Add Claude Managed Agents and Cursor SDK providers for trading agents

[anthonyrs06]

Context

TickEngine currently supports two LLM provider families via agent_key: ACP subprocess agents (claude-code, gemini, ...) that rebuild the full prompt each tick, and PydanticAI direct-API models. Neither offers a persistent conversation — every tick starts from a cold context, so the agent has no native memory of earlier ticks beyond what the journal re-injects.

Change

Two new opt-in providers with persistent sessions and local tool execution:

claude-managed[:<model>] — runs the brain on Anthropic's hosted Managed Agents harness:

• One hosted conversation per Condor session; ticks are follow-up messages (harness handles compaction/caching), so only dynamic state is sent per tick (build_managed_tick_prompt)
• Cross-session memory store mounted at /mnt/memory in the agent sandbox — the self-learning loop survives restarts
• Trading tools exposed as custom tools through a new McpToolBridge (condor/acp/managed_tools.py): the hosted agent emits tool requests, Condor executes them locally through the same MCP servers, gated by the existing auto_approve_with_risk_check callback. Credentials never leave the machine; dry-run/risk blocking behaves identically to other providers
• Wedged hosted sessions (which otherwise dead-loop silently) are rotated automatically after 3 consecutive tick timeouts, with user notification

cursor-managed[:<model>] — runs the brain on the Cursor SDK local runtime:

• Persistent local conversation per session; filesystem memory under the agent directory (memory/playbooks.md, mistakes.md, regimes.md)
• Same MCP bridge + risk gating as above
• Auto-recovery: send() retries once after resetting the SDK bridge and rotating the persisted agent on InternalServerError/ConnectError; the engine rotates after 2 consecutive provider errors; new sessions always provision a fresh remote conversation; starting a second cursor-managed session for the same agent returns 409

Risks

• Zero impact when unused — both providers are opt-in via agent_key; existing ACP/PydanticAI paths are untouched apart from a shared prompt-branch condition
• New deps (anthropic>=0.109.1, cursor-sdk>=0.1.8) are import-guarded inside the provider modules
• All recovery paths are covered by unit tests with faked SDK clients — no network in CI (29 tests added)

Rollout

• Docs: provider sections + troubleshooting table in condor/trading_agent/README.md
• Requires ANTHROPIC_API_KEY (claude-managed) / CURSOR_API_KEY (cursor-managed)

Test plan

• pytest tests/ — 29 passed (managed tool bridge, session rotation, cursor client event mapping, permission blocking, bridge recovery, engine routing)
• import condor.trading_agent.engine / condor.web.routes.agents clean
• Live validation on a fork: both providers ran multi-day BTC-perp paper/live sessions (60+ ticks per session) including automatic recovery from a wedged Cursor bridge and a stale remote agent

Made with Cursor