Build a GPU Technician Thermal Fault Maintenance Dossier Generator — implementation

## Build
Ingest normalized read-only DCGM Prometheus counters to detect throttled accelerators. Synthesize physical inspection packets and remedial instructions for datacenter technicians using hosted Claude.

## User
Datacenter Hardware Technician

## Workflow
1. Parse normalized DCGM thermal and power exporter metrics from Prometheus time-series logs.
2. Detect thermal throttling event markers where device clock speed drops below standard operating baselines.
3. Call hosted Claude to generate an actionable step-by-step physical inspection checklist for the specific server chassis.
4. Export a technician maintenance packet with rack identifier, slot index, and verified heatsink diagnostic steps.

## Interface
Terminal dispatch console generating Markdown and JSON maintenance work packets for field engineers.

## Data
Simplified Prometheus metric text files containing normalized DCGM counters like DCGM_FI_DEV_GPU_TEMP and DCGM_FI_DEV_CLOCK_THROTTLE_REASONS.
Starter examples below are synthetic. They are not downloaded records.

## Source
[NVIDIA DCGM exporter](https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/README.md)
Extracts standard NVIDIA DCGM exporter metric schemas including temperature and clock throttle indicators to detect degraded hardware.

## Tools
langgraph, anthropic.
LangGraph manages stateful transitions between Prometheus metric ingestion, throttle classification, and Claude dispatch generation.
Hosted a currently available hosted Claude model via official Anthropic Python client; no hardware control commands emitted.

## Deliverables
Submit runnable source, pinned dependencies, setup commands, a working interface, example outputs, and a recorded end-to-end demo. Include tests and an action log.

## Acceptance
- Parses normalized DCGM Prometheus metric formats correctly.
- Identifies active throttle flags from clock reason masks.
- Generates complete technician checklists including chassis slot and heatsink verification procedures.

## Starter Examples
```json
[
  {
    "input": {
      "chassis_id": "rack-04-node-02",
      "gpu_index": 3,
      "gpu_temp_celsius": 89.5,
      "throttle_reason_bitmask": 8,
      "power_draw_w": 280
    },
    "expected_output": {
      "maintenance_required": true,
      "priority": "HIGH",
      "throttle_type": "Thermal",
      "inspection_dossier": "Verify slot 3 airflow shroud seating, inspect thermal interface material, and check fan duct velocity."
    }
  },
  {
    "input": {
      "chassis_id": "rack-01-node-05",
      "gpu_index": 0,
      "gpu_temp_celsius": 62,
      "throttle_reason_bitmask": 0,
      "power_draw_w": 250
    },
    "expected_output": {
      "maintenance_required": false,
      "priority": "NONE",
      "throttle_type": "None",
      "inspection_dossier": "Chassis operates within standard nominal parameters. No maintenance action required."
    }
  }
]
```

## Safety
Read-only metric ingestion; outputs advisory maintenance checklists without issuing hardware IPMI or power cycling commands.
Use mock external writes. Require human review before any real deployment.