{
  "id": "page:docs-development-02-atlas-graph-and-agent-catalog",
  "_kind": "Page",
  "_file": "wiki/docs/development/02-atlas-graph-and-agent-catalog.md",
  "_cluster": "wiki",
  "attributes": {
    "nodeKind": "Page",
    "sourcePath": "docs/development/02-atlas-graph-and-agent-catalog.md",
    "sourceKind": "repo-docs",
    "title": "Atlas Graph & Agent Catalog",
    "displayName": "Atlas Graph & Agent Catalog",
    "slug": "docs/development/02-atlas-graph-and-agent-catalog",
    "articlePath": "wiki/docs/development/02-atlas-graph-and-agent-catalog.md",
    "article": "\n# Atlas Graph & Agent Catalog\n\nThe knowledge layer that drives all runtime behavior.\n\n## Atlas Graph\n\nThe atlas is a YAML-defined knowledge graph at `packages/atlas/graph/`. It models:\n\n### Node Kinds\n\n| Kind | Example | What it represents |\n|------|---------|-------------------|\n| `AgentVersion` | `agentVersion:claude:ge-1-0-0` | A harness at a version range |\n| `AgentProduct` | `agent:claude-code` | The product identity |\n| `Capability` | `capability:supports-mcp` | An abstract capability |\n| `CapabilitySupport` | `capSupport:claude:mcp` | Claim that an agent has a capability |\n| `ModelVersion` | `modelVersion:gpt-5.5` | A model with pricing/limits |\n| `ModelProviderVersion` | `providerVersion:foundry` | A provider endpoint |\n| `TransportProtocol` | `transportProtocol:anthropic` | An API wire format |\n| `HookDescriptor` | `hook:sessionStart` | A lifecycle hook |\n| `PluginTarget` | `pluginTarget:claude-code` | A plugin installation target |\n| `LaunchConfig` | `launchConfig:claude.yolo` | A launch recipe |\n| `Evidence` | `evidence:claude-mcp-docs` | A capability evidence source |\n\n### Edge Kinds\n\n| Edge | From → To | Meaning |\n|------|-----------|---------|\n| `supports` | AgentVersion → Capability | Agent has this capability |\n| `version_of` | AgentVersion → AgentProduct | Version belongs to product |\n| `composed_of` | AgentVersion → CoreImpl, RuntimeImpl, etc. | Architecture layers |\n| `speaks` | CoreImpl → TransportProtocol | Agent uses this wire format |\n| `defaults_to_model` | AgentVersion → ModelVersion | Default model selection |\n\n### `adapterMetadata` Block\n\nEach `AgentVersion` YAML includes an `adapterMetadata` section with runtime data:\n\n```yaml\nadapterMetadata:\n  authMethods:\n    - type: api_key\n      name: \"API Key\"\n      envVars: [ANTHROPIC_API_KEY]\n  authFiles: [\"~/.claude.json\"]\n  hostEnvSignals: [CLAUDECODE, CLAUDE_CODE_SESSION_ID]\n  sessionDir: \"~/.claude/projects\"\n  sessionPersistence: file\n  automationEnv:\n    GEMINI_CLI_TRUST_WORKSPACE: \"true\"\n  approvalModes: [yolo, prompt, deny]\n  capabilityFlags:\n    canResume: true\n    supportsMultiTurn: true\n    supportsMCP: true\n    supportsThinking: true\n    # ... 30+ flags\n  runtimeHooks:\n    preToolUse: blocking\n    stop: nonblocking\n  configSchema:\n    configFormat: json\n    configFilePaths: [\"~/.claude/settings.json\"]\n  displayName: \"Claude Code\"\n  defaultModelId: \"claude-sonnet-4-20250514\"\n```\n\n## Atlas Catalog\n\n`packages/atlas/src/catalog` provides a typed query API over the atlas graph\nthrough the `@a5c-ai/atlas/catalog` export:\n\n### Core Queries\n\n```typescript\nimport {\n  getAgentVersion,\n  getCapabilityFlags,\n  getAdapterMetadata,\n  getInstallMethods,\n  getAutomationEnv,\n  getHostEnvSignals,\n  getSessionConfig,\n  getRuntimeHooks,\n  getConfigSchema,\n  getDisplayName,\n  getDefaultModelId,\n  getBridgeCapabilities,\n  getHookSupport,\n  getYoloLaunchArgs,\n  getTransportCodecCapabilities,\n} from '@a5c-ai/atlas/catalog';\n```\n\n### How Adapters Use It\n\nAdapters are thin wrappers — all data comes from the graph:\n\n```typescript\nclass ClaudeAdapter extends BaseAgentAdapter {\n  get displayName() { return getDisplayName(this.agent); }\n  get hostEnvSignals() { return getHostEnvSignals(this.agent); }\n  get capabilities() {\n    const flags = getCapabilityFlags(this.agent);\n    const hooks = getRuntimeHooks(this.agent);\n    return {\n      agent: this.agent,\n      canResume: Boolean(flags.canResume),\n      supportsMCP: Boolean(flags.supportsMCP),\n      runtimeHooks: {\n        preToolUse: hooks.preToolUse ?? 'nonblocking',\n        // ...\n      },\n      installMethods: getInstallMethods(this.agent).map(m => ({\n        platform: 'all', type: m.type, command: m.command,\n      })),\n    };\n  }\n}\n```\n\n### Evidence System\n\nThe graph tracks evidence quality for each capability claim:\n\n| Level | Meaning |\n|-------|---------|\n| `corroborated` | Multiple independent sources confirm |\n| `partial` | Some evidence but gaps remain |\n| `inferred` | Derived from related evidence |\n| `unresolved-gap` | No evidence found yet |\n\nEvidence is linked via `sourced_from:` edges from capability support nodes to evidence records.\n\n### Graph Build Pipeline\n\n```\npackages/atlas/graph/**/*.yaml\n    → atlas indexer (scripts/build-index.mjs)\n    → packages/atlas/src/index.json (37MB)\n    → imported by agent-catalog at build time\n    → query functions available at runtime\n```\n",
    "documents": []
  },
  "outgoingEdges": [],
  "incomingEdges": [
    {
      "from": "page:docs-development",
      "to": "page:docs-development-02-atlas-graph-and-agent-catalog",
      "kind": "contains_page"
    }
  ]
}