503 lines
16 KiB
Python
503 lines
16 KiB
Python
import json
|
|
from pathlib import Path
|
|
from types import SimpleNamespace
|
|
from typing import Any, cast
|
|
|
|
import pytest
|
|
from pydantic import BaseModel
|
|
|
|
from agentci.engine.model import Job, JobKind, Workflow, WorkflowKind, WorkflowStatus
|
|
from agentci.engine.run import JobRun
|
|
from agentci.integrations.gitea.models import CommentInfo, IssueInfo, PullRequestInfo
|
|
from agentci.workflows.model import AgentResult, ReviewReport
|
|
from agentci.workflows.pull_request import fix_pull_request, iterate_implementation
|
|
from agentci.workflows.render import JobRejected
|
|
from agentci.workflows.services import WorkflowServices
|
|
|
|
|
|
class RecordingRun(JobRun):
|
|
def __init__(self) -> None:
|
|
self.stages: list[str] = []
|
|
self.linked_workflows: list[tuple[str, str]] = []
|
|
self.linked_sessions: list[str] = []
|
|
|
|
async def stage(self, stage: str) -> None:
|
|
self.stages.append(stage)
|
|
|
|
async def link_workflow(self, workflow_id: str, stage: str) -> None:
|
|
self.linked_workflows.append((workflow_id, stage))
|
|
|
|
async def link_session(self, session_id: str) -> None:
|
|
self.linked_sessions.append(session_id)
|
|
|
|
|
|
class FakeRepository:
|
|
def __init__(
|
|
self,
|
|
*,
|
|
workflow: Workflow | None = None,
|
|
plan: Workflow | None = None,
|
|
) -> None:
|
|
self.workflow = workflow
|
|
self.plan = plan
|
|
self.saved_workflows: list[Workflow] = []
|
|
|
|
async def workflow_for_pr(self, *_args: object) -> Workflow | None:
|
|
return self.workflow
|
|
|
|
async def latest_workflow(self, *_args: object) -> Workflow | None:
|
|
return self.plan
|
|
|
|
async def operational_comment_ids(self, *_args: object) -> set[int]:
|
|
return set()
|
|
|
|
async def save_workflow(self, workflow: Workflow) -> None:
|
|
self.workflow = workflow
|
|
self.saved_workflows.append(workflow)
|
|
|
|
|
|
class FakeGitea:
|
|
def __init__(self, pull: PullRequestInfo) -> None:
|
|
self.pull = pull
|
|
self.pull_calls: list[tuple[str, str, int]] = []
|
|
|
|
async def pull_request(self, owner: str, repo: str, number: int) -> PullRequestInfo:
|
|
self.pull_calls.append((owner, repo, number))
|
|
return self.pull
|
|
|
|
async def issue_comments(self, *_args: object) -> list[CommentInfo]:
|
|
return [CommentInfo(1, "alice", "Please add coverage.", "2026-07-01")]
|
|
|
|
async def pull_reviews(self, *_args: object) -> list[dict[str, Any]]:
|
|
return [
|
|
{
|
|
"id": 3,
|
|
"user": {"login": "bob"},
|
|
"state": "REQUEST_CHANGES",
|
|
"body": "Handle empty input.",
|
|
}
|
|
]
|
|
|
|
async def review_comments(self, *_args: object) -> list[dict[str, Any]]:
|
|
return [{"path": "src/widget.py", "new_position": 8, "body": "Add a guard."}]
|
|
|
|
async def pull_commits(self, *_args: object) -> list[dict[str, Any]]:
|
|
return [{"sha": "abcdef1234567890", "commit": {"message": "Initial change"}}]
|
|
|
|
async def issue(self, *_args: object) -> IssueInfo:
|
|
return IssueInfo(7, "Fix widget", "The widget is broken.", "open")
|
|
|
|
|
|
class RecordingGit:
|
|
def __init__(self, *, changed: bool = True) -> None:
|
|
self.changed = changed
|
|
