@f5-sales-demo/xcsh 20.8.6 → 20.9.0

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@f5-sales-demo/xcsh",
-	"version": "20.8.6",
+	"version": "20.9.0",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://github.com/f5-sales-demo/xcsh",
 	"author": "Can Boluk",
@@ -40,6 +40,7 @@
 		"check:bundle": "bun scripts/check-bundle.ts",
 		"pretest": "bun ../../scripts/check-installed-dependencies.ts",
 		"test": "bun run generate-build-info && bun scripts/bun-test-guarded.ts --max-concurrency 1",
+		"bench:routing-matrix": "bun scripts/bench-routing-matrix.ts",
 		"fix": "biome check --write --unsafe . && bun run format-prompts && bun run generate-docs-index && bun run generate-api-spec-index && bun run generate-build-info",
 		"fmt": "biome format --write . && bun run format-prompts",
 		"format-prompts": "bun scripts/format-prompts.ts",
@@ -61,13 +62,13 @@
 	"dependencies": {
 		"@agentclientprotocol/sdk": "1.3.0",
 		"@mozilla/readability": "^0.6",
-		"@f5-sales-demo/xcsh-stats": "20.8.6",
-		"@f5-sales-demo/pi-agent-core": "20.8.6",
-		"@f5-sales-demo/pi-ai": "20.8.6",
-		"@f5-sales-demo/pi-natives": "20.8.6",
-		"@f5-sales-demo/pi-resource-management": "20.8.6",
-		"@f5-sales-demo/pi-tui": "20.8.6",
-		"@f5-sales-demo/pi-utils": "20.8.6",
+		"@f5-sales-demo/xcsh-stats": "20.9.0",
+		"@f5-sales-demo/pi-agent-core": "20.9.0",
+		"@f5-sales-demo/pi-ai": "20.9.0",
+		"@f5-sales-demo/pi-natives": "20.9.0",
+		"@f5-sales-demo/pi-resource-management": "20.9.0",
+		"@f5-sales-demo/pi-tui": "20.9.0",
+		"@f5-sales-demo/pi-utils": "20.9.0",
 		"@sinclair/typebox": "^0.34",
 		"@xterm/headless": "^6.0",
 		"ajv": "^8.20",

package/src/internal-urls/build-info.generated.ts

@@ -17,17 +17,17 @@ export interface BuildInfo {
 }
 
 export const BUILD_INFO: BuildInfo = {
-	"version": "20.8.6",
-	"commit": "ffac88b3a59886c6e6e579bc266e0a62e3d354d5",
-	"shortCommit": "ffac88b",
+	"version": "20.9.0",
+	"commit": "c120e6e1b6530d1ba068a89905dcbd1df72d7b2c",
+	"shortCommit": "c120e6e",
 	"branch": "main",
-	"tag": "v20.8.6",
-	"commitDate": "2026-08-10T05:40:49Z",
-	"buildDate": "2026-08-10T06:09:48.444Z",
+	"tag": "v20.9.0",
+	"commitDate": "2026-08-10T17:23:39Z",
+	"buildDate": "2026-08-10T17:52:55.552Z",
 	"dirty": true,
 	"prNumber": "",
 	"repoUrl": "https://github.com/f5-sales-demo/xcsh",
 	"repoSlug": "f5-sales-demo/xcsh",
-	"commitUrl": "https://github.com/f5-sales-demo/xcsh/commit/ffac88b3a59886c6e6e579bc266e0a62e3d354d5",
-	"releaseUrl": "https://github.com/f5-sales-demo/xcsh/releases/tag/v20.8.6"
+	"commitUrl": "https://github.com/f5-sales-demo/xcsh/commit/c120e6e1b6530d1ba068a89905dcbd1df72d7b2c",
+	"releaseUrl": "https://github.com/f5-sales-demo/xcsh/releases/tag/v20.9.0"
 };

