feat: Task Pipeline + Planner E2E 성공 — stdin기반 GeminiCaller 확정 #task-189 #task-190
This commit is contained in:
@@ -1,10 +1,10 @@
|
||||
"""GeminiCaller — gemini -p 역할별 headless 호출.
|
||||
"""GeminiCaller — gemini headless 호출.
|
||||
|
||||
stdin으로 시스템 프롬프트 + 컨텍스트를 직접 전달합니다.
|
||||
cmd /c 래핑으로 PowerShell 실행 정책 우회.
|
||||
"""
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
import time
|
||||
from pathlib import Path
|
||||
|
||||
@@ -21,12 +21,9 @@ class GeminiCaller:
|
||||
self.last_call_time = 0.0
|
||||
|
||||
async def call(self, role: str, context: str, timeout: int = 120) -> str:
|
||||
"""역할별 프롬프트로 gemini -p 호출.
|
||||
"""역할별 프롬프트로 gemini 호출.
|
||||
|
||||
Args:
|
||||
role: 프롬프트 파일명 (planner, coder, reviewer, tester)
|
||||
context: 전달할 컨텍스트
|
||||
timeout: 최대 대기 시간 (초)
|
||||
시스템 프롬프트와 컨텍스트를 하나로 합쳐 stdin으로 전달.
|
||||
"""
|
||||
# 시스템 프롬프트 로드
|
||||
prompt_file = ROLE_PROMPTS_DIR / f"{role}.md"
|
||||
@@ -35,31 +32,24 @@ class GeminiCaller:
|
||||
else:
|
||||
system_prompt = f"You are a {role}. Respond in Korean."
|
||||
|
||||
# cmd 구성
|
||||
cmd_parts = ["gemini", "-p", context]
|
||||
|
||||
if system_prompt:
|
||||
cmd_parts.extend(["--system", system_prompt])
|
||||
|
||||
cmd_parts.extend(["--approval-mode", "yolo"])
|
||||
|
||||
if self.project_path:
|
||||
cmd_parts.extend(["--include-directories", self.project_path])
|
||||
|
||||
# cmd /c 래핑 (PowerShell 실행 정책 우회)
|
||||
escaped_context = context.replace('"', '\\"')
|
||||
cmd_str = f'gemini -p "{escaped_context}" --approval-mode yolo'
|
||||
if self.project_path:
|
||||
cmd_str += f' --include-directories "{self.project_path}"'
|
||||
# 시스템 프롬프트 + 컨텍스트를 하나의 입력으로 합침
|
||||
full_input = (
|
||||
f"=== SYSTEM INSTRUCTIONS ===\n"
|
||||
f"{system_prompt}\n\n"
|
||||
f"=== USER INPUT ===\n"
|
||||
f"{context}"
|
||||
)
|
||||
|
||||
try:
|
||||
proc = await asyncio.create_subprocess_shell(
|
||||
f'cmd /c {cmd_str}',
|
||||
proc = await asyncio.create_subprocess_exec(
|
||||
"cmd", "/c", "gemini --approval-mode yolo",
|
||||
stdin=asyncio.subprocess.PIPE,
|
||||
stdout=asyncio.subprocess.PIPE,
|
||||
stderr=asyncio.subprocess.PIPE,
|
||||
)
|
||||
stdout, stderr = await asyncio.wait_for(
|
||||
proc.communicate(), timeout=timeout
|
||||
proc.communicate(input=full_input.encode("utf-8")),
|
||||
timeout=timeout
|
||||
)
|
||||
|
||||
self.call_count += 1
|
||||
@@ -85,3 +75,5 @@ class GeminiCaller:
|
||||
async def call_simple(self, prompt: str, timeout: int = 60) -> str:
|
||||
"""시스템 프롬프트 없이 단순 호출."""
|
||||
return await self.call("default", prompt, timeout)
|
||||
|
||||
|
||||
|
||||
141
core/task_pipeline.py
Normal file
141
core/task_pipeline.py
Normal file
@@ -0,0 +1,141 @@
|
||||
"""Task Pipeline — Plan → Code → Review → Ship.
|
||||
|
||||
E2E 파이프라인을 구성하고 실행합니다.
|
||||
"""
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
import re
|
||||
from pathlib import Path
|
||||
from core.project_indexer import ProjectIndex
|
||||
from core.context_manager import ContextManager
|
||||
from core.gemini_caller import GeminiCaller
|
||||
|
||||
|
||||
class TaskPipeline:
|
||||
"""작업 파이프라인: 사용자 요청을 분해하고 순차 실행합니다."""
