@wix/pathgrade 1.0.32 → 1.0.33

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (30) hide show
  1. package/dist/agents/codex-app-server/agent.js +24 -8
  2. package/dist/agents/codex-app-server/protocol/ClientRequest.d.ts +3 -2
  3. package/dist/agents/codex-app-server/protocol/GetAccountTokenUsageParams.d.ts +3 -0
  4. package/dist/agents/codex-app-server/protocol/GetAccountTokenUsageParams.js +4 -0
  5. package/dist/agents/codex-app-server/protocol/GetAccountTokenUsageResponse.d.ts +6 -0
  6. package/dist/agents/codex-app-server/protocol/GetAccountTokenUsageResponse.js +4 -0
  7. package/dist/agents/codex-app-server/protocol/ThreadTokenUsage.d.ts +6 -0
  8. package/dist/agents/codex-app-server/protocol/ThreadTokenUsage.js +4 -0
  9. package/dist/agents/codex-app-server/protocol/ThreadTokenUsageUpdatedNotification.d.ts +6 -0
  10. package/dist/agents/codex-app-server/protocol/ThreadTokenUsageUpdatedNotification.js +4 -0
  11. package/dist/agents/codex-app-server/protocol/ThreadUsage.d.ts +7 -0
  12. package/dist/agents/codex-app-server/protocol/ThreadUsage.js +4 -0
  13. package/dist/agents/codex-app-server/protocol/ThreadUsageBreakdownGroup.d.ts +11 -0
  14. package/dist/agents/codex-app-server/protocol/ThreadUsageBreakdownGroup.js +4 -0
  15. package/dist/agents/codex-app-server/protocol/TokenUsageBreakdown.d.ts +8 -0
  16. package/dist/agents/codex-app-server/protocol/TokenUsageBreakdown.js +4 -0
  17. package/dist/agents/codex-app-server/protocol/index.d.ts +7 -0
  18. package/dist/agents/codex-app-server/turn-notifications.js +4 -0
  19. package/dist/agents/codex-app-server/turn-state.d.ts +3 -0
  20. package/dist/agents/codex-app-server/usage-accounting.d.ts +18 -0
  21. package/dist/agents/codex-app-server/usage-accounting.js +121 -0
  22. package/dist/openai-oauth/chatgpt-oauth-llm.js +12 -1
  23. package/dist/sdk/agent.js +1 -3
  24. package/dist/sdk/evaluate.js +3 -2
  25. package/dist/sdk/lifecycle.js +4 -0
  26. package/dist/types.d.ts +2 -6
  27. package/dist/utils/llm-mocks.js +5 -0
  28. package/dist/utils/llm-types.d.ts +3 -1
  29. package/dist/utils/llm.js +5 -0
  30. package/package.json +2 -2
@@ -17,6 +17,7 @@ import { CodexItemLifecycle, } from './item-lifecycle.js';
17
17
  import { createCodexManagedAuth } from './managed-auth.js';
18
18
  import { clearTurnAuthorityDeadline, closeUncorrelatedTurnTransport, establishAuthoritativeTurn, isAbortError, isAuthoritativeTurnCompletion, rememberAuthoritativeTurnId, shouldRetireAuthoritativeTurnTransport, } from './turn-authority.js';
19
19
  import { dispatchCodexNotification, failTurn } from './turn-notifications.js';
20
+ import { CodexThreadCostAccounting, CodexTurnUsageAccounting, readCodexThreadCost, } from './usage-accounting.js';
20
21
  const TURN_COMPLETED_METHOD = 'turn/completed';
21
22
  function recordFromUnknown(value) {
22
23
  return value && typeof value === 'object' && !Array.isArray(value)
@@ -74,6 +75,7 @@ export class CodexAppServerAgent extends BaseAgent {
74
75
  let disposed = false;
75
76
  let sessionRetired = false;
76
77
  let codexUserAgent = 'unknown';
78
+ const costAccounting = new CodexThreadCostAccounting();
77
79
  const managedAuth = createCodexManagedAuth({
78
80
  enabled: !runtimeEnv.OPENAI_API_KEY && !runtimeEnv.OPENAI_BASE_URL
79
81
  && (this.deps.authBroker !== undefined || this.deps.createTransport === undefined),
@@ -165,6 +167,7 @@ export class CodexAppServerAgent extends BaseAgent {
165
167
  assistantMessageParts: [],
166
168
  turnFailed: false,
167
169
  pendingTurnCompletions: [],
170
+ usageAccounting: new CodexTurnUsageAccounting(),
168
