@oh-my-pi/pi-ai 15.10.1 → 15.10.3

package/CHANGELOG.md

@@ -2,6 +2,33 @@
 
 ## [Unreleased]
 
+## [15.10.3] - 2026-06-08
+
+### Removed
+
+- Removed the synthetic `<turn-aborted>` developer guidance note that `transformMessages` injected after an aborted/errored assistant turn (and its `turn-aborted-guidance.md` prompt). The per-call synthetic `"aborted"` tool results already tell the model the turn's tools were terminated, so the extra "verify current state before retrying" note was redundant — and it biased the model toward second-guessing a deliberate user interrupt when the turn was resumed.
+- Removed the legacy Anthropic first-user-message skip for `<system-reminder>` blocks now that synthetic reminders no longer travel as user messages.
+
+
+## [15.10.2] - 2026-06-08
+### Added
+
+- Added support for `impersonated_service_account` Application Default Credentials (ADC) in Vertex AI to enable chained impersonation without failing via 401 `invalid_client`.
+- Added `AuthStorage.getCredentialOrigin(provider)` (returning a structured `CredentialOrigin` / `CredentialOriginKind`) and `getEnvApiKeyName(provider)`, so callers can render where a provider's auth comes from — runtime override, config, stored OAuth/api-key, env var (with the backing variable name), or fallback resolver — without parsing the prose of `describeCredentialSource`.
+
+### Changed
+
+- Changed `onSseEvent` recording for OpenAI Responses, Azure OpenAI Responses, OpenAI Completions, and Anthropic stream providers to emit reconstructed SSE events from decoded SDK stream items instead of wrapping raw fetch responses
+- Changed OpenAI Completions SSE diagnostics to include `event: "chat.completion.chunk"` in `onSseEvent` records for chunked responses
+- Changed the default Anthropic model in `DEFAULT_MODEL_PER_PROVIDER` from `claude-sonnet-4-6` to `claude-opus-4-6`, so sessions that fall back to the provider default (no configured `default` role, no `--model`, no restored session) now start on Claude Opus 4.6.
+
+### Fixed
+
+- Fixed duplicate upstream `tool_call_id` values collapsing distinct tool calls during message transformation, preserving one call/result pairing per emitted tool call before provider replay and keeping generated duplicate IDs distinct after OpenAI/Mistral wire-length caps. ([#2055](https://github.com/can1357/oh-my-pi/issues/2055))
+- Fixed the Anthropic provider retrying persistent account usage/quota limits (e.g. `429 "This request would exceed your account's rate limit"`, `usage_limit_reached`) as if they were transient. Because the error text contains "rate limit", `isProviderRetryableError` matched it and the stream retry loop looped through its 2s/4s/8s backoff (then the `streamSimple` a/b/c policy re-minted the credential and ran the whole thing again) before surfacing the failure — even though the server's `retry-after` parked the account for minutes-to-hours. These errors are now recognized via `isUsageLimitError` and surfaced immediately to the credential-rotation layer, so e.g. `omp dry-balance --bench` reports a rate-limited account as failed at once instead of appearing to hang.
+- Fixed MiniMax-compatible OpenAI-completions hosts losing tool-call argument content when `function.arguments` is streamed as an object across more than one delta. The accumulator added in #1776 wrote `block.partialArgs = rawArgs` per chunk, so every chunk but the last was overwritten — for an `edit` call this surfaced as a tail-slice of the patch text being applied (e.g. a single-line `replace 91..91:` body extending the deletion across the surrounding rows). Chunks are now shallow-merged; for shared string keys, `startsWith` distinguishes cumulative restatements (take the latest) from per-chunk-delta fragments (concatenate). Per-chunk `toolcall_delta` emission for the object branch is suppressed (the previous code emitted `JSON.stringify(rawArgs)` per chunk, which fed downstream concat consumers — `packages/agent/src/proxy.ts`, `openai-chat-server`, `openai-responses-server`, `anthropic-messages-server` — an invalid sequence like `{"input":"a"}{"input":"b"}`); the merged object is flushed instead as a single concat-safe delta in `finishToolCallBlock` before `toolcall_end`, so accumulators reconstruct the args correctly. The single-chunk shape covered by the existing #1776 regression test stays correct end-to-end. ([#2080](https://github.com/can1357/oh-my-pi/issues/2080))
+- Fixed the OpenAI Responses compatibility server misrouting late `toolcall_delta` events for earlier parallel tool calls after a later `toolcall_start`. The encoder now keeps OpenFunctionCall state by content index, allocates output indexes at item start, and closes each tool item by its own `toolcall_end`, preserving deferred MiniMax object-argument flushes for the matching call. ([#2080](https://github.com/can1357/oh-my-pi/issues/2080))
+
 ## [15.10.1] - 2026-06-07
 
