README.md

---
title: OpenEnv Email Triage Environment
emoji: 📬
colorFrom: blue
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# OpenEnv Email Triage Environment

A real-world AI agent training environment that simulates professional email triage.
Built to the OpenEnv specification for standardized agent evaluation and benchmarking.

- **Status:** In Development
- **Domain:** Email Triage
- **Deployment:** Hugging Face Spaces (Docker)

---

## Table of Contents

- [What Is This?](#what-is-this)
- [Who Is This For?](#who-is-this-for)
- [Observation Space](#observation-space)
- [Action Space](#action-space)
- [Tasks](#tasks)
- [Reward Function](#reward-function)
- [Quick Start](#quick-start)
- [Running Inference](#running-inference)
- [Inference Architecture](#inference-architecture)
- [Score Table](#score-table)
- [Docker Deployment](#docker-deployment)
- [Hugging Face Space](#hugging-face-space)
- [Pre-Submission Validation](#pre-submission-validation)
- [API Reference](#api-reference)
- [Project Structure](#project-structure)
- [Known Limitations](#known-limitations)
- [Contributing](#contributing)
- [License](#license)

---

## What Is This?

This environment simulates a professional email inbox where an AI agent must:

1. Read incoming emails with realistic metadata (sender, subject, body, thread history).
2. Classify each email with the correct priority label.
3. Route each email to the appropriate department or person.
4. Summarize the email's key information.

Think of it as OpenAI Gym for office work. Instead of balancing a pole, the agent triages an
inbox. The environment provides structured observations, accepts structured actions, and
returns graded rewards with partial credit.

Every decision is scored by a deterministic programmatic grader: no LLM-as-judge,
no randomness, fully reproducible.

---

## Who Is This For?

| Audience | Use Case |
|---|---|
| AI Safety Researchers | Measure agent behavior on realistic tasks with known ground truth |
| LLM Agent Developers | Benchmark models and prompting strategies on real-world work |
| RL Researchers | Train agents with shaped rewards in a professional task environment |
| Companies | Evaluate LLM agents before deploying them to handle real email |

---

## Observation Space

What the agent sees at each step:

| Field | Type | Description |
|---|---|---|
| `email_id` | `str` | Unique identifier for this email |
| `subject` | `str` | Email subject line |
| `body` | `str` | Full email body text |
| `sender` | `str` | Sender's email address |
| `timestamp` | `str` | ISO 8601 timestamp of when the email was received |
| `thread_history` | `list[str]` | Previous messages in the email thread (may be empty) |
| `task_id` | `str` | Which task is currently active |
| `step_number` | `int` | Current step in the episode (0-indexed) |
| `total_emails` | `int` | Total number of emails to process in this task |

The observation never contains the correct answer. The agent must reason from email content.

---

## Action Space

What the agent must output at each step:

| Field | Type | Allowed Values | Description |
|---|---|---|---|
| `label` | `Literal` | `"urgent"`, `"normal"`, `"spam"`, `"archive"` | Priority classification |
| `summary` | `str` | Free text | Brief summary of the email's content and intent |
| `route_to` | `str` | Free text (`"billing"`, `"safety"`, `"engineering"`) | Department or person |

### Example action JSON

```json
{
  "label": "urgent",
  "summary": "Customer reports a safety issue with product overheating",
  "route_to": "safety"
}
```

---

## Tasks

Each task now contains multiple deterministic scenario variants. By default, `/reset`
cycles through the public scenario pool for the selected task.

Private evaluation split selection is controlled server-side via environment
configuration (`OPENENV_EVAL_SPLIT`), and client-side override can be disabled
to preserve benchmark integrity.

To keep private evaluation data out of source control, supply hidden scenarios at
runtime using `OPENENV_PRIVATE_SCENARIOS_JSON` (JSON object keyed by task id).

Example deployment configuration:

```bash
export OPENENV_EVAL_SPLIT="private_eval"
export OPENENV_ALLOW_CLIENT_EVAL_OVERRIDE="false"
export OPENENV_PRIVATE_SCENARIOS_JSON='{"task_easy":[{"scenario_id":"easy-private-001","emails":[{"email_id":"easy-p-001","subject":"Private billing exception","body":"Please correct invoice mismatch for contract addendum B-7 before end of day.","sender":"contracts@partner.example","timestamp":"2026-04-03T09:00:00Z","thread_history":["Customer requested corrected invoice reference."]}],"ground_truth":[{"label":"normal","route_to":"billing","priority_weight":1.0,"summary_keywords":["invoice mismatch","contract addendum","correct"]}]}],"task_medium":[],"task_hard":[]}'
```

Notes:

- Keep this value in deployment secrets or runtime environment config.
- Use valid JSON with double quotes only.
- You can provide multiple scenarios per task by adding more objects to each task list.

