code-yeongyu · eddieparc · Sep 2, 2026 · Sep 2, 2026 · Sep 2, 2026 · Sep 2, 2026
diff --git a/packages/ai/CHANGELOG.md b/packages/ai/CHANGELOG.md
@@ -10,6 +10,9 @@
 
 ### Fixed
 
+- Login no longer double-appends a provider-owned OAuth account pool as a fake `login-N` slot copied from the flat compatibility sentinel.
+- Claude Agent SDK `Lock file is already being held` is classified as a transient retryable error instead of an unknown/terminal failure.
+
 ### Removed
 
 ## [2026.9.2] - 2026-09-02

diff --git a/packages/ai/src/auth/pool/slots.ts b/packages/ai/src/auth/pool/slots.ts
@@ -146,6 +146,9 @@ function nextLoginSlotName(credential: PooledCredential): string {
  * user's stored bytes change until a second credential actually exists.
  */
 export function appendLoginSlot(current: PooledCredential | undefined, flat: Credential): Credential {
+	if ("accounts" in flat && Array.isArray(flat.accounts) && flat.accounts.length > 0) {
+		return flat;
+	}
 	if (!current || !Array.isArray(current.accounts) || current.accounts.length === 0) {
 		return flat;
 	}

diff --git a/packages/ai/src/changes.md b/packages/ai/src/changes.md
@@ -1,3 +1,41 @@
+## Classify Claude SDK session lock contention as retryable (2026-09-02)
+
+### What changed
+
+- `packages/ai/src/utils/retry.ts`: `RETRYABLE_PROVIDER_ERROR_PATTERN` matches `Lock file is already being held`.
+- `packages/ai/test/retry.test.ts`: pins that wording as a retryable assistant error.
+
+### Why
+
+- Claude Agent SDK session resume/stream hits proper-lockfile while a previous subprocess still holds `session.json`. The failure is local and transient; treating it as unknown/terminal made the coding-agent hard-error fallback hop providers.
+
+### Why an extension could not handle it
+
+- Retry classification lives in the shared `pi-ai` regexes used by every caller of `isRetryableAssistantError`.
+
+### Expected merge conflict zones
+
+- LOW: `RETRYABLE_PROVIDER_ERROR_PATTERN` in `retry.ts`.
+
+## Preserve provider-owned credential pools during login (2026-09-02)
+
+### What changed
+
+- `packages/ai/src/auth/pool/slots.ts`: `appendLoginSlot` now accepts an OAuth credential whose provider already returned a populated `accounts` pool as the complete post-login credential instead of appending its flat compatibility fields as another generated slot.
+- `packages/ai/test/credential-pool-resolve-slot.test.ts`: covers a provider-owned named account pool and still appends unnamed flat credentials as `login-N`.
+
+### Why
+
+- The shared login layer automatically appends ordinary flat credentials. The Claude SDK OAuth provider already returns its current pool plus the newly named account, so applying the generic append a second time stored the provider's managed top-level sentinel as a fake `login-2` account.
+
+### Why an extension could not handle it
+
+- The double append happens after the provider login returns, inside the shared credential-pool write path. Providers cannot prevent the runtime from reinterpreting their completed pool as a flat credential.
+
+### Expected merge conflict zones
+
+- LOW: `auth/pool/slots.ts` at the start of `appendLoginSlot`.
+
 ## Cursor conversation cache eviction cannot break a live request (2026-08-31)
 
 ### What changed

diff --git a/packages/ai/src/utils/retry.ts b/packages/ai/src/utils/retry.ts
@@ -130,6 +130,11 @@ const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
 
 	// gRPC based providers (e.g. NVIDIA NIM)
 	"ResourceExhausted",
+
+	// Claude Agent SDK session.json lock contention. A second stream/resume
+	// hits proper-lockfile while the previous subprocess still holds the file.
+	// Same-process retry recovers; hopping providers cannot release that lock.
+	"Lock file is already being held",
 ]);
 
