@openhands/extensions 0.22.1 → 0.24.0

package/tests/test_github_triage_delivery.py

@@ -41,6 +41,29 @@ def test_triage_submits_each_changed_issue_as_agent_work(tmp_path, monkeypatch):
     ]
     assert "Do not implement code" in submit.call_args_list[0].kwargs["prompt"]
     assert "GITHUB_PERSONAL_ACCESS_TOKEN" in submit.call_args_list[0].kwargs["prompt"]
+    assert (
+        "The following comments and acceptance criteria were added by the "
+        "OpenHands AI agent." in submit.call_args_list[0].kwargs["prompt"]
+    )
+    assert (
+        "closest existing or adjacent implementation"
+        in submit.call_args_list[0].kwargs["prompt"]
+    )
+    assert "Add `ready-for-dev` only after" in submit.call_args_list[0].kwargs["prompt"]
+    assert (
+        "Do not invent acceptance criteria around an arbitrary choice"
+        in submit.call_args_list[0].kwargs["prompt"]
+    )
+    assert (
+        "ask only the focused follow-up questions needed to resolve it"
+        in submit.call_args_list[0].kwargs["prompt"]
+    )
+    prompt = submit.call_args_list[0].kwargs["prompt"]
+    assert "<!-- openhands-ai-triage:start -->" in prompt
+    assert "<!-- openhands-ai-triage:end -->" in prompt
+    assert "Do not post a triage comment" in prompt
+    assert "DECISION NEEDED" in prompt
+    assert "Leave the human-owned issue body unchanged" in prompt
 
 
 def test_triage_continues_after_one_submission_fails(tmp_path, monkeypatch):
@@ -65,7 +88,44 @@ def test_triage_continues_after_one_submission_fails(tmp_path, monkeypatch):
     ]
 
 
-def test_triage_skips_ready_blocked_and_unchanged_issues(tmp_path, monkeypatch):
+def test_triage_event_submits_only_the_named_issue(tmp_path, monkeypatch):
+    issues = [
+        {"number": 1, "title": "First", "body": "A", "labels": []},
+        {"number": 2, "title": "Second", "body": "B", "labels": []},
+    ]
+    _module, run = _triage(tmp_path, monkeypatch, issues)
+    monkeypatch.setenv(
+        "AUTOMATION_EVENT_PAYLOAD",
+        '{"event":{"payload":{"repository":{"full_name":"owner/repo"},'
+        '"issue":{"number":2}}}}',
+    )
+    run.dispatcher.deliver.return_value = {
+        "disposition": "created",
+        "conversation_id": "second",
+    }
+
+    run.run()
+
+    assert run.dispatcher.deliver.call_args.kwargs["subject"] == "9876:issue:2"
+
+
+def test_triage_event_skips_other_repositories(tmp_path, monkeypatch):
+    issues = [{"number": 1, "title": "First", "body": "A", "labels": []}]
+    _module, run = _triage(tmp_path, monkeypatch, issues)
+    monkeypatch.setenv(
+        "AUTOMATION_EVENT_PAYLOAD",
+        '{"event":{"payload":{"repository":{"full_name":"other/repo"},'
+        '"issue":{"number":1}}}}',
+    )
+
+    run.run()
+
+    run.dispatcher.deliver.assert_not_called()
+
+
+def test_triage_checks_ready_issue_but_skips_blocked_and_unchanged_issues(
+    tmp_path, monkeypatch
+):
     issues = [
         {
             "number": 1,
@@ -78,13 +138,147 @@ def test_triage_skips_ready_blocked_and_unchanged_issues(tmp_path, monkeypatch):
     ]
     module, run = _triage(tmp_path, monkeypatch, issues)
     run.dependencies_complete = lambda issue: issue["number"] != 2
-    marker = module.hashlib.sha256(
-        module.json.dumps(["Unchanged", "", []], sort_keys=True).encode()
+    unchanged_marker = module.hashlib.sha256(
+        module.json.dumps(
+            [module.TRIAGE_FORMAT_VERSION, "Unchanged", "", []], sort_keys=True
+        ).encode()
     ).hexdigest()
-    run.gh_pages = lambda path: [{"id": 10, "body": f"<!-- triage-source:{marker} -->"}]
-    submit = Mock()
+    run.gh_pages = lambda path: (
+        [
+            {
+                "id": 10,
+                "body": f"<!-- triage-source:{unchanged_marker} -->",
+            }
+        ]
+        if path == "/issues/3/comments"
+        else []
+    )
+    submit = Mock(return_value={"disposition": "created", "conversation_id": "ready"})
     run.dispatcher.deliver = submit
 
     run.run()
 