package/src/routing/classifier.ts

@@ -86,7 +86,12 @@ export async function classifyTaskHybrid(options: HybridClassifierOptions): Prom
 		if (err instanceof Error && err.name === "AbortError") {
 			throw err;
 		}
-		const isSyntax = err instanceof SyntaxError || (err instanceof Error && err.message.includes("malformed"));
+		const isSyntax =
+			err instanceof SyntaxError ||
+			(err instanceof Error &&
+				(err.message.includes("malformed") ||
+					err.message.includes("JSON") ||
+					err.message.includes("Unexpected token")));
 		return {
 			...baseProfile,
 			reasons: [...baseProfile.reasons, isSyntax ? "classifier_fallback_malformed" : "classifier_fallback_timeout"],

package/src/routing/coordinator.ts

@@ -28,6 +28,8 @@ export interface EvaluateTurnOptions {
 	downshiftAfterTurns?: number;
 	getModelContextWindow?: (modelId: string) => number;
 	runRoutingClassifier?: (utilityModel: string, prompt: string) => Promise<string>;
+	signal?: AbortSignal;
+	ttftStartNs?: number;
 }
 
 export class RoutingCoordinator {

package/src/routing/delegation.ts

@@ -41,7 +41,10 @@ export function validateDelegationPlan(plan: ReadOnlyDelegationPlan, maxTasks =
 
 export async function executeReadOnlyDelegationPlan(
 	plan: ReadOnlyDelegationPlan,
-	executor: (subtaskPrompt: string, signal?: AbortSignal) => Promise<{ result: string; tokens: number }>,
+	executor: (
+		subtaskPrompt: string,
+		options?: { signal?: AbortSignal; allowedTools?: (toolName: string) => boolean },
+	) => Promise<{ result: string; tokens: number }>,
 	maxTasks = 3,
 	options?: { signal?: AbortSignal },
 ): Promise<{ results: Array<{ id: string; result: string }>; tokensUsed: number }> {
@@ -65,7 +68,10 @@ export async function executeReadOnlyDelegationPlan(
 			]
 				.filter(Boolean)
 				.join("\n");
-			const { result, tokens } = await executor(promptPayload, options?.signal);
+			const { result, tokens } = await executor(promptPayload, {
+				signal: options?.signal,
+				allowedTools: isDelegationAllowedTool,
+			});
 			return { id: task.id, result, tokens };
 		} catch (err) {
 			return { id: task.id, result: `Failed: ${String(err)}`, tokens: 0 };

package/src/session/agent-session.ts

@@ -43,7 +43,6 @@ import type {
 } from "@f5-sales-demo/pi-ai";
 import {
 	calculateRateLimitBackoffMs,
-	completeSimple,
 	getSupportedEfforts,
 	isContextOverflow,
 	isUsageLimitError,
@@ -445,6 +444,7 @@ export class AgentSession {
 	 * distinguishable from launch configuration (#2459). Defaults to "config": settings, a remembered
 	 * role, or a resumed session are all config, and only --model or a live switch is not.
 	 */
+	#initialModelResolutionSource: ModelResolutionSource = "config";
 	#modelResolutionSource: ModelResolutionSource = "config";
 	#promptTemplates: PromptTemplate[];
 	#slashCommands: FileSlashCommand[];
@@ -596,7 +596,8 @@ export class AgentSession {
 		this.#pythonKernelOwnerId = config.pythonKernelOwnerId ?? `agent-session:${Snowflake.next()}`;
 		this.#scopedModels = config.scopedModels ?? [];
 		this.#thinkingLevel = config.thinkingLevel;
-		this.#modelResolutionSource = config.modelResolutionSource ?? "config";
+		this.#initialModelResolutionSource = config.modelResolutionSource ?? "config";
+		this.#modelResolutionSource = this.#initialModelResolutionSource;
 		this.#promptTemplates = config.promptTemplates ?? [];
 		this.#slashCommands = config.slashCommands ?? [];
 		this.#extensionRunner = config.extensionRunner;
@@ -1348,7 +1349,6 @@ export class AgentSession {
 						}
 
 						if (!outcome.safeToContinue) {
-							this.#routingCoordinator.getStateMachine().setEscalationFloor(targetTier);
 							return;
 						}
 
