@agi-cli/server 0.1.98 → 0.1.99

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@agi-cli/server",
-	"version": "0.1.98",
+	"version": "0.1.99",
 	"description": "HTTP API server for AGI CLI",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -29,8 +29,8 @@
 		"typecheck": "tsc --noEmit"
 	},
 	"dependencies": {
-		"@agi-cli/sdk": "0.1.98",
-		"@agi-cli/database": "0.1.98",
+		"@agi-cli/sdk": "0.1.99",
+		"@agi-cli/database": "0.1.99",
 		"drizzle-orm": "^0.44.5",
 		"hono": "^4.9.9",
 		"zod": "^4.1.8"

package/src/runtime/db-operations.ts

@@ -76,12 +76,15 @@ export async function updateSessionTokensIncremental(
 				: priorCachedMsg;
 
 	// Compute deltas for this step; clamp to 0 in case provider reports smaller values
+	// Cached tokens reduce the billable input, so we subtract them from the delta
 	const deltaInput = Math.max(0, cumPrompt - priorPromptMsg);
 	const deltaOutput = Math.max(0, cumCompletion - priorCompletionMsg);
 	const deltaCached = Math.max(0, cumCached - priorCachedMsg);
 	const deltaReasoning = Math.max(0, cumReasoning - priorReasoningMsg);
 
-	const nextInputSess = priorInputSess + deltaInput;
+	// Session input should only count non-cached tokens
+	// Total cached tokens are tracked separately for reference
+	const nextInputSess = priorInputSess + deltaInput - deltaCached;
 	const nextOutputSess = priorOutputSess + deltaOutput;
 	const nextCachedSess = priorCachedSess + deltaCached;
 	const nextReasoningSess = priorReasoningSess + deltaReasoning;

package/src/tools/adapter.ts

@@ -220,41 +220,7 @@ export function adaptTools(
 							stepIndex: ctx.stepIndex,
 						},
 					});
-					// Optionally persist in the background without blocking ordering
-					(async () => {
-						try {
-							const index = await ctx.nextIndex();
-							await ctx.db.insert(messageParts).values({
-								id: callPartId,
-								messageId: ctx.messageId,
-								index,
-								stepIndex: ctx.stepIndex,
-								type: 'tool_call',
-								content: JSON.stringify({ name, args, callId }),
-								agent: ctx.agent,
-								provider: ctx.provider,
-								model: ctx.model,
-								startedAt: startTs,
-								toolName: name,
-								toolCallId: callId,
-							});
-						} catch {}
-					})();
-					if (typeof base.onInputAvailable === 'function') {
-						// biome-ignore lint/suspicious/noExplicitAny: AI SDK types are complex
-						await base.onInputAvailable(options as any);
-					}
-					return;
-				}
-
-				// Publish promptly so UI shows the call header before results
-				publish({
-					type: 'tool.call',
-					sessionId: ctx.sessionId,
-					payload: { name, args, callId, stepIndex: ctx.stepIndex },
-				});
-				// Persist best-effort in the background to avoid delaying output
-				(async () => {
+					// Persist synchronously to maintain correct ordering
 					try {
 						const index = await ctx.nextIndex();
 						await ctx.db.insert(messageParts).values({
@@ -272,7 +238,37 @@ export function adaptTools(
 							toolCallId: callId,
 						});
 					} catch {}
-				})();
+					if (typeof base.onInputAvailable === 'function') {
+						// biome-ignore lint/suspicious/noExplicitAny: AI SDK types are complex
+						await base.onInputAvailable(options as any);
+					}
+					return;
+				}
+
+				// Publish promptly so UI shows the call header before results
+				publish({
+					type: 'tool.call',
+					sessionId: ctx.sessionId,
+					payload: { name, args, callId, stepIndex: ctx.stepIndex },
+				});
+				// Persist synchronously to maintain correct ordering
+				try {
+					const index = await ctx.nextIndex();
+					await ctx.db.insert(messageParts).values({
+						id: callPartId,
+						messageId: ctx.messageId,
+						index,
+						stepIndex: ctx.stepIndex,
+						type: 'tool_call',
+						content: JSON.stringify({ name, args, callId }),
+						agent: ctx.agent,
+						provider: ctx.provider,
+						model: ctx.model,
+						startedAt: startTs,
+						toolName: name,
+						toolCallId: callId,
+					});
+				} catch {}
 				if (typeof base.onInputAvailable === 'function') {
 					// biome-ignore lint/suspicious/noExplicitAny: AI SDK types are complex
 					await base.onInputAvailable(options as any);