-    submit.assert_not_called()
+    assert [call.kwargs["subject"] for call in submit.call_args_list] == [
+        "9876:issue:1"
+    ]
+
+
+def test_triage_passes_human_discussion_and_stale_bot_comment(tmp_path, monkeypatch):
+    issue = {
+        "number": 7,
+        "title": "Improve behavior",
+        "body": "Original request",
+        "labels": [{"name": "enhancement"}],
+    }
+    _module, run = _triage(tmp_path, monkeypatch, [issue])
+    run.gh_pages = lambda path: [
+        {
+            "id": 101,
+            "user": {"login": "maintainer"},
+            "created_at": "2026-09-18T00:00:00Z",
+            "updated_at": "2026-09-18T00:00:00Z",
+            "body": "Keep compatibility with saved configurations.",
+        },
+        {
+            "id": 102,
+            "user": {"login": "all-hands-bot"},
+            "created_at": "2026-09-18T00:01:00Z",
+            "updated_at": "2026-09-18T00:01:00Z",
+            "body": "Old triage\n<!-- triage-source:old -->",
+        },
+    ]
+    run.dispatcher.deliver.return_value = {
+        "disposition": "resumed",
+        "conversation_id": "triage",
+    }
+
+    run.run()
+
+    prompt = run.dispatcher.deliver.call_args.kwargs["prompt"]
+    assert '"author": "maintainer"' in prompt
+    assert "Keep compatibility with saved configurations." in prompt
+    assert '"id": 102' in prompt
+    assert "Old triage" not in prompt
+    assert "Delete prior comments listed as automated triage comments" in prompt
+
+
+def test_triage_ignores_managed_body_when_computing_delivery(tmp_path, monkeypatch):
+    original = "Human request\n\n### Desired Behavior\nKeep the API stable."
+    issue = {
+        "number": 7,
+        "title": "Improve behavior",
+        "body": original,
+        "labels": [{"name": "enhancement"}],
+    }
+    module, run = _triage(tmp_path, monkeypatch, [issue])
+    digest = module.hashlib.sha256(
+        module.json.dumps(
+            [module.TRIAGE_FORMAT_VERSION, "Improve behavior", original, []],
+            sort_keys=True,
+        ).encode()
+    ).hexdigest()
+    issue["body"] += (
+        "\n\n<!-- openhands-ai-triage:start -->\n"
+        "---\n## OpenHands AI triage\n"
+        "### Acceptance Criteria\n- [ ] Preserve behavior\n"
+        f"<!-- triage-source:{digest} -->\n"
+        "<!-- openhands-ai-triage:end -->"
+    )
+
+    run.run()
+
+    run.dispatcher.deliver.assert_not_called()
+    assert module.author_body(issue["body"]) == original
+
+
+def test_triage_reprocesses_legacy_comment_digest(tmp_path, monkeypatch):
+    issue = {"number": 7, "title": "Improve behavior", "body": "Request", "labels": []}
+    module, run = _triage(tmp_path, monkeypatch, [issue])
+    legacy_digest = module.hashlib.sha256(
+        module.json.dumps(["Improve behavior", "Request", []], sort_keys=True).encode()
+    ).hexdigest()
+    run.gh_pages = lambda path: [
+        {"id": 10, "body": f"<!-- triage-source:{legacy_digest} -->"}
+    ]
+    run.dispatcher.deliver.return_value = {
+        "disposition": "created",
+        "conversation_id": "triage",
+    }
+
+    run.run()
+
+    run.dispatcher.deliver.assert_called_once()
+
+
+def test_triage_passes_only_human_owned_body_to_agent(tmp_path, monkeypatch):
+    issue = {
+        "number": 8,
+        "title": "Clarify behavior",
+        "body": (
+            "Original description\n\n"
+            "<!-- openhands-ai-triage:start -->\n"
+            "Old generated criteria\n"
+            "<!-- triage-source:old -->\n"
+            "<!-- openhands-ai-triage:end -->"
+        ),
+        "labels": [{"name": "enhancement"}],
+    }
+    _module, run = _triage(tmp_path, monkeypatch, [issue])
+    run.dispatcher.deliver.return_value = {
+        "disposition": "resumed",
+        "conversation_id": "triage",
+    }
+
+    run.run()
+
+    prompt = run.dispatcher.deliver.call_args.kwargs["prompt"]
+    assert '"author_body": "Original description"' in prompt
+    assert '"managed_triage_section_present": true' in prompt
+    assert "Old generated criteria" not in prompt
+
+
+def test_incomplete_marker_pair_remains_human_owned(tmp_path, monkeypatch):
+    module, _run = _triage(tmp_path, monkeypatch, [])
+    body = "Human text\n<!-- openhands-ai-triage:start -->\nUnclosed text"
+
+    assert module.author_body(body) == body

