GEWU_agent/CLAUDE.md

GEWU research agent — Claude harness

This file provides discovery and instructions for Claude (Claude Code, Claude Desktop, and Claude-based agent hosts) operating in a GEWU-enabled workspace.

Instruction routing

Follow the canonical GEWU research agent profile in AGENTS.md.

• Role and workflow routing: AGENTS.md
• Evidence and support rules: protocols/EVIDENCE.md
• Project and data boundaries: protocols/SCOPE.md
• MCP proposals, tools and resume: protocols/CONTRIBUTIONS.md

Available skills

When executing research requests, load and follow the corresponding skill and pipeline:

Intent	Skill	Pipeline
Process selected sources	kb-ingest	ingest
Query knowledge base with citations	kb-query	query
Audit freshness, contradictions, links	kb-curate	curate
Synthesize comparative findings	kb-synthesize	synthesize

Claude host configuration

Configure the selected host's project MCP entry for the authenticated Streamable HTTP endpoint, normally <service-url>/mcp. Supply the bearer token through the host's secure credential mechanism. Native host setup is not yet qualified. The local gewu-connector command transfers explicitly selected local files and does not provide the general GEWU tool catalog.

Workspace coexistence

When installed in a user project workspace alongside other agents:

1. Preserve existing persona: Do not replace an existing project-level CLAUDE.md or AGENTS.md. Keep GEWU's profile in .agents/gewu/AGENTS.md and reference it when research tasks are requested.
2. Skill namespacing: All GEWU skills are namespaced with kb-*. They live in .agents/gewu/skills/ without conflicting with existing domain or engineering skills.
3. Operational isolation: Source processing operates exclusively on files in Inbox/, unaccepted drafts remain in Drafts/, and execution checkpoints reside in Runs/. Never overwrite user code files directly; submit changes through typed MCP proposals.