Prepare Audita v4 release notes and retire completed plans

This commit is contained in:
2026-11-17 19:51:50 +00:00
parent 788f830b19
commit 8551121ef6
20 changed files with 92 additions and 6009 deletions

View File

@@ -1,5 +1,6 @@
# Release Notes
- [Audita v4.0.0](v4.0.0.md)
- [Audita v3.1.1](v3.1.1.md)
- [Audita v3.1.0](v3.1.0.md)
- [Audita v3.0.1](v3.0.1.md)

89
docs/releases/v4.0.0.md Normal file
View File

@@ -0,0 +1,89 @@
# Audita v4.0.0
Audita v4.0.0 adopts application configuration schema v3 and native library
commands, adds repair of rejected corrections, and improves processing
reliability and resource use. This major release requires migration from
Audita v3.1.x configurations and command invocations.
## Configuration and command migration
- Replace `pipeline.modules` with ordered `pipeline.stages` records. Each entry
can select its own profiles, glossary context, protected terms, validation
budget, and transcript description. Optional stage IDs identify repeated
modules in reports and debug evidence.
- Move standalone glossaries into named `glossary_collections`, composed through
ConfigKit imports and companion files. Select glossary context and protected
terms independently. Replace `--glossary` and `--modules` invocations with the
new configuration and collection-selection flags.
- Move profiles and backends into native PromptKit resources. Application
configuration discovery, overlays, imports, and prompt companions now follow
the adopted library conventions. Processing preflight checks the entire
selected pipeline before preparing outputs or invoking a model.
- Replace `config print-effective` and the curated profile summary with native
configuration and PromptKit inspection commands. Help lists the available
inspection, comparison, preview, validation, and explicit provider smoke
commands.
- Remove confidence thresholds and confidence fields from custom prompts and
response examples. Audita now decides whether to apply corrections through
validation without model-supplied confidence scores.
- Update automation for CommandKit reporting modes. Completion counts require
`--verbose` or `--debug`. Quiet processing requires a transcript file
destination or `/dev/null`. `--debug` enables retained diagnostic artifacts;
`--progress` shows stage, chunk, proposal, validation, and repair activity on
stderr.
Follow the [configuration migration guide](../migrations/v3.md) and
[maintained examples](../../examples/), then inspect the resolved configuration
and try a small transcript. The migration guide's v3 refers to the application
configuration schema. The [CLI reference](../cli.md) owns current commands and
exit semantics.
## Correction quality and recovery
Rejected proposals can be revised using validator feedback and the original
proposal conversation. Correction repair defaults to one attempt per proposal,
can use a separate profile, and can be disabled. Every revision passes through
validation again. Withdrawn, exhausted, and unavailable branches finish
independently while other approved corrections can proceed.
Prompts clarify module focus, incidental editorial cleanup, protected names,
span deletion, and repair outcomes. Missing required validator decisions continue
to affect only their candidates. Recoverable model failures can coexist with a
successful run and applied corrections. See [configuration](../config.md#settings)
for budgets and profile selection.
## Reliability and reports
Processing uses a bounded window of sections within each stage, preserves
deterministic correction identities, and commits stage text after successful
execution. Fatal failures and cancellation stop and join active workers.
Transcript decoding, normalization, chunking, glossary preparation, and evidence
retention use less repeated work and fewer unnecessary copies. An explicit
target section count now produces exactly that many nonempty sections when
feasible and reports infeasible requests as input errors.
Output publication uses checked destinations and retains processing and delivery
failure evidence. Ordinary reports retain schema version 1 and add stage IDs,
correction repair outcomes, and available peer interruption details. See the
[output contract](../integrations/output-files.md) and
[operations guide](../operations.md) for report interpretation and output
lifecycle.
Audita no longer performs application-wide credential substitution in diagnostic
messages or artifacts. Review diagnostic handling before deployment and protect
retained debug data according to the sensitivity of its transcript, prompt, and
response content. Native library inspection retains its own rendering behavior.
Debug artifacts have changed, including removal of confidence fields and the
Audita-owned utilization diagnostics file.
## Dependencies and source builds
Source builds require Go 1.27.0 or newer. Release builds and Woodpecker use
Go 1.27.4. Distributed Linux binaries require no Go installation.
This release selects ConfigKit v1.2.1, PromptKit v0.23.2, CommandKit v0.7.0,
ckconfig v0.9.0, ckprompt v0.10.0, ReleaseKit v0.10.0, and stable go-yaml v4.0.0.
Audita's direct YAML imports now use v4. Linux amd64 and arm64 remain the
production targets; macOS arm64 receives best-effort development support.
[COPYRIGHT](https://gitea.maximumdirect.net/eric/audita/src/tag/v4.0.0/COPYRIGHT) | [LICENSE](https://gitea.maximumdirect.net/eric/audita/src/tag/v4.0.0/LICENSE) | [ATTRIBUTION](https://gitea.maximumdirect.net/eric/audita/src/tag/v4.0.0/docs/oss/attribution.md)