self.calls: list[tuple[Any, ...]] = []
|
|
|
|
async def clone(self, owner: str, repo: str, branch: str, destination: Path) -> str:
|
|
self.calls.append(("clone", owner, repo, branch, destination))
|
|
return "head-sha"
|
|
|
|
async def sync_branch(self, workspace: Path, branch: str) -> str:
|
|
self.calls.append(("sync_branch", workspace, branch))
|
|
return "head-sha"
|
|
|
|
async def has_changes(self, workspace: Path) -> bool:
|
|
self.calls.append(("has_changes", workspace))
|
|
return self.changed
|
|
|
|
async def diff_check(self, workspace: Path) -> None:
|
|
self.calls.append(("diff_check", workspace))
|
|
|
|
async def commit(self, workspace: Path, message: str) -> str:
|
|
self.calls.append(("commit", workspace, message))
|
|
return "new-sha"
|
|
|
|
async def push(self, workspace: Path, branch: str, *, set_upstream: bool = False) -> None:
|
|
self.calls.append(("push", workspace, branch, set_upstream))
|
|
|
|
|
|
class RecordingDevelopment:
|
|
description = "Python 3.13"
|
|
|
|
def __init__(self) -> None:
|
|
self.workspaces: list[Path] = []
|
|
|
|
async def prepare(self, workspace: Path) -> None:
|
|
self.workspaces.append(workspace)
|
|
|
|
|
|
class RecordingPrompts:
|
|
def __init__(self) -> None:
|
|
self.calls: list[tuple[str, dict[str, str]]] = []
|
|
|
|
def render(self, name: str, **values: str) -> str:
|
|
self.calls.append((name, values))
|
|
return f"rendered {name}"
|
|
|
|
|
|
class RecordingOpenCode:
|
|
def __init__(self, responses: list[BaseModel] | None = None) -> None:
|
|
self.responses = list(responses or [])
|
|
self.created_sessions: list[tuple[Path, str]] = []
|
|
self.resume_calls: list[dict[str, Any]] = []
|
|
|
|
async def create_session(self, workspace: Path, title: str) -> str:
|
|
self.created_sessions.append((workspace, title))
|
|
return f"{title}-session"
|
|
|
|
async def resume(self, **values: Any) -> BaseModel:
|
|
self.resume_calls.append(values)
|
|
response = self.responses.pop(0)
|
|
assert isinstance(response, values["result_type"])
|
|
return response
|
|
|
|
|
|
def job(kind: JobKind, *, pr_number: int | None = 12) -> Job:
|
|
return Job(
|
|
id="job-pr",
|
|
kind=kind,
|
|
target_key="org/repo:pr:12",
|
|
repo_owner="org",
|
|
repo_name="repo",
|
|
issue_number=7,
|
|
pr_number=pr_number,
|
|
requester="alice",
|
|
comment_id=9,
|
|
delivery_id="delivery-pr",
|
|
receive_sequence=1,
|
|
command_body="/agent iterate",
|
|
message="Handle the review.",
|
|
)
|
|
|
|
|
|
def implementation_workflow(
|
|
*,
|
|
status: WorkflowStatus = WorkflowStatus.COMPLETED,
|
|
runtime: str = "opencode",
|
|
branch: str | None = "agent/issue-7",
|
|
primary_session_id: str | None = "primary-session",
|
|
reviewer_session_id: str | None = "reviewer-session",
|
|
) -> Workflow:
|
|
return Workflow(
|
|
id="implementation-flow",
|
|
kind=WorkflowKind.IMPLEMENT,
|
|
repo_owner="org",
|
|
repo_name="repo",
|
|
issue_number=7,
|
|
workspace_path=Path("/workspace/implementation"),
|
|
base_sha="base-sha",
|
|
runtime=runtime,
|
|
branch=branch,
|
|
pr_number=12,
|
|
primary_session_id=primary_session_id,
|
|
reviewer_session_id=reviewer_session_id,
|
|
artifact='{"summary_markdown":"Prior","tests":[]}',
