vllm-neuron-parity

{
  "name": "vllm-neuron-parity",
  "version": "1.6.0",
  "description": "Run evidence-backed campaigns that close approved vLLM-Neuron parity gaps against upstream GPU vLLM.",
  "author": {
    "name": "jinhuang12"
  },
  "keywords": [
    "vllm",
    "neuron",
    "trainium",
    "parity",
    "pave"
  ],
  "skills": "./skills/",
  "interface": {
    "displayName": "vLLM-Neuron Parity",
    "shortDescription": "Run evidence-backed vLLM-Neuron parity campaigns.",
    "longDescription": "Scan upstream gaps, approve closing routes, validate Neuron changes against a GPU baseline, and prepare evidence-backed fork pull requests.",
    "developerName": "jinhuang12",
    "category": "Developer Tools",
    "capabilities": [
      "Multi-agent orchestration",
      "Repository write",
      "Hooks"
    ],
    "defaultPrompt": [
      "Start a vLLM-Neuron parity run for these targets.",
      "Resume the active vLLM-Neuron parity run."
    ]
  }
}

{
  "name": "vllm-neuron-parity",
  "version": "1.6.0",
  "description": "Bring the vLLM-Neuron platform plugin fork to parity with upstream GPU vLLM: scan the upstream delta, cost each target's closing route, rank the backlog, and execute user-gated campaigns through correctness and performance gates against a GPU baseline into evidence-backed fork PRs. Ships the vllm-neuron-parity lead workflow skill plus its six dispatched role agent types.",
  "author": {
    "name": "jinhuang12"
  },
  "compatibility": "Requires Claude Code plugin hooks (PreToolUse, PostToolUse, Stop, SessionStart, SubagentStart) registered in hooks/hooks.json, the registered vllm-neuron-parity:* agent types, bash, and Python 3. The three blocking prohibition guards (P1 protected branches, P2 Neuron compile cache, P3 venv/opt), the graph edit guard (no direct edit of the live graph or revisions.yaml outside an apply step), the lead hook pair (stop guard with its audit-checkpoint branch, state-staleness reminder), the write log, the lead-only deny read from the graph's write_limits block and the hook-owned checkpoint files, the no-retry-copy rule, the re-entry dispatch advisory, the write-for-the-reader reminder, and the goal restatement at resume, compaction, and seat start all ride that hook runtime; without it the guards degrade to contract text and the lead's own resume, checkpoint, and write-for-the-reader duties, which is a recorded degradation, not a silent one. Graph validation (scripts/validate_pave.py) and the revision log tool (scripts/record_revision.py) additionally require pyyaml and jsonschema and fail closed without them; a successor revision also needs the pave-init plugin (2.6.0 or later) installed, because its pave-evolve seats draft and review every revision and its workflow-updater runs the audit checkpoint; run-state validation (scripts/validate_run_state.py) is full only with jsonschema and otherwise falls back to a labeled stdlib check that enforces the schema's declared keywords, including field length caps, and fails loudly on any keyword it cannot enforce."
}