Mourya234/openenv-ticket-triage

OpenEnv: Support Ticket Triage Environment

Submission links

• —GitHub repository: https://github.com/Mourya611/openEV_round1
• —Hugging Face Space: https://hf.co/spaces/Mourya234/openenv-ticket-triage
• —Hugging Face app URL: https://mourya234-openenv-ticket-triage.hf.space

This project implements a complete real-world OpenEnv-style environment where an agent learns to triage customer support tickets.

Why this environment

Support triage is a real operational workflow in SaaS teams. Agents must balance urgency, policy, customer context, and technical uncertainty. This environment captures that with deterministic scoring and progressive task difficulty.

Problem modeled

• —Domain: customer support operations
• —Goal: choose the right triage action for each ticket
• —Realism:
• —Different issue types (billing, security, integration, compliance, bug)
• —Tier-aware urgency context
• —Policy-sensitive escalations for security/compliance scenarios

OpenEnv interface

The environment exposes:

• —POST /reset -> starts a fresh episode (optionally by task name)
• —POST /step -> applies one agent action
• —GET /state -> returns full internal state
• —GET /tasks -> lists task metadata
• —GET /grade -> returns the current task score and grader breakdown

The environment core class is SupportTicketTriageEnv in envs/environment.py, with:

• —reset(task_name: Optional[str])
• —step(action: TriageAction)
• —state()
• —grade(task_name: Optional[str])

Typed models are implemented via Pydantic in envs/models.py.

Action space

TriageAction fields:

• —ticket_id: str
• —decision: resolve | escalate | request_info | defer
• —priority: low | medium | high | urgent
• —response_template: short | empathetic | technical | compliance
• —notes: str (required rationale text)

Observation space

ObservationModel includes:

• —task_name
• —objective
• —current_ticket (or null)
• —queue_remaining
• —processed_count
• —progress (reported inside (0, 1) to avoid boundary-value validator issues)
• —last_feedback
• —allowed decisions/priorities/templates

Reward design

Per-step reward is reported strictly inside (0, 1) and combines:

• —Local action quality from deterministic rubric matching:
• —decision correctness
• —priority correctness
• —response template correctness
• —keyword coverage in notes
• —Global trajectory progress bonus

Formula:

reward = 0.75 * action_quality + 0.25 * projected_progress, clamped into (0,1).

This provides dense, partial-progress signals and discourages random behavior.

Tasks and graders

Three deterministic tasks are included:

1. 1.ticket-triage-easy (2 tickets)
2. 2.ticket-triage-medium (3 tickets)
3. 3.ticket-triage-hard (4 tickets)

Task definitions: envs/tasks.py Graders: envs/graders.py Episode final score: normalized deterministic grade reported strictly inside (0, 1).

Inference baseline (required)

The required root script inference.py:

• —Uses OpenAI client for all LLM calls
• —Falls back to a deterministic built-in policy if the OpenAI client or token is unavailable
• —Reads the validator-required env vars API_BASE_URL, MODEL_NAME, and HF_TOKEN
• —Emits structured stdout logs:
• —[START]
• —[STEP]
• —[END]
• —Uses validator-compatible [END] success=... steps=... score=... rewards=... output
• —Formats scores and rewards to 2 decimal places and emits only validator-compatible line types

Environment variables

Create .env from .env.example:

API_BASE_URL=https://api.openai.com/v1
MODEL_NAME=gpt-5-mini
HF_TOKEN=your_hf_token
ENV_BASE_URL=http://localhost:7860

API_BASE_URL and MODEL_NAME include defaults in inference.py. If HF_TOKEN is not set, the script falls back to the deterministic local policy so the validator can still execute the run.

Local run

pip install -r requirements.txt
uvicorn app:app --host 0.0.0.0 --port 7860

In another terminal:

python inference.py

Docker

docker build -t openenv-ticket-triage .
docker run --rm -p 7860:7860 openenv-ticket-triage

Quick precheck

After starting the app:

powershell -ExecutionPolicy Bypass -File .\scripts\precheck.ps1

Hugging Face Spaces deployment

1. 1.Push this repository to GitHub.
2. 2.Create a Docker Space on Hugging Face.
3. 3.Connect the repo.
4. 4.Add secrets in Space settings:
5. 5.API_BASE_URL (optional if you want to override the default)
6. 6.MODEL_NAME
7. 7.HF_TOKEN
8. 8.ENV_BASE_URL (if your Space serves the env on a non-default URL)
9. 9.Ensure Space responds on /health, /reset, /step, /state, /tasks, and /grade.

Project structure

.
├── app.py
├── inference.py
├── openenv.yaml
├── Dockerfile
├── requirements.txt
├── .env.example
├── envs
│   ├── __init__.py
│   ├── models.py
│   ├── tasks.py
│   ├── graders.py
│   └── environment.py
└── scripts
    └── precheck.ps1