# OpenAI-Compatible LLM Integration

Audita uses the pinned [PromptKit](https://forge.maximumdirect.net/base-layer/promptkit/src/commit/171fb11f4e967a1da3af57843c6467954b14ba09)
built-in OpenAI-compatible transport. PromptKit owns chat-completions wire framing,
provider envelope parsing, transient retries, per-generation timeouts, and JSON
Schema repair. A backend's configured base endpoint must support the selected
profile's model and structured JSON-schema responses; local configuration
inspection does not contact the provider.

Audita selects role profiles independently for each pipeline entry and supplies
its domain response schema and per-operation budget. PromptKit owns backend and
credential selection, request preparation, transport, retries, repairs, and
admission. Audita executes the same prepared handle and retains its own domain
validation. Profiles may use native endpoint and credential settings; local
inspection does not resolve a credential value or call the provider.

`llm.transient_retries` counts additional attempts for eligible transient
failures; explicit zero makes one attempt. `llm.repair_attempts` counts
additional generations for schema-invalid model output. The operation timeout
covers preparation, capacity waiting, retries, generation, and repair. Caller
cancellation or an earlier deadline also applies. See [configuration](../config.md#settings)
for defaults and the [PromptKit consumer contract](https://forge.maximumdirect.net/base-layer/promptkit/src/tag/v0.21.0/docs/consumers/pkg-promptkit.md)
for protocol and retry details.

A model operation that exhausts structural repair, a live-parent local operation
timeout, capacity rejection, transport failure, malformed provider response,
or provider HTTP failure is recoverable for that operation. Proposal generation
then contributes no candidates from that response; a failed validation batch
marks its candidates unavailable while other batches and later modules can
continue. Caller cancellation, broken assets or configuration, and unknown
internal failures remain fatal and cancel peers. Audita decodes into domain
types only after PromptKit validates the JSON. Ordinary reports use safe reason
codes and counts rather than provider messages. Full request/response content
is written only with `--debug`; see [operations](../operations.md#debug-evidence).