@@ -1357,12 +1357,11 @@ export class AgentSession {
 						const previousModel = this.model;
 
 						try {
+							await this.setModelRoutingSwitch(targetModel);
 							this.agent.abort();
 
 							this.#routingCoordinator.getStateMachine().setEscalationFloor(targetTier);
 							this.#routingCoordinator.restoreState({ currentTier: targetTier });
-
-							await this.setModelRoutingSwitch(targetModel);
 							const effortMap = this.settings.get("routing.tierEffort") as Record<string, string> | undefined;
 							const { mapTierToEffort } = await import("../routing/effort");
 							const effort = mapTierToEffort(targetTier, effortMap);
@@ -2809,264 +2808,21 @@ export class AgentSession {
 
 		await this.#maybeRestoreRetryFallbackPrimary();
 
-		// Evaluate dynamic model routing if enabled
-		const routingMode = (this.settings.get("routing.mode") as RoutingMode) ?? "off";
-		if (routingMode !== "off" && this.model) {
-			if (this.#activeRetryFallback) {
-				this.#emitSessionEvent({
-					type: "routing_skipped",
-					epochId: `skip-${Date.now()}`,
-					reasons: ["retry_fallback"],
-				} as any).catch(() => {});
+		const delegationOutput = await this.#evaluateAndApplyRouting(
+			expandedText,
+			options?.images ? options.images.length > 0 : false,
+			{ signal: options?.signal },
+		);
+		if (delegationOutput) {
+			expandedText += delegationOutput;
+			const textBlock = userContent.find(c => c.type === "text") as Extract<
+				(typeof userContent)[0],
+				{ type: "text" }
+			>;
+			if (textBlock) {
+				textBlock.text += delegationOutput;
 			} else {
-				const anchorModel = `${this.model.provider}/${this.model.id}`;
-				const availableModels =
-					this.#scopedModels.length > 0
-						? this.#scopedModels.map(sm => `${sm.model.provider}/${sm.model.id}`)
-						: this.#modelRegistry.getAvailable().map(m => `${m.provider}/${m.id}`);
-				const startTime = Date.now();
-				const systemTokens = Math.round((this.systemPrompt?.length ?? 0) / 4);
-				const promptTokens = Math.round(expandedText.length / 4);
-				const toolsStr = JSON.stringify(Array.from(this.#toolRegistry.values()));
-				const toolsTokens = Math.round(toolsStr.length / 4);
-				const reserveTokens = (this.settings.get("compaction.reserveTokens") as number) ?? 0;
-				const contextEstimate = {
-					usedTokens: calculateUsedTokens(this.messages) + systemTokens + promptTokens + toolsTokens,
-					reserveTokens,
-					contextWindow: this.model.contextWindow ?? 128000,
-				};
-				const decision = await this.#routingCoordinator.evaluateTurn({
-					anchorModel,
-					mode: routingMode,
-					prompt: expandedText,
-					hasImages: options?.images && options.images.length > 0,
-					priorRejection: this.#routingCoordinator.getState().escalationFloor !== undefined,
-					availableModels,
-					customPools: validateCustomPools(this.settings.get("routing.pools")),
-					disabledPresets: (this.settings.get("routing.disabledPresets") as readonly string[]) ?? [],
-					familyPolicy: (this.settings.get("routing.familyPolicy") as "sticky" | "configured-mixed") ?? "sticky",
-					profilerMode: (this.settings.get("routing.profiler") as any) ?? "hybrid",
-					contextEstimate,
-					getModelContextWindow: (modelId: string) => {
-						const found = this.#modelRegistry
-							.getAll()
-							.find(m => `${m.provider}/${m.id}` === modelId || m.id === modelId);
-						return found?.contextWindow ?? 128000;
-					},
-					runRoutingClassifier: async (utilityModel: string, promptText: string) => {
-						const resolved = this.#modelRegistry
-							.getAll()
-							.find(m => `${m.provider}/${m.id}` === utilityModel || m.id === utilityModel);
-						if (!resolved) throw new Error("Classifier model not found");
-
-						let targetModel = resolved;
-						if (
-							resolved.provider === "openai" ||
-							(resolved.provider === "litellm" && resolved.id.includes("openai"))
-						) {
-							const internalUrl = this.settings.get("routing.internalOpenAiUrl") as string | undefined;
-							if (internalUrl) targetModel = { ...resolved, baseUrl: internalUrl };
-						} else if (
-							resolved.provider === "anthropic" ||
-							(resolved.provider === "litellm" &&