171
  };
169
172
  const turnStartController = new AbortController();
170
173
  let turnStartCanReuseTransport = false;
@@ -196,6 +199,7 @@ export class CodexAppServerAgent extends BaseAgent {
196
199
  throw new Error(`Codex app-server scripted MCP approval policy drift (userAgent=${codexUserAgent})`);
197
200
  }
198
201
  threadId = resp.thread.id;
202
+ costAccounting.beginThread(threadId);
199
203
  }
200
204
  const authoritativeThreadId = threadId;
201
205
  turn.authoritativeThreadId = authoritativeThreadId;
@@ -224,14 +228,20 @@ export class CodexAppServerAgent extends BaseAgent {
224
228
  const authoritativeTurnReady = new Promise((resolve) => {
225
229
  resolveAuthoritativeTurn = resolve;
226
230
  });
227
- const markAuthoritativeTurn = (candidateTurnId) => establishAuthoritativeTurn({
228
- turn,
229
- candidateTurnId,
230
- authoritativeThreadId,
231
- authoritativeTurnIds,
232
- correlator,
233
- onReady: () => resolveAuthoritativeTurn?.(),
234
- });
231
+ const markAuthoritativeTurn = (candidateTurnId) => {
232
+ const established = establishAuthoritativeTurn({
233
+ turn,
234
+ candidateTurnId,
235
+ authoritativeThreadId,
236
+ authoritativeTurnIds,
237
+ correlator,
238
+ onReady: () => resolveAuthoritativeTurn?.(),
239
+ });
240
+ if (established)
241
+ turn.usageAccounting.authorize(authoritativeThreadId, candidateTurnId);
242
+ return established;
243
+ };
244
+ let terminalCompletionAccepted = false;
235
245
  const turnCompletion = new Promise((resolve, reject) => {
236
246
  let settled = false;
237
247
  const off = t.onNotification((n) => {
@@ -253,6 +263,7 @@ export class CodexAppServerAgent extends BaseAgent {
253
263
  }))
254
264
  return;
255
265
  settled = true;
266
+ terminalCompletionAccepted = true;
256
267
  off();
257
268
  closeOff();
258
269
  const failure = extractTurnCompletionFailure(params);
@@ -350,6 +361,9 @@ export class CodexAppServerAgent extends BaseAgent {
350
361
  sensitiveValues, resolvedModel: model,
351
362
  });
352
363
  }
364
+ if (terminalCompletionAccepted) {
365
+ turn.costUsd = await readCodexThreadCost(t, authoritativeThreadId, costAccounting, options?.getRemainingMs);
366
+ }
353
367
  if (turn.turnFailed) {
354
368
  return assembleTurnResult({
355
369
  askBus,
@@ -592,6 +606,8 @@ function assembleTurnResult(args) {
592
606
  exitCode,
593
607
  traceOutput: rawOutput,
594
608
  toolEvents, resolvedModel,
609
+ ...activeTurn.usageAccounting.result(),
610
+ ...(activeTurn.costUsd === undefined ? {} : { costUsd: activeTurn.costUsd }),
595
611
  ...(exitCode !== 0
596
612
  ? {
597
613
  crashInfo: {
@@ -1,5 +1,5 @@
1
1
  /**
2
- * The 8 client-request methods the pathgrade driver may send to the Codex
2
+ * The 9 client-request methods the pathgrade driver may send to the Codex
3
3
  * app-server. Scoped to this driver's surface rather than the full upstream
4
4
  * protocol; phantom response-shaped entries are deliberately excluded because
5
5
  * JSON-RPC responses to server-initiated requests flow through
@@ -13,8 +13,9 @@
13
13
  * - thread/read
14
14
  * - thread/list
15
15
  * - review/start
16
+ * - account/usage/read
16
17
  */
17
- export type ClientRequestMethod = 'initialize' | 'thread/start' | 'turn/start' | 'turn/interrupt' | 'thread/inject_items' | 'thread/read' | 'thread/list' | 'review/start';
18
+ export type ClientRequestMethod = 'initialize' | 'thread/start' | 'turn/start' | 'turn/interrupt' | 'thread/inject_items' | 'thread/read' | 'thread/list' | 'review/start' | 'account/usage/read';
18
19
  /**
19
20
  * JSON-RPC 2.0 client→server request envelope the driver constructs.
