meta_hackathon

Support Triage OpenEnv

A complete, real-world OpenEnv environment for training/evaluating agents on customer support ticket triage. The environment simulates what support teams actually do: read inbox tickets, classify urgency/category, draft safe responses, and resolve the right ticket.

Why this environment

Most agent benchmarks under-model production support workflows. This environment focuses on practical support operations with:

• Multi-ticket inbox context selection
• Policy-compliant communication
• Priority + escalation decisions
• Deterministic graders and dense reward shaping

OpenEnv API compliance

The environment exposes:

• reset(task_id?: str) -> Observation
• step(action: Action) -> (Observation, Reward, done, info)
• state() -> dict

Typed Pydantic models:

• Observation: src/support_triage_openenv/models.py
• Action: src/support_triage_openenv/models.py
• Reward: src/support_triage_openenv/models.py

Metadata:

• openenv.yaml

Action space

Action model fields:

• action_type: one of read_ticket | classify_ticket | draft_reply | resolve_ticket
• ticket_id: required for read_ticket, classify_ticket, resolve_ticket
• priority: optional enum low | medium | high | urgent
• category: optional enum account | billing | technical | abuse | general
• needs_escalation: optional bool
• message: text for draft_reply

Observation space

Observation includes:

• task_id, objective, step_count, max_steps
• inbox: ticket metadata list (ticket_id, subject, tier, age, read flag)
• current_ticket_content: only visible after reading selected ticket
• latest_system_note: feedback from last step
• score_hint: partial grader components (read, classify, reply, resolve)

Tasks and difficulty

1. easy_password_reset (Easy)

• Correctly process account lockout and send secure reset guidance.

1. medium_billing_dispute (Medium)

• Investigate duplicate billing with context ticket and provide policy-compliant refund timeline.

1. hard_outage_incident (Hard)

• Handle a high-stakes outage report requiring multi-ticket context, urgent escalation, and careful incident messaging.

Each task has deterministic grading in support_triage_openenv.graders.grade_task, returning a score 0.0-1.0.

Reward design

Reward is shaped and meaningful across the trajectory:

• Positive dense signal from partial grader progress (read/context, classification fields, reply quality, resolve correctness)
• Penalties for invalid actions, repeated loops, and malformed steps
• Final step guarantees score alignment with deterministic grader output

Project structure

• src/support_triage_openenv/env.py - environment implementation
• src/support_triage_openenv/models.py - typed OpenEnv models
• src/support_triage_openenv/tasks.py - task specs (easy/medium/hard)
• src/support_triage_openenv/graders.py - deterministic grader logic
• scripts/run_baseline.py - OpenAI baseline inference runner
• scripts/validate_env.py - tests + optional openenv validate
• app.py - FastAPI app for HF Space runtime
• Dockerfile - containerized deployment

Setup

cd /home/ai24mtech14005/meta_hackathon
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run tests

python -m pytest -q

Run baseline

OpenAI model baseline:

export API_BASE_URL=https://your-openai-compatible-endpoint/v1
export MODEL_NAME=your-model-id
export HF_TOKEN=your-api-key
python inference.py --mode openai --output scores/inference_scores.json

Deterministic heuristic baseline:

python inference.py --mode heuristic --output scores/inference_scores.json

Outputs JSON report to scores/inference_scores.json and structured stdout logs with [START], [STEP], [END].

Run API locally

uvicorn app:app --host 0.0.0.0 --port 7860

Endpoints:

• GET /health
• POST /reset
• POST /step
• GET /state

Docker

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

Hugging Face Space deployment

• Create a Docker Space.
• Push this repository to the Space.
• Keep README.md frontmatter tags including openenv.
• Space serves the API on port 7860.

One-command remote bootstrap

If you want this local repo to automatically create and push to both GitHub + HF:

export GITHUB_USERNAME=your_github_user
export GITHUB_TOKEN=your_github_pat
export HF_USERNAME=your_hf_user
export HF_TOKEN=your_hf_token
bash scripts/bootstrap_remotes.sh support-triage-openenv

Baseline scores (heuristic reproducible)

Generated with:

python inference.py --mode heuristic --output scores/inference_scores.json

• easy_password_reset: grader 1.0, reward 1.0
• medium_billing_dispute: grader 1.0, reward 1.0
• hard_outage_incident: grader 1.0, reward 1.0
• Overall average grader score: 1.0
• Tracked reference artifact: baseline_expected_scores.json

Pre-submission validator

Run full strict validation (all disqualification gates):

python pre_submission_validate.py --space-url https://your-space-name.hf.space

Local-only run while iterating (skips Docker daemon + remote space ping):

python pre_submission_validate.py --skip-docker --skip-space

Run organizer-provided script directly (integrated path):

bash scripts/pre_validation_script.sh https://your-space-name.hf.space .

Notes:

• scripts/sample_inference_script.sh is kept as organizer reference.
• Root inference.py is aligned to the required [START], [STEP], [END] line format.