|
||||
|
||||
def __init__(self, project_path: str, token_budget: int = 50_000):
|
||||
self.project_path = project_path
|
||||
self.index = ProjectIndex(project_path)
|
||||
self.ctx = ContextManager(self.index, token_budget)
|
||||
self.gemini = GeminiCaller(project_path)
|
||||
self.log: list[dict] = []
|
||||
|
||||
def setup(self):
|
||||
"""프로젝트 인덱싱."""
|
||||
self.index.scan()
|
||||
return self
|
||||
|
||||
async def plan(self, user_request: str) -> dict:
|
||||
"""Planner로 작업 분해."""
|
||||
structure = self.index.get_structure_summary()
|
||||
prompt = (
|
||||
f"## User Request\n{user_request}\n\n"
|
||||
f"## Project Structure\n{structure}\n\n"
|
||||
f"Decompose this request into concrete tasks."
|
||||
)
|
||||
|
||||
response = await self.gemini.call("planner", prompt, timeout=90)
|
||||
self._log("plan", user_request, response)
|
||||
|
||||
# JSON 추출
|
||||
plan = self._extract_json(response)
|
||||
return plan or {"summary": response, "tasks": [], "raw": response}
|
||||
|
||||
async def code(self, task: dict) -> str:
|
||||
"""Coder로 코드 수정."""
|
||||
# 관련 파일 컨텍스트 수집
|
||||
context = self.ctx.gather(task.get("description", task.get("title", "")))
|
||||
|
||||
prompt = (
|
||||
f"## Task\n{json.dumps(task, ensure_ascii=False, indent=2)}\n\n"
|
||||
f"## Context\n{context}\n\n"
|
||||
f"Implement the changes described in the task."
|
||||
)
|
||||
|
||||
response = await self.gemini.call("coder", prompt, timeout=120)
|
||||
self._log("code", task.get("title", ""), response)
|
||||
return response
|
||||
|
||||
async def review(self, task: dict, code_output: str) -> dict:
|
||||
"""Reviewer로 코드 리뷰."""
|
||||
prompt = (
|
||||
f"## Task\n{json.dumps(task, ensure_ascii=False, indent=2)}\n\n"
|
||||
f"## Code Output\n{code_output}\n\n"
|
||||
f"Review the code changes."
|
||||
)
|
||||
|
||||
response = await self.gemini.call("reviewer", prompt, timeout=90)
|
||||
self._log("review", task.get("title", ""), response)
|
||||
|
||||
review = self._extract_json(response)
|
||||
return review or {"passed": True, "summary": response, "raw": response}
|
||||
|
||||
async def execute(self, user_request: str) -> dict:
|
||||
"""전체 파이프라인 실행."""
|
||||
result = {
|
||||
"request": user_request,
|
||||
"plan": None,
|
||||
"tasks_completed": [],
|
||||
"reviews": [],
|
||||
}
|
||||
|
||||
# 1. Plan
|
||||
plan = await self.plan(user_request)
|
||||
result["plan"] = plan
|
||||
|
||||
tasks = plan.get("tasks", [])
|
||||
if not tasks:
|
||||
result["error"] = "Planner returned no tasks"
|
||||
return result
|
||||
|
||||
# 2. Code + Review for each task
|
||||
for task in tasks:
|
||||
code_output = await self.code(task)
|
||||
|
||||
review = await self.review(task, code_output)
|
||||
|
||||
if not review.get("passed", True):
|
||||
# 리뷰 실패 시 한 번 재시도
|
||||
code_output = await self.code({
|
||||
**task,
|
||||
"description": task.get("description", "") +
|
||||
f"\n\n## Review Feedback\n{json.dumps(review.get('issues', []), ensure_ascii=False)}"
|
||||
})
|
||||
review = await self.review(task, code_output)
|
||||
|
||||
result["tasks_completed"].append({
|
||||
"task": task,
|
||||
"output": code_output[:500], # 요약
|
||||
"review": review,
|
||||
})
|
||||
|
||||
return result
|
||||
|
||||
def _extract_json(self, text: str) -> dict | None:
|
||||
"""텍스트에서 JSON 블록 추출."""
