@inflexa-ai/harness 0.14.1 → 0.14.2

package/dist/index.d.ts

@@ -39,7 +39,7 @@ export type { AgentSession, RequestSession, RunSession, Identity, Scope, Credent
 export { makeLocalAuth } from "./auth/local-auth-context.js";
 export { createAnthropicProvider } from "./providers/anthropic.js";
 export type { AnthropicProviderDeps } from "./providers/anthropic.js";
-export { createConfiguredAiSdkProvider } from "./providers/ai-sdk.js";
+export { createConfiguredAiSdkProvider, DEFAULT_MAX_OUTPUT_TOKENS } from "./providers/ai-sdk.js";
 export type { AiSdkProviderConfig, ConfiguredAiSdkProviderDeps } from "./providers/ai-sdk.js";
 export { createEmbeddingProvider } from "./providers/embedding.js";
 export type { EmbeddingProviderDeps } from "./providers/embedding.js";

package/dist/index.js

@@ -48,7 +48,7 @@ export { createAnthropicProvider } from "./providers/anthropic.js";
 // union discriminated over the `anthropic` and `openai-compatible` kinds — with
 // `ConfiguredAiSdkProviderDeps` as its argument shape. `createAnthropicProvider`
 // above is a convenience over this union's `anthropic` arm.
-export { createConfiguredAiSdkProvider } from "./providers/ai-sdk.js";
+export { createConfiguredAiSdkProvider, DEFAULT_MAX_OUTPUT_TOKENS } from "./providers/ai-sdk.js";
 export { createEmbeddingProvider } from "./providers/embedding.js";
 // Prompt caching. `PromptCachePolicy` is the harness's vendor-neutral cache
 // directive; a composition root sets it per run via `RunAgentOptions.promptCache`

package/dist/providers/ai-sdk.d.ts

@@ -13,11 +13,25 @@ export declare const RETRY_MAX_RETRIES = 10;
 export declare const RETRY_INITIAL_DELAY_MS = 2000;
 export declare const RETRY_BACKOFF_FACTOR = 2;
 export declare const RETRY_MAX_DELAY_MS = 30000;
+/**
+ * Output-token ceiling requested when a config names none — a floor against a
+ * provider default, not a target. `@ai-sdk/anthropic` resolves an unset value
+ * from a capability table that lags Anthropic's releases and falls back to 4096
+ * for ids it does not know, so the newest models silently truncate mid
+ * tool-call. The SDK clamps this down per known model (opus-4-x to 32k,
+ * claude-3-haiku to 4096), so it cannot over-request one it recognizes.
+ *
+ * Lower it per-config for servers that validate `prompt + max_tokens <= context`
+ * (vLLM and similar) against a small-context model.
+ */
+export declare const DEFAULT_MAX_OUTPUT_TOKENS = 64000;
 export interface AiSdkProviderDeps {
     readonly model: LanguageModel;
     readonly resolveBilling: ResolveBilling;
     readonly capabilities?: Partial<ProviderCapabilities>;
     readonly logger?: Logger;
+    /** Output-token ceiling per request. Defaults to {@link DEFAULT_MAX_OUTPUT_TOKENS}. */
+    readonly maxOutputTokens?: number;
 }
 /**
  * Connection + model configuration for one `ChatProvider`, discriminated by
@@ -36,6 +50,8 @@ export type AiSdkProviderConfig = {
     readonly model: string;
     readonly fetch?: FetchLike;
     readonly capabilities?: Partial<ProviderCapabilities>;
+    /** Output-token ceiling per request. Defaults to {@link DEFAULT_MAX_OUTPUT_TOKENS}. */
+    readonly maxOutputTokens?: number;
 } | {
     readonly kind: "openai-compatible";
     readonly name: string;
@@ -44,6 +60,8 @@ export type AiSdkProviderConfig = {
     readonly model: string;
     readonly fetch?: FetchLike;
     readonly capabilities?: Partial<ProviderCapabilities>;
+    /** Output-token ceiling per request. Defaults to {@link DEFAULT_MAX_OUTPUT_TOKENS}. */
+    readonly maxOutputTokens?: number;
 };
 export interface ConfiguredAiSdkProviderDeps {
     readonly config: AiSdkProviderConfig;