-								(resolved.id.includes("anthropic") || resolved.id.includes("claude")))
-						) {
-							const internalUrl = this.settings.get("routing.internalAnthropicUrl") as string | undefined;
-							if (internalUrl) targetModel = { ...resolved, baseUrl: internalUrl };
-						}
-
-						const apiKey = await this.#modelRegistry.getApiKey(targetModel, this.sessionId);
-						if (!apiKey) throw new Error("No API key for classifier model");
-
-						const systemInstruction = `You are a routing classifier. Analyze the user's prompt and output a JSON object: { "complexityScore": number, "confidence": number, "delegation"?: { "reason": string, "subtasks": { "id": string, "title": string, "description": string, "targetFilesOrPaths": string[], "desiredTier"?: string }[] } }. Score complexity from 0 to 100 based on required reasoning, context width, and coding effort. Return ONLY valid JSON without markdown blocks.`;
-
-						const res = await completeSimple(
-							targetModel,
-							{
-								systemPrompt: systemInstruction,
-								messages: [
-									{ role: "user", content: [{ type: "text", text: promptText }], timestamp: Date.now() },
-								],
-							},
-							{ apiKey, signal: AbortSignal.timeout(15000) },
-						);
-						const rawText = res.content
-							.filter(c => c.type === "text")
-							.map(c => (c as TextContent).text)
-							.join("");
-
-						const jsonMatch = rawText.match(/\{[\s\S]*\}/);
-						if (jsonMatch) {
-							const parsed = JSON.parse(jsonMatch[0]); // Check validity
-							if (res.usage) {
-								this.settings.getStorage()?.recordModelUsage(`${resolved.provider}/${resolved.id}`);
-								parsed.routingUsage = (res.usage as any).totalTokens ?? (res.usage as any).total ?? 0;
-							}
-							return JSON.stringify(parsed);
-						}
-						throw new Error("Classifier returned malformed JSON");
-					},
-					downshiftAfterTurns: (this.settings.get("routing.downshiftAfterTurns") as number) ?? 2,
-				});
-				const durationMs = Date.now() - startTime;
-
-				this.#emitSessionEvent({
-					type: decision.applied ? "routing_applied" : "routing_decision",
-					epochId: decision.epochId,
-					mode: decision.mode,
-					provider: this.model.provider,
-					poolId: decision.poolId,
-					effectiveTier: decision.effectiveTier,
-					selectedModel: decision.selectedModel,
-					reasons: decision.reasons,
-					delegated: !!decision.delegation,
-					escalated: decision.reasons.some(
-						r => r === "context_capacity_promotion" || r === "escalation_floor_active",
-					),
-					contextTokens: contextEstimate.usedTokens,
-					routingUsage: decision.routingUsage,
-					durationMs,
-					state: this.#routingCoordinator.getState(),
-				} as any).catch(() => {});
-
-				if (decision.selectedModel && decision.applied) {
-					const targetModel = this.#modelRegistry
-						.getAvailable()
-						.find(m => `${m.provider}/${m.id}` === decision.selectedModel || m.id === decision.selectedModel);
-					if (targetModel) {
-						const currentModelId = this.model ? `${this.model.provider}/${this.model.id}` : undefined;
-						const targetModelId = `${targetModel.provider}/${targetModel.id}`;
-						if (currentModelId !== targetModelId) {
-							await this.setModelRoutingSwitch(targetModel);
-						}
-						if (decision.effectiveTier) {
-							const effortMap = this.settings.get("routing.tierEffort") as Record<string, string> | undefined;
-							const { mapTierToEffort } = await import("../routing/effort");
-							const effort = mapTierToEffort(decision.effectiveTier, effortMap);
-							this.setThinkingLevel(effort as any);
-						}
-					}
-				}
-
-				if (
-					decision.applied &&
-					decision.delegation &&
-					decision.delegation.subtasks.length > 1 &&
-					(this.settings.get("routing.delegation") as string) === "read-only"
-				) {
-					let maxTasks = (this.settings.get("routing.delegationMaxTasks") as number) ?? 3;
-					maxTasks = Math.min(Math.max(1, maxTasks), 3);
-
-					const { results, tokensUsed } = await executeReadOnlyDelegationPlan(
-						decision.delegation,
-						async (subtaskPrompt, signal) => {
-							if (signal?.aborted) return { result: "Delegation failed: Aborted", tokens: 0 };