20
21
  * Generic over the params type for the chosen method.
@@ -0,0 +1,3 @@
1
+ export type GetAccountTokenUsageParams = {
2
+ threadId?: string | null;
3
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/GetAccountTokenUsageParams.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,6 @@
1
+ import type { ThreadUsage } from './ThreadUsage.js';
2
+ export type GetAccountTokenUsageResponse = {
3
+ summary: unknown;
4
+ dailyUsageBuckets: unknown[] | null;
5
+ threadUsage?: ThreadUsage | null;
6
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/GetAccountTokenUsageResponse.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,6 @@
1
+ import type { TokenUsageBreakdown } from './TokenUsageBreakdown.js';
2
+ export type ThreadTokenUsage = {
3
+ total: TokenUsageBreakdown;
4
+ last: TokenUsageBreakdown;
5
+ modelContextWindow: number | null;
6
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/ThreadTokenUsage.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,6 @@
1
+ import type { ThreadTokenUsage } from './ThreadTokenUsage.js';
2
+ export type ThreadTokenUsageUpdatedNotification = {
3
+ threadId: string;
4
+ turnId: string;
5
+ tokenUsage: ThreadTokenUsage;
6
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/ThreadTokenUsageUpdatedNotification.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,7 @@
1
+ import type { ThreadUsageBreakdownGroup } from './ThreadUsageBreakdownGroup.js';
2
+ export type ThreadUsage = {
3
+ threadId: string;
4
+ estimatedUsageCreditsMicros: bigint;
5
+ estimatedUsageUsdMicros: bigint | null;
6
+ groups: ThreadUsageBreakdownGroup[];
7
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/ThreadUsage.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,11 @@
1
+ export type ThreadUsageBreakdownGroup = {
2
+ model: string | null;
3
+ reasoningEffort: string | null;
4
+ speed: string | null;
5
+ estimatedUsageCreditsMicros: bigint;
6
+ netNewInputTokens: bigint | null;
7
+ cachedInputTokens: bigint | null;
8
+ inputTokens: bigint | null;
9
+ outputTokens: bigint | null;
10
+ totalTokens: bigint | null;
11
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/ThreadUsageBreakdownGroup.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -0,0 +1,8 @@
1
+ export type TokenUsageBreakdown = {
2
+ totalTokens: number;
3
+ inputTokens: number;
4
+ cachedInputTokens: number;
5
+ cacheWriteInputTokens: number;
6
+ outputTokens: number;
7
+ reasoningOutputTokens: number;
8
+ };
@@ -0,0 +1,4 @@
1
+ // Vendored from openai/codex@rust-v0.149.0-alpha.4.3
2
+ // Source: codex-rs/app-server-protocol/schema/typescript/v2/TokenUsageBreakdown.ts
3
+ // GENERATED CODE in upstream; do not modify locally either.
4
+ export {};
@@ -14,3 +14,10 @@ export type { McpElicitationRequestParams } from './McpElicitationRequestParams.
