Compare commits
17 Commits
main
...
codex-pr21
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
50bd61574f | ||
|
|
25fadb1426 | ||
|
|
2780db25f2 | ||
|
|
2ba6c1f2bb | ||
|
|
8596d2aedd | ||
|
|
043b3e1b66 | ||
|
|
b493fa7fdf | ||
|
|
16669e80d7 | ||
|
|
6caeb1bac5 | ||
|
|
10b03ca738 | ||
|
|
1303315cc9 | ||
|
|
9138ddf88a | ||
|
|
bb514e3ab5 | ||
|
|
8e9962cc78 | ||
|
|
c25367bbcf | ||
|
|
40e7ab30b8 | ||
|
|
6f27af2528 |
5
.changeset/verifier-trajectory-events.md
Normal file
5
.changeset/verifier-trajectory-events.md
Normal file
@ -0,0 +1,5 @@
|
||||
---
|
||||
"@browserbasehq/stagehand": patch
|
||||
---
|
||||
|
||||
Capture verifier trajectory evidence from agent evidence callbacks for offline scoring.
|
||||
73
packages/core/lib/v3/agent/utils/captureAriaTreeProbe.ts
Normal file
73
packages/core/lib/v3/agent/utils/captureAriaTreeProbe.ts
Normal file
@ -0,0 +1,73 @@
|
||||
/**
|
||||
* captureAriaTreeProbe — capture a truncated accessibility tree of the active
|
||||
* page for use as tier-2 evidence in the trajectory recorder.
|
||||
*
|
||||
* Shared by v3AgentHandler and v3CuaAgentHandler. Listener-gated by the
|
||||
* callers so ordinary agent runs (no TrajectoryRecorder attached) don't pay
|
||||
* the cost.
|
||||
*
|
||||
* The a11y tree is the same payload the agent's `ariaTree` tool sees, but
|
||||
* captured by the harness (not the agent) so the verifier has independent
|
||||
* textual ground truth for grounding non-visual claims — prices, names,
|
||||
* dates, list contents — without OCR'ing screenshots.
|
||||
*
|
||||
* Budget: defaults to ~8000 tokens (32k chars). Per-step a11y captures
|
||||
* across a ~30-step trajectory at that cap sum to ~240k tokens total,
|
||||
* which the verifier handles via per-criterion top-K selection. The cap
|
||||
* is configurable via VERIFIER_ARIATREE_TOKEN_BUDGET so consumers can
|
||||
* trade RAM/disk for fidelity. Truncated content is marked explicitly so
|
||||
* the verifier knows it was clipped.
|
||||
*/
|
||||
import type { V3 } from "../../v3.js";
|
||||
|
||||
const APPROX_CHARS_PER_TOKEN = 4;
|
||||
const DEFAULT_TOKEN_BUDGET = 8_000;
|
||||
const DEFAULT_TIMEOUT_MS = 5_000;
|
||||
|
||||
interface CaptureAriaTreeOptions {
|
||||
/** Soft cap on token count (chars/4 approximation). Default 8000. */
|
||||
tokenBudget?: number;
|
||||
/** Hard timeout on the capture. Default 5s. */
|
||||
timeoutMs?: number;
|
||||
}
|
||||
|
||||
/**
|
||||
* Returns the truncated a11y tree as a plain string, or undefined when
|
||||
* capture fails. Never throws — a11y capture is best-effort tier-2 evidence,
|
||||
* not a hard requirement, so failures are silently absorbed (the verifier
|
||||
* surfaces this via evidence_insufficient).
|
||||
*/
|
||||
export async function captureAriaTreeProbe(
|
||||
v3: V3,
|
||||
opts: CaptureAriaTreeOptions = {},
|
||||
): Promise<string | undefined> {
|
||||
const envBudget = parseInt(
|
||||
process.env.VERIFIER_ARIATREE_TOKEN_BUDGET ?? "",
|
||||
10,
|
||||
);
|
||||
const tokenBudget =
|
||||
opts.tokenBudget ??
|
||||
(Number.isFinite(envBudget) && envBudget > 0
|
||||
? envBudget
|
||||
: DEFAULT_TOKEN_BUDGET);
|
||||
const timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;
|
||||
const maxChars = tokenBudget * APPROX_CHARS_PER_TOKEN;
|
||||
|
||||
try {
|
||||
// v3.extract() without a schema returns { pageText } where pageText is the
|
||||
// rendered accessibility tree — same path the agent's ariaTree tool uses.
|
||||
const result = await v3.extract({ timeout: timeoutMs });
|
||||
const pageText = result?.pageText;
|
||||
if (typeof pageText !== "string" || pageText.length === 0) return undefined;
|
||||
|
||||
if (pageText.length > maxChars) {
|
||||
return (
|
||||
pageText.slice(0, maxChars) +
|
||||
`\n\n[CONTENT TRUNCATED at ~${tokenBudget} tokens — set VERIFIER_ARIATREE_TOKEN_BUDGET to raise]`
|
||||
);
|
||||
}
|
||||
return pageText;
|
||||
} catch {
|
||||
return undefined;
|
||||
}
|
||||
}
|
||||
55
packages/core/lib/v3/agent/utils/cuaEvidenceStepTracker.ts
Normal file
55
packages/core/lib/v3/agent/utils/cuaEvidenceStepTracker.ts
Normal file
@ -0,0 +1,55 @@
|
||||
import type { AgentScreenshotEvidenceEvent } from "../../types/public/agentEvidenceEvents.js";
|
||||
|
||||
export interface PairedCuaActionStep {
|
||||
stepIndex: number;
|
||||
replayScreenshot?: AgentScreenshotEvidenceEvent;
|
||||
}
|
||||
|
||||
export class CuaEvidenceStepTracker {
|
||||
private nextStepIndex = 0;
|
||||
private latestScreenshot?: AgentScreenshotEvidenceEvent;
|
||||
private latestScreenshotConsumed = true;
|
||||
|
||||
reset(): void {
|
||||
this.nextStepIndex = 0;
|
||||
this.latestScreenshot = undefined;
|
||||
this.latestScreenshotConsumed = true;
|
||||
}
|
||||
|
||||
recordScreenshot(
|
||||
screenshot: Buffer,
|
||||
url: string,
|
||||
): AgentScreenshotEvidenceEvent {
|
||||
const event: AgentScreenshotEvidenceEvent = {
|
||||
type: "screenshot",
|
||||
stepIndex: this.nextStepIndex++,
|
||||
screenshot,
|
||||
url,
|
||||
evidenceRole: "agent",
|
||||
};
|
||||
this.latestScreenshot = event;
|
||||
this.latestScreenshotConsumed = false;
|
||||
return event;
|
||||
}
|
||||
|
||||
pairAction(): PairedCuaActionStep {
|
||||
if (this.latestScreenshot && !this.latestScreenshotConsumed) {
|
||||
this.latestScreenshotConsumed = true;
|
||||
return { stepIndex: this.latestScreenshot.stepIndex };
|
||||
}
|
||||
|
||||
const stepIndex = this.nextStepIndex++;
|
||||
if (this.latestScreenshot) {
|
||||
return {
|
||||
stepIndex,
|
||||
replayScreenshot: { ...this.latestScreenshot, stepIndex },
|
||||
};
|
||||
}
|
||||
|
||||
return { stepIndex };
|
||||
}
|
||||
|
||||
get latestScreenshotUrl(): string | undefined {
|
||||
return this.latestScreenshot?.url;
|
||||
}
|
||||
}
|
||||
90
packages/core/lib/v3/agent/utils/postStepProbeEvidence.ts
Normal file
90
packages/core/lib/v3/agent/utils/postStepProbeEvidence.ts
Normal file
@ -0,0 +1,90 @@
|
||||
import type { AgentEvidenceCallback } from "../../types/public/agentEvidenceEvents.js";
|
||||
import type { LogLine } from "../../types/public/logs.js";
|
||||
import type { V3 } from "../../v3.js";
|
||||
import { captureAriaTreeProbe } from "./captureAriaTreeProbe.js";
|
||||
|
||||
interface CaptureProbeEvidenceOptions {
|
||||
v3: V3;
|
||||
url: string;
|
||||
logger: (message: LogLine) => void;
|
||||
warningMessage: string;
|
||||
}
|
||||
|
||||
interface EmitPostStepProbeEvidenceOptions extends CaptureProbeEvidenceOptions {
|
||||
stepIndices: number | number[];
|
||||
evidenceCallback?: AgentEvidenceCallback;
|
||||
}
|
||||
|
||||
function errorMessage(error: unknown): string {
|
||||