 /**

diff --git a/packages/ai/test/credential-pool-resolve-slot.test.ts b/packages/ai/test/credential-pool-resolve-slot.test.ts
@@ -1,7 +1,7 @@
 import { describe, expect, test } from "vitest";
 import { InMemoryCredentialStore } from "../src/auth/credential-store.ts";
 import { envApiKeyAuth } from "../src/auth/helpers.ts";
-import { listSlots, type PooledCredential } from "../src/auth/pool/slots.ts";
+import { appendLoginSlot, listSlots, type PooledCredential } from "../src/auth/pool/slots.ts";
 import { resolveProviderAuth } from "../src/auth/resolve.ts";
 import type { OAuthCredential } from "../src/auth/types.ts";
 import { createProvider, type Provider } from "../src/models.ts";
@@ -63,6 +63,35 @@ function oauthProvider(refreshed: (credential: OAuthCredential) => OAuthCredenti
 }
 
 describe("slot-scoped auth resolution", () => {
+	test("login preserves a provider-owned pooled credential instead of double-appending its flat sentinel", () => {
+		const current = pooledOAuthEntry();
+		const providerOwned: PooledCredential = {
+			...current,
+			accounts: [
+				...(current.accounts ?? []),
+				{ name: "work", access: "named-access", refresh: "r-named", expires: FUTURE, source: "login" },
+			],
+		};
+
+		expect(appendLoginSlot(current, providerOwned)).toEqual(providerOwned);
+	});
+
+	test("unnamed flat oauth still appends as login-N", () => {
+		const current = pooledOAuthEntry();
+		const next = appendLoginSlot(current, {
+			type: "oauth",
+			access: "new-access",
+			refresh: "r-new",
+			expires: FUTURE,
+		}) as PooledCredential;
+
+		expect(listSlots(next).map((slot) => slot.name)).toEqual(["default", "alt", "login-2"]);
+		expect(listSlots(next).find((slot) => slot.name === "login-2")).toMatchObject({
+			access: "new-access",
+			refresh: "r-new",
+		});
+	});
+
 	test("slotName resolves the named api_key slot instead of the flat projection", async () => {
 		const store = new InMemoryCredentialStore();
 		await store.modify("slottest", async () => pooledApiKeyEntry());

diff --git a/packages/ai/test/retry.test.ts b/packages/ai/test/retry.test.ts
@@ -242,6 +242,17 @@ describe("provider retry classification", () => {
 		).toBe(true);
 	});
 
+	it("matches Claude Agent SDK session lock contention", () => {
+		expect(
+			isRetryableAssistantError(
+				fauxAssistantMessage("", {
+					stopReason: "error",
+					errorMessage: "Lock file is already being held",
+				}),
+			),
+		).toBe(true);
+	});
+
 	it("matches upstream request buffer exhaustion wording", () => {
 		expect(
 			isRetryableAssistantError(

diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
@@ -6,6 +6,13 @@
 
 ### Fixed
 
+- Claude SDK OAuth readiness now treats a rotation-selected concrete OAuth slot as configured, so a second login no longer fails every request with `Provider is not configured: claude-sdk-oauth`.
+- `Provider is not configured:` is no longer a hard-error model fallback, so an auth miss on Claude SDK OAuth does not eject the turn onto another provider.
+- OAuth login no longer paints two live `>` prompts when the browser callback finishes before the paste-code field is submitted.
+- Claude Agent SDK `Lock file is already being held` retries on the same model and no longer hard-error-falls back onto another provider.
+- Claude SDK stream-start timeouts and a bare `invalid_request` remint the same model instead of hopping to an unauthenticated OpenGateway Anthropic route.
+- A persisted Claude SDK binding whose prompt/toolset drifted after a timeout now forks at the last assistant UUID instead of flattening megabytes of transcript.
+
 ### New Features
 
