1 Commits

Author SHA1 Message Date
jeremy 58a6e87d89 Add Opus 4.8 Terminal-Bench harness via native Anthropic API
- New harbor adapter (jcode_harbor_claude_agent.py) routing Opus 4.8 through
  anthropic-api (claude-opus-4-8), with OpenRouter and Claude OAuth fallbacks.
- New runner (run_terminal_bench_claude.sh) defaulting to TB 2.1 + Opus 4.8.
- Drop the /usr/lib/ssl/certs symlink that hijacked the container system cert
  dir and broke each task verifier's curl/uv bootstrap (reward forced to 0).
- Record CC+Opus4.8 baseline (351/445 = 78.9%) as the floor to beat.

Validated: regex-log pilot now scores reward 1.0 (verifier PASSED).
2026-07-01 13:20:17 -07:00