package/tests/test_maintainer_handoff.py

@@ -0,0 +1,198 @@
+"""Deterministic maintainer selection for positive automated reviews."""
+
+import importlib.util
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import Mock
+from urllib.error import HTTPError
+from urllib.parse import parse_qs, urlsplit
+
+_MODULE_PATH = (
+    Path(__file__).resolve().parents[1]
+    / "skills/github-pr-reviewer/scripts/maintainer_handoff.py"
+)
+_SPEC = importlib.util.spec_from_file_location("maintainer_handoff_test", _MODULE_PATH)
+assert _SPEC and _SPEC.loader
+maintainer_handoff = importlib.util.module_from_spec(_SPEC)
+_SPEC.loader.exec_module(maintainer_handoff)
+
+
+def _pr(**overrides):
+    result = {
+        "number": 7,
+        "head": {"sha": "head"},
+        "base": {
+            "ref": "main",
+            "repo": {"owner": {"type": "Organization"}},
+        },
+        "user": {"login": "author"},
+        "requested_reviewers": [],
+    }
+    result.update(overrides)
+    return result
+
+
+def _repository(commits=None, reviews=None, loads=None):
+    commits = commits or {}
+    loads = loads or {}
+
+    def gh_pages(path):
+        if path.endswith("/files"):
+            return [{"filename": name} for name in commits]
+        if path.endswith("/reviews"):
+            return reviews or []
+        raise AssertionError(path)
+
+    def gh(method, path, body=None):
+        if path.startswith("/commits?"):
+            filename = parse_qs(urlsplit(path).query)["path"][0]
+            return [{"author": {"login": login}} for login in commits[filename]]
+        if method == "POST" and path.endswith("/requested_reviewers"):
+            return {}
+        raise AssertionError((method, path, body))
+
+    def api(method, path, params=None, body=None):
+        assert (method, path) == ("GET", "/search/issues")
+        login = params["q"].rsplit(":", 1)[1]
+        return {"total_count": loads.get(login, 0)}
+
+    return SimpleNamespace(
+        repository="OpenHands/OpenHands",
+        token="token",
+        gh_pages=gh_pages,
+        gh=Mock(side_effect=gh),
+        api=Mock(side_effect=api),
+    )
+
+
+def test_code_proximity_precedes_review_load():
+    repo = _repository(
+        {"src/api.py": ["VascoSch92", "VascoSch92", "neubig"]},
+        loads={"VascoSch92": 8, "neubig": 0},
+    )
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo, _pr(), ["neubig", "VascoSch92"]
+    )
+
+    assert selected == "VascoSch92"
+    assert repo.gh.call_args_list[-1].args[:2] == (
+        "POST",
+        "/pulls/7/requested_reviewers",
+    )
+    assert repo.gh.call_args_list[-1].args[2] == {"reviewers": ["VascoSch92"]}
+
+
+def test_review_load_breaks_equal_proximity():
+    repo = _repository({"src/api.py": []}, loads={"VascoSch92": 1, "neubig": 4})
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo, _pr(), ["neubig", "VascoSch92"]
+    )
+
+    assert selected == "VascoSch92"
+
+
+def test_review_load_uses_user_qualifier_for_user_owned_repo():
+    repo = _repository({"src/api.py": []}, loads={"VascoSch92": 1, "neubig": 4})
+    repo.repository = "neubig/project"
+    pr = _pr(base={"ref": "main", "repo": {"owner": {"type": "User"}}})
+
+    maintainer_handoff.request_maintainer_review(
+        repo, pr, ["neubig", "VascoSch92"]
+    )
+
+    assert all(
+        "user:neubig" in call.kwargs["params"]["q"]
+        for call in repo.api.call_args_list
+    )
+
+
+def test_existing_request_is_idempotent():
+    repo = _repository()
+    pr = _pr(requested_reviewers=[{"login": "VascoSch92"}])
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo, pr, ["neubig", "VascoSch92"]
+    )
+
+    assert selected == "VascoSch92"
+    repo.api.assert_not_called()
+    repo.gh.assert_not_called()
+
+
+def test_prior_head_approval_satisfies_handoff():
+    repo = _repository(
+        reviews=[
+            {
+                "commit_id": "prior-head",
+                "state": "APPROVED",
+                "user": {"login": "VascoSch92"},
+            }
+        ]
+    )
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo, _pr(), ["neubig", "VascoSch92"]
+    )
+
+    assert selected == "VascoSch92"
+    repo.api.assert_not_called()
+    assert not any(call.args[0] == "POST" for call in repo.gh.call_args_list)
+
+
+def test_later_changes_request_supersedes_prior_approval():
+    repo = _repository(
+        {"src/api.py": []},
+        reviews=[
+            {
+                "commit_id": "prior-head",
+                "state": "APPROVED",
+                "user": {"login": "VascoSch92"},
+            },
+            {
+                "commit_id": "head",
+                "state": "CHANGES_REQUESTED",
+                "user": {"login": "VascoSch92"},
+            },
+        ],
+    )
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo, _pr(), ["neubig", "VascoSch92"]
+    )
+
+    assert selected == "VascoSch92"
+    assert not any(call.args[0] == "POST" for call in repo.gh.call_args_list)
+
+
+def test_author_is_excluded_and_empty_roster_is_compatible():
+    repo = _repository({"src/api.py": ["neubig"]})
+
+    selected = maintainer_handoff.request_maintainer_review(
+        repo,
+        _pr(user={"login": "neubig"}),
+        ["neubig", "VascoSch92"],
+    )
+
+    assert selected == "VascoSch92"
+    assert maintainer_handoff.request_maintainer_review(repo, _pr(), "") is None
+
+
+def test_unassignable_maintainer_is_a_configuration_error():
+    repo = _repository({"src/api.py": []})
+    original_gh = repo.gh.side_effect
+
+    def fail_assignment(method, path, body=None):
+        if method == "POST":
+            raise HTTPError("url", 422, "invalid reviewer", {}, None)
+        return original_gh(method, path, body)
+
+    repo.gh.side_effect = fail_assignment
+
+    try:
+        maintainer_handoff.request_maintainer_review(repo, _pr(), ["neubig"])
+    except maintainer_handoff.HandoffConfigurationError as exc:
+        assert "@neubig" in str(exc)
+    else:
+        raise AssertionError("expected HandoffConfigurationError")

