@oh-my-pi/pi-ai 18.3.2 → 18.3.4

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [18.3.4] - 2026-09-27
+
+### Fixed
+
+- Fixed Anthropic OAuth requests capping output at 64k tokens; they now request the model's full ceiling (128k on Opus 5.5), matching Claude Code and API-key requests
+
 ## [18.3.2] - 2026-09-25
 
 ### Fixed

package/dist/types/providers/claude-code-fingerprint.d.ts

@@ -36,5 +36,3 @@ export declare function adoptRequiredClaudeCodeVersion(error: unknown): boolean;
 export declare const claudeToolPrefix: string;
 /** Identity block prepended by Claude Code's CLI runtime. */
 export declare const claudeCodeSystemInstruction = "You are Claude Code, Anthropic's official CLI for Claude.";
-/** Claude Code's per-request output-token ceiling. */
-export declare const CLAUDE_CODE_MAX_OUTPUT_TOKENS = 64000;

package/dist/types/types.d.ts

@@ -27,8 +27,7 @@ export type { StopDetails } from "./providers/anthropic-wire.js";
 export type { AssistantMessageEventStream } from "./utils/event-stream.js";
 /**
  * Ceiling on the output-token count omp requests from any OpenAI-family endpoint
- * (openai-responses, azure/xai responses, and openai-completions). Mirrors
- * Anthropic's {@link CLAUDE_CODE_MAX_OUTPUT_TOKENS}.
+ * (openai-responses, azure/xai responses, and openai-completions).
  *
  * Catalog `maxTokens` frequently reflects a model's context window rather than a
  * given upstream's real per-request output cap. OpenRouter, for instance,

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-ai",
-	"version": "18.3.2",
+	"version": "18.3.4",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"keywords": [
 		"ai",
@@ -155,11 +155,11 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.3.2",
-		"@oh-my-pi/pi-catalog": "18.3.2",
-		"@oh-my-pi/pi-natives": "18.3.2",
-		"@oh-my-pi/pi-utils": "18.3.2",
-		"@oh-my-pi/pi-wire": "18.3.2"
+		"@oh-my-pi/omptype": "18.3.4",
+		"@oh-my-pi/pi-catalog": "18.3.4",
+		"@oh-my-pi/pi-natives": "18.3.4",
+		"@oh-my-pi/pi-utils": "18.3.4",
+		"@oh-my-pi/pi-wire": "18.3.4"
 	},
 	"devDependencies": {
 		"@types/bun": "^1.3.14"

package/src/providers/anthropic.ts

@@ -116,7 +116,6 @@ import {
 	type TextBlockParam,
 } from "./anthropic-wire";
 import {
-	CLAUDE_CODE_MAX_OUTPUT_TOKENS,
 	claudeCodeSdkVersion,
 	claudeCodeSystemInstruction,
 	adoptRequiredClaudeCodeVersion,
@@ -4255,6 +4254,9 @@ function extractClaudeCodeFirstUserMessageText(messages: readonly Message[]): st
  */
 const ANTHROPIC_CONTROL = Symbol("anthropicControl");
 
+/** `max_tokens` requested when the catalog has no output ceiling for the model. */
+const UNKNOWN_MODEL_MAX_OUTPUT_TOKENS = 64_000;
+
 /** One control message: tool changes (removals first) and an optional per-message effort. */
 type AnthropicControlSpec = { toolChanges: AnthropicToolChange[]; effort?: AnthropicOutputEffort };
 
@@ -4738,11 +4740,9 @@ function buildParams(
 	}
 	const outputConfig = Object.keys(outputConfigEntries).length ? outputConfigEntries : undefined;
 
-	// Claude Code requests at most 64k output tokens; clamp only OAuth requests,
-	// where the wire fingerprint must match. API-key callers keep the full model
-	// ceiling (e.g. 128k on Opus 4.8).
-	const modelMaxTokens = model.maxTokens ?? CLAUDE_CODE_MAX_OUTPUT_TOKENS;
-	const maxOutputTokens = isOAuthToken ? Math.min(CLAUDE_CODE_MAX_OUTPUT_TOKENS, modelMaxTokens) : modelMaxTokens;
+	// OAuth and API-key requests alike get the full model ceiling; Claude Code
+	// itself requests 128k on Opus 5.5.
+	const maxOutputTokens = model.maxTokens ?? UNKNOWN_MODEL_MAX_OUTPUT_TOKENS;
 
 	// A caller-owned client targets its own endpoint: route body betas by the
 	// client's URL when it exposes one, not the model's routing. Otherwise the
@@ -4773,7 +4773,7 @@ function buildParams(
 		...(systemBlocks && { system: systemBlocks }),
 		...(tools !== undefined && { tools }),
 		...(metadata && { metadata }),
-		max_tokens: Math.min(maxOutputTokens, options?.maxTokens ?? modelMaxTokens),
+		max_tokens: Math.min(maxOutputTokens, options?.maxTokens ?? maxOutputTokens),
 		...(thinking && { thinking }),
 		...(contextManagement && { context_management: contextManagement }),
 		...(compactionRequest && {

package/src/providers/claude-code-fingerprint.ts

@@ -70,5 +70,3 @@ export function adoptRequiredClaudeCodeVersion(error: unknown): boolean {
 export const claudeToolPrefix: string = "_";
 /** Identity block prepended by Claude Code's CLI runtime. */
 export const claudeCodeSystemInstruction = "You are Claude Code, Anthropic's official CLI for Claude.";
-/** Claude Code's per-request output-token ceiling. */
-export const CLAUDE_CODE_MAX_OUTPUT_TOKENS = 64000;

package/src/types.ts

@@ -60,8 +60,7 @@ export type { AssistantMessageEventStream } from "./utils/event-stream";
 
 /**
  * Ceiling on the output-token count omp requests from any OpenAI-family endpoint
- * (openai-responses, azure/xai responses, and openai-completions). Mirrors
- * Anthropic's {@link CLAUDE_CODE_MAX_OUTPUT_TOKENS}.
+ * (openai-responses, azure/xai responses, and openai-completions).
  *
  * Catalog `maxTokens` frequently reflects a model's context window rather than a
  * given upstream's real per-request output cap. OpenRouter, for instance,