v0.8.0 — Hardened agent loop, typed failures, real recovery

The 0.8.0 rewrite, hardened by fifteen review rounds: ~120 verified defects fixed, each pinned by a
regression test.

Highlights

• One execution path — run() and streaming can no longer diverge; every run ends with a typed
  StoppedEvent
• Typed failures — MODEL_ERROR / RUNTIME_ERROR / INTERNAL_ERROR are distinct: provider,
  runtime, or bug
• Real recovery — output truncation, mid-stream failures, and prompt overflow resume instead of
  corrupting history
• Robust parsing — line-anchored, case-insensitive fences; differentially tested against the 0.7.x
  extractor at every chunk size
• Preemptible runtimes — max_exec_timeout works on IPyKernelRuntime and is refused where it
  can't be honored
• Oversize output persists — large stdout becomes a runtime variable the model can slice, not a
  truncation
• Honest token accounting — measured per-script rates (CJK, Cyrillic, emoji…), API-anchored
  estimates, compaction that survives summarizer failures

Breaking changes

• Model subclasses implement _complete, not call
• interrupt() moved from Runtime to PreemptibleRuntime
• max_exec_timeout + in-process runtime now raises at construction
• persisted_output_variable → persisted_output_prefix
• Histories from ≤0.7.5 need a one-time migrate_legacy_summaries()
• utils.extract_first_code_block removed

Install: pip install cave-agent==0.8.0 · extras: [ipykernel] [openai] [litellm] [all]