docs: update HANDOFF for Phase 4 completion

- webapp.py lifespan now initializes json_logging, recovery, cost_guard, task_history
- Dispatcher receives cost_guard and task_history dependencies
- ContainerCleaner starts if Docker is available
- Added /health/costs endpoint for API cost monitoring
- Added tests/test_smoke.py with 2 tests for basic health and costs endpoint
- All existing health tests still pass (8 tests)

HANDOFF.md

@@ -72,12 +72,21 @@ galaxis-agent 리포에 16개 커밋, 40개 테스트 전부 통과, Gitea에 pu
 | Task 7: Health Check | ✅ COMPLETE | `d35efae` | /health, /health/gitea, /health/discord, /health/queue, 8 테스트 |
 | Task 8: 전체 검증 | ✅ COMPLETE | `a58bbca` | 통합 테스트 (webhook→queue→dispatcher), 107 테스트 통과, import 확인 |
 
-### Phase 4: 미작성
+### Phase 4: 안정화 & 자율 모드 — COMPLETE
 
-- CostGuard (일일/작업당 API 비용 제한)
-- 복구 메커니즘 (서버 재시작 시 미완료 작업 복구, 좀비 컨테이너 정리)
-- 자동 머지 모드 (autonomous 설정 + E2E 통과 조건)
-- 구조화 로깅, 작업 이력 DB, 스모크 테스트
+**실행 방식**: Subagent-Driven Development (5개 독립 Task 병렬 → 2개 순차)
+
+7개 커밋, **139개 테스트 통과** (107 Phase1-3 + 32 Phase4), Gitea에 push 완료.
+
+| Task | 상태 | 커밋 | 설명 |
+|------|------|------|------|
+| Task 1: CostGuard | ✅ COMPLETE | `edeb336` | API 비용 추적/제한, 일일/작업당 한도, 8 테스트 |
+| Task 2: TaskHistory | ✅ COMPLETE | `db6e9b4` | 완료 작업 이력 DB (SQLite), 4 테스트 |
+| Task 3: Dispatcher 연동 | ✅ COMPLETE | `c0cb4b7` | CostGuard+TaskHistory를 Dispatcher에 통합, 2 테스트 |
+| Task 4: JSON 로깅 | ✅ COMPLETE | `3f0d021` | 구조화 JSON 로깅, LOG_FORMAT 설정, 5 테스트 |
+| Task 5: Recovery | ✅ COMPLETE | `e82dfe1` | 서버 시작 시 복구, ContainerCleaner (30분 주기), 4 테스트 |
+| Task 6: AutoMerge | ✅ COMPLETE | `0c4c22b` | E2E 조건부 자동 머지, blocked_paths 보호, 7 테스트 |
+| Task 7: webapp 통합 | ✅ COMPLETE | `b2c52ab` | Lifespan에 전 컴포넌트 통합, /health/costs 엔드포인트, 2 테스트 |
 
 ## What Worked
 
@@ -121,18 +130,19 @@ result = await loop.run_in_executor(None, sandbox_backend.execute, cmd)
 
 ## Next Steps
 
-### Phase 4 플랜 작성 필요
+### Phase 5: 배포 & 모니터링
 
-Phase 3이 완료되었으므로, Phase 4 플랜을 작성해야 한다:
-- CostGuard (일일/작업당 API 비용 제한)
-- 복구 메커니즘 (서버 재시작 시 미완료 작업 복구, 좀비 컨테이너 정리)
-- 자동 머지 모드 (autonomous 설정 + E2E 통과 조건)
-- 구조화 로깅 (JSON 포맷), 작업 이력 DB, 스모크 테스트
+Phase 4 완료로 프로덕션 안정성 확보. Phase 5에서는:
+- Oracle VM 배포 자동화 (Ansible/Docker Compose)
+- 모니터링 대시보드 (Grafana + SQLite → metrics)
+- 알림 고도화 (Gitea PR 코멘트에 비용/소요시간 포함)
+- 멀티 리포 지원 (galaxis-po 외 다른 리포)
+- 1주일 conservative 운영 후 autonomous 전환
 
 ### 실행 방법
 
 ```bash
 cd ~/workspace/quant/galaxis-agent
 git log --oneline
-uv run pytest tests/ -v  # 107 테스트 통과 확인
+uv run pytest tests/ -v  # 139 테스트 통과 확인
 ```

agent/auto_merge.py

@@ -0,0 +1,61 @@
+"""E2E 조건부 자동 머지 로직.
+
+autonomous 모드에서 조건을 검증하여 PR을 자동 머지한다.
+"""
+from __future__ import annotations
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def should_auto_merge(
+    auto_merge: bool, require_e2e: bool, max_files_changed: int,
+    blocked_paths: list[str], changed_files: list[str],
+    tests_passed: bool, e2e_passed: bool,
+) -> bool:
+    if not auto_merge:
+        return False
+    if not tests_passed:
+        return False
+    if require_e2e and not e2e_passed:
+        return False
+    if len(changed_files) > max_files_changed:
+        return False
+    for f in changed_files:
+        for blocked in blocked_paths:
+            if f == blocked or f.endswith("/" + blocked):
+                return False
+    return True
+
+
+class AutoMergeChecker:
+    def __init__(
+        self, auto_merge: bool = False, require_e2e: bool = False,
+        max_files_changed: int = 10, blocked_paths: list[str] | None = None,
+    ):
+        self._auto_merge = auto_merge
+        self._require_e2e = require_e2e
+        self._max_files_changed = max_files_changed
+        self._blocked_paths = blocked_paths or []
+
+    async def try_merge(
+        self, owner: str, repo: str, pr_number: int,
+        changed_files: list[str], tests_passed: bool, e2e_passed: bool,
+    ) -> dict:
+        can_merge = should_auto_merge(
+            auto_merge=self._auto_merge, require_e2e=self._require_e2e,
+            max_files_changed=self._max_files_changed, blocked_paths=self._blocked_paths,
+            changed_files=changed_files, tests_passed=tests_passed, e2e_passed=e2e_passed,
+        )
+        if not can_merge:
+            return {"merged": False, "reason": "conditions not met"}
+        try:
+            from agent.utils.gitea_client import get_gitea_client
+            client = get_gitea_client()
+            await client.merge_pull_request(owner=owner, repo=repo, pr_number=pr_number)
+            logger.info("Auto-merged PR #%d on %s/%s", pr_number, owner, repo)
+            return {"merged": True, "reason": "all conditions met"}
+        except Exception as e:
+            logger.exception("Auto-merge failed for PR #%d", pr_number)
+            return {"merged": False, "reason": f"merge failed: {e}"}