package/tests/test_pr_review_prompt.py

@@ -67,6 +67,8 @@ def test_prompt_instructs_reading_repo_guidance():
     assert "AGENTS.md" in prompt
     assert "MUST read" in prompt
     assert "CONTRIBUTING.md" in prompt
+    assert "inspect how the affected feature or code path works" in prompt
+    assert "concrete failure on the current head" in prompt
 
 
 def test_format_prompt_omits_evidence_requirements_by_default():
@@ -129,14 +131,16 @@ def test_file_reviewer_skill_content():
     content = module.FILE_REVIEWER_SKILL
 
     assert "file-level code reviewer" in content
-    # Unified review style
-    assert "pragmatic" in content
+    # Concrete findings only, while retaining limited-impact defects.
+    assert "concrete failure" in content
     # JSON schema documented
     assert "path" in content
     assert "line" in content
     assert "severity" in content
     assert "body" in content
     assert "critical" in content
+    assert '`"minor"`' in content
+    assert '`"nit"`' not in content
     # Tool access documented
     assert "terminal" in content
     assert "file_editor" in content

package/tests/test_pr_review_review_context.py

@@ -212,6 +212,150 @@ def test_format_thread_includes_rendered_suggestion_text_in_review_context():
     assert "```suggestion" not in formatted
 