### Task 1 — Easy (`task_easy`)

Objective: Correctly classify a single unambiguous email.

Scoring:

- Correct label: 1.0
- Wrong label but correct routing: 0.3
- Everything wrong: 0.0

### Task 2 — Medium (`task_medium`)

Objective: Triage a queue of 5 emails with mixed priority signals.

Scoring:

- Each email scored individually
- Score = (correct labels / total emails) * priority weight factor
- Higher-priority misclassifications are penalized more heavily
- Final score = weighted mean of all individual scores

### Task 3 — Hard (`task_hard`)

Objective: Handle a complex complaint that crosses multiple categories.

Scoring:

- Escalated to safety: 0.4 weight
- Correct routing: 0.3 weight
- Marked as urgent: 0.3 weight
- Penalty: -0.2 if marked as spam
- Final score = weighted sum of sub-scores (clipped to 0.0 minimum)

### Task 4 — Production (`task_production`)

Objective: Simulate a production inbox with mixed operational load across safety,
engineering, billing, support, spam, and low-priority traffic.

Scoring:

- Per-email weighted scoring by business priority
- Route-noise penalty when actions route to too many teams
- Summary quality based on contextual evidence keywords and anti-stuffing rules
- Deterministic escalation follow-ups are inserted when critical triage is missed
- Runtime controls available via `/reset` payload for production simulations:
  - `production_profile`: `light` | `standard` | `heavy`
  - `business_hours_mode`: `true` | `false`
  - `escalation_mode`: `low` | `normal` | `high`

---

## Reward Function

The reward function provides dense training signal at every step, not just binary pass/fail.

### Formula

```text
final_reward = base_score - (step_count * 0.01) + trajectory_bonus - penalties
```

### Components

| Component | Value | Condition |
|---|---|---|
| Base score | 0.0-1.0 | Raw grader score for the current step |
| Step penalty | -0.01 per step | Encourages efficiency |
| Trajectory bonus | +0.2 | If all tasks completed with mean score > 0.8 |
| Destructive action penalty | -0.5 | Agent archives or deletes without reading |
| Loop detection penalty | -0.3 | Same action repeated 3+ times consecutively |

The final reward is clipped to [-1.0, 1.0] before being returned.

---

## Quick Start

### Prerequisites

- Python 3.11+
- API endpoint, model name, and token for inference

### Installation

```bash
pip install -r requirements.txt
export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="gpt-4o"
export HF_TOKEN="your-token-here"
```

### Run the environment locally

```bash
python server.py

curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task_id": "task_easy"}'

curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"label": "urgent", "summary": "Test", "route_to": "billing"}'

curl -X POST http://localhost:7860/state
```

---

## Running Inference

```bash
python inference.py --task all
python inference.py --task 1
python inference.py --task 4 --production-profile heavy --business-hours-mode --escalation-mode high
```

The script reads API settings from environment variables and uses fallback actions when
model output is unparseable, so episodes still complete.

---

## Inference Architecture

The inference script (inference.py) follows this loop:

```text
1. Initialize OpenAI client + environment
2. reset() to get first observation
3. Loop until done or MAX_STEPS:
  - Build prompt from observation + history
  - Call LLM with OpenAI client (catch request errors)
  - Parse response into action (fallback on parse failure)
  - env.step(action)
  - Record reward and history
4. Print score table
```

### Environment Variables Required

```bash
export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="gpt-4o"
export HF_TOKEN="your-token-here"
export INFERENCE_RUNTIME_BUDGET_SECONDS="1140"
export INFERENCE_REQUEST_TIMEOUT_SECONDS="12"
```

Runtime controls:

- `INFERENCE_RUNTIME_BUDGET_SECONDS` limits full-script wall-clock runtime (default 1140s, under 20 minutes).
- `INFERENCE_REQUEST_TIMEOUT_SECONDS` limits each LLM request timeout (default 12s).
- Equivalent CLI flags: `--runtime-budget-seconds` and `--request-timeout-seconds`.