return error instanceof Error ? error.message : String(error);
|
||||
}
|
||||
|
||||
export async function captureProbeEvidence({
|
||||
v3,
|
||||
url,
|
||||
logger,
|
||||
warningMessage,
|
||||
}: CaptureProbeEvidenceOptions): Promise<{
|
||||
url: string;
|
||||
screenshot?: Buffer;
|
||||
ariaTree?: string;
|
||||
}> {
|
||||
let probeUrl = url;
|
||||
let screenshot: Buffer | undefined;
|
||||
try {
|
||||
const page = await v3.context.awaitActivePage();
|
||||
probeUrl = page.url();
|
||||
screenshot = await page.screenshot({ fullPage: false });
|
||||
} catch (e) {
|
||||
logger({
|
||||
category: "agent",
|
||||
message: `${warningMessage}: ${errorMessage(e)}`,
|
||||
level: 1,
|
||||
});
|
||||
}
|
||||
|
||||
const ariaTree = await captureAriaTreeProbe(v3);
|
||||
return {
|
||||
url: probeUrl,
|
||||
...(screenshot ? { screenshot } : {}),
|
||||
...(ariaTree !== undefined ? { ariaTree } : {}),
|
||||
};
|
||||
}
|
||||
|
||||
export async function emitPostStepProbeEvidence({
|
||||
v3,
|
||||
stepIndices,
|
||||
url,
|
||||
evidenceCallback,
|
||||
logger,
|
||||
warningMessage,
|
||||
}: EmitPostStepProbeEvidenceOptions): Promise<void> {
|
||||
if (!evidenceCallback) return;
|
||||
|
||||
const indices = Array.isArray(stepIndices) ? stepIndices : [stepIndices];
|
||||
if (indices.length === 0) return;
|
||||
|
||||
const probe = await captureProbeEvidence({
|
||||
v3,
|
||||
url,
|
||||
logger,
|
||||
warningMessage,
|
||||
});
|
||||
for (const stepIndex of indices) {
|
||||
if (probe.screenshot) {
|
||||
await evidenceCallback({
|
||||
type: "screenshot",
|
||||
stepIndex,
|
||||
screenshot: probe.screenshot,
|
||||
url: probe.url,
|
||||
evidenceRole: "probe",
|
||||
});
|
||||
}
|
||||
await evidenceCallback({
|
||||
type: "step_observed",
|
||||
stepIndex,
|
||||
url: probe.url,
|
||||
ariaTree: probe.ariaTree,
|
||||
});
|
||||
}
|
||||
}
|
||||
76
packages/core/lib/v3/agent/utils/toolOutputEvidence.ts
Normal file
76
packages/core/lib/v3/agent/utils/toolOutputEvidence.ts
Normal file
@ -0,0 +1,76 @@
|
||||
import type { AgentStepFinishedEvent } from "../../types/public/agentEvidenceEvents.js";
|
||||
|
||||
const ERROR_STRING_LIMIT = 1000;
|
||||
|
||||
function isRecord(value: unknown): value is Record<string, unknown> {
|
||||
return value !== null && typeof value === "object" && !Array.isArray(value);
|
||||
}
|
||||
|
||||
function hasOwn(value: Record<string, unknown>, key: string): boolean {
|
||||
return Object.prototype.hasOwnProperty.call(value, key);
|
||||
}
|
||||
|
||||
function normalizeError(value: unknown): string | undefined {
|
||||
if (value === undefined || value === null || value === false) {
|
||||
return undefined;
|
||||
}
|
||||
if (value instanceof Error) {
|
||||
return value.message;
|
||||
}
|
||||
if (typeof value === "string") {
|
||||
return value;
|
||||
}
|
||||
if (
|
||||
typeof value === "number" ||
|
||||
typeof value === "boolean" ||
|
||||
typeof value === "bigint"
|
||||
) {
|
||||
return String(value);
|
||||
}
|
||||
|
||||
let serialized: string;
|
||||
try {
|
||||
serialized = JSON.stringify(value) ?? String(value);
|
||||
} catch {
|
||||
serialized = String(value);
|
||||
}
|
||||
if (serialized.length <= ERROR_STRING_LIMIT) {
|
||||
return serialized;
|
||||
}
|
||||
return `${serialized.slice(0, ERROR_STRING_LIMIT)}... [truncated]`;
|
||||
}
|
||||
|
||||
function statusCandidates(toolResult: unknown): Record<string, unknown>[] {
|
||||
if (!isRecord(toolResult)) {
|
||||
return [];
|
||||
}
|
||||
|
||||
const candidates = [toolResult];
|
||||
const output = toolResult.output;
|
||||
if (isRecord(output)) {
|
||||
candidates.push(output);
|
||||
}
|
||||
return candidates;
|
||||
}
|
||||
|
||||
export function inferToolOutput(
|
||||
toolResult: unknown,
|
||||
): AgentStepFinishedEvent["toolOutput"] {
|
||||
const candidates = statusCandidates(toolResult);
|
||||
const error = candidates
|
||||
.map((candidate) =>
|
||||
hasOwn(candidate, "error") ? normalizeError(candidate.error) : undefined,
|
||||
)
|
||||
.find((message): message is string => message !== undefined);
|
||||
|
||||
const successFalse = candidates.some(
|
||||
(candidate) => candidate.success === false,
|
||||
);
|
||||
const isError = candidates.some((candidate) => Boolean(candidate.isError));
|
||||
|
||||
return {
|
||||
ok: error === undefined && !isError && !successFalse,
|
||||
result: toolResult,
|
||||
error,
|
||||
};
|
||||
}
|
||||
@ -31,6 +31,7 @@ import {
|
||||
AgentModelConfig,
|
||||
Variables,
|
||||
} from "../types/public/agent.js";
|
||||
import type { AgentEvidenceCallback } from "../types/public/agentEvidenceEvents.js";
|
||||
import { HYBRID_CAPABLE_MODEL_PATTERNS } from "../types/private/agent.js";
|
||||
import { V3FunctionName } from "../types/public/methods.js";
|
||||
import { mapToolResultToActions } from "../agent/utils/actionMapping.js";
|
||||
@ -41,6 +42,11 @@ import {
|
||||
AgentAbortError,
|
||||
} from "../types/public/sdkErrors.js";
|
||||
import { handleDoneToolCall } from "../agent/utils/handleDoneToolCall.js";
|
||||
import {
|
||||
captureProbeEvidence,
|
||||
emitPostStepProbeEvidence,
|
||||
} from "../agent/utils/postStepProbeEvidence.js";
|
||||
import { inferToolOutput } from "../agent/utils/toolOutputEvidence.js";
|
||||
import {
|
||||
CaptchaSolver,
|
||||
CAPTCHA_SOLVED_MSG,
|
||||
@ -51,6 +57,19 @@ function getErrorMessage(error: unknown): string {
|
||||
return error instanceof Error ? error.message : String(error);
|
||||
}
|
||||
|
||||
type FinalAnswerDraft = {
|
||||
message: string;
|
||||
output?: Record<string, unknown>;
|
||||
};
|
||||
|
||||
interface StepHandlerOptions {
|
||||
userCallback?:
|
||||
| GenerateTextOnStepFinishCallback<ToolSet>
|
||||
| StreamTextOnStepFinishCallback<ToolSet>;
|
||||
evidenceCallback?: AgentEvidenceCallback;
|
||||
onFinalAnswer?: (answer: FinalAnswerDraft) => void;
|
||||
}
|
||||
|
||||
/**
|
||||
* Prepends a system message with cache control to the messages array.
|
||||
* The cache control providerOptions are used by Anthropic and ignored by other providers.
|
||||
@ -244,10 +263,16 @@ export class V3AgentHandler {
|
||||
|
||||
private createStepHandler(
|
||||
state: AgentState,
|
||||
userCallback?:
|
||||
| GenerateTextOnStepFinishCallback<ToolSet>
|
||||
| StreamTextOnStepFinishCallback<ToolSet>,
|
||||
{
|
||||
userCallback,
|
||||
evidenceCallback,
|
||||
onFinalAnswer,
|
||||
}: StepHandlerOptions,
|
||||
) {
|
||||
// Monotonic step counter scoped to this execute() call. Each tool call in
|
||||
// the agent loop becomes one trajectory step. The counter feeds stepIndex
|
||||
// on evidence callback events.