agent/config.py

@@ -62,4 +62,7 @@ class Settings(BaseSettings):
     DAILY_COST_LIMIT_USD: float = 10.0
     PER_TASK_COST_LIMIT_USD: float = 3.0
 
+    # Logging
+    LOG_FORMAT: str = "json"
+
     model_config = {"env_file": ".env", "extra": "ignore"}

agent/cost_guard.py

@@ -0,0 +1,125 @@
+"""API 비용 추적 및 제한.
+
+Anthropic API 응답의 usage 필드에서 토큰 비용을 추적하고,
+일일/작업당 한도를 초과하면 작업을 차단한다.
+"""
+from __future__ import annotations
+
+import json
+import logging
+import os
+from datetime import datetime, timezone, date
+
+import aiosqlite
+
+logger = logging.getLogger(__name__)
+
+_CREATE_TABLE = """
+CREATE TABLE IF NOT EXISTS cost_records (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    task_id TEXT NOT NULL,
+    tokens_input INTEGER NOT NULL,
+    tokens_output INTEGER NOT NULL,
+    cost_usd REAL NOT NULL,
+    recorded_at TEXT NOT NULL
+)
+"""
+
+DEFAULT_INPUT_COST_PER_TOKEN = 3.0 / 1_000_000
+DEFAULT_OUTPUT_COST_PER_TOKEN = 15.0 / 1_000_000
+
+
+class CostGuard:
+    """API 비용 추적 및 제한."""
+
+    def __init__(
+        self,
+        db_path: str = "/data/cost_guard.db",
+        daily_limit: float = 10.0,
+        per_task_limit: float = 3.0,
+        input_cost_per_token: float = DEFAULT_INPUT_COST_PER_TOKEN,
+        output_cost_per_token: float = DEFAULT_OUTPUT_COST_PER_TOKEN,
+    ):
+        self._db_path = db_path
+        self._daily_limit = daily_limit
+        self._per_task_limit = per_task_limit
+        self._input_cost = input_cost_per_token
+        self._output_cost = output_cost_per_token
+        self._db: aiosqlite.Connection | None = None
+
+    async def initialize(self) -> None:
+        self._db = await aiosqlite.connect(self._db_path)
+        await self._db.execute(_CREATE_TABLE)
+        await self._db.commit()
+
+    async def close(self) -> None:
+        if self._db:
+            await self._db.close()
+
+    def calculate_cost(self, tokens_input: int, tokens_output: int) -> float:
+        return tokens_input * self._input_cost + tokens_output * self._output_cost
+
+    async def record_usage(self, task_id: str, tokens_input: int, tokens_output: int) -> float:
+        cost = self.calculate_cost(tokens_input, tokens_output)
+        now = datetime.now(timezone.utc).isoformat()
+        await self._db.execute(
+            "INSERT INTO cost_records (task_id, tokens_input, tokens_output, cost_usd, recorded_at) VALUES (?, ?, ?, ?, ?)",
+            (task_id, tokens_input, tokens_output, cost, now),
+        )
+        await self._db.commit()
+        logger.info("Recorded cost $%.4f for task %s (in=%d, out=%d)", cost, task_id, tokens_input, tokens_output)
+        return cost
+
+    async def get_daily_cost(self) -> float:
+        today = date.today().isoformat()
+        cursor = await self._db.execute(
+            "SELECT COALESCE(SUM(cost_usd), 0) FROM cost_records WHERE recorded_at >= ?", (today,),
+        )
+        row = await cursor.fetchone()
+        return row[0]
+
+    async def get_task_cost(self, task_id: str) -> float:
+        cursor = await self._db.execute(
+            "SELECT COALESCE(SUM(cost_usd), 0) FROM cost_records WHERE task_id = ?", (task_id,),
+        )
+        row = await cursor.fetchone()
+        return row[0]
+
+    async def check_daily_limit(self) -> bool:
+        daily = await self.get_daily_cost()
+        return daily < self._daily_limit
+
+    async def check_task_limit(self, task_id: str) -> bool:
+        task_cost = await self.get_task_cost(task_id)
+        return task_cost < self._per_task_limit
+
+    async def get_daily_summary(self) -> dict:
+        today = date.today().isoformat()
+        cursor = await self._db.execute(
+            "SELECT COUNT(*), COALESCE(SUM(cost_usd), 0), COALESCE(SUM(tokens_input), 0), COALESCE(SUM(tokens_output), 0) FROM cost_records WHERE recorded_at >= ?",
+            (today,),
+        )
+        row = await cursor.fetchone()
+        total_cost = row[1]
+        return {
+            "record_count": row[0],
+            "total_cost_usd": round(total_cost, 4),
+            "daily_limit_usd": self._daily_limit,
+            "remaining_usd": round(max(0, self._daily_limit - total_cost), 4),
+            "total_tokens_input": row[2],
+            "total_tokens_output": row[3],
+        }
+
+
+_guard: CostGuard | None = None
+
+
+async def get_cost_guard() -> CostGuard:
+    global _guard
+    if _guard is None:
+        db_path = os.environ.get("COST_GUARD_DB", "/data/cost_guard.db")
+        daily_limit = float(os.environ.get("DAILY_COST_LIMIT_USD", "10.0"))
+        per_task_limit = float(os.environ.get("PER_TASK_COST_LIMIT_USD", "3.0"))
+        _guard = CostGuard(db_path=db_path, daily_limit=daily_limit, per_task_limit=per_task_limit)
+        await _guard.initialize()
+    return _guard

