0.9.0

0.9.0 (2026-08-04)

⚠ BREAKING CHANGES

• core: VendorCall gains a keyed_body field. Every in-tree constructor is updated; an out-of-tree one must supply it. Cache keys are unchanged for any provider that declares no request.body — the overlay is absent, so both bodies are the document that was keyed before.

Bug Fixes

• core: apply the request.body overlay on the judge and embeddings paths (#78) (ba479d6)
• runs: hide every fully cached run, whatever its verdict (f684322)

Documentation

• cache: correct why a replay's verdict does not save it (77e9662)

0.8.0

0.8.0 (2026-08-03)

⚠ BREAKING CHANGES

• core: base_url is no longer part of the cache key for the vendor providers. A gateway and a direct connection to the same API now share entries instead of paying for the same answers twice — the same call cache_key already made about the model a vendor reports having served. The address is recorded on each entry and a hit from a different one warns, exactly as program_digest reports a rebuilt exec program; cache_salt (new on these variants) separates two endpoints outright when a warning is not enough.

Features

• core: customize provider requests via request:, and stop keying base_url (#65) (56f9298)

0.7.1

0.7.1 (2026-08-03)

Features

• core: refresh pricing table to current-gen Anthropic and OpenAI models (645cc7e)

0.7.0

0.7.0 (2026-08-01)

⚠ BREAKING CHANGES

• ChatRole gains a Tool variant and ChatMessage.content is now MessageContent rather than String. Embedders matching exhaustively on either will need updating.

Features

• tool and content-block turns in per-case history, and validate warnings (#57) (8420ba3)
• web: make sets reachable, rows clickable, and the UI usable on a phone (#55) (66d5ad8)

0.6.2

0.6.2 (2026-07-31)

Features

• per-case conversation history, spliced at a prompt history marker (#47) (1ad6829)
• provenance: read the branch from CI, not from a detached HEAD (#44) (68b7868)

Bug Fixes

• ci: stop bash -e aborting the eval action's binary and summary steps (#46) (046f7d7)

0.6.1

0.6.1 (2026-07-31)

Features

• run-set access control and a browsable Sets hierarchy (#42) (270d27e)

Bug Fixes

• caching: resolve $digest: on a provider's cache_salt (#38) (e83e6eb)
• declared env keys against legacy stores, and the eval gate names the failure (#43) (bedbec3)

0.6.0

0.6.0 (2026-07-30)

⚠ BREAKING CHANGES

• core: let graded assertions see the model's tool calls (#36)

Features

• core: let graded assertions see the model's tool calls (#36) (e9f0e30)

0.5.1

0.5.1 (2026-07-30)

Bug Fixes

• release: unblock release-please, and guard the message format (#33) (2933573)

0.5.0

0.5.0 (2026-07-30)

⚠ BREAKING CHANGES

• caching: provider cache keys move. A store written by 0.4.x or earlier is adopted on first lookup and re-filed, so a warm store keeps serving; --no-cache-migration opts out and re-pays instead.

Features

• caching: one rule — key every outgoing request on its content (#29) (b914837)

Documentation

• overhaul the site — six-section IA, UI screenshots, local-LLM stack, examples 33-39 (#31) (e04c94f)

0.4.0

0.4.0 (2026-07-29)

⚠ BREAKING CHANGES

• caching: key on the request, not on the machine (#26)
• caching: exec provider and exec grader cache entries are re-keyed by the switch from mtime to content digests. Existing entries are not served and will age out; no action is needed beyond expecting one cold run per suite. Network providers are unaffected.

Features

• caching: key exec identity on content, correct the cache_salt docs (#24) (80479dc)
• caching: key on the request, not on the machine (#26) (c90b9f7)
• server: authenticated MCP endpoint, dual-era and read-only (#27) (88739d3)

Documentation

• publish a documentation site with 32 end-to-end tested examples (#28) (4d1bab7)