diff --git a/docs/releases/README.md b/docs/releases/README.md index 19478c3..a987cd1 100644 --- a/docs/releases/README.md +++ b/docs/releases/README.md @@ -1,5 +1,6 @@ # Release Notes +- [Audita v2.0.0](v2.0.0.md) - [Audita v1.1.0](v1.1.0.md) See the [release documentation requirements](../policy/release.md#release-documentation). diff --git a/docs/releases/v2.0.0.md b/docs/releases/v2.0.0.md new file mode 100644 index 0000000..20e7219 --- /dev/null +++ b/docs/releases/v2.0.0.md @@ -0,0 +1,38 @@ +# Audita v2.0.0 + +Audita v2 adopts ConfigKit for configuration and PromptKit for LLM execution, +adds resilient processing and focused reports, and improves correction prompts +and validation. This is an intentional breaking release; existing installations +must migrate their configuration and invocation settings. + +- Configuration uses `version: 2` YAML, named backends, PromptKit model profiles, + and optional named overlays. Offline configuration validation and inspection + expose effective settings, selected profiles, and provenance. +- Proposal and validation calls share per-backend capacity. PromptKit handles + transport retries and structural response repair within bounded operation + timeouts. +- Transcript chunks remain fixed through the LLM stages. Recoverable model + failures preserve progress and allow other work to continue. If every LLM + operation fails recoverably, processing still produces the deterministically + normalized transcript without LLM corrections. Cancellation, application + failures, and required output failures remain fatal. +- Ordinary output includes `errors.yml`, `recoveries.yml`, and `report.md`, plus + a brief completion summary. Detailed correction and LLM diagnostics are opt-in + with `--debug`; `--reporting=false` suppresses all three ordinary reports. +- Empty replacement text supports validated span deletion while preserving a + nonblank segment. Validator responses retain usable decisions despite missing, + duplicate, or extraneous entries; unavailable decisions never approve an edit. +- Prompts use short request-local references, consistent shared instructions, + and relevant neighboring/glossary context for validation. Spoken-word cleanup + can reduce accidental protected-name repetition subject to validation. + +Follow the [v2 migration guide](../migrations/v2.md) before upgrading. Version-1 +configuration, implicit environment-setting overlays, retired model/credential/ +concurrency flags, and the former JSON report interface have no compatibility +shim. Start from the maintained examples linked in that guide, configure profiles +and credential references, and validate the new configuration. Update report and +debug consumers for the [current output contract](../integrations/output-files.md). + +Linux amd64 and arm64 remain the production targets; macOS arm64 is best effort +for development. Binary and checksum delivery runs asynchronously after source/tag +publication, as described in the [release procedure](../release.md).