Fallback behavior when parsing fails:

```json
{"label": "normal", "summary": "Unable to parse response", "route_to": "general"}
```

---

## Score Table

Placeholder until inference is run.

| Model | Task 1 (Easy) | Task 2 (Medium) | Task 3 (Hard) | Mean |
|---|---|---|---|---|
| MODEL_NAME | TBD | TBD | TBD | TBD |

Expected rough ranges:

- GPT-4o: 0.8-1.0 on easy, 0.5-0.8 on medium, 0.4-0.7 on hard

---

## Docker Deployment

```bash
docker build -t email-triage-env .
docker run -p 7860:7860 email-triage-env

curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task_id": "task_easy"}'
```

For Apple Silicon:

```bash
docker build --platform linux/amd64 -t email-triage-env .
```

---

## Hugging Face Space

Live URL placeholder:

`https://huggingface.co/spaces/YOUR_USERNAME/email-triage-env`

The Space homepage (`/`) now serves a lightweight interactive triage console for
manual testing. Machine-readable service metadata is available at `GET /meta`.

Example interaction:

```bash
export SPACE_URL="https://YOUR_USERNAME-email-triage-env.hf.space"

curl -X POST "$SPACE_URL/reset" \
  -H "Content-Type: application/json" \
  -d '{"task_id": "task_easy"}'
```

---

## Pre-Submission Validation

Run the validator before submitting your environment.

```bash
chmod +x validate-submission.sh
./validate-submission.sh https://YOUR_USERNAME-email-triage-env.hf.space .
```

The script checks:

- HF Space `/reset` health (HTTP 200 expected)
- Docker build success
- `openenv validate` pass status

---

## API Reference

### POST /reset

Request:

```json
{"task_id": "task_easy"}
```

Response:

```json
{
  "observation": {
    "email_id": "easy-001",
    "subject": "Quarterly invoice available",
    "body": "...",
    "sender": "accounts@vendor-example.com",
    "timestamp": "2026-03-25T09:15:00Z",
    "thread_history": ["..."],
    "task_id": "task_easy",
    "step_number": 0,
    "total_emails": 1
  },
  "info": {"task_id": "task_easy", "step": 0}
}
```

### POST /step

Request:

```json
{
  "label": "urgent",
  "summary": "Customer needs immediate help",
  "route_to": "support"
}
```

Response:

```json
{
  "observation": {},
  "reward": 0.85,
  "done": false,
  "info": {"step": 1, "task_id": "task_easy"}
}
```

### POST /state

No request body required.

Response: `EnvironmentState` JSON object.

---

## Project Structure

```text
.
├── models.py
├── tasks.py
├── graders.py
├── environment.py
├── server.py
├── server/
│   └── app.py
├── inference.py
├── openenv.yaml
├── Dockerfile
├── requirements.txt
├── pyproject.toml
├── uv.lock
├── validate-submission.sh
├── README.md
└── RULES.md
```

---

## Known Limitations

| Limitation | Impact |
|---|---|
| Static scenario pools | No live inbox ingestion from production systems |
| Single-agent server instance | Concurrent agents can conflict |
| No live thread simulation | Thread history is static |
| English-only content | No multilingual coverage |
| No attachments | Text-only triage |
| Simplified routing | No org chart or availability modeling |
| Limited temporal dynamics | Production task can generate deterministic escalations, but not full live message streams |
| Rule-based grading edges | Equivalent decisions may score differently from humans |

What an agent cannot exploit:

- The correct answer is never present in observations
- The grader is a pure function and cannot be manipulated
- Step penalty cannot be bypassed except by efficient actions

---

## Summary of Revision 2 Changes

| What Changed | Before | After | Why |
|---|---|---|---|
| Return type of step() | tuple | StepResult object | Match sample result.observation pattern |
| Return type of reset() | EmailObservation | ResetResult object | Match sample result.observation pattern |
| New models | 4 models | 6 models (+StepResult, +ResetResult) | Match sample interface |
| API key reading | OPENAI_API_KEY style | HF_TOKEN or API_KEY via os.getenv | Match sample fallback pattern |
| Temperature guidance | 0 | 0.2 | Match sample behavior |
| Response parsing | JSON-only assumption | Text parsing with fallback action | Robustness to non-JSON model output |
| History tracking | Optional | Mandatory | Match sample architecture |
| Step cap | Not explicit | MAX_STEPS constant | Runtime safety and reproducibility |

---

## Contributing

Read `RULES.md` before contributing.

Key constraints:

- Type hints and Pydantic models required
- No extra dependencies without explicit approval
- No features beyond project brief
- Graders must remain deterministic pure functions

---

## License

MIT License.