agent/dispatcher.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 import asyncio
 import logging
 import os
+from datetime import datetime, timezone
 from typing import Any
 
 from agent.task_queue import PersistentTaskQueue
@@ -22,9 +23,13 @@ class Dispatcher:
         self,
         task_queue: PersistentTaskQueue,
         poll_interval: float = 2.0,
+        cost_guard: "CostGuard | None" = None,
+        task_history: "TaskHistory | None" = None,
     ):
         self._queue = task_queue
         self._poll_interval = poll_interval
+        self._cost_guard = cost_guard
+        self._task_history = task_history
         self._running = False
         self._task: asyncio.Task | None = None
 
@@ -56,21 +61,83 @@ class Dispatcher:
 
     async def _poll_once(self) -> None:
         """큐에서 작업을 하나 꺼내 처리한다."""
+        # Check daily limit before dequeuing
+        if self._cost_guard:
+            if not await self._cost_guard.check_daily_limit():
+                logger.warning("Daily cost limit exceeded, skipping task dequeue")
+                return
+
         task = await self._queue.dequeue()
         if not task:
             return
 
         logger.info("Processing task %s (thread %s)", task["id"], task["thread_id"])
 
+        start_time = datetime.now(timezone.utc)
         try:
             result = await self._run_agent_for_task(task)
             await self._queue.mark_completed(task["id"], result=result)
             logger.info("Task %s completed successfully", task["id"])
+
+            # Record cost and history after successful completion
+            tokens_input = result.get("tokens_input", 0)
+            tokens_output = result.get("tokens_output", 0)
+
+            if self._cost_guard:
+                await self._cost_guard.record_usage(
+                    task["id"],
+                    tokens_input=tokens_input,
+                    tokens_output=tokens_output,
+                )
+
+            if self._task_history:
+                end_time = datetime.now(timezone.utc)
+                duration_seconds = (end_time - start_time).total_seconds()
+                cost_usd = self._cost_guard.calculate_cost(tokens_input, tokens_output) if self._cost_guard else 0.0
+                payload = task["payload"]
+
+                await self._task_history.record(
+                    task_id=task["id"],
+                    thread_id=task["thread_id"],
+                    issue_number=payload.get("issue_number", 0),
+                    repo_name=payload.get("repo_name", ""),
+                    source=task["source"],
+                    status="completed",
+                    created_at=task["created_at"],
+                    completed_at=end_time.isoformat(),
+                    duration_seconds=duration_seconds,
+                    tokens_input=tokens_input,
+                    tokens_output=tokens_output,
+                    cost_usd=cost_usd,
+                )
+
         except Exception as e:
             logger.exception("Task %s failed", task["id"])
             await self._queue.mark_failed(task["id"], error=str(e))
             await self._notify_failure(task, str(e))
 
+            # Record history after failure
+            if self._task_history:
+                end_time = datetime.now(timezone.utc)
+                duration_seconds = (end_time - start_time).total_seconds()
+                payload = task["payload"]
+
+                await self._task_history.record(
+                    task_id=task["id"],
+                    thread_id=task["thread_id"],
+                    issue_number=payload.get("issue_number", 0),
+                    repo_name=payload.get("repo_name", ""),
+                    source=task["source"],
+                    status="failed",
+                    created_at=task["created_at"],
+                    completed_at=end_time.isoformat(),
+                    duration_seconds=duration_seconds,
+                    tokens_input=0,
+                    tokens_output=0,
+                    cost_usd=0.0,
+                    error_message=str(e),
+                )
+
     async def _run_agent_for_task(self, task: dict) -> dict[str, Any]:
         """작업에 대해 에이전트를 실행한다."""
         from agent.server import get_agent

agent/json_logging.py

@@ -0,0 +1,55 @@
+"""구조화 JSON 로깅.
+
+LOG_FORMAT 환경변수로 json | text 선택 가능.
+"""
+from __future__ import annotations
+
+import json
+import logging
+import traceback
+from datetime import datetime, timezone
+
+_BUILTIN_ATTRS = {
+    "args", "created", "exc_info", "exc_text", "filename", "funcName",
+    "levelname", "levelno", "lineno", "module", "msecs", "message", "msg",
+    "name", "pathname", "process", "processName", "relativeCreated",
+    "stack_info", "thread", "threadName", "taskName",
+}
+
+
+class JsonFormatter(logging.Formatter):
+    def format(self, record: logging.LogRecord) -> str:
+        log_data = {
+            "timestamp": datetime.fromtimestamp(record.created, tz=timezone.utc).isoformat(),
+            "level": record.levelname,
+            "logger": record.name,
+            "message": record.getMessage(),
+        }
+        for key, value in record.__dict__.items():
+            if key not in _BUILTIN_ATTRS and not key.startswith("_"):
+                try:
+                    json.dumps(value)
+                    log_data[key] = value
+                except (TypeError, ValueError):
+                    log_data[key] = str(value)
+        if record.exc_info and record.exc_info[0] is not None:
+            log_data["exception"] = "".join(traceback.format_exception(*record.exc_info))
+        return json.dumps(log_data, ensure_ascii=False)
+
+
+def setup_logging(
+    log_format: str = "json",
+    level: int = logging.INFO,
+    logger: logging.Logger | None = None,
+) -> None:
+    target = logger or logging.getLogger()
+    if log_format == "json":
+        formatter = JsonFormatter()
+    else:
+        formatter = logging.Formatter(
+            "%(asctime)s %(levelname)s %(name)s: %(message)s",
+            datefmt="%Y-%m-%dT%H:%M:%S",
+        )
+    for handler in target.handlers:
+        handler.setFormatter(formatter)
+    target.setLevel(level)