 ### Breaking Changes

diff --git a/packages/coding-agent/src/core/agent-session.ts b/packages/coding-agent/src/core/agent-session.ts
@@ -2354,10 +2354,11 @@ export class AgentSession {
 			const hardErrorFallbackEligible = this._isHardErrorFallbackEligible(msg);
 			const cursorZeroTokenRe = isCursorZeroTokenResourceExhausted(msg);
 			const cursorQuotaRe = isCursorQuotaResourceExhausted(msg, this.model?.contextWindow ?? 0);
+			const claudeSdkSameModelRemint = this._isClaudeSdkSameModelRemintError(msg);
 			const retryCanAdmitProvider =
 				!userAbortSuppressedQueuedContinuation &&
 				this.settingsManager.getRetrySettings().enabled &&
-				(retryableError || hardErrorFallbackEligible || cursorZeroTokenRe || cursorQuotaRe);
+				(retryableError || hardErrorFallbackEligible || cursorZeroTokenRe || cursorQuotaRe || claudeSdkSameModelRemint);
 			let compactedBeforeRetry = false;
 			if (
 				retryCanAdmitProvider &&
@@ -2381,6 +2382,8 @@ export class AgentSession {
 					// failed assistant before provider fallback so replay stays valid.
 					this._retireFailedRetryAssistant(msg);
 					retryOutcome = await this._handleRetryableError(msg, { hardErrorFallback: true });
+				} else if (claudeSdkSameModelRemint) {
+					retryOutcome = await this._handleRetryableError(msg, { sameModelRemint: true });
 				} else if (retryableError) {
 					retryOutcome = await this._handleRetryableError(msg);
 				} else if (hardErrorFallbackEligible) {
@@ -7375,9 +7378,27 @@ export class AgentSession {
 		return true;
 	}
 
+	private _isClaudeSdkSessionLockError(message: AssistantMessage): boolean {
+		return (message.errorMessage ?? "").includes("Lock file is already being held");
+	}
+
+	private _isClaudeSdkInvalidRequestError(message: AssistantMessage): boolean {
+		return message.errorMessage === "invalid_request";
+	}
+
+	private _isClaudeSdkSameModelRemintError(message: AssistantMessage): boolean {
+		return (
+			this._isClaudeSdkSessionLockError(message) ||
+			this._isClaudeSdkInvalidRequestError(message) ||
+			isProviderStreamStallError(message)
+		);
+	}
+
 	private _isHardErrorFallbackEligible(message: AssistantMessage): boolean {
 		return (
 			!message.errorMessage?.startsWith(TURN_RETRY_SUPPRESSION_PREFIX) &&
+			!message.errorMessage?.startsWith("Provider is not configured:") &&
+			!this._isClaudeSdkSameModelRemintError(message) &&
 			message.stopReason === "error" &&
 			!isContextOverflow(message, this.model?.contextWindow ?? 0) &&
 			!this._isCursorPayloadOverflow(message) &&

diff --git a/packages/coding-agent/src/core/changes.md b/packages/coding-agent/src/core/changes.md
@@ -1,5 +1,62 @@
 # changes
 