 
+def test_get_linked_issues_uses_github_closing_issue_relation(monkeypatch):
+    module = _load_agent_script_module()
+    monkeypatch.setenv("REPO_NAME", "OpenHands/extensions")
+    calls = []
+
+    def fake_paginate(**kwargs):
+        calls.append(kwargs)
+        return [{"number": 12, "title": "criteria", "body": "- [ ] works"}]
+
+    monkeypatch.setattr(module, "_paginate_graphql", fake_paginate)
+
+    assert module.get_linked_issues("34") == [
+        {"number": 12, "title": "criteria", "body": "- [ ] works"}
+    ]
+    assert calls[0]["query"] == module.LINKED_ISSUES_QUERY
+    assert calls[0]["variables"] == {
+        "owner": "OpenHands",
+        "repo": "extensions",
+        "pr_number": 34,
+        "count": 5,
+    }
+
+
+def test_get_pr_issue_comments_fetches_latest_graphql_page(monkeypatch):
+    module = _load_agent_script_module()
+    monkeypatch.setenv("REPO_NAME", "OpenHands/extensions")
+    calls = []
+
+    def fake_paginate(**kwargs):
+        calls.append(kwargs)
+        return [
+            {
+                "id": "new",
+                "author": {"login": "last", "__typename": "User"},
+                "authorAssociation": "MEMBER",
+                "body": "latest concern",
+                "createdAt": "2026-01-02T00:00:00Z",
+            },
+            {
+                "id": "old",
+                "author": {"login": "first", "__typename": "User"},
+                "authorAssociation": "MEMBER",
+                "body": "older concern",
+                "createdAt": "2026-01-01T00:00:00Z",
+            },
+        ]
+
+    monkeypatch.setattr(module, "_paginate_graphql", fake_paginate)
+
+    comments = module.get_pr_issue_comments("612", max_comments=25)
+
+    assert [comment["body"] for comment in comments] == [
+        "older concern",
+        "latest concern",
+    ]
+    assert calls[0]["query"] == module.ISSUE_COMMENTS_QUERY
+    assert calls[0]["variables"]["count"] == 25
+    assert calls[0]["path_to_nodes"] == ["pullRequest", "comments"]
+    assert calls[0]["max_items"] == 25
+
+
+def test_review_context_prioritizes_issue_and_top_level_discussion():
+    module = _load_agent_script_module()
+
+    context = module.format_review_context(
+        reviews=[
+            {
+                "user": {"login": "all-hands-bot"},
+                "state": "APPROVED",
+                "body": "Looks good",
+            }
+        ],
+        threads=[],
+        issue_comments=[
+            {
+                "user": {"login": "maintainer", "type": "User"},
+                "author_association": "MEMBER",
+                "body": "The credential must never be persisted in plaintext.",
+            }
+        ],
+        linked_issues=[
+            {
+                "number": 42,
+                "title": "Protect stored credentials",
+                "body": "Acceptance: encrypt credentials outside the user home.",
+            }
+        ],
+    )
+
+    assert "### Linked Issues and Acceptance Criteria" in context
+    assert "### Top-level PR Discussion" in context
+    assert "**maintainer** (MEMBER)" in context
+    assert "credential must never be persisted" in context
+    assert context.index("Linked Issues") < context.index("Previous Review Decisions")
+
+
+def test_review_context_omits_bot_status_comments():
+    module = _load_agent_script_module()
+
+    context = module.format_review_context(
+        reviews=[],
+        threads=[],
+        issue_comments=[
+            {
+                "user": {"login": "github-actions", "type": "Bot"},
+                "body": "Coverage report",
+            },
+            {
+                "user": {"login": "all-hands-bot", "type": "User"},
+                "body": "OpenHands gave up on this review",
+            },
+            {
+                "user": {"login": "maintainer", "type": "User"},
+                "author_association": "MEMBER",
+                "body": "Check the cancellation path.",
+            },
+        ],
+    )
+
+    assert "Coverage report" not in context
+    assert "OpenHands gave up" not in context
+    assert "Check the cancellation path." in context
+
+
+def test_review_context_keeps_latest_twelve_human_comments():
+    module = _load_agent_script_module()
+    comments = [
+        {
+            "user": {"login": "maintainer", "type": "User"},
+            "author_association": "MEMBER",
+            "body": f"Concern-{index:02d}",
+            "created_at": f"2026-01-{index + 1:02d}T00:00:00Z",
+        }
+        for index in range(13)
+    ]
+
+    context = module.format_review_context([], [], issue_comments=comments)
+
+    assert "Concern-00" not in context
+    assert "Concern-01" in context
+    assert "Concern-12" in context
+    assert context.index("Concern-01") < context.index("Concern-12")
+
+
 def test_register_sub_agents_completes_without_error():
     """Smoke test: _register_sub_agents() runs without raising."""
     module = _load_agent_script_module()