agent/recovery.py

@@ -0,0 +1,90 @@
+"""서버 시작 시 복구 + 좀비 컨테이너 정리."""
+from __future__ import annotations
+
+import asyncio
+import logging
+from datetime import datetime, timezone
+
+from agent.task_queue import PersistentTaskQueue
+
+logger = logging.getLogger(__name__)
+
+
+async def recover_on_startup(task_queue: PersistentTaskQueue) -> None:
+    reset_count = await task_queue.reset_running_to_pending()
+    if reset_count:
+        logger.info("Recovery: reset %d running task(s) to pending", reset_count)
+    await _cleanup_zombie_containers()
+
+
+async def _cleanup_zombie_containers() -> int:
+    try:
+        import docker
+        client = docker.from_env()
+        containers = client.containers.list(
+            filters={"label": "galaxis-agent-sandbox"}, all=True,
+        )
+        cleaned = 0
+        for container in containers:
+            try:
+                container.stop(timeout=10)
+                container.remove(force=True)
+                cleaned += 1
+                logger.info("Recovery: removed zombie container %s", container.name)
+            except Exception:
+                logger.warning("Recovery: failed to remove container %s", container.name)
+        return cleaned
+    except Exception:
+        logger.debug("Recovery: Docker not available, skipping container cleanup")
+        return 0
+
+
+class ContainerCleaner:
+    def __init__(self, docker_client=None, max_age_seconds: int = 1200, interval_seconds: int = 1800):
+        self._docker = docker_client
+        self._max_age = max_age_seconds
+        self._interval = interval_seconds
+        self._running = False
+        self._task: asyncio.Task | None = None
+
+    async def start(self) -> None:
+        self._running = True
+        self._task = asyncio.create_task(self._loop())
+
+    async def stop(self) -> None:
+        self._running = False
+        if self._task:
+            self._task.cancel()
+            try:
+                await self._task
+            except asyncio.CancelledError:
+                pass
+
+    async def _loop(self) -> None:
+        while self._running:
+            try:
+                await self.cleanup_once()
+            except Exception:
+                logger.exception("ContainerCleaner error")
+            await asyncio.sleep(self._interval)
+
+    async def cleanup_once(self) -> int:
+        if not self._docker:
+            return 0
+        now = datetime.now(timezone.utc)
+        containers = self._docker.containers.list(
+            filters={"label": "galaxis-agent-sandbox"}, all=True,
+        )
+        removed = 0
+        for container in containers:
+            created_str = container.attrs.get("Created", "")
+            try:
+                created = datetime.fromisoformat(created_str.replace("Z", "+00:00"))
+                age = (now - created).total_seconds()
+                if age > self._max_age:
+                    container.stop(timeout=10)
+                    container.remove(force=True)
+                    removed += 1
+            except Exception:
+                logger.debug("Failed to check/remove container %s", getattr(container, "name", "unknown"))
+        return removed

agent/task_history.py

@@ -0,0 +1,84 @@
+"""완료 작업 이력 DB.
+
+작업의 비용, 소요시간, 토큰 사용량을 SQLite에 기록한다.
+"""
+from __future__ import annotations
+
+import logging
+import os
+
+import aiosqlite
+
+logger = logging.getLogger(__name__)
+
+_CREATE_TABLE = """
+CREATE TABLE IF NOT EXISTS task_history (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    task_id TEXT UNIQUE NOT NULL,
+    thread_id TEXT NOT NULL,
+    issue_number INTEGER NOT NULL DEFAULT 0,
+    repo_name TEXT NOT NULL DEFAULT '',
+    source TEXT NOT NULL DEFAULT '',
+    status TEXT NOT NULL,
+    created_at TEXT NOT NULL,
+    completed_at TEXT NOT NULL,
+    duration_seconds REAL NOT NULL DEFAULT 0,
+    tokens_input INTEGER NOT NULL DEFAULT 0,
+    tokens_output INTEGER NOT NULL DEFAULT 0,
+    cost_usd REAL NOT NULL DEFAULT 0,
+    error_message TEXT NOT NULL DEFAULT ''
+)
+"""
+
+
+class TaskHistory:
+    def __init__(self, db_path: str = "/data/task_history.db"):
+        self._db_path = db_path
+        self._db: aiosqlite.Connection | None = None
+
+    async def initialize(self) -> None:
+        self._db = await aiosqlite.connect(self._db_path)
+        self._db.row_factory = aiosqlite.Row
+        await self._db.execute(_CREATE_TABLE)
+        await self._db.commit()
+
+    async def close(self) -> None:
+        if self._db:
+            await self._db.close()
+
+    async def record(
+        self, task_id: str, thread_id: str, issue_number: int, repo_name: str,
+        source: str, status: str, created_at: str, completed_at: str,
+        duration_seconds: float, tokens_input: int, tokens_output: int,
+        cost_usd: float, error_message: str = "",
+    ) -> None:
+        await self._db.execute(
+            "INSERT OR REPLACE INTO task_history "
+            "(task_id, thread_id, issue_number, repo_name, source, status, "
+            "created_at, completed_at, duration_seconds, tokens_input, tokens_output, "
+            "cost_usd, error_message) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
+            (task_id, thread_id, issue_number, repo_name, source, status,
+             created_at, completed_at, duration_seconds, tokens_input, tokens_output,
+             cost_usd, error_message),
+        )
+        await self._db.commit()
+        logger.info("Recorded history: task=%s status=%s cost=$%.4f", task_id, status, cost_usd)
+
+    async def get_recent(self, limit: int = 20) -> list[dict]:
+        cursor = await self._db.execute(
+            "SELECT * FROM task_history ORDER BY completed_at DESC LIMIT ?", (limit,),
+        )
+        rows = await cursor.fetchall()
+        return [dict(row) for row in rows]
+
+
+_history: TaskHistory | None = None
+
+
+async def get_task_history() -> TaskHistory:
+    global _history
+    if _history is None:
+        db_path = os.environ.get("TASK_HISTORY_DB", "/data/task_history.db")
+        _history = TaskHistory(db_path=db_path)
+        await _history.initialize()
+    return _history