+## 2026-09-02 - Keep Claude SDK stalls and invalid_request on the same model
+
+### What changed
+
+- `packages/coding-agent/src/core/agent-session.ts`: `Provider stream start timed out after Nms` and a bare `invalid_request` remint the same model. They are not hard-error provider hops.
+- `packages/coding-agent/test/suite/retry-fallback-hard-error.test.ts`: both recover on faux-1 and never apply a fallback chain.
+
+### Why
+
+- After a 1.7MB flatten the SDK timed out, then returned `invalid_request`. Hard-error fallback switched onto `opengateway/anthropic/claude-opus-4-8` which has no key and 401-looped until the goal continuation cap fired.
+
+### Why an extension could not handle it
+
+- Hard-error vs same-model retry is decided in `AgentSession` before extension failover runs.
+
+### Expected merge conflict zones
+
+- LOW: `_isHardErrorFallbackEligible` and the `agent_end` remint branch in `agent-session.ts`.
+
+## 2026-09-02 - Retry Claude SDK session locks on the same model
+
+### What changed
+
+- `packages/coding-agent/src/core/agent-session.ts`: `Lock file is already being held` is same-model remint, not hard-error provider fallback.
+- `packages/coding-agent/test/suite/retry-fallback-hard-error.test.ts`: lock recovers on the original model and never hops after budget exhaustion.
+
+### Why
+
+- A held Claude Agent SDK session lock cannot be released by switching to OpenGateway or another provider. Immediate hard-error fallback produced 401 storms and `resume_initialization_aborted` resends.
+
+### Why an extension could not handle it
+
+- Hard-error vs same-model retry is decided in `AgentSession` before extension failover runs.
+
+### Expected merge conflict zones
+
+- LOW: `_isHardErrorFallbackEligible` and the `agent_end` retry branch in `agent-session.ts`.
+
+## 2026-09-02 - Do not hard-fallback a provider-not-configured auth miss
+
+### What changed
+
+- `packages/coding-agent/src/core/agent-session.ts`: `_isHardErrorFallbackEligible` no longer treats `Provider is not configured:` as a model hard-error that ejects onto another provider.
+- `packages/coding-agent/test/suite/retry-fallback-hard-error.test.ts`: a configured fallback chain stays unused when the current model fails with that auth miss.
+
+### Why
+
+- Auth wiring failures were classified as hard-error and immediately switched `claude-sdk-oauth/claude-opus-5` onto a different provider (for example `opengateway/anthropic/claude-opus-5`) instead of staying on Claude SDK OAuth or its sibling accounts.
+
+### Why an extension could not handle it
+
+- Hard-error fallback eligibility is decided in `AgentSession` before extension failover runs.
+
+### Expected merge conflict zones
+
+- LOW: `_isHardErrorFallbackEligible` in `agent-session.ts`.
+
 ## 2026-08-31 - Session activity contract for host occupancy decisions
 
 ### What changed

diff --git a/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/changes.md b/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/changes.md
@@ -1,5 +1,43 @@
 # claude-sdk-oauth
 
+## 2026-09-02 - Fork persisted bindings on options drift instead of flattening
+
+### What changed
+
+- `session-continuity.ts`: a persisted binding whose account/model/prompt/toolset drifted now forks at the last assistant UUID instead of flattening the whole transcript.
+- `test/claude-sdk-oauth-restored-security.test.ts` and `test/claude-sdk-oauth-continuity-retry-checkpoint.test.ts`: drift after a timeout teardown is a fork, not a 1.7MB flatten.
+
+### Why
+
+- Stream-start timeouts tear down the live registry entry. The next attempt only had the sidecar binding, and `options_changed` (reload / cache-warm / toolset hash) flattened hundreds of messages. That megabyte resend then timed out again and hopped providers.
+
+### Why an extension could not handle it
+
+- Continuity decisions live inside the Claude SDK OAuth resident lane.
+
+### Expected merge conflict zones
+
+- LOW: `decideFromBinding` identity-drift branch in `session-continuity.ts`.
+
+## 2026-09-02 - Preserve selected OAuth slots during provider preflight
+
+### What changed
+
+- `oauth-login.ts`: readiness now recognizes a concrete OAuth credential selected by the shared credential-rotation layer, while still excluding the provider's synthetic managed sentinel.
+- `test/claude-sdk-oauth-login.test.ts`: projected non-sentinel slots pass `check`; projected sentinels do not.
+
+### Why
+
+- With two or more Claude logins, shared credential rotation passes one selected OAuth slot to provider auth resolution. The Claude readiness predicate counted only the parent credential's `accounts` array, so the selected slot appeared empty and the request failed with `Provider is not configured: claude-sdk-oauth` even though both accounts were valid.
+
+### Why an extension could not handle it
+
+- Readiness is this provider's `oauth.check` predicate. An external extension cannot change how a selected slot is counted once rotation has already projected it.
+
+### Expected merge conflict zones
+
+- LOW in `oauth-login.ts` `configuredFor` account counting.
+
 ## 2026-08-21 - Cache provider settings loads by mtime+size to cut lock convoy
 
