check"
+    )
+    run = check_step["run"]
+
+    # The window between the PR-author resolution and the engagement heuristic
+    # is the only place the gate can sit and still satisfy all three orderings.
+    gate = run[run.index("PR_AUTHOR=$(gh pr view") : run.index("BOT_REVIEWS=$(")]
+    assert '[ "$KIND" = "pull_request_review" ]' in gate
+    assert '[ "$REVIEW_AUTHOR" = "test-bot" ]' in gate
+
+    # Keyed on author alone. Reusing #747's state/body clauses here would let
+    # every review in #915's evidence through, since they are non-empty and
+    # mostly COMMENTED.
+    assert '[ "$REVIEW_STATE"' not in gate
+    assert '[ -z "$COMMENT_BODY" ]' not in gate
+
+    # The bot's own PR still reaches the actionable short-circuit, because that
+    # check exits before the gate is read.
+    assert run.index('[ "$PR_AUTHOR" = "test-bot" ]') < run.index(
+        '[ "$KIND" = "pull_request_review" ]', run.index("PR_AUTHOR=$(gh pr view")
+    )
+
+    # And an @-mention inside the review — body or inline comment — still wins.
+    # Anchor the right-hand side to the gate itself, not to the heuristic below
+    # it: the inline fetch has always preceded BOT_REVIEWS, so that comparison
+    # would hold on the pre-change template and pin nothing.
+    assert run.index("reviews/$PAYLOAD_ID/comments") < run.index(
+        '[ "$REVIEW_AUTHOR" = "test-bot" ]', run.index("PR_AUTHOR=$(gh pr view")
+    )
+
+
 # ---------------------------------------------------------------------------
 # Fork guard
 # ---------------------------------------------------------------------------