agent/task_queue.py

@@ -133,6 +133,24 @@ class PersistentTaskQueue:
         row = await cursor.fetchone()
         return row["cnt"] > 0
 
+    async def reset_running_to_pending(self) -> int:
+        """running 상태 작업을 pending으로 리셋한다 (복구용).
+
+        Returns:
+            리셋된 작업 수.
+        """
+        cursor = await self._db.execute(
+            "SELECT COUNT(*) as cnt FROM tasks WHERE status = 'running'"
+        )
+        row = await cursor.fetchone()
+        count = row["cnt"]
+        if count:
+            await self._db.execute(
+                "UPDATE tasks SET status = 'pending', started_at = NULL WHERE status = 'running'"
+            )
+            await self._db.commit()
+        return count
+
 
 # 지연 초기화 싱글턴
 _queue: PersistentTaskQueue | None = None

agent/webapp.py

@@ -24,14 +24,43 @@ async def lifespan(app: FastAPI):
     from agent.message_store import get_message_store
     from agent.dispatcher import Dispatcher
     from agent.integrations.discord_handler import DiscordHandler
+    from agent.json_logging import setup_logging
+    from agent.recovery import recover_on_startup, ContainerCleaner
+    from agent.cost_guard import get_cost_guard
+    from agent.task_history import get_task_history
+
+    # 구조화 로깅 설정
+    setup_logging(log_format=os.environ.get("LOG_FORMAT", "json"))
 
     task_queue = await get_task_queue()
     message_store = await get_message_store()
 
-    dispatcher = Dispatcher(task_queue=task_queue)
+    # 서버 시작 시 복구
+    await recover_on_startup(task_queue)
+
+    # CostGuard + TaskHistory 초기화
+    cost_guard = await get_cost_guard()
+    task_history = await get_task_history()
+
+    # Dispatcher에 CostGuard + TaskHistory 주입
+    dispatcher = Dispatcher(task_queue=task_queue, cost_guard=cost_guard, task_history=task_history)
     await dispatcher.start()
     app.state.dispatcher = dispatcher
 
+    # ContainerCleaner 시작
+    container_cleaner = None
+    try:
+        import docker
+        docker_client = docker.from_env()
+        sandbox_timeout = int(os.environ.get("SANDBOX_TIMEOUT", "600"))
+        container_cleaner = ContainerCleaner(
+            docker_client=docker_client,
+            max_age_seconds=sandbox_timeout * 2,
+        )
+        await container_cleaner.start()
+    except Exception:
+        logger.debug("Docker not available, container cleanup disabled")
+
     discord_token = os.environ.get("DISCORD_TOKEN", "")
     discord_handler = None
     if discord_token:
@@ -43,8 +72,12 @@ async def lifespan(app: FastAPI):
     yield
 
     await dispatcher.stop()
+    if container_cleaner:
+        await container_cleaner.stop()
     if discord_handler:
         await discord_handler.close()
+    await cost_guard.close()
+    await task_history.close()
     await task_queue.close()
     await message_store.close()
     logger.info("Application shutdown complete")
@@ -170,6 +203,14 @@ async def health_queue():
     }
 
 
+@app.get("/health/costs")
+async def health_costs():
+    """API 비용 현황을 반환한다."""
+    from agent.cost_guard import get_cost_guard
+    guard = await get_cost_guard()
+    return await guard.get_daily_summary()
+
+
 @app.post("/webhooks/gitea")
 @limiter.limit("10/minute")
 async def gitea_webhook(request: Request):