|
||||
let stepCounter = 0;
|
||||
return async (event: StepResult<ToolSet>) => {
|
||||
this.logger({
|
||||
category: "agent",
|
||||
@ -255,6 +280,9 @@ export class V3AgentHandler {
|
||||
level: 2,
|
||||
});
|
||||
|
||||
const stepIndicesInTurn: number[] = [];
|
||||
let lastFinalAnswer: FinalAnswerDraft | undefined;
|
||||
|
||||
if (event.toolCalls && event.toolCalls.length > 0) {
|
||||
for (let i = 0; i < event.toolCalls.length; i++) {
|
||||
const toolCall = event.toolCalls[i];
|
||||
@ -279,6 +307,13 @@ export class V3AgentHandler {
|
||||
? `${allReasoning} ${doneReasoning}`.trim()
|
||||
: allReasoning || "Task completed successfully";
|
||||
}
|
||||
lastFinalAnswer = {
|
||||
message: state.finalMessage,
|
||||
output:
|
||||
typeof args?.output === "object" && args?.output !== null
|
||||
? (args.output as Record<string, unknown>)
|
||||
: undefined,
|
||||
};
|
||||
}
|
||||
const mappedActions = mapToolResultToActions({
|
||||
toolCallName: toolCall.toolName,
|
||||
@ -292,8 +327,42 @@ export class V3AgentHandler {
|
||||
action.timestamp = Date.now();
|
||||
state.actions.push(action);
|
||||
}
|
||||
|
||||
const stepIndex = stepCounter++;
|
||||
stepIndicesInTurn.push(stepIndex);
|
||||
const finishedAt = new Date().toISOString();
|
||||
await evidenceCallback?.({
|
||||
type: "step_finished",
|
||||
stepIndex,
|
||||
actionName: toolCall.toolName,
|
||||
actionArgs:
|
||||
typeof args === "object" && args !== null
|
||||
? (args as Record<string, unknown>)
|
||||
: {},
|
||||
reasoning: event.text ?? "",
|
||||
toolOutput: inferToolOutput(toolResult),
|
||||
finishedAt,
|
||||
});
|
||||
}
|
||||
state.currentPageUrl = (await this.v3.context.awaitActivePage()).url();
|
||||
|
||||
// Harness probe — take a single screenshot / a11y snapshot per AI SDK
|
||||
// step and attach it to every tool call in that turn. The observation
|
||||
// reflects the settled page state after the batch of tool calls; this
|
||||
// is more faithful than dropping probe evidence for all but the last
|
||||
// tool call, while still avoiding per-tool screenshot overhead.
|
||||
await emitPostStepProbeEvidence({
|
||||
v3: this.v3,
|
||||
stepIndices: stepIndicesInTurn,
|
||||
url: state.currentPageUrl,
|
||||
evidenceCallback,
|
||||
logger: this.logger,
|
||||
warningMessage: "Warning: harness probe failed",
|
||||
});
|
||||
}
|
||||
|
||||
if (lastFinalAnswer) {
|
||||
onFinalAnswer?.(lastFinalAnswer);
|
||||
}
|
||||
|
||||
if (userCallback) {
|
||||
@ -321,6 +390,7 @@ export class V3AgentHandler {
|
||||
completed: false,
|
||||
currentPageUrl: "",
|
||||
};
|
||||
let finalAnswerFromDoneTool: FinalAnswerDraft | undefined;
|
||||
|
||||
let messages: ModelMessage[] = [];
|
||||
let captchaSolver: CaptchaSolver | undefined;
|
||||
@ -379,7 +449,16 @@ export class V3AgentHandler {
|
||||
callbacks?.prepareStep,
|
||||
captchaSolver,
|
||||
),
|
||||
onStepFinish: this.createStepHandler(state, callbacks?.onStepFinish),
|
||||
onStepFinish: this.createStepHandler(
|
||||
state,
|
||||
{
|
||||
userCallback: callbacks?.onStepFinish,
|
||||
evidenceCallback: callbacks?.onEvidence,
|
||||
onFinalAnswer: (answer) => {
|
||||
finalAnswerFromDoneTool = answer;
|
||||
},
|
||||
},
|
||||
),
|
||||
abortSignal: preparedOptions.signal,
|
||||
providerOptions: {
|
||||
google: { mediaResolution: "MEDIA_RESOLUTION_HIGH" },
|
||||
@ -396,6 +475,15 @@ export class V3AgentHandler {
|
||||
preparedOptions.output,
|
||||
this.logger,
|
||||
);
|
||||
const output = doneResult.output ?? finalAnswerFromDoneTool?.output;
|
||||
await this.emitFinalEvidence(
|
||||
state,
|
||||
{
|
||||
message: state.finalMessage,
|
||||
output,
|
||||
},
|
||||
callbacks?.onEvidence,
|
||||
);
|
||||
|
||||
return this.consolidateMetricsAndResult(
|
||||
startTime,
|
||||
@ -403,7 +491,7 @@ export class V3AgentHandler {
|
||||
doneResult.messages,
|
||||
result,
|
||||
maxSteps,
|
||||
doneResult.output,
|
||||
output,
|
||||
);
|
||||
} catch (error) {
|
||||
// Re-throw validation errors that should propagate to the caller
|
||||
@ -449,6 +537,7 @@ export class V3AgentHandler {
|
||||
// Highlight cursor defaults to true for hybrid mode, can be overridden
|
||||
const shouldHighlightCursor =
|
||||
streamOptions?.highlightCursor ?? this.mode === "hybrid";
|
||||
let finalAnswerFromDoneTool: FinalAnswerDraft | undefined;
|
||||
|
||||
const {
|
||||
options,
|
||||
@ -515,7 +604,16 @@ export class V3AgentHandler {
|
||||
callbacks?.prepareStep,
|
||||
captchaSolver,
|
||||
),
|
||||
onStepFinish: this.createStepHandler(state, callbacks?.onStepFinish),
|
||||
onStepFinish: this.createStepHandler(
|
||||
state,
|
||||
{
|
||||
userCallback: callbacks?.onStepFinish,
|
||||
evidenceCallback: callbacks?.onEvidence,
|
||||
onFinalAnswer: (answer) => {
|
||||
finalAnswerFromDoneTool = answer;
|
||||
},
|
||||
},
|
||||
),
|
||||
onError: (event) => {
|
||||
captchaSolver?.dispose();
|
||||
if (callbacks?.onError) {
|
||||
@ -541,17 +639,29 @@ export class V3AgentHandler {
|
||||
options.instruction,
|
||||
options.output,
|
||||
this.logger,
|
||||
).then((doneResult) => {
|
||||
const result = this.consolidateMetricsAndResult(
|
||||
startTime,
|
||||
state,
|
||||
doneResult.messages,
|
||||
event,
|
||||
maxSteps,
|
||||
doneResult.output,
|
||||
);
|
||||
resolveResult(result);
|
||||
});
|
||||
)
|
||||
.then(async (doneResult) => {
|
||||
const output =
|
||||
doneResult.output ?? finalAnswerFromDoneTool?.output;
|
||||
await this.emitFinalEvidence(
|
||||
state,
|
||||
{
|
||||
message: state.finalMessage,
|
||||
output,
|
||||
},
|
||||
callbacks?.onEvidence,
|
||||
);
|
||||
const result = this.consolidateMetricsAndResult(
|
||||
startTime,
|
||||
state,
|
||||
doneResult.messages,
|
||||
event,
|
||||
maxSteps,
|
||||
output,
|
||||
);
|
||||
resolveResult(result);
|
||||
})
|
||||
.catch(handleError);
|
||||
},
|
||||
onAbort: (event) => {
|
||||
captchaSolver?.dispose();
|
||||
@ -580,6 +690,26 @@ export class V3AgentHandler {
|
||||
return agentStreamResult;
|
||||
}
|
||||
|
||||
private async emitFinalEvidence(
|
||||
state: AgentState,
|
||||
finalAnswer: { message: string; output?: Record<string, unknown> },
|
||||
evidenceCallback?: AgentEvidenceCallback,
|
||||
): Promise<void> {
|
||||
if (!evidenceCallback) return;
|
||||
|
||||
const observation = await captureProbeEvidence({
|
||||
v3: this.v3,
|
||||
url: state.currentPageUrl,
|
||||
logger: this.logger,
|
||||
warningMessage: "Warning: final harness probe failed",
|
||||
});
|
||||
await evidenceCallback({
|
||||
type: "final_answer",
|
||||
...finalAnswer,
|
||||
observation,
|
||||
});
|
||||
}
|
||||
|
||||