|
||||
# ```json ... ``` 패턴
|
||||
match = re.search(r"```json\s*\n(.*?)\n\s*```", text, re.DOTALL)
|
||||
if match:
|
||||
try:
|
||||
return json.loads(match.group(1))
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
# { ... } 직접 찾기
|
||||
match = re.search(r"\{.*\}", text, re.DOTALL)
|
||||
if match:
|
||||
try:
|
||||
return json.loads(match.group(0))
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
|
||||
return None
|
||||
|
||||
def _log(self, phase: str, input_summary: str, output: str):
|
||||
self.log.append({
|
||||
"phase": phase,
|
||||
"input": input_summary[:200],
|
||||
"output": output[:500],
|
||||
})
|
||||
29
prompts/coder.md
Normal file
29
prompts/coder.md
Normal file
@@ -0,0 +1,29 @@
|
||||
You are a **Coder** — an AI that implements specific code changes.
|
||||
|
||||
## Your Role
|
||||
|
||||
Given a task description and relevant source files, you MUST:
|
||||
|
||||
1. **Read** the provided source files carefully
|
||||
2. **Implement** the requested changes
|
||||
3. **Output** the complete modified file(s)
|
||||
|
||||
## Output Format
|
||||
|
||||
For each modified file, output:
|
||||
|
||||
```
|
||||
=== FILE: path/to/file.py ===
|
||||
(complete file content here)
|
||||
=== END FILE ===
|
||||
```
|
||||
|
||||
## Rules
|
||||
|
||||
- Output the COMPLETE file, not just the diff
|
||||
- Preserve existing code structure and style
|
||||
- Add comments for non-obvious changes
|
||||
- Handle edge cases
|
||||
- Do NOT add unnecessary changes beyond what was requested
|
||||
- If creating a new file, include proper docstring/header
|
||||
- Respond in the same language as the existing code comments
|
||||
38
prompts/planner.md
Normal file
38
prompts/planner.md
Normal file
@@ -0,0 +1,38 @@
|
||||
You are a **Planner** — an AI that decomposes user requests into concrete, actionable tasks.
|
||||
|
||||
## Your Role
|
||||
|
||||
Given a user request and project context, you MUST:
|
||||
|
||||
1. **Analyze** the request and understand what needs to change
|
||||
2. **Identify** which files are affected
|
||||
3. **Decompose** into specific tasks, each modifying 1-3 files
|
||||
|
||||
## Output Format
|
||||
|
||||
Respond in this exact JSON format:
|
||||
|
||||
```json
|
||||
{
|
||||
"summary": "Brief summary of what needs to be done",
|
||||
"tasks": [
|
||||
{
|
||||
"id": 1,
|
||||
"title": "Task title",
|
||||
"files": ["path/to/file.py"],
|
||||
"description": "Specific changes to make",
|
||||
"type": "create|modify|delete"
|
||||
}
|
||||
],
|
||||
"risk": "low|medium|high",
|
||||
"estimated_calls": 3
|
||||
}
|
||||
```
|
||||
|
||||
## Rules
|
||||
|
||||
- Keep task count minimal (1-5 tasks)
|
||||
- Each task should be independently executable
|
||||
- Order tasks by dependency (dependencies first)
|
||||
- Be specific about what to change, not vague
|
||||
- Respond in Korean for summaries, English for code references
|
||||
37
prompts/reviewer.md
Normal file
37
prompts/reviewer.md
Normal file
@@ -0,0 +1,37 @@
|
||||
You are a **Reviewer** — an AI that reviews code changes for correctness and quality.
|
||||
|
||||
## Your Role
|
||||
|
||||
Given original and modified code, you MUST:
|
||||
|
||||
1. **Compare** the original and modified versions
|
||||
2. **Check** for correctness, bugs, edge cases
|
||||
3. **Verify** the changes match the task requirements
|
||||
4. **Assess** code quality and style
|
||||
|
||||
## Output Format
|
||||
|
||||
```json
|
||||
{
|
||||
"passed": true|false,
|
||||
"issues": [
|
||||
{
|
||||
"severity": "critical|warning|info",
|
||||
"file": "path/to/file.py",
|
||||
"line": 42,
|
||||
"description": "Issue description"
|
||||
}
|
||||
],
|
||||
"suggestions": ["Optional improvement suggestions"],
|
||||
"summary": "Review summary in Korean"
|
||||
}
|
||||
```
|
||||
|
||||
## Rules
|
||||
|
||||
- Be thorough but not pedantic
|
||||
- Focus on correctness over style
|
||||
- Critical issues = must fix before merge
|
||||
- Warnings = should fix, won't block
|
||||
- Info = nice to have improvements
|
||||
- If no issues found, set passed=true with empty issues array
|
||||
56
tests/test_pipeline_e2e.py
Normal file
56
tests/test_pipeline_e2e.py
Normal file
@@ -0,0 +1,56 @@
|
||||
"""E2E Test: Task Pipeline with real Gemini CLI.
