[FEATURE] Cross-Project Inter-Session Communication (Agent Lobby)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Developers working on ecosystems of interconnected projects (libraries, services, infrastructure tools, documentation systems) run multiple Claude Code sessions simultaneously. Each session is fully isolated. When cross-project knowledge is needed — "what's your API schema?", "how did you set up routing?", "I'm adding an endpoint consumers should know about" — the human is the sole bridge, manually copying context between terminals.
This bottleneck grows with the number of active projects and undermines the productivity gains Claude Code provides within a single project.

Proposed Solution

A lightweight local agent lobby where sessions register and communicate via message passing.
Key constraint: communication, not access. Agent A cannot touch Project B's filesystem. It can only send a natural language request to Agent B, who decides what to share from its own context. The trust model is conversational — agents are peers, not privileged accessors.

# Terminal 1 — project-api
> /connect
Connected to agent lobby. You are: project-api (Python API service)
Lobby members: project-frontend, project-docs

# Terminal 2 — project-frontend
> I need to call the new auth endpoint — ask project-api what the schema looks like
# Agent messages project-api, gets the answer, continues working

Alternative Solutions

Manual context shuttling: Copying architecture details, API schemas, and error messages between terminals by hand. This is the primary workaround and the core bottleneck this feature would eliminate.
MCP-based knowledge systems: Using tools like a knowledge base MCP server to store and retrieve project documentation that multiple sessions can query. This provides async, indirect "communication" but lacks real-time awareness — Agent A can't ask Agent B a question about what it's doing right now.
Shared CLAUDE.md / MEMORY.md conventions: Documenting cross-project conventions in project-level markdown files. Static and quickly outdated — it captures what was true when written, not the current state of a live session.

Why these fall short: All of these are asynchronous and human-mediated. None of them give agents real-time peer awareness or the ability to have a direct conversation with another active session.

Priority

High - Significant impact on productivity

Feature Category

Interactive mode (TUI)

Use Case Example

• Coordination: Agent A adds an endpoint to a shared library and notifies consumer project agents of the change
• Consultation: A fresh project agent asks a mature project agent for setup guidance, conventions, and pitfalls — leveraging the deep context it has built up over a working session
• Awareness: A documentation agent asks another project agent to summarize its core data models, instead of the developer copying definitions between terminals
• Knowledge dissemination: Agents accumulate expertise through deep engagement with their project. Other agents can tap into that expertise through conversation, like a senior dev onboarding a junior one

Additional Context

This is related to but distinct from several existing requests:

#4993 — Agent-to-agent communication (closed by stale bot) — focuses on collaborative workflows within a single project
#21277 — Cross-context communication (closed as duplicate) — same-codebase session coordination
#20559 — Cross-repository session communication
#30140 — Shared channel for agent teams — persistent communication within a team

What's different here: The above issues focus on agents collaborating on the same task or codebase. This request is about independent sessions across separate projects that can consult each other as peers — not share access or co-own work.