MCP and Dream Chat tool parity

apps/lantern-garage/lib/tool-runner.js is the canonical source for shared tool names, schemas, policies, execution, and receipt shape.

Dream Chat exposes the manifest at GET /api/dream/tools and executes tools directly through that runner. MCP invokes scripts/tool-runner-bridge.js through src/mcp_server/shared_tool_bridge.py; Python does not contain a second schema or policy table.

manifests/tool-capability-manifest-v1.json is a generated fallback for MCP discovery when Node is unavailable. Regenerate it with:


node scripts/tool-runner-bridge.js generate-manifest

Contract tests compare it to the live JS registry so drift fails CI.

Shared tools:

read: Read, LS, Glob, Grep, web_search, web_fetch
operator-only shell: Bash, PowerShell
operator-only mutation: Write, Edit

MCP tools/list includes the same manifest as sharedToolManifest. Each shared tool carries _meta.lantern.kind = shared_capability. Queue, task, fleet, GitHub, local-runner, and convergence tools remain available but are labeled mcp_specific_operational.

Gates

CHAT_TOOL_EXEC=1 enables execution on both surfaces. When absent, tools

remain discoverable and return status=unavailable with reason_code=chat_tool_exec_disabled.

MCP_SHARED_TOOL_OPERATOR=1 authorizes operator-only tools for the MCP

process. It is off by default.

Repo sandboxing, the shared command allowlist, safe execution, and web-fetch

private-host blocking remain authoritative in the Node runner.

Every result uses:


{

  "status": "executed | denied | unavailable | blocked",

  "tool": "Read",

  "reason_code": null,

  "receipt": {

    "schema_version": 1

  }

}

The focused contract checks are:


node tests/test_tool_capability_manifest.js

python -m pytest tests/test_mcp_tool_parity.py -q