Plain Completion Joins the Named-Configuration Path

Context

The three-tier model (Provider → Model → Configuration, ADR-001) reaches every chat-shaped and embedding capability through a *ForConfiguration entry point: chatWithConfiguration(), streamChatWithConfiguration(), chatWithToolsForConfiguration() and — since ADR-055 — embedForConfiguration() all resolve the adapter from a DB-backed LlmConfiguration and run through the middleware pipeline, so budgets are enforced and cost is attributed per configuration.

The high-level CompletionService did not. Its complete() family resolves only the instance-default configuration (chat() picks the active default, ADR-034). A consumer that needs several distinct named text configurations — a summariser, a classifier and a chatbot in one extension — could not target them by identifier; every plain completion went to the single default. The low-level LlmServiceManager::completeWithConfiguration() existed but routes through the provider's raw complete operation (no system-prompt shaping, no response-format normalisation, no per-user budget metadata), so it is not a drop-in for the message-based complete() path.

Decision

Plain completion joins the configuration path. LlmServiceManager::completeForConfiguration(string $prompt, LlmConfiguration $configuration, ?ChatOptions $options = null) mirrors chat()'s default-configuration branch — it builds the system/user ChatMessage value objects, injects the configuration's skills, and threads the per-user budget and idempotency metadata from the options — but against the caller's chosen configuration instead of the resolved default. A pinned provider on the options is irrelevant on the configuration path and is dropped, exactly as chat() does. The method takes a typed ChatOptions rather than the low-level completeWithConfiguration() metadata/override arrays, so budget and idempotency parity is handled once inside the manager.

The high-level feature service follows. CompletionServiceInterface gains the named-configuration counterparts of its whole family: completeForConfiguration(), completeJsonForConfiguration(), completeMarkdownForConfiguration(), completeFactualForConfiguration() and completeCreativeForConfiguration(). Each applies the same option transforms as its instance-default twin (response-format normalisation, Markdown system-prompt augmentation, factual/creative presets) before delegating to the manager — the shared transforms are extracted into private helpers so the two paths cannot drift.

Method-based, not a ChatOptions field. Targeting a configuration is expressed as an explicit LlmConfiguration parameter, consistent with the entire *ForConfiguration family, keeping ChatOptions a pure scalar readonly DTO.

Consequences

• Completion consumers select a backend-managed configuration by identifier instead of relying on the single instance default; per-configuration budgets and cost attribution apply to plain completion like to every other capability.

• Behaviour parity holds: the JSON/Markdown/factual/creative variants apply identical transforms on the configuration path as on the instance-default path, guaranteed by shared private helpers rather than duplicated logic.

• CompletionServiceInterface and LlmServiceManagerInterface gained methods — implementers outside this repo must add them. In-repo the concrete services and the hand-written StaticCompletionService test double are updated.

• The low-level completeWithConfiguration() (raw complete operation) is unchanged and remains available for callers that deliberately want the non-chat completion endpoint.