-
-							const { resolveModelPool } = await import("../routing/presets");
-							const pool = resolveModelPool(
-								decision.poolId ?? (this.model ? `${this.model.provider}/${this.model.id}` : ""),
-								this.settings.get("routing.pools") as any,
-								this.settings.get("routing.disabledPresets") as readonly string[],
-								this.settings.get("routing.familyPolicy") as any,
-							);
-
-							let resolvedUtility = this.model;
-							if (pool?.tiers?.utility) {
-								let uId = pool.tiers.utility;
-								if (!uId.includes("/") && pool.provider) {
-									uId = `${pool.provider}/${uId}`;
-								}
-								const found = this.#modelRegistry
-									.getAvailable()
-									.find(m => `${m.provider}/${m.id}` === uId || m.id === uId);
-								if (found) resolvedUtility = found;
-							}
-
-							if (!resolvedUtility) return { result: "Delegation failed: no model.", tokens: 0 };
-
-							const { createAgentSession } = await import("../sdk");
-							const { SessionManager: SDKSessionManager } = await import("./session-manager");
-							const childSettings = await this.settings.cloneForCwd(process.cwd());
-							childSettings.set("routing.delegation", "off");
-
-							// Fix abort race by checking right before creation
-							if (signal?.aborted) return { result: "Delegation failed: Aborted", tokens: 0 };
-
-							const { session: childSession } = await createAgentSession({
-								model: resolvedUtility,
-								sessionManager: SDKSessionManager.inMemory(),
-								settings: childSettings,
-								authStorage: this.#modelRegistry.authStorage,
-								modelRegistry: this.#modelRegistry,
-								toolNames: Array.from(this.#toolRegistry.values())
-									.map(t => t.name)
-									.filter(isDelegationAllowedTool),
-								enableLsp: false,
-								enableMCP: false,
-							});
-
-							try {
-								const onAbort = () => childSession.agent.abort();
-								signal?.addEventListener("abort", onAbort);
-								if (signal?.aborted) {
-									onAbort();
-									return { result: "Delegation failed: Aborted", tokens: 0 };
-								}
-								try {
-									await childSession.prompt(subtaskPrompt);
-									if (signal?.aborted) return { result: "Delegation failed: Aborted", tokens: 0 };
-									const resultStr = childSession
-										.buildDisplaySessionContext()
-										.messages.filter(m => m.role === "assistant")
-										.map(m =>
-											m.content
-												.filter((c: any) => c.type === "text")
-												.map((c: any) => c.text)
-												.join(""),
-										)
-										.join("\n");
-									const tokens = childSession.buildDisplaySessionContext().usedTokens || 0;
-									return { result: resultStr, tokens };
-								} finally {
-									signal?.removeEventListener("abort", onAbort);
-								}
-							} catch (err) {
-								if (signal?.aborted) return { result: "Delegation failed: Aborted", tokens: 0 };
-								return { result: `Delegation failed: ${String(err)}`, tokens: 0 };
-							} finally {
-								await childSession.dispose();
-							}
-						},
-						maxTasks,
-						{ signal: options?.signal },
-					);
-
-					if (results.length > 0) {
-						let resultsString = JSON.stringify(results, null, 2);
-						if (resultsString.length > 8000) {
-							resultsString = `${resultsString.substring(0, 8000)}\n...[truncated]`;
-						}
-						const delegationOutput = `\n\n<delegation_results>\n${resultsString}\n</delegation_results>`;
-						expandedText += delegationOutput;
-						const textBlock = userContent.find(c => c.type === "text") as Extract<
-							(typeof userContent)[0],
-							{ type: "text" }
-						>;
-						if (textBlock) {
-							textBlock.text += delegationOutput;
-						} else {
-							userContent.push({ type: "text", text: delegationOutput });
-						}
-
-						this.#emitSessionEvent({
-							type: "routing_delegated",
-							epochId: decision.epochId,
-							tasks: decision.delegation.subtasks.length,
-							completed: results.length,
-							tokensUsed,
-						} as any).catch(() => {});
-					}
-				}
+				userContent.push({ type: "text", text: delegationOutput });
 			}
 		}
 