 ### Breaking Changes
@@ -3007,4 +3034,4 @@ _Dedicated to Peter's shoulder ([@steipete](https://twitter.com/steipete))_
 
 ## [0.9.4] - 2025-11-26
 
-Initial release with multi-provider LLM support.
+Initial release with multi-provider LLM support.

package/dist/types/auth-storage.d.ts

@@ -22,6 +22,21 @@ export type OAuthCredential = {
 export type AuthCredential = ApiKeyCredential | OAuthCredential;
 export type AuthCredentialEntry = AuthCredential | AuthCredential[];
 export type AuthStorageData = Record<string, AuthCredentialEntry>;
+/**
+ * Cascade leg that supplies a provider's active credential, highest precedence
+ * first — mirrors {@link AuthStorage.getApiKey}'s resolution order.
+ */
+export type CredentialOriginKind = "runtime" | "config" | "oauth" | "api_key" | "env" | "fallback";
+/**
+ * Structured provenance for a provider's auth, for UI that needs a machine
+ * tag (the `/login` provider list) rather than the prose of
+ * {@link AuthStorage.describeCredentialSource}.
+ */
+export interface CredentialOrigin {
+    kind: CredentialOriginKind;
+    /** Env var name when `kind === "env"` and a single named variable backs it. */
+    envVar?: string;
+}
 /**
  * Serialized representation of AuthStorage for passing to subagent workers.
  * Contains only the essential credential data, not runtime state.
@@ -525,6 +540,15 @@ export declare class AuthStorage {
      * silently satisfies xai-oauth and routes around `providers.xai.baseUrl`.
      */
     hasNonEnvCredential(provider: string): boolean;
+    /**
+     * Classify where a provider's auth comes from, following the same precedence
+     * as {@link AuthStorage.getApiKey}: runtime override → config override →
+     * stored credential (api_key before oauth, matching getApiKey) → env var →
+     * fallback resolver. Returns undefined when no auth is configured.
+     *
+     * Compact, structured counterpart to {@link describeCredentialSource}.
+     */
+    getCredentialOrigin(provider: string): CredentialOrigin | undefined;
     /**
      * Check if OAuth credentials are configured for a provider.
      */

package/dist/types/providers/__tests__/google-auth.test.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/types/providers/anthropic.d.ts

@@ -117,7 +117,6 @@ export type AnthropicClientOptionsArgs = {
     hasTools?: boolean;
     thinkingEnabled?: boolean;
     thinkingDisplay?: AnthropicThinkingDisplay;
-    onSseEvent?: AnthropicOptions["onSseEvent"];
     fetch?: FetchImpl;
     claudeCodeSessionId?: string;
 };