|
|
review_json='{"summary":"Prior review","findings":[]}',
|
|
status=status,
|
|
)
|
|
|
|
|
|
def plan_workflow() -> Workflow:
|
|
return Workflow(
|
|
id="plan-flow",
|
|
kind=WorkflowKind.PLAN,
|
|
repo_owner="org",
|
|
repo_name="repo",
|
|
issue_number=7,
|
|
workspace_path=Path("/workspace/plan"),
|
|
base_sha="base",
|
|
artifact="Canonical plan",
|
|
status=WorkflowStatus.COMPLETED,
|
|
)
|
|
|
|
|
|
def pull(
|
|
*,
|
|
state: str = "open",
|
|
merged: bool = False,
|
|
branch: str = "agent/issue-7",
|
|
) -> PullRequestInfo:
|
|
return PullRequestInfo(
|
|
number=12,
|
|
title="Fix widget",
|
|
body="Implementation body",
|
|
state=state,
|
|
merged=merged,
|
|
base_branch="main",
|
|
head_branch=branch,
|
|
head_sha="abcdef1234567890",
|
|
head_owner="contributor",
|
|
head_repo="fork",
|
|
)
|
|
|
|
|
|
def make_services(
|
|
tmp_path: Path,
|
|
*,
|
|
workflow: Workflow | None = None,
|
|
pull_info: PullRequestInfo | None = None,
|
|
plan: Workflow | None = None,
|
|
changed: bool = True,
|
|
responses: list[BaseModel] | None = None,
|
|
) -> tuple[
|
|
WorkflowServices,
|
|
FakeRepository,
|
|
FakeGitea,
|
|
RecordingGit,
|
|
RecordingDevelopment,
|
|
RecordingPrompts,
|
|
RecordingOpenCode,
|
|
]:
|
|
repository = FakeRepository(workflow=workflow, plan=plan)
|
|
gitea = FakeGitea(pull_info or pull())
|
|
git = RecordingGit(changed=changed)
|
|
development = RecordingDevelopment()
|
|
prompts = RecordingPrompts()
|
|
opencode = RecordingOpenCode(responses)
|
|
services = cast(
|
|
WorkflowServices,
|
|
SimpleNamespace(
|
|
settings=SimpleNamespace(
|
|
workspaces_dir=tmp_path / "workspaces",
|
|
implement_model="provider/model",
|
|
implement_variant="high",
|
|
),
|
|
repository=repository,
|
|
gitea=gitea,
|
|
git=git,
|
|
development=development,
|
|
prompts=prompts,
|
|
opencode=opencode,
|
|
),
|
|
)
|
|
return services, repository, gitea, git, development, prompts, opencode
|
|
|
|
|
|
def result(summary: str = "# Refine widget") -> AgentResult:
|
|
return AgentResult(summary_markdown=summary, tests=["pytest: passed"])
|
|
|
|
|
|
async def test_iterate_implementation_reuses_sessions_reviews_commits_and_pushes(
|
|
tmp_path: Path,
|
|
) -> None:
|
|
existing = implementation_workflow()
|
|
services, repository, _, git, development, prompts, opencode = make_services(
|
|
tmp_path,
|
|
workflow=existing,
|
|
plan=plan_workflow(),
|
|
responses=[result(), ReviewReport(summary="Ready", findings=[])],
|
|
)
|
|
run = RecordingRun()
|
|
|
|
body = await iterate_implementation(job(JobKind.ITERATE_IMPLEMENT), run, services)
|
|
|
|
assert run.linked_workflows == [(existing.id, "synchronizing branch")]
|
|
assert development.workspaces == [existing.workspace_path]
|
|
assert opencode.created_sessions == []
|
|
assert [call["session_id"] for call in opencode.resume_calls] == [
|
|
"primary-session",
|
|
"reviewer-session",
|
|
]
|
|
assert [call["result_type"] for call in opencode.resume_calls] == [
|
|
AgentResult,
|
|
ReviewReport,
|
|
]
|
|
iterate_prompt = prompts.calls[0]
|
|
assert iterate_prompt[0] == "implementation_iterate"
|
|
assert iterate_prompt[1]["message"] == "Handle the review."