private consolidateMetricsAndResult(
|
||||
startTime: number,
|
||||
state: AgentState,
|
||||
|
||||
@ -7,6 +7,12 @@ import { GoogleCUAClient } from "../agent/GoogleCUAClient.js";
|
||||
import { OpenAICUAClient } from "../agent/OpenAICUAClient.js";
|
||||
import { mapKeyToPlaywright } from "../agent/utils/cuaKeyMapping.js";
|
||||
import { ensureXPath } from "../agent/utils/xpath.js";
|
||||
import {
|
||||
captureProbeEvidence,
|
||||
emitPostStepProbeEvidence,
|
||||
} from "../agent/utils/postStepProbeEvidence.js";
|
||||
import { CuaEvidenceStepTracker } from "../agent/utils/cuaEvidenceStepTracker.js";
|
||||
import { inferToolOutput } from "../agent/utils/toolOutputEvidence.js";
|
||||
import {
|
||||
ActionExecutionResult,
|
||||
AgentAction,
|
||||
@ -16,6 +22,7 @@ import {
|
||||
SafetyConfirmationHandler,
|
||||
} from "../types/public/agent.js";
|
||||
import { LogLine } from "../types/public/logs.js";
|
||||
import type { AgentEvidenceCallback } from "../types/public/agentEvidenceEvents.js";
|
||||
import { type Action, V3FunctionName } from "../types/public/methods.js";
|
||||
import { FlowLogger } from "../flowlogger/FlowLogger.js";
|
||||
import { toTitleCase } from "../../utils.js";
|
||||
@ -37,6 +44,8 @@ export class V3CuaAgentHandler {
|
||||
private captchaSolver: CaptchaSolver | null = null;
|
||||
private captchaClickGuardRemaining = 0;
|
||||
private currentInstruction = "";
|
||||
private readonly cuaEvidenceSteps = new CuaEvidenceStepTracker();
|
||||
private evidenceCallback?: AgentEvidenceCallback;
|
||||
|
||||
constructor(
|
||||
v3: V3,
|
||||
@ -76,6 +85,9 @@ export class V3CuaAgentHandler {
|
||||
this.ensureNotClosed();
|
||||
const page = await this.v3.context.awaitActivePage();
|
||||
const screenshotBuffer = await page.screenshot({ fullPage: false });
|
||||
|
||||
await this.emitCuaScreenshotNonFatal(screenshotBuffer, page.url());
|
||||
|
||||
return screenshotBuffer.toString("base64"); // base64 png
|
||||
});
|
||||
|
||||
@ -120,6 +132,8 @@ export class V3CuaAgentHandler {
|
||||
(this.options.clientOptions?.waitBetweenActions as number) ||
|
||||
defaultDelay;
|
||||
try {
|
||||
let executionResult: ActionExecutionResult | undefined;
|
||||
const startedAt = new Date().toISOString();
|
||||
// Try to inject cursor before each action if enabled
|
||||
if (this.highlightCursor) {
|
||||
try {
|
||||
@ -133,7 +147,7 @@ export class V3CuaAgentHandler {
|
||||
// takes its own screenshot via screenshotProvider between API turns.
|
||||
const shouldLog = action.type !== "screenshot";
|
||||
if (shouldLog) {
|
||||
await FlowLogger.runWithLogging(
|
||||
executionResult = await FlowLogger.runWithLogging(
|
||||
{
|
||||
eventType: `V3Cua${toTitleCase(action.type)}`, // e.g. "V3CuaClick"
|
||||
data: {
|
||||
@ -145,10 +159,13 @@ export class V3CuaAgentHandler {
|
||||
[action],
|
||||
);
|
||||
} else {
|
||||
await this.executeAction(action);
|
||||
executionResult = await this.executeAction(action);
|
||||
}
|
||||
|
||||
action.timestamp = Date.now();
|
||||
if (shouldLog) {
|
||||
await this.emitCuaActionStep(action, executionResult, startedAt);
|
||||
}
|
||||
|
||||
await new Promise((r) => setTimeout(r, waitBetween));
|
||||
} catch (error) {
|
||||
@ -184,6 +201,8 @@ export class V3CuaAgentHandler {
|
||||
: optionsOrInstruction;
|
||||
|
||||
this.setSafetyConfirmationHandler(options.callbacks?.onSafetyConfirmation);
|
||||
this.evidenceCallback = options.callbacks?.onEvidence;
|
||||
this.cuaEvidenceSteps.reset();
|
||||
|
||||
this.highlightCursor = options.highlightCursor !== false;
|
||||
this.currentInstruction = options.instruction;
|
||||
@ -239,7 +258,28 @@ export class V3CuaAgentHandler {
|
||||
let result: AgentResult;
|
||||
try {
|
||||
result = await this.agent.execute({ options, logger: this.logger });
|
||||
if (this.evidenceCallback) {
|
||||
let finalUrl = "";
|
||||
try {
|
||||
finalUrl = (await this.v3.context.awaitActivePage()).url();
|
||||
} catch {
|
||||
finalUrl = this.cuaEvidenceSteps.latestScreenshotUrl ?? "";
|
||||
}
|
||||
const observation = await captureProbeEvidence({
|
||||
v3: this.v3,
|
||||
url: finalUrl,
|
||||
logger: this.logger,
|
||||
warningMessage: "Warning: CUA final probe failed",
|
||||
});
|
||||
await this.evidenceCallback({
|
||||
type: "final_answer",
|
||||
message: result.message,
|
||||
output: result.output,
|
||||
observation,
|
||||
});
|
||||
}
|
||||
} finally {
|
||||
this.evidenceCallback = undefined;
|
||||
this.captchaSolver?.dispose();
|
||||
this.captchaSolver = null;
|
||||
}
|
||||
@ -658,6 +698,10 @@ export class V3CuaAgentHandler {
|
||||
const screenshotBuffer = await page.screenshot({ fullPage: false });
|
||||
|
||||
const currentUrl = page.url();
|
||||
|
||||
// Mirror the same buffer the CUA client receives as agent evidence.
|
||||
await this.emitCuaScreenshotNonFatal(screenshotBuffer, currentUrl);
|
||||
|
||||
return await this.agentClient.captureScreenshot({
|
||||
base64Image: screenshotBuffer.toString("base64"),
|
||||
currentUrl,
|
||||
@ -767,6 +811,94 @@ export class V3CuaAgentHandler {
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Emit a pre-action CUA screenshot — the exact buffer the model received
|
||||
* as input. Tier-1 evidence (agent-mirrored); the tier-2 probe is taken
|
||||
* separately in emitCuaActionStep after the action runs, so the recorder
|
||||
* can compare what the model saw against what the page actually showed
|
||||
* once the keystrokes/clicks landed.
|
||||
*/
|
||||
private async emitCuaScreenshot(
|
||||
screenshot: Buffer,
|
||||
url: string,
|
||||
): Promise<void> {
|
||||
await this.evidenceCallback?.(
|
||||
this.cuaEvidenceSteps.recordScreenshot(screenshot, url),
|
||||
);
|
||||
}
|
||||
|
||||
private async emitCuaScreenshotNonFatal(
|
||||
screenshot: Buffer,
|
||||
url: string,
|
||||
): Promise<void> {
|
||||
try {
|
||||
await this.emitCuaScreenshot(screenshot, url);
|
||||
} catch (e) {
|
||||
this.logger({
|
||||
category: "agent",
|
||||
message: `Warning: CUA screenshot evidence callback failed: ${
|
||||
e instanceof Error ? e.message : String(e)
|
||||
}`,
|
||||
level: 1,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
private async emitCuaActionStep(
|
||||
action: AgentAction,
|
||||
result: ActionExecutionResult | undefined,
|
||||
startedAt: string,
|
||||
): Promise<void> {
|
||||
let pageUrl =
|
||||
typeof action.pageUrl === "string"
|
||||
? action.pageUrl
|
||||
: (this.cuaEvidenceSteps.latestScreenshotUrl ?? "");
|
||||
try {
|
||||
pageUrl = (await this.v3.context.awaitActivePage()).url();
|
||||
} catch {
|
||||
// Keep the best pre-action URL fallback.
