[emanuilo]

Stacked on #12; merge that first. This PR's base then switches to main.

Summary

The Cursor backend gets native, persistent conversations like Claude's in #12: one bridge process, one open agent per chat, resumed after a restart.

 run_turn(chat, context, message)
-  launch Node bridge                          # every turn
-  create agent; send system prompt + last 10 messages + message
-  close agent and bridge                      # transcript in memory only, lost on restart
+  bridge = shared bridge ?? launch(cwd = memory_dir)          # lazy; agents stored under ~/.cursor, keyed by cwd
+  agent  = open agent for chat
+        ?? agents.resume(session_id, model)                    # force on first send: take over a run that died with the old process
+        ?? create agent                                        # system prompt goes in its first message
+  agent.send(context + message, model, effort param)
+  failure → drop the agent (and the bridge if unreachable); resume next turn

backends/
├── model_selection.py   # ModelSelectionMixin: model/effort, model-list lookups, strict id check,
│                        #   saving to .env — shared by Claude and Cursor (was duplicated)
└── cursor.py
      run_turn
        _ensure_agent → "open" | "resumed" | "created"   # system prompt only when created,
        _send                                            #   force only when resumed

+ /new              closes the chat's agent
+ /model, /effort   models from client.models.list, same button menus as Claude;
+                   /effort maps to a model's reasoning/effort parameter when it has one
+                   saved as CURSOR_MODEL / CURSOR_EFFORT, applied per send (no reconnect)
  consolidation     reuses the shared bridge

Evidence

• Before: pytest on feat: persistent Claude sessions, /new, /model, /effort #12 → 317 passed
  After: pytest → 341 passed, ruff clean. Cursor tests cover:

  bridge launched once, with cwd = memory dir; closed on shutdown
  agent reused across turns; system prompt only in a new agent's first message
  restart → agents.resume(session_id) with model and force; resume failure → fresh agent
  failed turn drops the agent / bridge
  /model and /effort lists, selection, persistence; an id is refused when
    the list can't be loaded (as for Claude)
  bridge robustness: stderr lines > 64 KiB, timeouts don't drop the bridge,
    an exited bridge is relaunched, shutdown with a session added meanwhile
  

• Bridge launch verified live (working dir, clean shutdown, ~0.7 s).
• Not verified live: a full Cursor turn. The only available key is on Cursor's free plan, which the SDK rejects ("Cloud Agent is not available for free users"). Resume and model handling are checked against the SDK source and with mocks.
• Merged with fix: memory system bugs + end-to-end validation suite #14 in a scratch worktree: all 400 unit tests pass.

Merge Danger

Door: two-way

Revert returns Cursor to per-turn bridges and the in-memory transcript. Saved Cursor session ids are then ignored.

Blast Radius: cursor-backend

• Claude users are unaffected.
• Conversation history now lives in Cursor's local agent store under ~/.cursor, tied to the memory dir path. Moving the memory dir starts fresh conversations.
• Conflicts with fix: memory system bugs + end-to-end validation suite #14 in docs/memory-architecture.md only (the session bullets): keep fix: memory system bugs + end-to-end validation suite #14's Claude line and add this PR's Cursor line after it.

🤖 Generated with Claude Code