feat(knowledge): Hub agent packaging — gaia-agent.yaml, CLI entry point, MCP server

[kovtcharov-amd]

Phase 1 — Foundation

Scaffold the Knowledge Agent as a Hub agent with dual-layer Tavily integration (connector for auth + Python SDK for runtime).

Agent class

```python
class KnowledgeAgent(Agent):
name = "knowledge"
description = "Build local knowledge bases from web + local docs, query with local LLM"
default_model = "Qwen3.5-35B-A3B-GGUF"

REQUIRED_CONNECTORS = [
    ConnectorRequirement(
        connector_id="tavily",
        scopes=[],
        reason="Crawl websites and search the web for your knowledge base.",
    ),
]

```

Package structure

```
hub/agents/python/knowledge/
├── gaia-agent.yaml # connectors: [{id: tavily}]
├── pyproject.toml # amd-gaia + tavily-python
├── gaia_agent_knowledge/
│ ├── agent.py # KnowledgeAgent with REQUIRED_CONNECTORS
│ ├── cli.py # gaia-knowledge index|ask|search|...
│ ├── tavily_client.py # Reads key from keyring, wraps Python SDK
│ ├── crawler.py # TavilyClient.crawl → markdown files
│ ├── indexer.py # Markdown → RAG index
│ ├── knowledge_base.py # KB lifecycle management
│ ├── search.py # TavilyClient.search + cache
│ ├── cache.py # SQLite result cache
│ ├── credits.py # Credit tracking
│ ├── config.py # Preferences
│ ├── planner.py # Agent: research planning
│ ├── synthesizer.py # Agent: cross-source synthesis
│ └── tools/ # @tool wrappers for agent's MCP server
├── tests/
└── README.md
```

Deliverables

• `gaia-agent.yaml` manifest with `connectors: [{id: tavily}]` and `tavily-python` dep
• `KnowledgeAgent(Agent)` with `REQUIRED_CONNECTORS`
• CLI with subcommands: index, ask, search, extract, list, credits, config
• `--mcp` flag for MCP stdio server (via `AgentMCPServer`)
• Graceful startup when Tavily connector not configured (warn + enable fallback)
• Installable via `pip install -e hub/agents/python/knowledge`

Plan: `docs/plans/knowledge-agent.mdx` § "Packaging & Distribution"

[github-actions]

Tracking this as part of the knowledge-agent track (#1141–#1148). Phase 1 here lands cleanly only after the Agent Hub foundation is in place — flagging the dependency chain and one missing reference so this doesn't get picked up out of order.

Hard dependencies (need to merge first):

• Agent Hub: Restructure — move production agents to hub/agents/ #1102 — creates hub/agents/python/ (this issue's package lives here)
• Agent Hub: gaia-agent.yaml manifest parser + validator #1091 — gaia-agent.yaml manifest parser + schema (this issue's gaia-agent.yaml is validated by it)
• Agent Hub: gaia agent init scaffolding with proper package structure #1098 — gaia agent init scaffolding (the canonical way to bootstrap this package structure)
• Agent Hub: API server + MCP server interface standard for agent packages #1101 — API + MCP server interface standard (the --mcp flag contract)
• Agent Hub: Python wheel packaging per agent #1093 — Python wheel packaging per agent (how pip install -e hub/agents/python/knowledge actually works)

Plan reference doesn't resolve: the body points to docs/plans/research-agent.mdx § "Packaging & Distribution", but that file doesn't exist in the repo. docs/plans/agent-hub.mdx is the closest match — is this issue meant to point there, or is there an unmerged plan doc to land first? Worth resolving before someone starts implementing, so the 5-interface standard and the KnowledgeAgent(Agent) API surface have an authoritative source.

Suggested ordering: #1091 + #1098 + #1101 land → use gaia agent init knowledge to scaffold → this issue becomes "fill in the agent + CLI bodies" rather than hand-rolling the package layout from scratch. That keeps the knowledge agent aligned with whatever third-party authors will use.

Happy to draft the KnowledgeAgent(Agent) skeleton + CLI subcommand wiring once the plan ref is sorted and the hub scaffolding is available.