|
||||
}
|
||||
const { stepIndex, replayScreenshot } = this.cuaEvidenceSteps.pairAction();
|
||||
if (replayScreenshot) {
|
||||
await this.evidenceCallback?.(replayScreenshot);
|
||||
}
|
||||
|
||||
const actionArgs = Object.fromEntries(
|
||||
Object.entries(action).filter(([key]) => key !== "screenshot"),
|
||||
);
|
||||
const reasoning =
|
||||
typeof action.reasoning === "string"
|
||||
? action.reasoning
|
||||
: typeof action.action === "string"
|
||||
? action.action
|
||||
: "";
|
||||
|
||||
await this.evidenceCallback?.({
|
||||
type: "step_finished",
|
||||
stepIndex,
|
||||
actionName: String(action.type),
|
||||
actionArgs,
|
||||
reasoning,
|
||||
toolOutput: inferToolOutput(result ?? { success: true }),
|
||||
startedAt,
|
||||
finishedAt: new Date().toISOString(),
|
||||
});
|
||||
|
||||
// Post-action tier-2 probe. The pre-action screenshot from
|
||||
// screenshotProvider is what the model SAW; this one shows what the
|
||||
// page actually LOOKS LIKE after the action ran. Without this the
|
||||
// verifier has no visual evidence that keystrokes/clicks landed, and
|
||||
// has to trust the action history alone.
|
||||
await emitPostStepProbeEvidence({
|
||||
v3: this.v3,
|
||||
stepIndices: stepIndex,
|
||||
url: pageUrl,
|
||||
evidenceCallback: this.evidenceCallback,
|
||||
logger: this.logger,
|
||||
warningMessage: "Warning: CUA post-action probe failed",
|
||||
});
|
||||
}
|
||||
|
||||
private async injectCursor(): Promise<void> {
|
||||
try {
|
||||
const page = await this.v3.context.awaitActivePage();
|
||||
|
||||
@ -25,9 +25,14 @@ import { getAISDKLanguageModel } from "./llm/LLMProvider.js";
|
||||
import { __internalCreateInMemoryAgentCacheHandle } from "./cache/serverAgentCache.js";
|
||||
import { maybeRunShutdownSupervisorFromArgv } from "./shutdown/supervisor.js";
|
||||
import {
|
||||
buildAgentEvidenceFromStepFinished,
|
||||
loadTrajectoryFromDisk,
|
||||
mergeAgentEvidence,
|
||||
nextResultFilename,
|
||||
normalizeRubric,
|
||||
redactInlineImagePayloads,
|
||||
shouldPersistTrajectory,
|
||||
writeTrajectoryDir,
|
||||
} from "./verifier/index.js";
|
||||
|
||||
export { V3 } from "./v3.js";
|
||||
@ -90,9 +95,14 @@ export type {
|
||||
VerifierRawSteps,
|
||||
} from "./verifier/index.js";
|
||||
export {
|
||||
buildAgentEvidenceFromStepFinished,
|
||||
loadTrajectoryFromDisk,
|
||||
mergeAgentEvidence,
|
||||
nextResultFilename,
|
||||
normalizeRubric,
|
||||
redactInlineImagePayloads,
|
||||
shouldPersistTrajectory,
|
||||
writeTrajectoryDir,
|
||||
} from "./verifier/index.js";
|
||||
export { tool } from "ai";
|
||||
export { getAISDKLanguageModel } from "./llm/LLMProvider.js";
|
||||
@ -144,9 +154,14 @@ const StagehandDefault = {
|
||||
toJsonSchema,
|
||||
connectToMCPServer,
|
||||
V3Evaluator,
|
||||
buildAgentEvidenceFromStepFinished,
|
||||
loadTrajectoryFromDisk,
|
||||
mergeAgentEvidence,
|
||||
nextResultFilename,
|
||||
normalizeRubric,
|
||||
redactInlineImagePayloads,
|
||||
shouldPersistTrajectory,
|
||||
writeTrajectoryDir,
|
||||
tool,
|
||||
getAISDKLanguageModel,
|
||||
__internalCreateInMemoryAgentCacheHandle,
|
||||
|
||||
@ -15,6 +15,7 @@ import {
|
||||
import { LogLine } from "./logs.js";
|
||||
import { ClientOptions } from "./model.js";
|
||||
import { StagehandZodObject } from "../../zodCompat.js";
|
||||
import type { AgentEvidenceCallback } from "./agentEvidenceEvents.js";
|
||||
|
||||
// Re-export ModelMessage for consumers who want to use it for conversation continuation
|
||||
export type { ModelMessage } from "ai";
|
||||
@ -136,6 +137,11 @@ export interface AgentCallbacks {
|
||||
onStepFinish?:
|
||||
| GenerateTextOnStepFinishCallback<ToolSet>
|
||||
| StreamTextOnStepFinishCallback<ToolSet>;
|
||||
/**
|
||||
* Callback called when Stagehand captures agent-run evidence such as
|
||||
* screenshots, completed tool/action steps, or post-action observations.
|
||||
*/
|
||||
onEvidence?: AgentEvidenceCallback;
|
||||
}
|
||||
|
||||
/**
|
||||
|
||||
105
packages/core/lib/v3/types/public/agentEvidenceEvents.ts
Normal file
105
packages/core/lib/v3/types/public/agentEvidenceEvents.ts
Normal file
@ -0,0 +1,105 @@
|
||||
/**
|
||||
* Evidence events emitted through AgentExecuteOptions.callbacks.onEvidence.
|
||||
*
|
||||
* These events describe observations made by Stagehand during an agent run.
|
||||
* They are intentionally transport-level callback payloads; verifier-specific
|
||||
* storage and normalization live in the evals/verifier layers.
|
||||
*/
|
||||
|
||||
export type AgentEvidenceRole = "probe" | "agent";
|
||||
|
||||
export type AgentEvidenceEvent =
|
||||
| AgentScreenshotEvidenceEvent
|
||||
| AgentStepFinishedEvent
|
||||
| AgentStepObservedEvent
|
||||
| AgentFinalAnswerEvent;
|
||||
|
||||
/**
|
||||
* Screenshot captured during an agent run.
|
||||
*
|
||||
* In DOM/hybrid mode, post-tool screenshots are probe evidence. In CUA mode,
|
||||
* screenshots captured by the screenshot provider are agent evidence because
|
||||
* they are the exact bytes sent to the provider.
|
||||
*/
|
||||
export interface AgentScreenshotEvidenceEvent {
|
||||
type: "screenshot";
|
||||
/** Zero-based index of the step this screenshot corresponds to. */
|
||||
stepIndex: number;
|
||||
/** PNG bytes from page.screenshot(). */
|
||||
screenshot: Buffer;
|
||||
/** Page URL at the time of capture. */
|
||||
url: string;
|
||||
/** Role this screenshot plays in downstream evidence collection. */
|
||||
evidenceRole?: AgentEvidenceRole;
|
||||
}
|
||||
|
||||
/**
|
||||
* One completed agent tool/action step.
|
||||
*/
|
||||
export interface AgentStepFinishedEvent {
|
||||
type: "step_finished";
|
||||
stepIndex: number;
|
||||
/** Name of the tool/action that ran, e.g. "act", "extract", "click". */
|
||||
actionName: string;
|
||||
/** Arguments passed to the tool/action. */
|
||||
actionArgs: Record<string, unknown>;
|
||||
/** Agent textual reasoning for the step, when available. */
|
||||
reasoning: string;
|
||||
/** Outcome of the tool/action as seen by Stagehand. */
|
||||
toolOutput: {
|
||||
ok: boolean;
|
||||
/** Native return value from the tool/action. */
|
||||
result: unknown;
|
||||
error?: string;
|
||||
};
|
||||
/** ISO 8601 timestamp at which the step's tool execution started, when available. */
|
||||
startedAt?: string;
|
||||
/** ISO 8601 timestamp at which the step finished. */
|
||||
finishedAt: string;
|
||||
}
|
||||
|
||||
/**
|
||||
* Independent post-step browser observation.
|
||||
*/
|
||||
export interface AgentStepObservedEvent {
|
||||
type: "step_observed";
|
||||
stepIndex: number;
|
||||
/** Page URL after the step's tool/action execution. */
|
||||
url: string;
|
||||
/** Accessibility tree snapshot, when captured. */
|
||||
ariaTree?: string;
|
||||
/** Viewport scroll context, when captured. */
|
||||
scroll?: { top: number; pageHeight: number };
|
||||
}
|
||||
|
||||
export interface AgentFinalObservation {
|
||||
/** Page URL at the time of terminal capture. */
|
||||
url: string;
|
||||
/** PNG bytes from page.screenshot(), when capture succeeds. */
|
||||
screenshot?: Buffer;
|
||||
/** Accessibility tree snapshot, when captured. */
|
||||
ariaTree?: string;
|
||||
/** Viewport scroll context, when captured. */
|
||||
scroll?: { top: number; pageHeight: number };
|
||||
}
|
||||
|
||||
/** Final answer emitted by the agent, when available. */
|
||||
export interface AgentFinalAnswerEvent {
|
||||
type: "final_answer";
|
||||
/** The agent's final summary message. */
|
||||
message: string;
|
||||
/** Optional structured output if the agent's output schema was set. */
|
||||
output?: Record<string, unknown>;
|
||||
/**
|
||||
* Independent terminal browser observation captured after the agent finishes.
