routine: opus-4-8-effort (2026-05-28)

[jim4226]

Summary

Upgrades CSIS's builder checkpoint from claude-opus-4-7 to claude-opus-4-8 and exposes the new effort parameter so the Coordinator can tune reasoning depth per-call.

Source

• URL: https://www.anthropic.com/news/claude-opus-4-8
• Published: 2026-05-28
• Key quote: "Opus 4.8 is roughly four times less likely than Opus 4.7 to overlook code flaws, actively flagging uncertainties about its work rather than making unsupported claims."

Theme

Theme 6 — Substrate / capability boundaries (also Theme 2 — Trust + verification). CSIS's Builder and Researcher agents run on the alpha checkpoint. Upgrading to a model that is 4× less likely to miss code flaws directly improves artifact quality before it reaches the Verifier. The effort parameter maps to CSIS's existing cost-vs-thoroughness design axis (Theme 5: self-improvement loops) — the same axis that informed CriticEffortLevel in PR #13.

What changed

File	Change
csis/backends/anthropic.py	_DEFAULT_MODEL_MAP "alpha"/"mock-alpha": claude-opus-4-7 → claude-opus-4-8; complete() passes effort to API when LLMRequest.effort is non-None
csis/backends/base.py	LLMRequest gains `effort: str
tests/test_backends.py	7 new regression tests: effort field defaults, round-trips, model-map correctness, API pass-through when set, API omission when None

No cycle-9 chokepoints touched

_DEFAULT_MODEL_MAP and LLMRequest are leaf declarations. Coordinator.__init__, _BackendTracker, writer_iteration_id, and the promotion CAS are all untouched. The default effort=None means existing code paths pass no effort argument to the API — model's own default ("high" for Opus 4.8) applies, so runtime behaviour is identical to the previous model version until a caller explicitly sets effort.

Test plan

python -m pytest tests/test_backends.py -v   # 7 passed (all new)
python -m pytest tests/ -q                   # 257 passed, 0 failed

---

Generated by Claude Code