AgentStack

With only one tool there is no cross-tool confusion, but the single tool is an overloaded mega-action bundling auth, planning, work-graph traversal, and prompt retrieval behind opaque sub-commands like plan_get, work_next, and packet.next_action. An agent cannot infer which internal behavior applies without parsing dense prose, so intent boundaries are blurred rather than distinct.

The lone name 'agentstack.execute' follows a readable dot-namespaced, verb-based convention, and with a single tool consistency is trivially satisfied. The verb 'execute' is generic, however, and there is no set of names through which to demonstrate a predictable pattern.

The description implies a broad surface (authentication choices, project creation, plan lifecycle, work-graph navigation, budget prompts) yet exposes exactly one tool. That monolith forces every operation through one entry point with undocumented sub-actions, an extreme mismatch between stated scope and exposed surface.

The prose names many operations (auth.get_profile, create_project_anonymous, plan_execute, discovery.search, session_setup), but none are individually addressable, and it is unclear which operations actually exist versus are references to prompts or internal RPCs. Coverage of the workflow is suggested but not verifiable or usable as a discrete surface.