feat(monitor): backend-driven SWE-bench evaluation with thread-native traces

[shuxueshuxue]

Summary

• make thread/run the only trace ownership model in monitor backend
• remove session-trace API and keep session page focused on session metadata + terminal commands
• render trace on thread page via canonical thread trace endpoint
• keep Evaluation as links to evaluation threads/runs (not a fake local-only animation)

Why

• aligns data model with product semantics: trace belongs to thread/run
• avoids confusion from session-level trace ownership

Validation

• backend route check: only GET /api/monitor/thread/{thread_id}/trace exists
• GET /api/monitor/session/{session_id}/trace now returns 404
• frontend build: npm run build in frontend/monitor
• backend tests: .venv/bin/pytest -q tests/test_integration_new_arch.py (23 passed)
• Playwright manual check:
  • thread page shows Thread Trace Conversation
  • session page no longer shows trace panel

Screenshots

• /home/ubuntu/specops0/artifacts/playwright/thread-trace-thread-page.png
• /home/ubuntu/specops0/artifacts/playwright/thread-trace-session-page-no-trace.png

Follow-up Update

• switched /api/monitor/evaluations to backend direct SWE runner (no control prompt sent to agent)
• removed Terminal Commands from session API response and session UI
• added checkpoint-based fallback in GET /api/monitor/thread/{thread_id}/trace so SWE-bench threads still render tool_call/tool_result style trace when run_events are absent

Relationship to PR #88

• This PR does not replace feat(eval): case-driven SWE-bench A/B loop + self-contained report #88.
• PR feat(eval): case-driven SWE-bench A/B loop + self-contained report #88 scope: SWE-bench slice tooling, case reports, and eval docs/artifacts.
• PR feat(monitor): backend-driven SWE-bench evaluation with thread-native traces #93 scope: monitor integration (backend-driven evaluation trigger, thread-bound trace rendering, session page simplification).
• File overlap between feat(eval): case-driven SWE-bench A/B loop + self-contained report #88 and feat(monitor): backend-driven SWE-bench evaluation with thread-native traces #93 is zero, so they should be reviewed/merged independently.

[shuxueshuxue]

superseded by #182