Document Audita v2 operator migration and examples

This commit is contained in:
2026-09-25 16:00:27 +00:00
parent 2d167d0d67
commit e5c1029165
9 changed files with 201 additions and 87 deletions

View File

@@ -1,69 +1,54 @@
# Integration: Audita
## Purpose
Define the Audita adapter contract used by the `polish` stage.
Narratio invokes Audita v2's `process INPUT --glossary PATH --output PATH`
subprocess for `polish`. The configured timeout and parent cancellation bound
execution. Narratio passes explicit `--config`, `--overlay`,
`--proposal-profile`, `--validation-profile`, `--transcript-description`,
`--output-schema`, and `--modules` only when selected. An omitted module list
leaves Audita's default sequence in place; `modules: []` passes an explicit
empty argument and runs no correction modules. Order and repetitions are kept.
Narratio passes no debug, work-directory, or reporting flags, leaving those
choices to Audita's own configuration.
## External Boundary
Audita owns its ConfigKit file, backend and profile definitions, native config
discovery, and their validation. Narratio does not create an Audita
configuration or infer downstream overlay/profile names from its own selected
pipeline overlay. An empty Narratio Audita selector adds no CLI override.
The [configuration reference](../config.md#audita-selection-and-credentials)
owns Narratio's fields and defaults. Audita's
[v2 migration guide](https://gitea.maximumdirect.net/eric/audita/src/tag/v2.0.0/docs/migrations/v2.md)
owns its downstream file format.
Narratio invokes `audita process` as a subprocess for each polish operation.
The configured timeout and parent cancellation bound the invocation. Internal
runner composition is documented in
[the adapter implementation guide](../internal/adapters.md).
Each name in `pipeline.audita.credentials` is a ConfigKit secret reference
resolved when Narratio loads the pipeline. Narratio forwards only the selected
names and their captured, nonempty values under those exact names, in the
minimal child environment; it never maps them to a fixed alias or rereads the
parent environment. Every name is registered for captured stdout/stderr and
failure-diagnostic redaction. Audita resolves credentials for all configured
backends, including unselected ones, so operators must list every backend
credential that its selected configuration needs. The downstream files and
provider output remain Audita's responsibility, outside Narratio's
captured-stream redaction boundary.
## Request Contract
`PolishRequest` carries:
- required transcript/glossary/output/work-dir paths;
- optional report path (required when report mode is enabled);
- generated config and stdout/stderr log paths;
- optional per-invocation module override.
The corrected transcript is the only canonical polish output. A successful
invocation must produce a valid JSON transcript with a top-level `segments`
array within Narratio's 64 MiB bound. Exit zero is accepted even if Audita's
ordinary reports describe recoverable model errors or no corrections. Narratio
does not parse report status, counts, or Markdown. It discovers only
`errors.yml`, `recoveries.yml`, and `report.md` next to the requested output,
each subject to a 16 MiB read bound. Missing reports are allowed, for example
when Audita reporting is disabled externally. Available reports are diagnostic
references, not canonical artifact outputs. See Audita's
[ordinary output contract](https://gitea.maximumdirect.net/eric/audita/src/tag/v2.0.0/docs/integrations/output-files.md)
for their schemas, and [operations](../operations.md) for placement and
publication.
The constructed runner owns static Audita settings: binary, timeout,
credentials, default modules, model and endpoint settings, validation and output
settings, report mode, and concurrency. The `polish` stage supplies only
invocation-specific paths and may override modules for that invocation.
On nonzero exit, Narratio preserves the original subprocess error and
available run-local evidence; it never promotes a transcript merely because
the file exists. The adapter returns safe paths, exit details, and selection
metadata on failure. Narratio's `audita.generated.v2` YAML records the
invocation, including arguments, paths, and credential names but never values;
it is a diagnostic record, not a generated Audita config file.
## Result Contract
`PolishResult` returns:
- processed transcript path;
- optional report path;
- work dir and generated-config/log paths;
- exit code, duration, binary provenance;
- adapter metadata map.
## Validation and Failure Semantics
Construction fails for invalid static config values, including:
- empty binary;
- non-positive timeout;
- invalid base URL;
- invalid output schema;
- invalid work-dir retention value;
- invalid concurrency values.
Run fails for:
- missing required request paths;
- subprocess execution failure;
- invalid processed transcript JSON (`segments` array required);
- invalid report JSON when reporting is enabled.
Processed transcript JSON is limited to 64 MiB and optional report JSON to 16
MiB. Both must be regular files without symlinked path components.
Failure results still include output/log/config/exit metadata for diagnostics.
## Deterministic Behavior
- CLI args are built from runner config + request in a fixed order.
- A configured typed `llm_api_key` reference is resolved during pipeline load;
the captured value is supplied only as the child's `AUDITA_LLM_API_KEY`
override and is never re-read from the parent environment.
- Generated invocation YAML (`audita.generated.v1`) is emitted when requested.
- Manifest writes are stage-owned; adapter itself is stateless.
## Configuration
Operator-selected values are defined under `pipeline.audita.*` in the
[configuration reference](../config.md#pipeline).
Maintained example with Audita config:
- [Full annotated pipeline](../../examples/pipeline.full.annotated.yml)
- [Production-shaped pipeline](../../examples/pipeline.production.yml)
Maintained Narratio examples: [full annotated pipeline](../../examples/pipeline.full.annotated.yml)
and [production-shaped pipeline](../../examples/pipeline.production.yml).