0.8.97

Added

• The @job one-shot driver can now register embedder-defined host builtins (#3207.) New
  harn_serve::run_job_once_with(.., configure: impl FnOnce(&mut Vm)) hands the embedder the
  fully-built job VM after the standard stdlib + store/metadata registration and before the job
  entrypoint runs, so a host builtin (e.g. a sandbox_exec bridge) can coexist with — or override —
  the standard ones and be callable from the @job .harn code. run_job_once is unchanged and
  delegates to the new variant with a no-op closure. Fixes #3207.
  Added std/eval/agreement: deterministic, I/O-free cross-checked-success math for eval ledgers — the reusable
  counterpart to std/eval/stats. Exposes agreement_decision (the ">=2 independent judges must agree, with at least one
  independent re-execution among them" rule) and cohen_kappa (inter-judge agreement statistic), so eval clients can drop
  their own hand-rolled agreement math.
  Added estimate_cost_usd and realized_trial_cost_usd to std/eval/stats: cache-aware token→USD cost estimation
  (cache-read/write tokens are billed at their own rates and not re-charged at the full input rate) plus cached-replay
  realized-cost accounting. Lets eval harnesses drop their own hand-rolled LLM cost math.

Fixed

Release tooling no longer reflows already-published CHANGELOG.md sections.
make lint-md previously linted the assembled CHANGELOG.md (and CHANGELOG-pre-*.md
archives) under MD013 line-length, so long lines in published ## vX.Y.Z sections were
flagged and rewrapped during a release — tripping the retroactive-edit guard. Those
machine-assembled, append-only files are now excluded from markdownlint; the
changelog.d/* fragments are still linted at the source.

• Agent-loop no-progress feedback now respects native tool mode. Native-tool
  runs are nudged to use the provider tool channel instead of receiving
  text-mode <tool_call> and <user_response> syntax.

Install / Upgrade

cargo install harn-cli

Full Changelog: v0.8.96...v0.8.97