14
14
  export type { ThreadStartParams } from './ThreadStartParams.js';
15
15
  export type { SandboxMode } from './SandboxMode.js';
16
16
  export type { TurnCompletedNotification } from './TurnCompletedNotification.js';
17
+ export type { TokenUsageBreakdown } from './TokenUsageBreakdown.js';
18
+ export type { ThreadTokenUsage } from './ThreadTokenUsage.js';
19
+ export type { ThreadTokenUsageUpdatedNotification } from './ThreadTokenUsageUpdatedNotification.js';
20
+ export type { ThreadUsageBreakdownGroup } from './ThreadUsageBreakdownGroup.js';
21
+ export type { ThreadUsage } from './ThreadUsage.js';
22
+ export type { GetAccountTokenUsageParams } from './GetAccountTokenUsageParams.js';
23
+ export type { GetAccountTokenUsageResponse } from './GetAccountTokenUsageResponse.js';
@@ -12,6 +12,10 @@ export function dispatchCodexNotification(notification, turn, sensitiveValues) {
12
12
  }
13
13
  if (!turn)
14
14
  return;
15
+ if (notification.method === 'thread/tokenUsage/updated') {
16
+ turn.usageAccounting.receive(notification.params, turn.authoritativeThreadId, turn.authoritativeTurnId);
17
+ return;
18
+ }
15
19
  const params = notification.params;
16
20
  if (notification.method === 'item/started') {
17
21
  if (params)
@@ -1,6 +1,7 @@
1
1
  import type { ToolEvent } from '../../tool-events.js';
2
2
  import type { PendingMcpDenial } from './mcp-approval-correlator.js';
3
3
  import type { CodexItemLifecycle } from './item-lifecycle.js';
4
+ import type { CodexTurnUsageAccounting } from './usage-accounting.js';
4
5
  export interface ActiveTurnState {
5
6
  turnNumber: number;
6
7
  askBatchIds: string[];
@@ -14,6 +15,8 @@ export interface ActiveTurnState {
14
15
  authoritativeThreadId?: string;
15
16
  authoritativeTurnId?: string;
16
17
  pendingTurnCompletions: unknown[];
18
+ usageAccounting: CodexTurnUsageAccounting;
19
+ costUsd?: number;
17
20
  acceptTurnCompletion?: (params: unknown) => void;
18
21
  authorityDeadlineTimer?: ReturnType<typeof setTimeout>;
19
22
  }
@@ -0,0 +1,18 @@
1
+ import type { AgentTurnResult } from '../../types.js';
2
+ import type { AppServerTransport } from './transport.js';
3
+ export declare const WIX_CODEX_USD_PER_CREDIT = 0.034;
4
+ export declare class CodexTurnUsageAccounting {
5
+ private baseline?;
6
+ private latest?;
7
+ private pending;
8
+ authorize(threadId: string, turnId: string): void;
9
+ receive(value: unknown, threadId: string | undefined, turnId: string | undefined): void;
10
+ result(): Pick<AgentTurnResult, 'inputTokens' | 'outputTokens' | 'cacheReadInputTokens' | 'cacheCreationInputTokens'>;
11
+ }
12
+ export declare class CodexThreadCostAccounting {
13
+ private threadId?;
14
+ private cumulativeUsd;
15
+ beginThread(threadId: string): void;
16
+ receive(value: unknown, threadId: string): number | undefined;
17
+ }
18
+ export declare function readCodexThreadCost(transport: AppServerTransport, threadId: string, accounting: CodexThreadCostAccounting, getRemainingMs?: () => number): Promise<number | undefined>;
@@ -0,0 +1,121 @@
1
+ const TOKEN_KEYS = [
2
+ 'totalTokens', 'inputTokens', 'cachedInputTokens',
3
+ 'cacheWriteInputTokens', 'outputTokens', 'reasoningOutputTokens',
4
+ ];
5
+ export const WIX_CODEX_USD_PER_CREDIT = 0.034;
6
+ const USAGE_READ_FALLBACK_TIMEOUT_MS = 5_000;
7
+ export class CodexTurnUsageAccounting {
8
+ baseline;
9
+ latest;
10
+ pending = [];
11
+ authorize(threadId, turnId) {
12
+ for (const value of this.pending.splice(0))
13
+ this.receive(value, threadId, turnId);
14
+ }
15
+ receive(value, threadId, turnId) {
16
+ const params = record(value);
17
+ if (threadId && !turnId) {
18
+ if (this.pending.length < 64)
19
+ this.pending.push(value);
20
+ return;
21
+ }
22
+ if (!threadId || !turnId || params?.threadId !== threadId || params.turnId !== turnId)