|
||||
*
|
||||
* This preserves the legacy verifier behavior of evaluating against a final
|
||||
* page screenshot even when the last agent output is a final answer rather
|
||||
* than a browser action.
|
||||
*/
|
||||
observation?: AgentFinalObservation;
|
||||
}
|
||||
|
||||
export type AgentEvidenceCallback = (
|
||||
event: AgentEvidenceEvent,
|
||||
) => PromiseLike<void> | void;
|
||||
@ -1,4 +1,5 @@
|
||||
export * from "./agent.js";
|
||||
export * from "./agentEvidenceEvents.js";
|
||||
// Export api.ts under namespace to avoid conflicts with methods.ts types
|
||||
export * as Api from "./api.js";
|
||||
// Also export BrowserbaseRegion directly for convenience
|
||||
|
||||
118
packages/core/lib/v3/verifier/evidenceNormalization.ts
Normal file
118
packages/core/lib/v3/verifier/evidenceNormalization.ts
Normal file
@ -0,0 +1,118 @@
|
||||
import type { AgentStepFinishedEvent } from "../types/public/agentEvidenceEvents.js";
|
||||
import type { AgentEvidence } from "./types.js";
|
||||
|
||||
export const REDACTED_INLINE_IMAGE = "[redacted inline image payload]";
|
||||
|
||||
const INLINE_IMAGE_KEYS = new Set(["screenshotBase64"]);
|
||||
|
||||
function shouldRedactBase64Key(key: string, actionName?: string): boolean {
|
||||
return (
|
||||
INLINE_IMAGE_KEYS.has(key) ||
|
||||
(actionName === "screenshot" && key === "base64")
|
||||
);
|
||||
}
|
||||
|
||||
export function collectInlineImagePayloads(
|
||||
value: unknown,
|
||||
actionName?: string,
|
||||
out: string[] = [],
|
||||
): string[] {
|
||||
if (!value || typeof value !== "object") return out;
|
||||
if (Buffer.isBuffer(value)) return out;
|
||||
|
||||
if (Array.isArray(value)) {
|
||||
for (const item of value) {
|
||||
collectInlineImagePayloads(item, actionName, out);
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
for (const [key, nested] of Object.entries(value)) {
|
||||
if (shouldRedactBase64Key(key, actionName) && typeof nested === "string") {
|
||||
out.push(nested);
|
||||
continue;
|
||||
}
|
||||
collectInlineImagePayloads(nested, actionName, out);
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
export function redactInlineImagePayloads(
|
||||
value: unknown,
|
||||
actionName?: string,
|
||||
): unknown {
|
||||
if (!value || typeof value !== "object") return value;
|
||||
if (Buffer.isBuffer(value)) return value;
|
||||
|
||||
if (Array.isArray(value)) {
|
||||
return value.map((item) => redactInlineImagePayloads(item, actionName));
|
||||
}
|
||||
|
||||
const out: Record<string, unknown> = {};
|
||||
for (const [key, nested] of Object.entries(value)) {
|
||||
out[key] =
|
||||
shouldRedactBase64Key(key, actionName) && typeof nested === "string"
|
||||
? REDACTED_INLINE_IMAGE
|
||||
: redactInlineImagePayloads(nested, actionName);
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
export function mergeAgentEvidence(
|
||||
...parts: Array<AgentEvidence | undefined>
|
||||
): AgentEvidence {
|
||||
return {
|
||||
modalities: parts.flatMap((p) => p?.modalities ?? []),
|
||||
};
|
||||
}
|
||||
|
||||
export function buildAgentEvidenceFromStepFinished(
|
||||
event: AgentStepFinishedEvent,
|
||||
): AgentEvidence {
|
||||
const modalities: AgentEvidence["modalities"] = [];
|
||||
if (event.reasoning) {
|
||||
modalities.push({ type: "text", content: event.reasoning });
|
||||
}
|
||||
|
||||
const result = event.toolOutput.result;
|
||||
if (result === undefined || result === null) {
|
||||
return { modalities };
|
||||
}
|
||||
|
||||
if (typeof result === "string") {
|
||||
modalities.push({ type: "text", content: result });
|
||||
} else if (
|
||||
typeof result === "number" ||
|
||||
typeof result === "boolean" ||
|
||||
typeof result === "bigint"
|
||||
) {
|
||||
modalities.push({ type: "text", content: String(result) });
|
||||
} else if (Buffer.isBuffer(result)) {
|
||||
modalities.push({
|
||||
type: "image",
|
||||
bytes: result,
|
||||
mediaType: "image/png",
|
||||
});
|
||||
} else if (typeof result === "object") {
|
||||
for (const imageBase64 of collectInlineImagePayloads(
|
||||
result,
|
||||
event.actionName,
|
||||
)) {
|
||||
try {
|
||||
modalities.push({
|
||||
type: "image",
|
||||
bytes: Buffer.from(imageBase64, "base64"),
|
||||
mediaType: "image/png",
|
||||
});
|
||||
} catch {
|
||||
// Malformed base64; skip the image and keep the JSON modality.
|
||||
}
|
||||
}
|
||||
modalities.push({
|
||||
type: "json",
|
||||
content: redactInlineImagePayloads(result, event.actionName),
|
||||
});
|
||||
}
|
||||
|
||||
return { modalities };
|
||||
}
|
||||
@ -21,8 +21,17 @@ export type {
|
||||
VerifierFinding,
|
||||
VerifierRawSteps,
|
||||
} from "./types.js";
|
||||
export {
|
||||
buildAgentEvidenceFromStepFinished,
|
||||
collectInlineImagePayloads,
|
||||
mergeAgentEvidence,
|
||||
redactInlineImagePayloads,
|
||||
REDACTED_INLINE_IMAGE,
|
||||
} from "./evidenceNormalization.js";
|
||||
export {
|
||||
loadTrajectoryFromDisk,
|
||||
nextResultFilename,
|
||||
normalizeRubric,
|
||||
shouldPersistTrajectory,
|
||||
writeTrajectoryDir,
|
||||
} from "./trajectory.js";
|
||||
|
||||
@ -1,3 +1,5 @@
|
||||
import fs from "node:fs/promises";
|
||||
import path from "node:path";
|
||||
import type {
|
||||
AgentEvidenceModality,
|
||||
ProbeEvidence,
|
||||
@ -5,6 +7,7 @@ import type {
|
||||
Trajectory,
|
||||
TrajectoryStep,
|
||||
} from "./types.js";
|
||||
import { redactInlineImagePayloads } from "./evidenceNormalization.js";
|
||||
|
||||
type RawRubricCriterion = {
|
||||
criterion: unknown;
|
||||
@ -18,6 +21,10 @@ type RawRubric = {
|
||||
items?: unknown;
|
||||
};
|
||||
|
||||
type PersistedProbeEvidence = ProbeEvidence & {
|
||||
screenshotPath?: string;
|
||||
};
|
||||
|
||||
/**
|
||||
* Convert dataset or generated rubric JSON into the public Stagehand shape.
|
||||
* Snake-case dataset fields are accepted here so serialized quirks do not leak
|
||||
@ -92,8 +99,9 @@ function normalizeResultLabel(label?: string): string {
|
||||
*
|
||||
* Reverses the recorder's serialization tweaks:
|
||||
* - `probeEvidence.screenshotPath` → read file into `probeEvidence.screenshot`.
|
||||
* - Image modalities in `agentEvidence.modalities` carry `bytesBase64` on
|
||||
* disk (human-readable JSON) instead of raw Buffer; we decode back.
|
||||
* - Image modalities in `agentEvidence.modalities` carry `imagePath` on
|
||||
* disk instead of raw Buffer; legacy `bytesBase64` fixtures are also
|
||||
* accepted.
|
||||
*
|
||||
* @param dir absolute or cwd-relative path to a `<run-id>/<task-id>/` directory.
