llm-relay 0.77.0 → 0.78.0

package/dist/backend.d.ts

@@ -11,6 +11,7 @@ import type { ResolvedAttempt } from "./resolved-attempt.js";
 import { type DialectRefusalSignal } from "./tool-dialects.js";
 import { type RecoveredOpenAiChatProcessor } from "./openai-dialect.js";
 import { type UsageAccumulator } from "./usage-observer.js";
+import type { ThoughtSignatureMode, ToolCallIdMode } from "./config-types.js";
 /**
  * A tool-call envelope was present in the response text but could not be parsed — truncated, or a
  * dialect variant we do not model. Distinct from a mapper defect: the relay's translation is fine,
@@ -137,6 +138,21 @@ export declare const QUOTA_DEMOTED_HEADER = "x-llm-relay-quota-demoted";
  * Nothing secret, no credential values.
  */
 export declare const LATENCY_DEMOTED_HEADER = "x-llm-relay-latency-demoted";
+/**
+ * The serving candidate was placed by the PROBATION band: a free deployment with fewer than
+ * `minSamples` served-request samples that the relay deliberately put first to gather data on
+ * it (`routing.probation`, default ON).
+ *
+ * Fourth member of the `DEGRADED_HEADER` family, and it exists for the same reason: an
+ * automatic reorder is acceptable only because it is announced. Unlike its three siblings it
+ * does not state that the first choice was displaced — the probation member usually IS the
+ * first choice — but that the answer came from a deployment with almost no served-traffic
+ * evidence behind it.
+ *
+ * Value is one bounded line, e.g. `opencode/muse-spark-1.3-contributor-free (0 of 5 request
+ * samples)` — the spec and the sample count behind the banding. Nothing secret.
+ */
+export declare const PROBATION_HEADER = "x-llm-relay-probation";
 /**
  * A HEDGE ran: a slow in-flight attempt had the next candidate started beside it, rather than
  * after it. Fourth member of the `DEGRADED_HEADER` family, and the announcement half of the
@@ -287,6 +303,26 @@ export interface FetchBackendArgs {
     signal: AbortSignal;
     onEgress?: OnEgress;
 }
+export interface AnthropicToOpenAiResponsesOptions {
+    /** The resolved deployment's model id. Absent => no `model` key, same convention as `openai-request.ts`. */
+    model?: string | undefined;
+    /** Whether THIS hop streams — a relay decision, not the caller's. Falls back to the body. */
+    stream?: boolean | undefined;
+    /** Same resolved mode `anthropicRequestToOpenAi` receives; see that module for the provenance. */
+    toolCallIds?: ToolCallIdMode | undefined;
+    onToolCallIdsRewritten?: ((count: number) => void) | undefined;
+    thoughtSignature?: ThoughtSignatureMode | undefined;
+    onThoughtSignatureSentinels?: ((count: number) => void) | undefined;
+}
+/**
+ * Translate one Anthropic Messages request body into an OpenAI Responses request body.
+ *
+ * Item order is preserved exactly. Unknown top-level fields are not forwarded — a translation
+ * between two contracts, not a passthrough, the same rule `anthropicRequestToOpenAi` states.
+ *
+ * @throws {RequestMappingError} for a block or declaration that cannot be represented.
+ */
+export declare function anthropicRequestToOpenAiResponses(reqJson: unknown, opts?: AnthropicToOpenAiResponsesOptions): Record<string, unknown>;
 /**
  * Perform one upstream inference call against the resolved attempt.
  *
@@ -363,6 +399,18 @@ export interface FetchOpenAiFrontArgs {
     processRecoveredChat?: RecoveredOpenAiChatProcessor;
     usage?: UsageAccumulator;
     onEgress?: OnEgress;
+    /**
+     * Catalog max-output lookup for the Responses front's `anthropic`-kind fallback (2026-09-09).
+     * Anthropic's Messages API REQUIRES `max_tokens`, and `openaiResponsesRequestToAnthropic` no
+     * longer invents a value when the caller stated none (see that module's header) — so a
+     * passthrough target still needs one resolved, and the catalog's published figure is the
+     * second-strongest evidence after the deployment's own learned `max-output` fact. Optional:
+     * absent (or a provider/model the catalog holds nothing for) falls straight through to
+     * `DEFAULT_RESPONSES_MAX_TOKENS`.
+     */
+    catalogLimits?: (provider: string, model: string) => {
+        maxOutputTokens: number | null;
+    } | null | undefined;
 }
 /**
  * OpenAI-compatible FRONT: an OpenAI Chat Completions or Responses request comes in, its