feat: redesign agent harness with modular architecture + Responses API

[cristipufu]

Summary

• Modular agent architecture: Replaces monolithic agent_service.py (873 lines) with agent/ package — events, session, provider, tools, context, loop, service modules
• OpenAI Responses API: Reasoning models (gpt-5.x, o-series) use the Responses API with URL rewrite transport, enabling reasoning summary display in the UI
• Streaming responses: Text tokens stream to the UI in real-time via TextDelta events
• Parallel tool execution: Multiple tool calls execute concurrently with asyncio.gather
• Context management: Automatic compaction at 75% context window + stale tool result pruning
• Retry with backoff: Transient API errors (rate limits, 5xx) retry up to 3 times
• Sub-agent dispatch: dispatch_agent tool spawns isolated child loops for research tasks
• Token tracking: Per-turn and cumulative usage emitted to frontend
• Thinking UI: Collapsible "Reasoning" card in the chat for model thinking content
• Reasoning model support: Skips temperature param for gpt-5.x/o-series, prefix-based context window lookup

Test plan

• Start dev server → open agent chat
• Send message with gpt-5.x model → verify streaming tokens appear in real-time
• Verify "Thinking..." indicator hidden while streaming, shown during tool execution
• Verify tool approval flow works (approve/deny)
• Verify plan bullets update to "completed" before final response
• Test with non-reasoning model (gpt-4o) → verify Chat Completions path still works
• Test stop button cancels running session
• npm run build in frontend directory succeeds
• uv run ruff check src/ tests/ + uv run mypy src/ pass clean

🤖 Generated with Claude Code