Local regression measurement

The Kernel PerformanceCapacityBaselineTest measures bounded patch preparation, chunk planning, finding merge and rendering with generated worst-case inputs. It reads the included Dockerfile's explicit 128M worker memory limit. Its generous assertions detect regressions; they are not deployment capacity or throughput claims. Provider network latency and real language-sidecar execution are excluded. Run the full Kernel suite with the documented explicit 3600-second test timeout.

Generated-output capacity

Direct requests include an integer server-owned generated-output limit. Set ai_code_review_server.llm_gateway_payload.maximum_output_tokens in deployment service parameters; the default is 8192, and the separately reviewed maximum is 128000. Missing, non-integer, non-positive and above-policy values are configuration errors before credential opening or HTTP. Choose a value within the actual provider model's supported capacity; this module-wide ceiling does not discover model-specific capabilities.

The existing reviewed request-shape metadata supplies the wire field name: OpenAI Responses uses max_output_tokens. The direct transport verifies shapes against their required policy, and repository planning includes the same integer field when measuring the exact encoded body. Caller review requests still select only a registered provider_profile_id; they cannot set a model, endpoint, token limit or wire field. APISIX-managed requests keep the gateway's separately deployed output settings and unchanged body.

OpenAI's reasoning documentation explains that the limit includes reasoning and that reaching it can produce an incomplete response before visible output. Incomplete responses remain failed reviews, without automatic retries or a silent limit increase. Response byte and character ceilings remain separate protections.

A generated-output bound helps reserve an evaluation budget; it is not an exact monetary cap. Input tokens, prices, processing tier, call count and provider-side spending controls must also be part of an approved plan. Changing this setting does not approve a paid call, register a profile or promote a model.