tests/test_auto_merge.py

@@ -0,0 +1,70 @@
+import pytest
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from agent.auto_merge import should_auto_merge, AutoMergeChecker
+
+
+def test_should_not_merge_when_disabled():
+    result = should_auto_merge(
+        auto_merge=False, require_e2e=False, max_files_changed=10,
+        blocked_paths=[".env"], changed_files=["backend/app/main.py"],
+        tests_passed=True, e2e_passed=True,
+    )
+    assert result is False
+
+
+def test_should_merge_when_all_conditions_met():
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=True, max_files_changed=10,
+        blocked_paths=[".env", "quant.md"],
+        changed_files=["backend/app/main.py", "backend/tests/test_main.py"],
+        tests_passed=True, e2e_passed=True,
+    )
+    assert result is True
+
+
+def test_should_not_merge_when_tests_fail():
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=False, max_files_changed=10,
+        blocked_paths=[], changed_files=["a.py"],
+        tests_passed=False, e2e_passed=False,
+    )
+    assert result is False
+
+
+def test_should_not_merge_when_e2e_required_but_failed():
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=True, max_files_changed=10,
+        blocked_paths=[], changed_files=["a.py"],
+        tests_passed=True, e2e_passed=False,
+    )
+    assert result is False
+
+
+def test_should_not_merge_when_too_many_files():
+    files = [f"file{i}.py" for i in range(15)]
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=False, max_files_changed=10,
+        blocked_paths=[], changed_files=files,
+        tests_passed=True, e2e_passed=True,
+    )
+    assert result is False
+
+
+def test_should_not_merge_when_blocked_path_modified():
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=False, max_files_changed=10,
+        blocked_paths=[".env", "quant.md"],
+        changed_files=["backend/app/main.py", ".env"],
+        tests_passed=True, e2e_passed=True,
+    )
+    assert result is False
+
+
+def test_should_merge_without_e2e_when_not_required():
+    result = should_auto_merge(
+        auto_merge=True, require_e2e=False, max_files_changed=10,
+        blocked_paths=[], changed_files=["a.py"],
+        tests_passed=True, e2e_passed=False,
+    )
+    assert result is True

tests/test_cost_guard.py

@@ -0,0 +1,91 @@
+import pytest
+import os
+import tempfile
+from datetime import datetime, timezone
+
+from agent.cost_guard import CostGuard
+
+
+@pytest.fixture
+async def cost_guard():
+    fd, db_path = tempfile.mkstemp(suffix=".db")
+    os.close(fd)
+    guard = CostGuard(db_path=db_path, daily_limit=10.0, per_task_limit=3.0)
+    await guard.initialize()
+    yield guard
+    await guard.close()
+    os.unlink(db_path)
+
+
+@pytest.mark.asyncio
+async def test_record_usage(cost_guard):
+    """API 사용량을 기록한다."""
+    await cost_guard.record_usage(
+        task_id="task-1",
+        tokens_input=1000,
+        tokens_output=500,
+    )
+    daily = await cost_guard.get_daily_cost()
+    assert daily > 0
+
+
+@pytest.mark.asyncio
+async def test_calculate_cost(cost_guard):
+    """토큰에서 비용을 계산한다."""
+    cost = cost_guard.calculate_cost(tokens_input=1_000_000, tokens_output=1_000_000)
+    # input: $3/MTok + output: $15/MTok = $18
+    assert cost == pytest.approx(18.0, rel=0.01)
+
+
+@pytest.mark.asyncio
+async def test_check_daily_limit_ok(cost_guard):
+    """일일 한도 내에서 True를 반환한다."""
+    result = await cost_guard.check_daily_limit()
+    assert result is True
+
+
+@pytest.mark.asyncio
+async def test_check_daily_limit_exceeded(cost_guard):
+    """일일 한도 초과 시 False를 반환한다."""
+    for i in range(5):
+        await cost_guard.record_usage(f"task-{i}", tokens_input=1_000_000, tokens_output=200_000)
+    result = await cost_guard.check_daily_limit()
+    assert result is False
+
+
+@pytest.mark.asyncio
+async def test_check_task_limit_ok(cost_guard):
+    """작업당 한도 내에서 True를 반환한다."""
+    await cost_guard.record_usage("task-1", tokens_input=100, tokens_output=50)
+    result = await cost_guard.check_task_limit("task-1")
+    assert result is True
+
+
+@pytest.mark.asyncio
+async def test_check_task_limit_exceeded(cost_guard):
+    """작업당 한도 초과 시 False를 반환한다."""
+    await cost_guard.record_usage("task-1", tokens_input=1_000_000, tokens_output=100_000)
+    result = await cost_guard.check_task_limit("task-1")
+    assert result is False
+
+
+@pytest.mark.asyncio
+async def test_get_task_cost(cost_guard):
+    """특정 작업의 누적 비용을 반환한다."""
+    await cost_guard.record_usage("task-1", tokens_input=1000, tokens_output=500)
+    await cost_guard.record_usage("task-1", tokens_input=2000, tokens_output=1000)
+    cost = await cost_guard.get_task_cost("task-1")
+    assert cost > 0
+    other_cost = await cost_guard.get_task_cost("task-2")
+    assert other_cost == 0.0
+
+
+@pytest.mark.asyncio
+async def test_get_daily_summary(cost_guard):
+    """일일 요약 정보를 반환한다."""
+    await cost_guard.record_usage("task-1", tokens_input=1000, tokens_output=500)
+    summary = await cost_guard.get_daily_summary()
+    assert "total_cost_usd" in summary
+    assert "daily_limit_usd" in summary
+    assert "remaining_usd" in summary
+    assert summary["record_count"] == 1

