File size: 5,394 Bytes
e4accbb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
# Support Queue OpenEnv

A real-world OpenEnv benchmark for **SaaS support triage**.

Agents must read incoming support tickets, assign the right priority, route the case to the correct internal queue, choose the next action, and draft a safe first reply. The benchmark is designed to feel like an actual support operations workflow rather than a toy task.

## Why This Environment

Real support teams repeatedly solve the same high-value triage problems:

- decide how urgent a ticket is
- route it to the right team
- avoid unsafe or misleading replies
- handle ambiguous requests without over-escalating

This makes support triage a strong RL and agent-evaluation environment because success is measurable, partial credit is meaningful, and mistakes are easy to interpret.

## What The Agent Does

For each ticket, the agent must produce a `SupportQueueAction` with:

- `priority`: `P1 | P2 | P3 | P4`
- `queue`: `billing | security | technical | success | trust_safety`
- `disposition`: `respond | request_info | escalate | close`
- `summary`: short internal triage note
- `response`: first customer-facing reply
- `confidence`: float in `[0.0, 1.0]`

## Observation Space

Each `reset()` and `step()` returns a typed `SupportQueueObservation` containing:

| Field | Meaning |
| --- | --- |
| `task_id`, `task_title`, `difficulty` | Active benchmark task metadata |
| `instructions` | Task-specific operating guidance |
| `current_index`, `total_tickets` | Episode progress |
| `ticket` | Current customer ticket payload |
| `allowed_priorities`, `allowed_queues`, `allowed_dispositions` | Valid discrete actions |
| `scoring_weights` | Reward decomposition |
| `last_feedback` | Previous grader output |
| `reward`, `cumulative_reward`, `done` | Episode feedback |
| `info` | Extra metadata such as `episode_id` |

The ticket payload includes:

- `ticket_id`
- `subject`
- `body`
- `customer_tier`
- `product_area`
- `sla_hours`
- `recent_events`

## State Space

`state()` returns a typed `SupportQueueState` with:

- active task card
- current cursor
- cumulative and average reward
- processed ticket ids
- full action history
- full per-ticket grading history

## Tasks

The benchmark includes three deterministic tasks with increasing difficulty.

| Task ID | Difficulty | Tickets | Description |
| --- | --- | ---: | --- |
| `easy_inbox_cleanup` | Easy | 2 | Straightforward access and billing tickets |
| `medium_sla_defense` | Medium | 3 | Mix of phishing escalation, webhook failure, and billing ambiguity |
| `hard_exec_escalations` | Hard | 4 | Executive-pressure tickets spanning production, security, commercial, and retention workflows |

## Reward Design

Each processed ticket gets a reward in `[0.0, 1.0]`.

Reward components:

| Component | Weight |
| --- | ---: |
| Priority accuracy | `0.30` |
| Queue accuracy | `0.25` |
| Disposition accuracy | `0.20` |
| Summary keyword coverage | `0.15` |
| Response keyword coverage | `0.10` |
| Unsafe reply penalty | `-0.10` |

This gives useful partial progress signals. An agent can still earn reward for a good route or good reply even if one part of the triage decision is wrong.

## API Surface

The environment server exposes:

- `POST /reset`
- `POST /step`
- `GET /state`
- `GET /tasks`
- `GET /health`
- `GET /`

Example reset payload:

```json
{
  "task_id": "easy_inbox_cleanup"
}
```

## Project Structure

```text
support_queue_env/
  client.py
  grading.py
  models.py
  tasks.py
  server/
    app.py
    openenv_compat.py
    support_queue_environment.py
Dockerfile
openenv.yaml
inference.py
```

## Running Locally

### Python

```bash
pip install -r requirements.txt
uvicorn support_queue_env.server.app:app --host 0.0.0.0 --port 8000
```

### Docker

```bash
docker build -t support-queue-openenv .
docker run --rm -p 8000:8000 support-queue-openenv
```

## Baseline Inference

The required inference script is [inference.py](./inference.py).

It:

- uses the OpenAI Python client
- reads `API_BASE_URL`, `MODEL_NAME`, `HF_TOKEN`, and optional `LOCAL_IMAGE_NAME`
- emits structured `[START]`, `[STEP]`, and `[END]` logs
- writes `inference_results.json`

Set environment variables:

```bash
API_BASE_URL=https://api.openai.com/v1
MODEL_NAME=gpt-4o-mini
HF_TOKEN=your_token
LOCAL_IMAGE_NAME=
```

Then run:

```bash
python inference.py
```

## Baseline Scores

Expected deterministic baseline scores from the bundled heuristic policy:

| Task | Score |
| --- | ---: |
| `easy_inbox_cleanup` | `1.00` |
| `medium_sla_defense` | `0.98` |
| `hard_exec_escalations` | `0.97` |
| Average | `0.98` |

## Hugging Face Space

This repository is configured for a **Docker Space**.

- front matter in `README.md` sets `sdk: docker`
- app serves on port `8000`
- `GET /health` and `POST /reset` support deployment checks

## OpenEnv Files

Core submission files:

- [openenv.yaml](./openenv.yaml)
- [inference.py](./inference.py)
- [Dockerfile](./Dockerfile)
- [support_queue_env/models.py](./support_queue_env/models.py)
- [support_queue_env/server/support_queue_environment.py](./support_queue_env/server/support_queue_environment.py)

## Submission Checklist

- typed action, observation, and state models included
- `reset()`, `step()`, and `state()` implemented
- three graded tasks included
- reward bounded to `[0.0, 1.0]`
- Dockerfile included
- Hugging Face Docker Space compatible
- root `inference.py` included