package/dist/types/providers/transform-messages.d.ts

@@ -7,6 +7,5 @@ import type { Api, AssistantMessage, Message, Model } from "../types";
  * For aborted/errored turns, this function:
  * - Preserves tool call structure (unlike converting to text summaries)
  * - Injects synthetic "aborted" tool results
- * - Adds a <turn-aborted> guidance marker for the model
  */
-export declare function transformMessages<TApi extends Api>(messages: Message[], model: Model<TApi>, normalizeToolCallId?: (id: string, model: Model<TApi>, source: AssistantMessage) => string): Message[];
+export declare function transformMessages<TApi extends Api>(messages: Message[], model: Model<TApi>, normalizeToolCallId?: (id: string, model: Model<TApi>, source: AssistantMessage) => string, maxNormalizedToolCallIdLength?: number, duplicateToolCallIdSuffixPrefix?: string): Message[];

package/dist/types/stream.d.ts

@@ -12,6 +12,14 @@ import { AssistantMessageEventStream } from "./utils/event-stream";
  * Checks Bun.env, then cwd/.env, then ~/.env.
  */
 export declare function getEnvApiKey(provider: string): string | undefined;
+/**
+ * Name of the environment variable that backs `getEnvApiKey` for a provider,
+ * when that provider maps to a single named variable (e.g. `github-copilot` →
+ * `COPILOT_GITHUB_TOKEN`). Returns undefined for providers whose env fallback
+ * is computed (multi-var pickers, Vertex ADC / Bedrock probes, …) since no
+ * single variable name describes the source.
+ */
+export declare function getEnvApiKeyName(provider: string): string | undefined;
 /**
  * Enumerate every provider that has an env-var fallback for `getEnvApiKey`.
  * Used by `omp auth-broker migrate --include-env` to discover env-sourced keys

package/dist/types/utils/sse-debug.d.ts

@@ -1,10 +1,5 @@
 import type { ServerSentEvent } from "@oh-my-pi/pi-utils";
 import type { RawSseEvent } from "../types";
-type FetchFunction = (input: string | URL | Request, init?: RequestInit) => Promise<Response>;
-type FetchWithPreconnect = FetchFunction & {
-    preconnect?: typeof fetch.preconnect;
-};
 type RawSseObserver = (event: RawSseEvent) => void;
 export declare function notifyRawSseEvent(observer: RawSseObserver | undefined, event: ServerSentEvent | RawSseEvent): void;
-export declare function wrapFetchForSseDebug(fetchImpl: FetchWithPreconnect, observer: RawSseObserver | undefined): FetchWithPreconnect;
 export {};

package/dist/types/utils/stream-markup-healing.d.ts

@@ -75,6 +75,8 @@ export declare class StreamMarkupHealing {
 export declare function modelMayLeakKimiToolCalls(provider: string, modelId: string): boolean;
 /** Cheap model/provider gate for DeepSeek DSML envelope leaks. */
 export declare function modelMayLeakDsmlToolCalls(provider: string, modelId: string): boolean;
+/** Cheap model/provider gate for MiniMax plain thinking tag leaks. */
+export declare function modelMayLeakThinkingTags(provider: string, modelId: string): boolean;
 export declare function getStreamMarkupHealingPattern(provider: string, modelId: string, options?: {
     readonly parseThinkingTags?: boolean;
 }): StreamMarkupHealingPattern | undefined;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "15.10.1",
+	"version": "15.10.3",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -39,7 +39,7 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "15.10.1",
+		"@oh-my-pi/pi-utils": "15.10.3",
 		"openai": "^6.39.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/auth-storage.ts

@@ -13,7 +13,7 @@ import * as path from "node:path";
 import { getAgentDbPath, logger } from "@oh-my-pi/pi-utils";
 import type { ApiKeyResolver } from "./auth-retry";
 import { isUsageLimitError } from "./rate-limit-utils";
-import { getEnvApiKey } from "./stream";
+import { getEnvApiKey, getEnvApiKeyName } from "./stream";
 import type { Provider } from "./types";
 import type {
 	CredentialRankingStrategy,
@@ -59,6 +59,23 @@ export type AuthCredentialEntry = AuthCredential | AuthCredential[];
 
 export type AuthStorageData = Record<string, AuthCredentialEntry>;
 