|
||||
*/
|
||||
@ -105,6 +113,7 @@ export async function loadTrajectoryFromDisk(dir: string): Promise<Trajectory> {
|
||||
const trajectoryPath = path.join(trajectoryDir, "trajectory.json");
|
||||
const raw = await fs.readFile(trajectoryPath, "utf8");
|
||||
const parsed = JSON.parse(raw) as Trajectory & {
|
||||
finalObservation?: PersistedProbeEvidence;
|
||||
steps: Array<
|
||||
TrajectoryStep & {
|
||||
agentEvidence: {
|
||||
@ -113,20 +122,24 @@ export async function loadTrajectoryFromDisk(dir: string): Promise<Trajectory> {
|
||||
| {
|
||||
type: "image";
|
||||
mediaType: string;
|
||||
// On-disk form (recorder writes base64); accept either to
|
||||
// tolerate hand-edited fixtures.
|
||||
// On-disk forms. Current writer externalizes bytes to
|
||||
// imagePath; bytesBase64 is accepted for older fixtures.
|
||||
bytes?: unknown;
|
||||
bytesBase64?: string;
|
||||
imagePath?: string;
|
||||
}
|
||||
| { type: "json"; content: unknown }
|
||||
>;
|
||||
};
|
||||
probeEvidence: ProbeEvidence;
|
||||
probeEvidence: PersistedProbeEvidence;
|
||||
}
|
||||
>;
|
||||
};
|
||||
|
||||
const resolveWithinTrajectoryDir = (candidate: string): string => {
|
||||
const resolveWithinTrajectoryDir = (
|
||||
candidate: string,
|
||||
fieldName = "screenshotPath",
|
||||
): string => {
|
||||
const resolved = path.resolve(trajectoryDir, candidate);
|
||||
const relative = path.relative(trajectoryDir, resolved);
|
||||
const outside =
|
||||
@ -136,16 +149,16 @@ export async function loadTrajectoryFromDisk(dir: string): Promise<Trajectory> {
|
||||
|
||||
if (outside) {
|
||||
throw new Error(
|
||||
`Trajectory screenshotPath escapes trajectory directory: ${candidate}`,
|
||||
`Trajectory ${fieldName} escapes trajectory directory: ${candidate}`,
|
||||
);
|
||||
}
|
||||
|
||||
return resolved;
|
||||
};
|
||||
|
||||
for (const step of parsed.steps) {
|
||||
// Rehydrate tier-2 probe screenshot from its on-disk file reference.
|
||||
const probe = step.probeEvidence;
|
||||
const hydrateProbeScreenshot = async (
|
||||
probe: PersistedProbeEvidence | undefined,
|
||||
): Promise<void> => {
|
||||
if (probe?.screenshotPath && !probe.screenshot) {
|
||||
const resolved = resolveWithinTrajectoryDir(probe.screenshotPath);
|
||||
try {
|
||||
@ -155,25 +168,55 @@ export async function loadTrajectoryFromDisk(dir: string): Promise<Trajectory> {
|
||||
// evidence_insufficient path will handle it.
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
// Decode image modalities from base64 back to Buffer.
|
||||
for (const step of parsed.steps) {
|
||||
// Rehydrate tier-2 probe screenshot from its on-disk file reference.
|
||||
await hydrateProbeScreenshot(step.probeEvidence);
|
||||
|
||||
// Decode image modalities from disk references back to Buffer.
|
||||
if (step.agentEvidence?.modalities) {
|
||||
step.agentEvidence.modalities = step.agentEvidence.modalities.map((m) => {
|
||||
// The on-disk shape carries bytesBase64 instead of bytes, so we look
|
||||
// through `unknown` here rather than rely on the typed union.
|
||||
const raw = m as unknown as { bytesBase64?: string };
|
||||
const modalities: AgentEvidenceModality[] = [];
|
||||
for (const m of step.agentEvidence.modalities) {
|
||||
// The on-disk shape carries imagePath/bytesBase64 instead of bytes,
|
||||
// so we look through `unknown` rather than rely on the typed union.
|
||||
const raw = m as unknown as {
|
||||
bytesBase64?: string;
|
||||
imagePath?: string;
|
||||
};
|
||||
if (m.type === "image" && typeof raw.bytesBase64 === "string") {
|
||||
return {
|
||||
modalities.push({
|
||||
type: "image" as const,
|
||||
bytes: Buffer.from(raw.bytesBase64, "base64"),
|
||||
mediaType: m.mediaType,
|
||||
};
|
||||
});
|
||||
continue;
|
||||
}
|
||||
return m as AgentEvidenceModality;
|
||||
});
|
||||
if (m.type === "image" && typeof raw.imagePath === "string") {
|
||||
const resolved = resolveWithinTrajectoryDir(
|
||||
raw.imagePath,
|
||||
"imagePath",
|
||||
);
|
||||
try {
|
||||
modalities.push({
|
||||
type: "image" as const,
|
||||
bytes: await fs.readFile(resolved),
|
||||
mediaType: m.mediaType,
|
||||
});
|
||||
} catch {
|
||||
// Missing agent image file: omit that image modality. The
|
||||
// verifier's evidence_insufficient path will handle missing bytes.
|
||||
}
|
||||
continue;
|
||||
}
|
||||
modalities.push(m as AgentEvidenceModality);
|
||||
}
|
||||
step.agentEvidence.modalities = modalities;
|
||||
}
|
||||
}
|
||||
|
||||
await hydrateProbeScreenshot(parsed.finalObservation);
|
||||
|
||||
return parsed;
|
||||
}
|
||||
|
||||
@ -187,3 +230,164 @@ export async function loadTrajectoryFromDisk(dir: string): Promise<Trajectory> {
|
||||
export function nextResultFilename(label?: string): string {
|
||||
return `result_${normalizeResultLabel(label)}.json`;
|
||||
}
|
||||
|
||||
/**
|
||||
* Default persistence policy: explicit override, then env, then "on unless CI".
|
||||
*/
|
||||
export function shouldPersistTrajectory(
|
||||
override: boolean | undefined,
|
||||
): boolean {
|
||||
if (override !== undefined) return override;
|
||||
const env = process.env.VERIFIER_PERSIST_TRAJECTORIES?.toLowerCase();
|
||||
if (env === "1" || env === "true") return true;
|
||||
if (env === "0" || env === "false") return false;
|
||||
return !process.env.CI;
|
||||
}
|
||||
|
||||
/**
|
||||
* Write the on-disk trajectory layout under `dir`:
|
||||
*
|
||||
* <dir>/
|
||||
* ├── task_data.json
|
||||
* ├── trajectory.json (screenshots referenced by path)
|
||||
* ├── screenshots/
|
||||
* │ ├── probe/<N>.png
|
||||
* │ └── agent/<N>[_M].png
|
||||
* ├── times.json
|
||||
* ├── scores/ (empty; populated separately)
|
||||
* └── core.log
|
||||
*
|
||||
* Image bytes are externalized to PNG files; the in-memory Trajectory is left
|
||||
* untouched so callers can keep using it after persistence.
|
||||
*/
|
||||
export async function writeTrajectoryDir(
|
||||
dir: string,
|
||||
trajectory: Trajectory,
|
||||
): Promise<void> {
|
||||
await fs.mkdir(dir, { recursive: true });
|
||||
await fs.mkdir(path.join(dir, "screenshots", "probe"), { recursive: true });
|
||||
await fs.mkdir(path.join(dir, "screenshots", "agent"), { recursive: true });
|
||||
|
||||
const serializableSteps: unknown[] = [];
|
||||
for (const step of trajectory.steps) {
|
||||
const probe: ProbeEvidence = { ...step.probeEvidence };
|
||||
if (probe.screenshot) {
|
||||
const relPath = `screenshots/probe/${step.index + 1}.png`;
|
||||
await fs.writeFile(path.join(dir, relPath), probe.screenshot);
|
||||
probe.screenshotPath = relPath;
|
||||
delete probe.screenshot;
|
||||
}
|
||||
|
||||
const imageModalities = step.agentEvidence.modalities.filter(
|
||||
(m) => m.type === "image",
|
||||
);
|
||||
const multipleImages = imageModalities.length > 1;
|
||||
let imageSeq = 0;
|
||||
const modalities: unknown[] = [];
|
||||
for (const m of step.agentEvidence.modalities) {
|
||||
if (m.type !== "image") {
|
||||
modalities.push(
|
||||
m.type === "json"
|
||||
? {
|
||||
...m,
|
||||
content: redactInlineImagePayloads(m.content, step.actionName),
|
||||
}
|
||||
: m,
|
||||
);
|
||||
continue;
|
||||
}
|
||||
const suffix = multipleImages ? `_${imageSeq}` : "";
|
||||
const relPath = `screenshots/agent/${step.index + 1}${suffix}.png`;
|
||||
await fs.writeFile(path.join(dir, relPath), m.bytes);
|
||||
modalities.push({
|
||||
type: "image",
|
||||
imagePath: relPath,
|
||||
mediaType: m.mediaType,
|
||||
});
|
||||
imageSeq += 1;
|
||||
}
|
||||
serializableSteps.push({
|
||||
...step,
|
||||
probeEvidence: probe,
|
||||
agentEvidence: { modalities },
|
||||
toolOutput: {
|
||||
...step.toolOutput,
|
||||
result: redactInlineImagePayloads(
|
||||
step.toolOutput.result,
|
||||
step.actionName,
|
||||
),
|
||||
},
|
||||
});
|
||||
}
|
||||
|
||||
const finalObservation: ProbeEvidence | undefined =
|
||||
trajectory.finalObservation === undefined
|
||||
? undefined
|
||||
: { ...trajectory.finalObservation };
|
||||
if (finalObservation?.screenshot) {
|
||||
const relPath = "screenshots/probe/final.png";
|
||||
await fs.writeFile(path.join(dir, relPath), finalObservation.screenshot);
|
||||
finalObservation.screenshotPath = relPath;
|
||||
delete finalObservation.screenshot;
|
||||
}
|
||||
|
||||
// Image modalities carry imagePath instead of raw bytes on disk; cast
|
||||
// through unknown rather than widen Trajectory's type contract.