|
|
assert "Handle empty input." in iterate_prompt[1]["context"]
|
|
review_prompt = prompts.calls[1]
|
|
assert review_prompt[0] == "implementation_review"
|
|
assert review_prompt[1]["artifact"] == "Canonical plan"
|
|
assert "The widget is broken." in review_prompt[1]["issue_context"]
|
|
|
|
assert git.calls == [
|
|
("sync_branch", existing.workspace_path, "agent/issue-7"),
|
|
("has_changes", existing.workspace_path),
|
|
("diff_check", existing.workspace_path),
|
|
("commit", existing.workspace_path, "agent iterate: Refine widget"),
|
|
("push", existing.workspace_path, "agent/issue-7", False),
|
|
]
|
|
saved = repository.saved_workflows[-1]
|
|
assert AgentResult.model_validate_json(saved.artifact or "") == result()
|
|
assert json.loads(saved.review_json or "") == {
|
|
"summary": "Ready",
|
|
"findings": [],
|
|
}
|
|
assert body == (
|
|
"<!-- agentci:iteration workflow=implementation-flow -->\n"
|
|
"## Agent result\n\n# Refine widget\n\n"
|
|
"## Validation\n\n- pytest: passed\n\nCommit: `new-sha`"
|
|
)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("pr_number", "existing", "message"),
|
|
[
|
|
(None, implementation_workflow(), "This command requires a pull request."),
|
|
(
|
|
12,
|
|
None,
|
|
"This is not an open agent-created implementation PR. Use `/agent fix`.",
|
|
),
|
|
(
|
|
12,
|
|
implementation_workflow(status=WorkflowStatus.ACTIVE),
|
|
"This is not an open agent-created implementation PR. Use `/agent fix`.",
|
|
),
|
|
(
|
|
12,
|
|
implementation_workflow(runtime="codex"),
|
|
"The implementation predates OpenCode and cannot be resumed.",
|
|
),
|
|
(
|
|
12,
|
|
implementation_workflow(primary_session_id=None),
|
|
"The implementation sessions cannot be resumed.",
|
|
),
|
|
(
|
|
12,
|
|
implementation_workflow(reviewer_session_id=None),
|
|
"The implementation sessions cannot be resumed.",
|
|
),
|
|
],
|
|
)
|
|
async def test_iterate_implementation_rejects_missing_stale_or_incompatible_workflow(
|
|
tmp_path: Path,
|
|
pr_number: int | None,
|
|
existing: Workflow | None,
|
|
message: str,
|
|
) -> None:
|
|
services, repository, gitea, git, development, _, opencode = make_services(
|
|
tmp_path, workflow=existing
|
|
)
|
|
|
|
with pytest.raises(JobRejected) as error:
|
|
await iterate_implementation(
|
|
job(JobKind.ITERATE_IMPLEMENT, pr_number=pr_number),
|
|
RecordingRun(),
|
|
services,
|
|
)
|
|
|
|
assert str(error.value) == message
|
|
assert repository.saved_workflows == []
|
|
assert gitea.pull_calls == []
|
|
assert git.calls == []
|
|
assert development.workspaces == []
|
|
assert opencode.resume_calls == []
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("pull_info", "existing", "message"),
|
|
[
|
|
(
|
|
pull(state="closed"),
|
|
implementation_workflow(),
|
|
"Implementation iteration requires an open pull request.",
|
|
),
|
|
(
|
|
pull(branch="renamed-branch"),
|
|
implementation_workflow(),
|
|
"The pull request head branch no longer matches its workflow.",
|
|
),
|
|
],
|
|
ids=["closed", "branch-mismatch"],
|
|
)
|
|
async def test_iterate_implementation_rejects_closed_or_stale_branch_before_checkout(
|
|
tmp_path: Path,
|
|
pull_info: PullRequestInfo,
|
|
existing: Workflow,
|
|
message: str,
|
|
) -> None:
|
|
services, repository, _, git, development, _, opencode = make_services(