23
+ return;
24
+ const tokenUsage = record(params.tokenUsage);
25
+ const total = counters(tokenUsage?.total);
26
+ const last = counters(tokenUsage?.last);
27
+ if (!total || !last || !atLeast(total, last))
28
+ return;
29
+ if (!this.baseline) {
30
+ this.baseline = subtract(total, last);
31
+ this.latest = total;
32
+ return;
33
+ }
34
+ if (this.latest && !atLeast(total, this.latest))
35
+ return;
36
+ this.latest = total;
37
+ }
38
+ result() {
39
+ if (!this.baseline || !this.latest)
40
+ return {};
41
+ return {
42
+ inputTokens: this.latest.inputTokens - this.baseline.inputTokens,
43
+ outputTokens: this.latest.outputTokens - this.baseline.outputTokens,
44
+ cacheReadInputTokens: this.latest.cachedInputTokens - this.baseline.cachedInputTokens,
45
+ cacheCreationInputTokens: this.latest.cacheWriteInputTokens - this.baseline.cacheWriteInputTokens,
46
+ };
47
+ }
48
+ }
49
+ export class CodexThreadCostAccounting {
50
+ threadId;
51
+ cumulativeUsd = 0;
52
+ beginThread(threadId) {
53
+ if (this.threadId === threadId)
54
+ return;
55
+ this.threadId = threadId;
56
+ this.cumulativeUsd = 0;
57
+ }
58
+ receive(value, threadId) {
59
+ const response = record(value);
60
+ const usage = record(response?.threadUsage);
61
+ if (!usage || usage.threadId !== threadId || this.threadId !== threadId)
62
+ return undefined;
63
+ const usdValue = usage.estimatedUsageUsdMicros;
64
+ const current = usdValue !== undefined && usdValue !== null
65
+ ? microsToUsd(usdValue)
66
+ : creditsToUsd(usage.estimatedUsageCreditsMicros);
67
+ if (current === undefined || current < this.cumulativeUsd)
68
+ return undefined;
69
+ const delta = current - this.cumulativeUsd;
70
+ this.cumulativeUsd = current;
71
+ return delta;
72
+ }
73
+ }
74
+ export async function readCodexThreadCost(transport, threadId, accounting, getRemainingMs) {
75
+ const remainingMs = getRemainingMs?.() ?? USAGE_READ_FALLBACK_TIMEOUT_MS;
76
+ if (remainingMs <= 0)
77
+ return undefined;
78
+ const controller = new AbortController();
79
+ const timer = Number.isFinite(remainingMs) ? setTimeout(() => controller.abort(), remainingMs) : undefined;
80
+ try {
81
+ const response = await transport.sendRequest('account/usage/read', { threadId }, {
82
+ signal: controller.signal,
83
+ });
84
+ return accounting.receive(response, threadId);
85
+ }
86
+ catch {
87
+ return undefined;
88
+ }
89
+ finally {
90
+ if (timer)
91
+ clearTimeout(timer);
92
+ }
93
+ }
94
+ function record(value) {
95
+ return value !== null && typeof value === 'object' && !Array.isArray(value)
96
+ ? value
97
+ : undefined;
98
+ }
99
+ function counters(value) {
100
+ const source = record(value);
101
+ if (!source || !TOKEN_KEYS.every(key => Number.isSafeInteger(source[key]) && source[key] >= 0))
102
+ return undefined;
103
+ return Object.fromEntries(TOKEN_KEYS.map(key => [key, source[key]]));
104
+ }
105
+ function atLeast(value, floor) {
106
+ return TOKEN_KEYS.every(key => value[key] >= floor[key]);
107
+ }
108
+ function subtract(value, amount) {
109
+ return Object.fromEntries(TOKEN_KEYS.map(key => [key, value[key] - amount[key]]));
110
+ }
111
+ function optionalMicros(value) {
112
+ return Number.isSafeInteger(value) && value >= 0 ? value : undefined;
113
+ }
114
+ function microsToUsd(value) {
115
+ const micros = optionalMicros(value);
116
+ return micros === undefined ? undefined : micros / 1_000_000;
117
+ }
118
+ function creditsToUsd(value) {
119
+ const micros = optionalMicros(value);
120
+ return micros === undefined ? undefined : micros / 1_000_000 * WIX_CODEX_USD_PER_CREDIT;
121
+ }
@@ -342,12 +342,23 @@ export function createChatGptOAuthJudgeLLMWithDependencies(options, dependencies
342
342
  throw oauthError('OAUTH_PROTOCOL_RESPONSE_INVALID');
343
343