tests/test_dispatcher_cost.py

@@ -0,0 +1,80 @@
+import pytest
+import os
+import tempfile
+from unittest.mock import AsyncMock, patch
+
+from agent.task_queue import PersistentTaskQueue
+from agent.cost_guard import CostGuard
+from agent.task_history import TaskHistory
+from agent.dispatcher import Dispatcher
+
+
+@pytest.fixture
+async def resources():
+    paths = []
+    for _ in range(3):
+        fd, p = tempfile.mkstemp(suffix=".db")
+        os.close(fd)
+        paths.append(p)
+
+    queue = PersistentTaskQueue(db_path=paths[0])
+    await queue.initialize()
+    guard = CostGuard(db_path=paths[1], daily_limit=10.0, per_task_limit=3.0)
+    await guard.initialize()
+    history = TaskHistory(db_path=paths[2])
+    await history.initialize()
+
+    yield queue, guard, history
+
+    await queue.close()
+    await guard.close()
+    await history.close()
+    for p in paths:
+        os.unlink(p)
+
+
+@pytest.mark.asyncio
+async def test_dispatcher_records_cost(resources):
+    queue, guard, history = resources
+
+    await queue.enqueue("thread-1", "gitea", {
+        "issue_number": 42, "repo_owner": "quant",
+        "repo_name": "galaxis-po", "message": "Fix",
+    })
+
+    mock_run = AsyncMock(return_value={
+        "status": "completed", "tokens_input": 5000, "tokens_output": 2000,
+    })
+
+    dispatcher = Dispatcher(task_queue=queue, cost_guard=guard, task_history=history)
+    dispatcher._run_agent_for_task = mock_run
+
+    await dispatcher._poll_once()
+
+    daily = await guard.get_daily_cost()
+    assert daily > 0
+
+    records = await history.get_recent()
+    assert len(records) == 1
+    assert records[0]["status"] == "completed"
+
+
+@pytest.mark.asyncio
+async def test_dispatcher_blocks_when_daily_limit_exceeded(resources):
+    queue, guard, history = resources
+
+    for i in range(5):
+        await guard.record_usage(f"prev-{i}", tokens_input=1_000_000, tokens_output=200_000)
+
+    await queue.enqueue("thread-1", "gitea", {"message": "Should be blocked"})
+
+    mock_run = AsyncMock()
+    dispatcher = Dispatcher(task_queue=queue, cost_guard=guard, task_history=history)
+    dispatcher._run_agent_for_task = mock_run
+
+    await dispatcher._poll_once()
+
+    mock_run.assert_not_called()
+
+    pending = await queue.get_pending()
+    assert len(pending) == 1

tests/test_json_logging.py

@@ -0,0 +1,80 @@
+import json
+import logging
+import io
+
+from agent.json_logging import JsonFormatter, setup_logging
+
+
+def test_json_formatter_basic():
+    formatter = JsonFormatter()
+    record = logging.LogRecord(
+        name="test", level=logging.INFO, pathname="test.py",
+        lineno=1, msg="테스트 메시지", args=(), exc_info=None,
+    )
+    output = formatter.format(record)
+    parsed = json.loads(output)
+    assert parsed["message"] == "테스트 메시지"
+    assert parsed["level"] == "INFO"
+    assert "timestamp" in parsed
+
+
+def test_json_formatter_with_extra():
+    formatter = JsonFormatter()
+    record = logging.LogRecord(
+        name="test", level=logging.INFO, pathname="test.py",
+        lineno=1, msg="작업 시작", args=(), exc_info=None,
+    )
+    record.thread_id = "uuid-123"
+    record.issue = 42
+    output = formatter.format(record)
+    parsed = json.loads(output)
+    assert parsed["thread_id"] == "uuid-123"
+    assert parsed["issue"] == 42
+
+
+def test_json_formatter_with_exception():
+    formatter = JsonFormatter()
+    try:
+        raise ValueError("test error")
+    except ValueError:
+        import sys
+        record = logging.LogRecord(
+            name="test", level=logging.ERROR, pathname="test.py",
+            lineno=1, msg="에러 발생", args=(), exc_info=sys.exc_info(),
+        )
+    output = formatter.format(record)
+    parsed = json.loads(output)
+    assert "exception" in parsed
+    assert "ValueError" in parsed["exception"]
+
+
+def test_setup_logging_json():
+    test_logger = logging.getLogger("test_json_setup")
+    test_logger.handlers.clear()
+    test_logger.setLevel(logging.DEBUG)
+    stream = io.StringIO()
+    handler = logging.StreamHandler(stream)
+    test_logger.addHandler(handler)
+    setup_logging(log_format="json", logger=test_logger)
+    test_logger.info("hello")
+    output = stream.getvalue().strip()
+    parsed = json.loads(output)
+    assert parsed["message"] == "hello"
+
+
+def test_setup_logging_text():
+    test_logger = logging.getLogger("test_text_setup")
+    test_logger.handlers.clear()
+    test_logger.setLevel(logging.DEBUG)
+    stream = io.StringIO()
+    handler = logging.StreamHandler(stream)
+    test_logger.addHandler(handler)
+    setup_logging(log_format="text", logger=test_logger)
+    test_logger.info("hello")
+    output = stream.getvalue().strip()
+    assert "hello" in output
+    try:
+        json.loads(output)
+        assert False, "Should not be valid JSON"
+    except json.JSONDecodeError:
+        pass