+/**
+ * Cascade leg that supplies a provider's active credential, highest precedence
+ * first — mirrors {@link AuthStorage.getApiKey}'s resolution order.
+ */
+export type CredentialOriginKind = "runtime" | "config" | "oauth" | "api_key" | "env" | "fallback";
+
+/**
+ * Structured provenance for a provider's auth, for UI that needs a machine
+ * tag (the `/login` provider list) rather than the prose of
+ * {@link AuthStorage.describeCredentialSource}.
+ */
+export interface CredentialOrigin {
+	kind: CredentialOriginKind;
+	/** Env var name when `kind === "env"` and a single named variable backs it. */
+	envVar?: string;
+}
+
 /**
  * Serialized representation of AuthStorage for passing to subagent workers.
  * Contains only the essential credential data, not runtime state.
@@ -1430,6 +1447,26 @@ export class AuthStorage {
 		return false;
 	}
 
+	/**
+	 * Classify where a provider's auth comes from, following the same precedence
+	 * as {@link AuthStorage.getApiKey}: runtime override → config override →
+	 * stored credential (api_key before oauth, matching getApiKey) → env var →
+	 * fallback resolver. Returns undefined when no auth is configured.
+	 *
+	 * Compact, structured counterpart to {@link describeCredentialSource}.
+	 */
+	getCredentialOrigin(provider: string): CredentialOrigin | undefined {
+		if (this.#runtimeOverrides.has(provider)) return { kind: "runtime" };
+		if (this.#configOverrides.has(provider)) return { kind: "config" };
+		const stored = this.#getCredentialsForProvider(provider);
+		if (stored.length > 0) {
+			return { kind: stored.some(credential => credential.type === "api_key") ? "api_key" : "oauth" };
+		}
+		if (getEnvApiKey(provider)) return { kind: "env", envVar: getEnvApiKeyName(provider) };
+		if (this.#fallbackResolver?.(provider)) return { kind: "fallback" };
+		return undefined;
+	}
+
 	/**
 	 * Check if OAuth credentials are configured for a provider.
 	 */

package/src/models.json

@@ -13270,7 +13270,7 @@
 		},
 		"minimax/minimax-m3": {
 			"id": "minimax/minimax-m3",
-			"name": "MiniMax: MiniMax M3 (new)",
+			"name": "MiniMax: MiniMax M3",
 			"api": "openai-completions",
 			"provider": "kilo",
 			"baseUrl": "https://api.kilo.ai/api/gateway",
@@ -41929,8 +41929,8 @@
 				"image"
 			],
 			"cost": {
-				"input": 0.04,
-				"output": 0.13,
+				"input": 0.049999999999999996,
+				"output": 0.15,
 				"cacheRead": 0,
 				"cacheWrite": 0
 			},
@@ -42490,7 +42490,7 @@
 				"image"
 			],
 			"cost": {
-				"input": 0.08,
+				"input": 0.09999999999999999,
 				"output": 0.3,
 				"cacheRead": 0,
 				"cacheWrite": 0
@@ -43439,7 +43439,7 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.09999999999999999,
+				"input": 0.39999999999999997,
 				"output": 0.39999999999999997,
 				"cacheRead": 0,
 				"cacheWrite": 0
@@ -45516,7 +45516,7 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.071,
+				"input": 0.09,
 				"output": 0.09999999999999999,
 				"cacheRead": 0,
 				"cacheWrite": 0
@@ -45564,8 +45564,8 @@
 				"text"
 			],
 			"cost": {
-				"input": 0.09,
-				"output": 0.44999999999999996,
+				"input": 0.12,
+				"output": 0.5,
 				"cacheRead": 0,
 				"cacheWrite": 0
 			},
@@ -46226,13 +46226,13 @@
 				"image"
 			],
 			"cost": {
-				"input": 0.04,
+				"input": 0.09999999999999999,
 				"output": 0.15,
 				"cacheRead": 0,
 				"cacheWrite": 0
 			},
 			"contextWindow": 262144,
-			"maxTokens": 81920,
+			"maxTokens": 262144,
 			"thinking": {
 				"mode": "effort",
 				"minLevel": "minimal",
@@ -49976,7 +49976,7 @@
 				"cacheRead": 0,
 				"cacheWrite": 0
 			},
-			"contextWindow": 222222,
+			"contextWindow": 256000,
 			"maxTokens": 8888,
 			"compat": {
 				"supportsUsageInStreaming": false