@@ -3087,7 +2843,7 @@ export class AgentSession {
 
 	async promptCustomMessage<T = unknown>(
 		message: Pick<CustomMessage<T>, "customType" | "content" | "display" | "details" | "attribution">,
-		options?: Pick<PromptOptions, "streamingBehavior" | "toolChoice">,
+		options?: Pick<PromptOptions, "streamingBehavior" | "toolChoice" | "signal">,
 	): Promise<void> {
 		const textContent =
 			typeof message.content === "string"
@@ -3122,6 +2878,7 @@ export class AgentSession {
 		const delegationOutput = await this.#evaluateAndApplyRouting(
 			textContent,
 			Array.isArray(message.content) && message.content.some((c: any) => c.type === "image"),
+			{ signal: options?.signal },
 		);
 		if (delegationOutput) {
 			if (typeof customMessage.content === "string") {
@@ -3184,6 +2941,7 @@ export class AgentSession {
 			familyPolicy: (this.settings.get("routing.familyPolicy") as "sticky" | "configured-mixed") ?? "sticky",
 			profilerMode: (this.settings.get("routing.profiler") as any) ?? "hybrid",
 			contextEstimate,
+			signal: options?.signal,
 			getModelContextWindow: (modelId: string) => {
 				const m = this.#modelRegistry
 					.getAvailable()
@@ -3218,7 +2976,8 @@ export class AgentSession {
 				if (!needsSwitch) {
 					if (
 						targetModel.provider === "openai" ||
-						(targetModel.provider === "litellm" && targetModel.id.includes("openai"))
+						(targetModel.provider === "litellm" &&
+							(targetModel.id.includes("openai") || /gpt-/.test(targetModel.id)))
 					) {
 						const internalUrl = this.settings.get("routing.internalOpenAiUrl") as string | undefined;
 						if ((internalUrl || undefined) !== this.model.baseUrl) {
@@ -3253,7 +3012,9 @@ export class AgentSession {
 
 			const { results, tokensUsed } = await executeReadOnlyDelegationPlan(
 				decision.delegation,
-				async (subtaskPrompt, signal) => {
+				async (subtaskPrompt, options) => {
+					const signal = options?.signal;
+					const allowedTools = options?.allowedTools ?? isDelegationAllowedTool;
 					if (signal?.aborted) return { result: "Delegation failed: Aborted", tokens: 0 };
 
 					const { resolveModelPool } = await import("../routing/presets");
@@ -3280,7 +3041,8 @@ export class AgentSession {
 
