feat: add Cursor Agent CLI as new agent provider (Phase 1)

[cloud1305]

Summary

Add cursor-agent as the 5th agent provider in OpenCove, integrating the official Cursor Agent CLI (agent command) via PTY-based terminal sessions. This is Phase 1 (PTY-only) of the proposal documented in docs/proposals/cursor-agent-support.md.

Motivation

• User value: Cursor is one of the most popular AI-native IDEs. Its standalone CLI Agent allows OpenCove users to orchestrate Cursor Agent directly from the canvas workspace.
• Architecture fit: Cursor's ACP (Agent Client Protocol) aligns well with OpenCove's TerminalSessionState model. Phase 1 uses PTY-only; Phase 2 will leverage ACP for structured state observation.
• Model ecosystem: Cursor Agent aggregates models from multiple providers, accessible via agent models.

Changes

Phase 1 scope (this PR)

Subsystem	Files	Description
Type registration	agent.ts, agentSettings.ts, providerMeta.ts	Add cursor-agent to AgentProviderId, AGENT_PROVIDERS, provider metadata (label, capabilities)
IPC validation	normalize.ts	Add cursor-agent to the provider whitelist
CLI detection	AgentCliAvailability.ts	Two-step detection: which agent + agent --help fingerprint (contains "cursor")
Command factory	AgentCommandFactory.ts	Build CLI args: --yolo, --model, --resume, --continue, prompt
Model service	AgentModelService.ts	Parse agent models output (id - DisplayName (current) format), ANSI cleanup, cache with TTL, auto-resolve current model for title display
Session stubs	SessionFileResolver.ts, SessionTurnStateDetector.ts, SessionLastAssistantMessage.extractors.ts, AgentSessionLocator.ts	Explicit null returns to prevent fallthrough into gemini/codex default branches
UI helpers	helpers.ts	Add cursor-agent to providerTitlePrefix
IPC launch	register.ts	Auto-resolve default model display name when no model specified

Proposal document

docs/proposals/cursor-agent-support.md — comprehensive design document covering feasibility analysis, architecture, state ownership table, invariants, phased execution plan, risk analysis, and acceptance criteria.

Key design decisions

Decision	Choice	Rationale
Provider ID	cursor-agent (not cursor)	Avoid conflict with existing Cursor IDE opener semantics
runtimeObservation	none (Phase 1)	PTY-only for fast validation; ACP bridge deferred to Phase 2
CLI detection	--help fingerprint	agent --version output doesn't contain "cursor"; --help output does ("Start the Cursor Agent")
experimental flag	true	New provider needs opt-in validation before promotion
Default model display	Auto-resolve from agent models	Query (current) model to show actual model name instead of "default model"

Test coverage

• Unit tests: 15+ new test cases across 7 spec files covering command factory, model service, session locator, file resolver, state detector, last message extractor, and settings normalization
• All 421 unit tests pass, 122 test files
• E2E: 166/176 passed, 9 skipped (platform-specific), 1 pre-existing flaky test (minimap hover opacity — unrelated to this PR)

Verification

• pnpm pre-commit: lint ✅, format ✅, type-check ✅, unit tests ✅, E2E ✅ (1 pre-existing flaky failure)
• Manual testing: settings panel shows "Cursor Agent (Experimental)", model list displays correctly, PTY session launches and interacts successfully, window title shows actual current model name

Invariants maintained

1. All CLI command assembly goes through AgentCommandFactory.buildAgentLaunchCommand — no hardcoded args elsewhere
2. CLI detection requires two-step verification — which agent alone is insufficient
3. All session/watcher paths explicitly short-circuit to null — no accidental fallthrough into other providers' logic
4. TerminalSessionState push path (ptyState IPC channel) remains consistent across all providers
5. normalizeAgentSettings correctly handles old data without cursor-agent keys via AGENT_PROVIDERS.reduce

Phase 2 roadmap (not in this PR)

• ACP Bridge (agent acp + JSON-RPC client) for structured state observation
• session/update → working/standby mapping
• session/request_permission → UI approval dialog
• Remove experimental flag

[DeadWaveWave]

建议在本 PR 完成 Phase 2 的内容，并提供正确运行时的截图验证后再尝试提交