@zvada/agent-server 0.3.7 → 0.3.8

package/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.3.8
+
+- Report each Claude turn's cost as the increase in the SDK's cumulative query cost. Preserve the native total in raw diagnostics, reset the baseline with a fresh query or conversation reset, and keep unknown cost distinct from reported zero.
+
 ## 0.3.7
 
 - Retain Claude's reported 5-minute and 1-hour cache-creation token counts in the existing token usage contract. Missing durations stay absent and reported zeroes are preserved.

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@zvada/agent-server",
-	"version": "0.3.7",
+	"version": "0.3.8",
 	"description": "Harness-agnostic agent execution engine: run Claude Code, Codex (SDK/CLI + app-server), and any ACP agent behind one interface with a normalized event stream, multi-turn sessions, and resume. Root export is the wire contract; /core, /server, /client are the seats.",
 	"license": "MIT",
 	"type": "module",

package/src/core/agents/claude-code/adapter.ts

@@ -93,6 +93,8 @@ type ClaudeMessage =
 			usage?: RawUsage;
 			modelUsage?: Record<string, { contextWindow?: number }>;
 			total_cost_usd?: number;
+			/** Per-turn delta from the generator tap; null means unreported. */
+			turn_cost_usd?: number | null;
 			stop_reason?: string | null;
 			is_error?: boolean;
 	  };
@@ -519,7 +521,7 @@ export class ClaudeCodeTransformer implements EventTransformer<unknown> {
 				},
 			};
 		}
-		if (typeof msg.total_cost_usd === "number") this.cost = msg.total_cost_usd;
+		this.cost = "turn_cost_usd" in msg ? (msg.turn_cost_usd ?? undefined) : msg.total_cost_usd;
 		this.finishReason = msg.stop_reason ?? msg.subtype ?? this.finishReason;
 		this.stopReason = mapClaudeStopReason(msg.stop_reason, msg.subtype);
 		if (msg.subtype && msg.subtype !== "success") {

package/src/core/agents/claude-code/generator-session.ts

@@ -100,6 +100,8 @@ export class ClaudeGeneratorSession {
 	/** Resolved once at spawn; re-merged into every hot-swap. */
 	private sdkServers?: SdkMcpServers;
 	private nativeSessionId: string | null = null;
+	/** SDK cost is cumulative within this query; token usage is already per turn. */
+	private cumulativeCost: number | undefined = 0;
 
 	private currentTap: AsyncQueue<SDKMessage> | null = null;
 	private readonly idleWaiters: Array<() => void> = [];
@@ -284,8 +286,26 @@ export class ClaudeGeneratorSession {
 				if (event.type === "system" && event.subtype === "init" && event.session_id) {
 					this.nativeSessionId = event.session_id;
 				}
-				this.currentTap?.push(event);
-				if (event.type === "result") this.onTurnComplete();
+				if (event.type === "conversation_reset") this.cumulativeCost = 0;
+				if (event.type === "result") {
+					// Keep the native total intact for raw diagnostics. A crash can
+					// report a zeroed total; that does not establish this turn's cost.
+					const result = {
+						...event,
+						turn_cost_usd:
+							this.cumulativeCost !== undefined && event.total_cost_usd >= this.cumulativeCost
+								? event.total_cost_usd - this.cumulativeCost
+								: null,
+					};
+					this.cumulativeCost =
+						this.cumulativeCost !== undefined && event.total_cost_usd < this.cumulativeCost
+							? undefined
+							: event.total_cost_usd;
+					this.currentTap?.push(result);
+					this.onTurnComplete();
+				} else {
+					this.currentTap?.push(event);
+				}
 			}
 		} catch (err) {
 			this.currentTap?.fail(err);