TASKS.md

TASKS.md — Environment Task Definition

Senior OpenEnv Engineer Rule: Real-world task. No toys. No games.

---

🎯 Chosen Environment: Code Review Environment

Name: code_review_env
Domain: Software Engineering / Developer Tooling
Task: The agent acts as a code reviewer. Given a code diff or snippet, the agent must produce a structured, high-quality review identifying bugs, style issues, and improvement suggestions.

---

Why This Is Real-World ✅

• Code review is a high-value, daily engineering task
• Clear, measurable correctness signals (bug found / not found, severity match)
• Rich feedback loop: agent learns what good reviews look like
• Direct production utility — can be deployed in CI/CD pipelines

---

Episode Structure

reset()
  │
  └── Agent receives: code snippet + task context
        (e.g., language, PR description, critical path flag)

step(action)
  │
  └── Agent sends: structured review
        (issues: List[Issue], summary: str, severity: Severity)

  └── Environment returns:
        reward (float), feedback (str), done (bool)

---

Action Space

@dataclass
class CodeReviewAction(Action):
    issues: List[str]          # List of identified issues
    summary: str               # Overall review summary
    severity: str              # "low" | "medium" | "high" | "critical"
    metadata: Dict[str, Any]   # Optional extra context

---

Observation Space

@dataclass
class CodeReviewObservation(Observation):
    done: bool
    reward: float
    code_snippet: str          # Code to review (current step)
    language: str              # e.g., "python", "javascript"
    context: str               # PR description or task context
    ground_truth_issues: List[str]   # Hidden during training rollout
    feedback: str              # Human-readable feedback on last action
    step_number: int

---

State

@dataclass
class CodeReviewState(State):
    episode_id: Optional[str]
    step_count: int
    total_snippets: int        # How many snippets in this episode
    cumulative_reward: float
    language: str

---

Episode Flow

Step	Agent Receives	Agent Sends	Env Returns
1	Code snippet #1 + context	Structured review	Reward + feedback
2	Code snippet #2 (harder)	Structured review	Reward + feedback
…	…	…	…
N	Final snippet	Final review	Terminal reward, done=True

---

Data Sources

• CodeSearchNet — multi-language code samples
• Synthetic bug injection (off-by-one, null dereference, SQL injection, etc.)
• Human-curated review gold standards (severity labels)

---

Difficulty Levels

Level	Description
Easy	Obvious syntax error or unused variable
Medium	Logic bug, missing edge case handling
Hard	Security vulnerability, concurrency issue
Critical	Data corruption / memory leak pattern