 					if (
 						resolvedUtility.provider === "openai" ||
-						(resolvedUtility.provider === "litellm" && resolvedUtility.id.includes("openai"))
+						(resolvedUtility.provider === "litellm" &&
+							(resolvedUtility.id.includes("openai") || /gpt-/.test(resolvedUtility.id)))
 					) {
 						const internalUrl = this.settings.get("routing.internalOpenAiUrl") as string | undefined;
 						if (internalUrl) resolvedUtility = { ...resolvedUtility, baseUrl: internalUrl } as any;
@@ -3308,7 +3070,7 @@ export class AgentSession {
 						modelRegistry: this.#modelRegistry,
 						toolNames: Array.from(this.#toolRegistry.values())
 							.map(t => t.name)
-							.filter(isDelegationAllowedTool),
+							.filter(allowedTools),
 						enableLsp: false,
 						enableMCP: false,
 					});
@@ -3333,7 +3095,7 @@ export class AgentSession {
 										.join(""),
 								)
 								.join("\n");
-							const tUsage = ((childSession as any).agent?.lastUsage?.totalTokens ?? 0) as number;
+							const tUsage = childSession.buildDisplaySessionContext().usedTokens || 0;
 							return { result: resultStr || "No output", tokens: tUsage };
 						} finally {
 							signal?.removeEventListener("abort", onAbort);
@@ -3822,7 +3584,7 @@ export class AgentSession {
 	 */
 	async sendCustomMessage<T = unknown>(
 		message: Pick<CustomMessage<T>, "customType" | "content" | "display" | "details" | "attribution">,
-		options?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn" },
+		options?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn"; signal?: AbortSignal },
 	): Promise<void> {
 		const clonedContent = Array.isArray(message.content)
 			? message.content.map((c: any) => ({ ...c }))
@@ -4367,7 +4129,10 @@ export class AgentSession {
 		const previousEditMode = this.#resolveActiveEditMode();
 
 		let targetModel = model;
-		if (model.provider === "openai" || (model.provider === "litellm" && model.id.includes("openai"))) {
+		if (
+			model.provider === "openai" ||
+			(model.provider === "litellm" && (model.id.includes("openai") || /gpt-/.test(model.id)))
+		) {
 			const internalUrl = this.settings.get("routing.internalOpenAiUrl") as string | undefined;
 			if (internalUrl) targetModel = { ...model, baseUrl: internalUrl };
 		} else if (
@@ -6981,8 +6746,10 @@ export class AgentSession {
 	#syncRoutingStateFromBranch() {
 		this.#routingCoordinator.reset();
 		const entries = this.sessionManager.getBranch();
+		let foundRoutingState = false;
 		for (const entry of entries) {
 			if (entry.type === "custom" && entry.customType === "routing_event") {
+				foundRoutingState = true;
 				const event = entry.data as any;
 				if (event.state) {
 					this.#routingCoordinator.restoreState(event.state);
@@ -7001,6 +6768,16 @@ export class AgentSession {
 				}
 			}
 		}
+		if (!foundRoutingState) {
+			this.#modelResolutionSource = this.#initialModelResolutionSource;
+			const hasServiceTierEntry = this.sessionManager
+				.getBranch()
+				.some(entry => entry.type === "service_tier_change");
+			if (!hasServiceTierEntry) {
+				const configuredServiceTier = this.settings?.get("serviceTier");
+				this.agent.serviceTier = configuredServiceTier === "none" ? undefined : configuredServiceTier;
+			}
+		}
 	}
 
 	/**
@@ -7054,6 +6831,7 @@ export class AgentSession {
 			this.sessionManager.createBranchedSession(selectedEntry.parentId);
 		}
 		this.#syncTodoPhasesFromBranch();
+		this.#syncRoutingStateFromBranch();
 		this.agent.sessionId = this.sessionManager.getSessionId();
 
 		// Reload messages from entries (works for both file and in-memory mode)
@@ -7600,6 +7378,9 @@ export class AgentSession {
 export function calculateUsedTokens(messages: any[]): number {
 	const countTextLength = (content: any): number => {
 		if (typeof content === "string") {
+			if (content.startsWith("data:image/") || content.startsWith("base64,")) {
+				return 0;
+			}
 			return content.length;
 		}
 		if (Array.isArray(content)) {