 ### What changed

diff --git a/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/oauth-login.ts b/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/oauth-login.ts
@@ -109,9 +109,13 @@ export function createOAuthConfig(deps: {
 		environment?: Record<string, string>,
 	): Promise<boolean> => {
 		const storedAccounts = stored?.type === "oauth" && Array.isArray(stored.accounts) ? stored.accounts : [];
+		const selectedStoredAccount =
+			stored?.type === "oauth" &&
+			stored.access !== SENTINEL_OAUTH_FIELDS.access &&
+			stored.refresh !== SENTINEL_OAUTH_FIELDS.refresh;
 		const effectiveEnvironment = environment ?? (await claudeEnvironment(ctx));
 		const environmentTokenCount = Object.values(effectiveEnvironment).filter(Boolean).length;
-		const accountCount = storedAccounts.length + environmentTokenCount;
+		const accountCount = storedAccounts.length + (selectedStoredAccount ? 1 : 0) + environmentTokenCount;
 		const settings = deps.readSettings?.();
 		const lane = settings?.tokenInjection ?? (accountCount > 0 ? "oauth-slots" : "ambient");
 		if (lane === "ambient") {

diff --git a/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/session-continuity.ts b/packages/coding-agent/src/core/extensions/builtin/claude-sdk-oauth/session-continuity.ts
@@ -132,10 +132,24 @@ function retryCheckpointDecision(
 	};
 }
 
+function forkBindingOrFlatten(
+	binding: ContinuityBindingSnapshot,
+	reason: ContinuityReason,
+): ContinuityDecision {
+	if (!binding.lastAssistantUuid) return { kind: "flatten", reason };
+	return {
+		kind: "fork",
+		sdkSessionId: binding.sdkSessionId,
+		atUuid: binding.lastAssistantUuid,
+		from: binding.sentCount,
+		reason,
+	};
+}
+
 function decideFromBinding(input: ContinuityDecisionInput, binding: ContinuityBindingSnapshot): ContinuityDecision {
 	if (!input.transcriptAvailable) return { kind: "flatten", reason: "transcript_missing" };
 	const drift = identityDrift(input, binding);
-	if (drift) return { kind: "flatten", reason: drift };
+	if (drift) return forkBindingOrFlatten(binding, drift);
 	const retry = retryCheckpointDecision(input, binding);
 	if (retry) return retry;
 	if (binding.sentPrefixHash !== undefined) {

diff --git a/packages/coding-agent/src/modes/interactive/changes.md b/packages/coding-agent/src/modes/interactive/changes.md
@@ -1,5 +1,24 @@
 # changes
 
+## 2026-09-02 - Do not paint two live login inputs
+
+### What changed
+
+- `packages/coding-agent/src/modes/interactive/components/login-dialog.ts`: `showManualInput` and `showPrompt` remount the single Input widget instead of adding it twice, so a browser-callback login no longer shows two stacked `>` prompts.
+- `packages/coding-agent/test/suite/regressions/5433-extension-oauth-prompt-input.test.ts`: covers an unsubmitted paste-code prompt followed by the account-name prompt.
+
+### Why
+
+- Anthropic OAuth completes via localhost callback while the paste-code input is still mounted. The name prompt then added the same Input child again, and the TUI painted two live `>` rows.
+
+### Why an extension could not handle it
+
+- Login chrome is the interactive LoginDialogComponent, not an extension surface.
+
+### Expected merge conflict zones
+
+- LOW: `showManualInput` / `showPrompt` in `login-dialog.ts`.
+
 ## 2026-09-01 - Never swallow an interactive quit request
 
 ### What changed