|
|
tmp_path, workflow=existing, pull_info=pull_info
|
|
)
|
|
|
|
with pytest.raises(JobRejected) as error:
|
|
await iterate_implementation(job(JobKind.ITERATE_IMPLEMENT), RecordingRun(), services)
|
|
|
|
assert str(error.value) == message
|
|
assert repository.saved_workflows == []
|
|
assert git.calls == []
|
|
assert development.workspaces == []
|
|
assert opencode.resume_calls == []
|
|
|
|
|
|
async def test_fix_pull_request_clones_head_runs_agent_commits_and_pushes(
|
|
tmp_path: Path,
|
|
) -> None:
|
|
services, repository, _, git, development, prompts, opencode = make_services(
|
|
tmp_path,
|
|
responses=[result("Fix empty input")],
|
|
)
|
|
run = RecordingRun()
|
|
|
|
body = await fix_pull_request(job(JobKind.FIX), run, services)
|
|
|
|
workspace = tmp_path / "workspaces" / "fix-job-pr" / "repo"
|
|
assert git.calls == [
|
|
("clone", "contributor", "fork", "agent/issue-7", workspace),
|
|
("has_changes", workspace),
|
|
("diff_check", workspace),
|
|
("commit", workspace, "agent fix: Fix empty input"),
|
|
("push", workspace, "agent/issue-7", False),
|
|
]
|
|
assert development.workspaces == [workspace]
|
|
assert opencode.created_sessions == [(workspace, "fix")]
|
|
assert run.linked_sessions == ["fix-session"]
|
|
assert opencode.resume_calls[0]["session_id"] == "fix-session"
|
|
assert opencode.resume_calls[0]["result_type"] is AgentResult
|
|
assert prompts.calls[0][0] == "fix"
|
|
assert "`src/widget.py:8`: Add a guard." in prompts.calls[0][1]["context"]
|
|
assert repository.saved_workflows == []
|
|
assert body == (
|
|
"<!-- agentci:fix workflow=job-pr -->\n"
|
|
"## Agent result\n\nFix empty input\n\n"
|
|
"## Validation\n\n- pytest: passed\n\nCommit: `new-sha`"
|
|
)
|
|
|
|
|
|
async def test_fix_pull_request_rejects_clean_worktree_before_push(
|
|
tmp_path: Path,
|
|
) -> None:
|
|
services, _, _, git, _, _, _ = make_services(
|
|
tmp_path,
|
|
changed=False,
|
|
responses=[result("No changes required")],
|
|
)
|
|
|
|
with pytest.raises(
|
|
JobRejected,
|
|
match="OpenCode completed without producing any file changes",
|
|
):
|
|
await fix_pull_request(job(JobKind.FIX), RecordingRun(), services)
|
|
|
|
assert [call[0] for call in git.calls] == ["clone", "has_changes"]
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
("pr_number", "pull_info", "message"),
|
|
[
|
|
(None, pull(), "This command requires a pull request."),
|
|
(12, pull(state="closed"), "Fixes require an open pull request."),
|
|
(12, pull(state="closed", merged=True), "Fixes require an open pull request."),
|
|
],
|
|
ids=["missing-pr", "closed", "merged"],
|
|
)
|
|
async def test_fix_pull_request_requires_open_pr_before_clone(
|
|
tmp_path: Path,
|
|
pr_number: int | None,
|
|
pull_info: PullRequestInfo,
|
|
message: str,
|
|
) -> None:
|
|
services, _, gitea, git, development, _, opencode = make_services(tmp_path, pull_info=pull_info)
|
|
|
|
with pytest.raises(JobRejected) as error:
|
|
await fix_pull_request(job(JobKind.FIX, pr_number=pr_number), RecordingRun(), services)
|
|
|
|
assert str(error.value) == message
|
|
assert git.calls == []
|
|
assert development.workspaces == []
|
|
assert opencode.resume_calls == []
|
|
assert gitea.pull_calls == ([] if pr_number is None else [("org", "repo", 12)])
|