  }
344
344
  };
345
+ let costUsd = 0;
346
+ let costObserved = false;
345
347
  const port = {
346
348
  get tokenUsage() { return { inputTokens, outputTokens }; },
349
+ get costUsd() { return costUsd; },
350
+ get costObserved() { return costObserved; },
351
+ addCost(usd) {
352
+ if (!Number.isFinite(usd) || usd < 0)
353
+ return;
354
+ costUsd += usd;
355
+ costObserved = true;
356
+ },
347
357
  get supportsToolUse() { return true; },
348
358
  async measure(fn) {
349
359
  const beforeInput = inputTokens;
350
360
  const beforeOutput = outputTokens;
361
+ const beforeCost = costUsd;
351
362
  const result = await fn();
352
363
  return {
353
364
  result,
@@ -355,7 +366,7 @@ export function createChatGptOAuthJudgeLLMWithDependencies(options, dependencies
355
366
  inputTokens: inputTokens - beforeInput,
356
367
  outputTokens: outputTokens - beforeOutput,
357
368
  },
358
- costUsd: 0,
369
+ costUsd: costUsd - beforeCost,
359
370
  };
360
371
  },
361
372
  async call(prompt, callOptions = {}) {
package/dist/sdk/agent.js CHANGED
@@ -261,9 +261,7 @@ class AgentImpl {
261
261
  const sendTurn = async (message) => {
262
262
  const turnResult = await ms.executeTurn(message);
263
263
  turnNumber++;
264
- // Accumulate agent turn tokens (from CLI stream-json) on the shared tracker.
265
- // Cost is populated by Claude SDK (from `total_cost_usd`); Codex/Cursor leave
266
- // it undefined, in which case this is a no-op.
264
+ // Accumulate provider-reported agent turn tokens and cost on the shared tracker.
267
265
  this.accumulateTurnUsage(turnResult);
268
266
  for (const toolEvent of turnResult.toolEvents) {
269
267
  this._log.push(buildToolEventLogEntry(toolEvent, new Date().toISOString()));
@@ -68,8 +68,9 @@ function makeEvaluateAgent() {
68
68
  : undefined;
69
69
  // Agent usage is independent of the optional scorer LLM. Same
70
70
  // first-eval attribution rule for cost; omit it when unavailable.
71
- // no conversation cost was captured (Codex / Cursor today).
72
- const conversationCost = isFirstEval && beforeCostUsd > 0
71
+ // Preserve an authoritative zero while retaining compatibility with
72
+ // older/custom ports whose positive total is the only presence signal.
73
+ const conversationCost = isFirstEval && (agent.llm.costObserved === true || beforeCostUsd > 0)
73
74
  ? { conversation_cost_usd: beforeCostUsd }
74
75
  : undefined;
75
76
  if (isFirstEval)
@@ -140,6 +140,7 @@ function synthesizeTrialFromAgent(agent) {
140
140
  return null;
141
141
  const nCommands = countShellCommandsFromLog(agent.log);
142
142
  const tokenUsage = agent.llm.tokenUsage;
143
+ const conversationCostUsd = agent.llm.costUsd ?? 0;
143
144
  const conversationEnd = [...agent.log].reverse().find((entry) => entry.type === 'conversation_end');
144
145
  const completionReason = conversationEnd?.completion_reason ?? (agent.log.some((entry) => entry.type === 'agent_result') ? 'completed' : undefined);
145
146
  return {
@@ -160,6 +161,9 @@ function synthesizeTrialFromAgent(agent) {
160
161
  conversation_output_tokens: tokenUsage.outputTokens,
161
162
  }
162
163
  : {}),
164
+ ...(agent.llm.costObserved === true || conversationCostUsd > 0
165
+ ? { conversation_cost_usd: conversationCostUsd }
166
+ : {}),
163
167
  session_log: [...agent.log],
164
168
  },
165
169
  diagnostics: buildDiagnosticsReport({
package/dist/types.d.ts CHANGED
@@ -85,8 +85,7 @@ export interface LogEntry {
85
85
  blocked_prompt_tool_use_id?: string;
86
86
  runtime_policies_applied?: RuntimePolicyDescriptor[];
87
87
  /**
88
- * Per-turn agent cost in USD when the upstream provider reports it (the
89
- * Claude Agent SDK populates `total_cost_usd` on result messages).
88
+ * Per-turn agent cost in USD when the executing agent reports or estimates it.