package/src/provider-models/descriptors.ts

@@ -132,7 +132,7 @@ function catalogDescriptor(
  * openai-codex) are handled separately because they require different config shapes.
  */
 export const PROVIDER_DESCRIPTORS: readonly ProviderDescriptor[] = [
-	descriptor("anthropic", "claude-sonnet-4-6", config => anthropicModelManagerOptions(config)),
+	descriptor("anthropic", "claude-opus-4-6", config => anthropicModelManagerOptions(config)),
 	catalogDescriptor(
 		"alibaba-coding-plan",
 		"qwen3.5-plus",

package/src/providers/__tests__/google-auth.test.ts

@@ -0,0 +1,144 @@
+import { afterEach, beforeEach, describe, expect, it } from "bun:test";
+import { Buffer } from "node:buffer";
+import * as fs from "node:fs/promises";
+import * as os from "node:os";
+import * as path from "node:path";
+import type { FetchImpl } from "../../types";
+import { __resetVertexTokenCache, getVertexAccessToken } from "../google-auth";
+
+const CLOUD_PLATFORM_SCOPE = "https://www.googleapis.com/auth/cloud-platform";
+const JWT_BEARER_GRANT = "urn:ietf:params:oauth:grant-type:jwt-bearer";
+
+/** Generate a real RS256 private key so signJwtRs256 / pemToPkcs8 run for real. */
+async function generateServiceAccountPem(): Promise<string> {
+	const keyPair = (await globalThis.crypto.subtle.generateKey(
+		{ name: "RSASSA-PKCS1-v1_5", modulusLength: 2048, publicExponent: new Uint8Array([1, 0, 1]), hash: "SHA-256" },
+		true,
+		["sign", "verify"],
+	)) as CryptoKeyPair;
+	const pkcs8 = new Uint8Array(await globalThis.crypto.subtle.exportKey("pkcs8", keyPair.privateKey));
+	const body = (
+		Buffer.from(pkcs8)
+			.toString("base64")
+			.match(/.{1,64}/g) ?? []
+	).join("\n");
+	return `-----BEGIN PRIVATE KEY-----\n${body}\n-----END PRIVATE KEY-----\n`;
+}
+
+function urlOf(input: string | URL | Request): string {
+	if (typeof input === "string") return input;
+	if (input instanceof URL) return input.toString();
+	return input.url;
+}
+
+describe("getVertexAccessToken impersonated_service_account ADC", () => {
+	let tmpDir: string;
+	let originalGac: string | undefined;
+
+	beforeEach(async () => {
+		__resetVertexTokenCache();
+		originalGac = Bun.env.GOOGLE_APPLICATION_CREDENTIALS;
+		tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), "omp-vertex-adc-"));
+	});
+
+	afterEach(async () => {
+		__resetVertexTokenCache();
+		if (originalGac === undefined) delete Bun.env.GOOGLE_APPLICATION_CREDENTIALS;
+		else Bun.env.GOOGLE_APPLICATION_CREDENTIALS = originalGac;
+		await fs.rm(tmpDir, { recursive: true, force: true });
+	});
+
+	it("rejects a malformed service_account_impersonation_url before any network call", async () => {
+		const adcPath = path.join(tmpDir, "impersonated-bad-url.json");
+		await Bun.write(
+			adcPath,
+			JSON.stringify({
+				type: "impersonated_service_account",
+				// Missing the trailing ":generateAccessToken" the principal parser requires.
+				service_account_impersonation_url:
+					"https://iamcredentials.googleapis.com/v1/projects/-/serviceAccounts/target@project.iam.gserviceaccount.com",
+				source_credentials: {
+					type: "authorized_user",
+					client_id: "client-id",
+					client_secret: "client-secret",