tests/test_recovery.py

@@ -0,0 +1,79 @@
+import pytest
+import os
+import tempfile
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from agent.task_queue import PersistentTaskQueue
+from agent.recovery import recover_on_startup, ContainerCleaner
+
+
+@pytest.fixture
+async def task_queue():
+    fd, db_path = tempfile.mkstemp(suffix=".db")
+    os.close(fd)
+    queue = PersistentTaskQueue(db_path=db_path)
+    await queue.initialize()
+    yield queue
+    await queue.close()
+    os.unlink(db_path)
+
+
+@pytest.mark.asyncio
+async def test_recover_resets_running_to_pending(task_queue):
+    await task_queue.enqueue("thread-1", "gitea", {"msg": "interrupted"})
+    await task_queue.dequeue()  # → running
+    assert await task_queue.has_running_task("thread-1") is True
+
+    with patch("agent.recovery._cleanup_zombie_containers", new_callable=AsyncMock):
+        await recover_on_startup(task_queue)
+
+    assert await task_queue.has_running_task("thread-1") is False
+    pending = await task_queue.get_pending()
+    assert len(pending) == 1
+
+
+@pytest.mark.asyncio
+async def test_recover_no_running_tasks(task_queue):
+    with patch("agent.recovery._cleanup_zombie_containers", new_callable=AsyncMock):
+        await recover_on_startup(task_queue)
+    pending = await task_queue.get_pending()
+    assert len(pending) == 0
+
+
+@pytest.mark.asyncio
+async def test_container_cleaner_removes_old():
+    mock_container = MagicMock()
+    mock_container.name = "galaxis-sandbox-old"
+    mock_container.labels = {"galaxis-agent-sandbox": "true"}
+    mock_container.attrs = {"Created": "2026-03-19T00:00:00Z"}
+    mock_container.stop = MagicMock()
+    mock_container.remove = MagicMock()
+
+    mock_docker = MagicMock()
+    mock_docker.containers.list.return_value = [mock_container]
+
+    cleaner = ContainerCleaner(docker_client=mock_docker, max_age_seconds=600)
+    removed = await cleaner.cleanup_once()
+
+    assert removed == 1
+    mock_container.stop.assert_called_once()
+    mock_container.remove.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_container_cleaner_keeps_recent():
+    from datetime import datetime, timezone
+    now = datetime.now(timezone.utc).isoformat()
+
+    mock_container = MagicMock()
+    mock_container.labels = {"galaxis-agent-sandbox": "true"}
+    mock_container.attrs = {"Created": now}
+
+    mock_docker = MagicMock()
+    mock_docker.containers.list.return_value = [mock_container]
+
+    cleaner = ContainerCleaner(docker_client=mock_docker, max_age_seconds=3600)
+    removed = await cleaner.cleanup_once()
+
+    assert removed == 0
+    mock_container.stop.assert_not_called()

tests/test_smoke.py

@@ -0,0 +1,44 @@
+import pytest
+from contextlib import asynccontextmanager
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import httpx
+
+
+@asynccontextmanager
+async def mock_lifespan(app):
+    yield
+
+
+@pytest.mark.asyncio
+async def test_smoke_health():
+    from agent.webapp import app
+    app.router.lifespan_context = mock_lifespan
+    async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client:
+        resp = await client.get("/health")
+        assert resp.status_code == 200
+        assert resp.json()["status"] == "ok"
+
+
+@pytest.mark.asyncio
+async def test_smoke_health_costs():
+    from agent.webapp import app
+    app.router.lifespan_context = mock_lifespan
+
+    mock_guard = MagicMock()
+    mock_guard.get_daily_summary = AsyncMock(return_value={
+        "total_cost_usd": 1.5,
+        "daily_limit_usd": 10.0,
+        "remaining_usd": 8.5,
+        "record_count": 3,
+        "total_tokens_input": 50000,
+        "total_tokens_output": 20000,
+    })
+
+    with patch("agent.cost_guard.get_cost_guard", new_callable=AsyncMock, return_value=mock_guard):
+        async with httpx.AsyncClient(transport=httpx.ASGITransport(app=app), base_url="http://test") as client:
+            resp = await client.get("/health/costs")
+            assert resp.status_code == 200
+            data = resp.json()
+            assert data["total_cost_usd"] == 1.5
+            assert data["daily_limit_usd"] == 10.0

tests/test_task_history.py

@@ -0,0 +1,69 @@
+import pytest
+import os
+import tempfile
+
+from agent.task_history import TaskHistory
+
+
+@pytest.fixture
+async def history():
+    fd, db_path = tempfile.mkstemp(suffix=".db")
+    os.close(fd)
+    h = TaskHistory(db_path=db_path)
+    await h.initialize()
+    yield h
+    await h.close()
+    os.unlink(db_path)
+
+
+@pytest.mark.asyncio
+async def test_record_completed(history):
+    await history.record(
+        task_id="task-1", thread_id="thread-1", issue_number=42,
+        repo_name="galaxis-po", source="gitea", status="completed",
+        created_at="2026-03-20T10:00:00Z", completed_at="2026-03-20T10:05:00Z",
+        duration_seconds=300.0, tokens_input=5000, tokens_output=2000, cost_usd=0.045,
+    )
+    records = await history.get_recent(limit=10)
+    assert len(records) == 1
+    assert records[0]["task_id"] == "task-1"
+    assert records[0]["status"] == "completed"
+
+
+@pytest.mark.asyncio
+async def test_record_failed(history):
+    await history.record(
+        task_id="task-2", thread_id="thread-2", issue_number=10,
+        repo_name="galaxis-po", source="discord", status="failed",
+        created_at="2026-03-20T11:00:00Z", completed_at="2026-03-20T11:01:00Z",
+        duration_seconds=60.0, tokens_input=1000, tokens_output=500, cost_usd=0.01,
+        error_message="Agent crashed",
+    )
+    records = await history.get_recent(limit=10)
+    assert len(records) == 1
+    assert records[0]["error_message"] == "Agent crashed"
+
+
+@pytest.mark.asyncio
+async def test_get_recent_ordered(history):
+    await history.record(
+        task_id="task-1", thread_id="t1", issue_number=1, repo_name="r",
+        source="gitea", status="completed", created_at="2026-03-20T10:00:00Z",
+        completed_at="2026-03-20T10:05:00Z", duration_seconds=300,
+        tokens_input=100, tokens_output=50, cost_usd=0.001,
+    )
+    await history.record(
+        task_id="task-2", thread_id="t2", issue_number=2, repo_name="r",
+        source="gitea", status="completed", created_at="2026-03-20T11:00:00Z",
+        completed_at="2026-03-20T11:05:00Z", duration_seconds=300,
+        tokens_input=200, tokens_output=100, cost_usd=0.002,
+    )
+    records = await history.get_recent(limit=10)
+    assert len(records) == 2
+    assert records[0]["task_id"] == "task-2"
+
+
+@pytest.mark.asyncio
+async def test_empty_history(history):
+    records = await history.get_recent(limit=10)
+    assert records == []