90
89
  * Written by `buildModelAgentResultLogEntry` only when the turn result
91
90
  * carries a `costUsd` value; absent for agents that do not expose cost.
92
91
  */
@@ -332,10 +331,7 @@ export interface AgentTurnResult {
332
331
  */
333
332
  cacheReadInputTokens?: number;
334
333
  /**
335
- * Optional turn cost in USD reported by providers that expose exact
336
- * pricing. The Claude Agent SDK populates this from the result
337
- * message's `total_cost_usd`. Other drivers (Codex, Cursor) leave it
338
- * undefined until their providers expose comparable metadata.
334
+ * Optional turn cost in USD reported or estimated by the executing agent.
339
335
  */
340
336
  costUsd?: number;
341
337
  /**
@@ -12,6 +12,7 @@ export function createMockLLM(opts = {}) {
12
12
  let inputTokens = 0;
13
13
  let outputTokens = 0;
14
14
  let costUsd = 0;
15
+ let costObserved = false;
15
16
  let callCount = 0;
16
17
  const take = () => {
17
18
  callCount++;
@@ -29,6 +30,9 @@ export function createMockLLM(opts = {}) {
29
30
  get costUsd() {
30
31
  return costUsd;
31
32
  },
33
+ get costObserved() {
34
+ return costObserved;
35
+ },
32
36
  get supportsToolUse() {
33
37
  return true;
34
38
  },
@@ -43,6 +47,7 @@ export function createMockLLM(opts = {}) {
43
47
  if (!Number.isFinite(usd) || usd < 0)
44
48
  return;
45
49
  costUsd += usd;
50
+ costObserved = true;
46
51
  },
47
52
  async measure(fn) {
48
53
  const inBefore = inputTokens;
@@ -32,9 +32,11 @@ export interface LLMPort {
32
32
  * comparable metadata.
33
33
  */
34
34
  readonly costUsd?: number;
35
+ /** True after any valid cost observation, including an observed zero. */
36
+ readonly costObserved?: boolean;
35
37
  /** Add externally-observed tokens (e.g. from agent CLI turns). */
36
38
  addTokens?(input: number, output: number): void;
37
- /** Add externally-observed cost (e.g. from Claude SDK `total_cost_usd`). */
39
+ /** Add externally-observed agent cost. */
38
40
  addCost?(usd: number): void;
39
41
  /** Run `fn` and return its token + cost delta. */
40
42
  measure?<T>(fn: () => Promise<T>): Promise<{
package/dist/utils/llm.js CHANGED
@@ -45,6 +45,7 @@ export function createLLMClient(arg, legacyAgentName) {
45
45
  let inputTokens = 0;
46
46
  let outputTokens = 0;
47
47
  let costUsd = 0;
48
+ let costObserved = false;
48
49
  let lastProvider;
49
50
  const warnFallthrough = (from, to, reason) => {
50
51
  if (silent)
@@ -67,6 +68,9 @@ export function createLLMClient(arg, legacyAgentName) {
67
68
  get costUsd() {
68
69
  return costUsd;
69
70
  },
71
+ get costObserved() {
72
+ return costObserved;
73
+ },
70
74
  get supportsToolUse() {
71
75
  return !!toolCapable;
72
76
  },
@@ -84,6 +88,7 @@ export function createLLMClient(arg, legacyAgentName) {
84
88
  if (!Number.isFinite(usd) || usd < 0)
85
89
  return;
86
90
  costUsd += usd;
91
+ costObserved = true;
87
92
  },
88
93
  async measure(fn) {
89
94
  const inBefore = inputTokens;
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@wix/pathgrade",
3
- "version": "1.0.32",
3
+ "version": "1.0.33",
4
4
  "packageManager": "yarn@4.12.0",
5
5
  "description": "Evaluate whether AI agents discover and use your skills correctly",
6
6
  "exports": {
@@ -140,5 +140,5 @@
140
140
  "typescript": "^5.9.3",
141
141
  "zod": "4.3.6"
142
142
  },
143
- "falconPackageHash": "c7a69bbfe306a62f3d89b21d92a09f18996516cd374f1772d1181d3f"
143
+ "falconPackageHash": "044358f11c1feb83b7e0a5b40972dd14eaf2c5b0d1261ba1339cf9d2"
144
144
  }