feat: multi-session MCP-over-HTTP — one connection, N tole sessions (D3 lanjutan)

[ajianaz]

What problem does this solve?

tole's tools are currently reachable via three hosts (CLI, tole mcp stdio, tole serve REST) — but all require the client to be on the same machine. MCP Streamable HTTP would make tole reachable from any MCP client on the network, and the multi-session routing design (one MCP connection → N tole sessions) is the flexible model: the agent opens/closes/switches sessions without reconnecting.

Proposed solution

A tole serve --transport mcp mode (or tole mcp-http) that serves the session host as MCP tools over Streamable HTTP (rmcp transport-streamable-http-server):

MCP tools exposed

Tool	Params	What it does
tole_session_new	{cwd}	Open a durable tole session, return session_id
tole_session_prompt	{session_id, text}	Run one tole turn; returns stop_reason + final_text
tole_session_status	{session_id}	entries count, busy flag
tole_session_list	—	all active session ids
(plus the registry tools: read_file, write_file, git, etc. — routed to the session's workspace)

Design decisions

• Multi-session routing by session_id in tool arguments (approved 2026-09-28): one MCP connection authenticates once (bearer token), then the agent addresses sessions explicitly — no hidden per-connection state
• Sessions are durable server-side (JSONL in the session's workspace jail)
• The tool registry is cached per session (jail = session's workspace)
• Auth: static bearer token (same as D3 v1); TLS documented as reverse-proxy concern
• Busy session → 409-equivalent MCP error (turn serialized per session)

Implementation

• rmcp transport-streamable-http-server feature (pulls axum/hyper — accepted dependency decision)
• Reuse RegistryServer handler (from D1 feat: tole acp — Agent Client Protocol host (D2, issue #95) #130) + session_host machinery (from feat: tole serve — token-authenticated HTTP daemon (D3 v1, issue #96) #131)
• Auth middleware: bearer token check in axum layer

Acceptance criteria

• MCP client connects over HTTP with bearer token, lists tools, calls tole_session_new + tole_session_prompt
• Multiple sessions concurrently from one client
• Durable sessions survive server restart (JSONL on disk)
• Token auth enforced on all MCP endpoints
• Integration test: MCP client over HTTP → full round-trip

Estimate

~1 day (rmcp server-side HTTP is well-supported; the session_host machinery is already tested).