package/dist/providers/ai-sdk.js

@@ -18,6 +18,38 @@ export const RETRY_MAX_RETRIES = 10;
 export const RETRY_INITIAL_DELAY_MS = 2_000;
 export const RETRY_BACKOFF_FACTOR = 2;
 export const RETRY_MAX_DELAY_MS = 30_000;
+/**
+ * Output-token ceiling requested when a config names none — a floor against a
+ * provider default, not a target. `@ai-sdk/anthropic` resolves an unset value
+ * from a capability table that lags Anthropic's releases and falls back to 4096
+ * for ids it does not know, so the newest models silently truncate mid
+ * tool-call. The SDK clamps this down per known model (opus-4-x to 32k,
+ * claude-3-haiku to 4096), so it cannot over-request one it recognizes.
+ *
+ * Lower it per-config for servers that validate `prompt + max_tokens <= context`
+ * (vLLM and similar) against a small-context model.
+ */
+export const DEFAULT_MAX_OUTPUT_TOKENS = 64_000;
+/**
+ * Point the AI SDK's warning logger at the harness `Logger`.
+ *
+ * The SDK writes warnings itself — `process.emitWarning`, else `console.warn` —
+ * and offers only this process-global hook to redirect them. Left alone it
+ * bypasses the Logger seam entirely and lands on the stdout a TUI host owns;
+ * `maxOutputTokens` makes that routine, since the SDK warns on every request
+ * whose ceiling it clamps. Installed once: the handler is global, and every
+ * provider over a boot shares one sink.
+ */
+function routeSdkWarningsTo(logger) {
+    const g = globalThis;
+    if (g.AI_SDK_LOG_WARNINGS !== undefined)
+        return;
+    g.AI_SDK_LOG_WARNINGS = (options) => {
+        for (const warning of options.warnings) {
+            logger.warn("ai sdk warning", { provider: options.provider, model: options.model, warning });
+        }
+    };
+}
 function workloadOf(session) {
     return `${session.scope.kind}:${scopeWorkloadId(session.scope)}`;
 }
@@ -246,6 +278,8 @@ function sanitizeMessages(messages) {
 export function createAiSdkProvider(deps) {
     const capabilities = { toolCalling: deps.capabilities?.toolCalling ?? true };
     const logger = (deps.logger ?? createNoopLogger()).named("providers.ai-sdk");
+    const maxOutputTokens = deps.maxOutputTokens ?? DEFAULT_MAX_OUTPUT_TOKENS;
+    routeSdkWarningsTo(logger);
     function chat(req, session, signal) {
         const run = async () => {
             const retry = createRetry(signal, logger);
@@ -260,6 +294,7 @@ export function createAiSdkProvider(deps) {
                         messages: sanitizeMessages(req.messages),
                         tools: req.tools,
                         toolChoice: req.toolChoice ?? "auto",
+                        maxOutputTokens,
                         stopWhen: [],
                         // The harness envelope owns retries; leaving the SDK default in
                         // place would multiply attempts (10 outer × 3 inner).
@@ -298,6 +333,7 @@ export function createAiSdkProvider(deps) {
                     messages: sanitizeMessages(req.messages),
                     tools: req.tools,
                     toolChoice: req.toolChoice ?? "auto",
+                    maxOutputTokens,
                     stopWhen: [],
                     maxRetries: 0,
                     headers,
@@ -372,6 +408,7 @@ export function createConfiguredAiSdkProvider(deps) {
             resolveBilling: deps.resolveBilling,
             capabilities: config.capabilities,
             logger: deps.logger,
+            ...(config.maxOutputTokens !== undefined ? { maxOutputTokens: config.maxOutputTokens } : {}),
         });
     }
     const provider = createOpenAICompatible({
@@ -385,5 +422,6 @@ export function createConfiguredAiSdkProvider(deps) {
         resolveBilling: deps.resolveBilling,
         capabilities: config.capabilities,
         logger: deps.logger,
+        ...(config.maxOutputTokens !== undefined ? { maxOutputTokens: config.maxOutputTokens } : {}),
     });
 }

package/dist/providers/anthropic.d.ts

@@ -14,6 +14,8 @@ export interface AnthropicProviderDeps {
     readonly resolveBilling: ResolveBilling;
     readonly fetch?: FetchLike;
     readonly logger?: Logger;
+    /** Output-token ceiling per request. Defaults to `DEFAULT_MAX_OUTPUT_TOKENS`. */
+    readonly maxOutputTokens?: number;
 }
 /**
  * Convenience over the `anthropic` arm of the public `AiSdkProviderConfig`

package/dist/providers/anthropic.js

@@ -22,6 +22,7 @@ export function createAnthropicProvider(deps) {
             model: deps.model,
             fetch: deps.fetch,
             capabilities: { toolCalling: true },
+            ...(deps.maxOutputTokens !== undefined ? { maxOutputTokens: deps.maxOutputTokens } : {}),
         },
     });
 }

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@inflexa-ai/harness",
-    "version": "0.14.1",
+    "version": "0.14.2",
     "description": "Host-agnostic bioinformatics agent harness \u2014 hand-rolled agent loop, DBOS-durable workflows, sandboxed compute.",
     "license": "Apache-2.0",
     "type": "module",
@@ -44,7 +44,7 @@
         "typecheck": "tsc --noEmit"
     },
     "dependencies": {
-        "@ai-sdk/anthropic": "^4.0.15",
+        "@ai-sdk/anthropic": "^4.0.20",
         "@ai-sdk/openai-compatible": "^3.0.10",
         "@ai-sdk/provider": "^4.0.3",
         "@ai-sdk/provider-utils": "5.0.10",