|
||||
|
||||
Tests Planner phase against the variet-agent project.
|
||||
"""
|
||||
|
||||
import sys
|
||||
import io
|
||||
import asyncio
|
||||
import json
|
||||
|
||||
if sys.stdout.encoding != "utf-8":
|
||||
sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8", errors="replace")
|
||||
sys.path.insert(0, r"C:\Users\CafeVariet-GL552VW\Desktop\source_diff\variet-agent")
|
||||
|
||||
from core.task_pipeline import TaskPipeline
|
||||
|
||||
PROJECT = r"C:\Users\CafeVariet-GL552VW\Desktop\source_diff\variet-agent"
|
||||
|
||||
|
||||
async def test_planner():
|
||||
print("=" * 60)
|
||||
print("E2E TEST: Planner")
|
||||
print("=" * 60)
|
||||
|
||||
pipeline = TaskPipeline(PROJECT, token_budget=30_000)
|
||||
pipeline.setup()
|
||||
|
||||
plan = await pipeline.plan(
|
||||
"project_indexer.py의 find_relevant 함수가 공백이 포함된 쿼리를 처리하지 못합니다. "
|
||||
"'gemini caller'로 검색하면 gemini_caller.py를 찾지 못합니다. "
|
||||
"밑줄과 공백을 동일하게 처리하도록 개선해주세요."
|
||||
)
|
||||
|
||||
print(f"\n📋 Plan result:")
|
||||
print(json.dumps(plan, ensure_ascii=False, indent=2))
|
||||
|
||||
if plan.get("tasks"):
|
||||
print(f"\n✅ Planner returned {len(plan['tasks'])} tasks")
|
||||
for t in plan["tasks"]:
|
||||
print(f" - {t.get('title', t.get('id', '?'))}: {t.get('description', '')[:80]}")
|
||||
else:
|
||||
print(f"\n⚠️ No structured tasks, raw response:")
|
||||
print(plan.get("raw", plan.get("summary", ""))[:500])
|
||||
|
||||
return plan
|
||||
|
||||
|
||||
async def main():
|
||||
plan = await test_planner()
|
||||
print(f"\n{'=' * 60}")
|
||||
print(f"Gemini calls: {1}")
|
||||
print(f"✅ E2E Planner test complete!")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
asyncio.run(main())
|
||||
38
tests/test_quick.py
Normal file
38
tests/test_quick.py
Normal file
@@ -0,0 +1,38 @@
|
||||
"""Debug: test --system flag with stdin."""
|
||||
import asyncio, tempfile, os
|
||||
|
||||
async def test():
|
||||
# Write system prompt to temp file
|
||||
sys_file = tempfile.NamedTemporaryFile(mode="w", suffix=".md", delete=False, encoding="utf-8")
|
||||
sys_file.write("You are a planner. Respond with a JSON object containing: summary, tasks array.")
|
||||
sys_file.close()
|
||||
|
||||
ctx = "List 2 tasks to improve a Python project. Respond in JSON format."
|
||||
|
||||
print(f"System file: {sys_file.name}")
|
||||
print(f"Context length: {len(ctx)}")
|
||||
|
||||
# Test 1: with --system
|
||||
proc = await asyncio.create_subprocess_exec(
|
||||
"cmd", "/c",
|
||||
f"gemini --system {sys_file.name} --approval-mode yolo",
|
||||
stdin=asyncio.subprocess.PIPE,
|
||||
stdout=asyncio.subprocess.PIPE,
|
||||
stderr=asyncio.subprocess.PIPE,
|
||||
)
|
||||
stdout, stderr = await asyncio.wait_for(
|
||||
proc.communicate(input=ctx.encode("utf-8")),
|
||||
timeout=90
|
||||
)
|
||||
|
||||
out = stdout.decode('utf-8', errors='replace')
|
||||
err = stderr.decode('utf-8', errors='replace')
|
||||
|
||||
lines = [l for l in out.splitlines() if "YOLO" not in l and "Loaded cached" not in l]
|
||||
print(f"\nSTDOUT ({len(out)} bytes): {out[:500]}")
|
||||
print(f"STDERR ({len(err)} bytes): {err[:500]}")
|
||||
print(f"CLEANED: {'|'.join(lines[:10])}")
|
||||
|
||||
os.unlink(sys_file.name)
|
||||
|
||||
asyncio.run(test())
|
||||
Reference in New Issue
Block a user