+					refresh_token: "refresh-token",
+				},
+			}),
+		);
+		Bun.env.GOOGLE_APPLICATION_CREDENTIALS = adcPath;
+
+		const calls: string[] = [];
+		const fetchImpl: FetchImpl = async input => {
+			calls.push(urlOf(input));
+			return new Response("{}");
+		};
+
+		// The principal is parsed before the source exchange, so a bad URL must fail
+		// up front rather than after burning a source-token round trip.
+		await expect(getVertexAccessToken({ fetch: fetchImpl })).rejects.toBeInstanceOf(RangeError);
+		expect(calls).toEqual([]);
+	});
+
+	it("signs an RS256 JWT for a service_account source and reconstructs the IAM URL", async () => {
+		const pem = await generateServiceAccountPem();
+		const adcPath = path.join(tmpDir, "impersonated-sa.json");
+		await Bun.write(
+			adcPath,
+			JSON.stringify({
+				type: "impersonated_service_account",
+				// Non-canonical project segment proves the request URL is rebuilt, not echoed.
+				service_account_impersonation_url:
+					"https://iamcredentials.googleapis.com/v1/projects/explicit-proj/serviceAccounts/target@project.iam.gserviceaccount.com:generateAccessToken",
+				source_credentials: {
+					type: "service_account",
+					client_email: "source@project.iam.gserviceaccount.com",
+					private_key: pem,
+					private_key_id: "key-1",
+				},
+				// delegates intentionally omitted — the IAM body must default to [].
+			}),
+		);
+		Bun.env.GOOGLE_APPLICATION_CREDENTIALS = adcPath;
+
+		const calls: { url: string; init?: RequestInit }[] = [];
+		const fetchImpl: FetchImpl = async (input, init) => {
+			const url = urlOf(input);
+			calls.push({ url, init });
+			if (url === "https://oauth2.googleapis.com/token") {
+				return new Response(JSON.stringify({ access_token: "sa-source-token", expires_in: 3600 }));
+			}
+			if (url.startsWith("https://iamcredentials.googleapis.com/")) {
+				return new Response(
+					JSON.stringify({
+						accessToken: "impersonated-token",
+						expireTime: new Date(Date.now() + 3_600_000).toISOString(),
+					}),
+				);
+			}
+			return new Response("unexpected", { status: 404 });
+		};
+
+		const token = await getVertexAccessToken({ fetch: fetchImpl });
+		expect(token).toBe("impersonated-token");
+
+		// Source JWT exchange happens first, then the impersonation exchange.
+		expect(calls.map(c => c.url)).toEqual([
+			"https://oauth2.googleapis.com/token",
+			"https://iamcredentials.googleapis.com/v1/projects/-/serviceAccounts/target@project.iam.gserviceaccount.com:generateAccessToken",
+		]);
+
+		// Source credential is exchanged via a signed JWT bearer assertion, not a refresh grant.
+		const sourceBody = new URLSearchParams(String(calls[0].init?.body));
+		expect(sourceBody.get("grant_type")).toBe(JWT_BEARER_GRANT);
+		expect((sourceBody.get("assertion") ?? "").split(".")).toHaveLength(3);
+
+		// The IAM call carries the source-derived bearer token and defaults delegates to [].
+		const iamHeaders = calls[1].init?.headers as Record<string, string>;
+		expect(iamHeaders.Authorization).toBe("Bearer sa-source-token");
+		expect(JSON.parse(String(calls[1].init?.body))).toEqual({
+			delegates: [],
+			scope: [CLOUD_PLATFORM_SCOPE],
+			lifetime: "3600s",
+		});
+	});
+});