4.3 KiB
Audita LLM Runtime
llm.NewRuntime assembles one captured PromptKit engine using its default
catalog, Audita's embedded prompts and schemas, and optional operator sources
anchored to the recorded application main context. Captured host sources and
message files are passed to output collision checks. An embedded-only diagnostic
construction distinguishes shipped asset failures from operator-source failures;
it never replaces the requested engine.
Each resolved pipeline entry receives role-bound clients with its selected
profiles and effective budgets. Clients share engine-local backend admission.
The adapter uses ordinary Run with OperationTimeout and OnPrepared, passes
Audita's per-call schema and repair budget, and decodes only validated results.
The operation budget includes preparation, admission, retries, and structural
repair. An expired parent context stops processing and retains any native cause.
With an active parent, PromptKit's operation deadline remains recoverable even
when a preparation error wraps it. A native transport timeout affects the HTTP
attempt and is recorded as a transport failure. The prepared notification
supplies prompt metadata to observers. With debug capture enabled, it also
supplies the actual request messages when preparation reaches that point.
PromptKit owns backend and credential selection. When a required credential is absent,
the CLI identifies the role, profile, and environment variable using the native
diagnostic. See the integration contract
for failure behavior and the configuration reference for sources.
When correction repair is enabled, the runner checks the module's repair prompt
and coordinator before submitting proposal work. The proposal call captures a
native immutable PromptKit exchange. Each actionable correction starts an independent
Engine.Coordinate branch from that exchange. PromptKit assembles continuation
history, owns the additional-turn budget, and delivers the last permitted response
for assessment. Follow-ups select the repair prompt, profile, and output schema
directly; Audita does not rerender the original proposal or assemble replay messages.
An opaque exchange in the structured-output contract keeps provider state outside
modules and validators. The adapter shares request construction, notifications,
observers, and once-only completion between ordinary and coordinated calls.
Completion attaches terminal counts when available and records one logical finish.
Prepared requests and terminal output travel in the returned response for an
optional interaction artifact. Logical call notifications carry metadata.
Audita owns actionable rejection, revision decoding, withdrawal, current previews, and full revalidation. Each branch returns its attempt facts and final approval to the section runner. Accepted siblings remain independent. Requesting more work at the budget boundary produces exhaustion without allocating another logical call. A canceled, undelivered response supplies diagnostic evidence only; it cannot enter the approved set. Repair disabled in configuration bypasses coordination entirely. The runner supplements each repair instruction with its current proposal preview as described in the module guide. The bundled prompts separate durable system principles from the initial user assignment so later repair instructions can allow acceptable minor cleanup; see prompt instruction ownership.
The observation collector tracks logical calls and execution events, and enrolls
selected runs in a native ExecutionGroup. Proposal, repair, and nested validator
clients explicitly share the invocation observer; history reuse adds no extra run.
Native summaries deduplicate terminal totals by run ID and retain terminal coverage
separately from usage presence and completeness. Logical calls and run membership
can differ, especially on failures before execution.
Processing operations retain the failure class, reason, and underlying error.
When a client omits the reason, the operation uses internal_failure and keeps
the original class and error chain. Ordinary reports use compact outcomes.
Request and response evidence is optional debug material. The runner preserves
ordered pipeline traversal and deterministic section collection while canceling
and joining peers on fatal errors.