|
||||
const serialized = {
|
||||
...trajectory,
|
||||
steps: serializableSteps,
|
||||
...(finalObservation ? { finalObservation } : {}),
|
||||
} as unknown;
|
||||
|
||||
await fs.writeFile(
|
||||
path.join(dir, "trajectory.json"),
|
||||
JSON.stringify(serialized, null, 2),
|
||||
);
|
||||
|
||||
await fs.writeFile(
|
||||
path.join(dir, "task_data.json"),
|
||||
JSON.stringify(
|
||||
{
|
||||
task: trajectory.task,
|
||||
status: trajectory.status,
|
||||
finalAnswer: trajectory.finalAnswer ?? null,
|
||||
},
|
||||
null,
|
||||
2,
|
||||
),
|
||||
);
|
||||
|
||||
await fs.writeFile(
|
||||
path.join(dir, "times.json"),
|
||||
JSON.stringify(
|
||||
{
|
||||
timing: trajectory.timing,
|
||||
usage: trajectory.usage,
|
||||
stepCount: trajectory.steps.length,
|
||||
},
|
||||
null,
|
||||
2,
|
||||
),
|
||||
);
|
||||
|
||||
await fs.mkdir(path.join(dir, "scores"), { recursive: true });
|
||||
await fs.writeFile(path.join(dir, "core.log"), coreLog(trajectory));
|
||||
}
|
||||
|
||||
function coreLog(trajectory: Trajectory): string {
|
||||
return (
|
||||
trajectory.steps
|
||||
.map((step) =>
|
||||
JSON.stringify({
|
||||
step: step.index,
|
||||
action: step.actionName,
|
||||
url: step.probeEvidence.url ?? null,
|
||||
ok: step.toolOutput.ok,
|
||||
reasoning: step.reasoning || undefined,
|
||||
startedAt: step.startedAt,
|
||||
finishedAt: step.finishedAt,
|
||||
}),
|
||||
)
|
||||
.join("\n") + "\n"
|
||||
);
|
||||
}
|
||||
|
||||
@ -136,7 +136,7 @@ export type TrajectoryStatus = "complete" | "aborted" | "stalled" | "error";
|
||||
* .trajectories/<run-id>/<task-id>/
|
||||
* ├── task_data.json — TaskSpec + result metadata
|
||||
* ├── trajectory.json — this object, with screenshotPath instead of bytes
|
||||
* ├── screenshot_1.png — probeEvidence.screenshot for step 1, etc.
|
||||
* ├── screenshots/ — step probe/agent images plus final observation
|
||||
* ├── scores/
|
||||
* │ └── result.json — Result from V3Evaluator.verify()
|
||||
* ├── core.log — captured action log
|
||||
@ -146,6 +146,8 @@ export interface Trajectory {
|
||||
task: TaskSpec;
|
||||
steps: TrajectoryStep[];
|
||||
finalAnswer?: string;
|
||||
/** Terminal page observation captured after the agent finishes. */
|
||||
finalObservation?: ProbeEvidence;
|
||||
status: TrajectoryStatus;
|
||||
usage: TrajectoryUsage;
|
||||
timing: { startedAt: string; endedAt: string };
|
||||
|
||||
@ -227,6 +227,10 @@ function collectLegacyScreenshots(trajectory: Trajectory): Buffer[] {
|
||||
}
|
||||
}
|
||||
|
||||
if (Buffer.isBuffer(trajectory.finalObservation?.screenshot)) {
|
||||
screenshots.push(trajectory.finalObservation.screenshot);
|
||||
}
|
||||
|
||||
return screenshots;
|
||||
}
|
||||
|
||||
@ -234,12 +238,14 @@ function renderLegacyAgentReasoning(
|
||||
trajectory: Trajectory,
|
||||
): string | undefined {
|
||||
const stepLines = (trajectory.steps ?? []).map((step) => {
|
||||
const status = step.toolOutput?.ok === false ? "Tool status: failed" : "";
|
||||
const output = step.toolOutput?.error
|
||||
? `Tool error: ${step.toolOutput.error}`
|
||||
: `Tool output: ${stringifyForPrompt(step.toolOutput?.result)}`;
|
||||
return [
|
||||
`Step ${step.index}: ${step.actionName}`,
|
||||
step.reasoning ? `Reasoning: ${step.reasoning}` : undefined,
|
||||
status || undefined,
|
||||
output,
|
||||
]
|
||||
.filter(Boolean)
|
||||
|
||||
@ -60,6 +60,7 @@ class FakeCuaClient {
|
||||
public contextNotes: string[] = [];
|
||||
public preStepHook?: () => Promise<void>;
|
||||
public actionHandler?: (action: Record<string, unknown>) => Promise<void>;
|
||||
public screenshotProvider?: () => Promise<string>;
|
||||
public executeImpl = vi.fn(async (options: unknown) => {
|
||||
void options;
|
||||
return {
|
||||
@ -72,7 +73,9 @@ class FakeCuaClient {
|
||||
public captureScreenshot = vi.fn(async () => null);
|
||||
public setViewport = vi.fn();
|
||||
public setCurrentUrl = vi.fn();
|
||||
public setScreenshotProvider = vi.fn();
|
||||
public setScreenshotProvider = vi.fn((provider: () => Promise<string>) => {
|
||||
this.screenshotProvider = provider;
|
||||
});
|
||||
public setSafetyConfirmationHandler = vi.fn();
|
||||
|
||||
setActionHandler(
|
||||
@ -247,7 +250,6 @@ describe("agent captcha hooks", () => {
|
||||
context: {
|
||||
awaitActivePage: async () => page,
|
||||
},
|
||||
bus: { emit: vi.fn() },
|
||||
isCaptchaAutoSolveEnabled: true,
|
||||
isAdvancedStealth: false,
|
||||
configuredViewport: { width: 1288, height: 711 },
|
||||
@ -316,7 +318,6 @@ describe("agent captcha hooks", () => {
|
||||
context: {
|
||||
awaitActivePage: async () => page,
|
||||
},
|
||||
bus: { emit: vi.fn() },
|
||||
isCaptchaAutoSolveEnabled: true,
|
||||
isAdvancedStealth: false,
|
||||
configuredViewport: { width: 1288, height: 711 },
|
||||
@ -392,7 +393,6 @@ describe("agent captcha hooks", () => {
|
||||
context: {
|
||||
awaitActivePage: async () => page,
|
||||
},
|
||||
bus: { emit: vi.fn() },
|
||||
isCaptchaAutoSolveEnabled: true,
|
||||
isAdvancedStealth: false,
|
||||
configuredViewport: { width: 1288, height: 711 },
|
||||
@ -474,7 +474,6 @@ describe("v3 cua handler screenshot behavior", () => {
|
||||
context: {
|
||||
awaitActivePage: async () => page,
|
||||
},
|
||||
bus: { emit: vi.fn() },
|
||||
isCaptchaAutoSolveEnabled: false,
|
||||
isAdvancedStealth: false,
|
||||
configuredViewport: { width: 1288, height: 711 },
|
||||
@ -504,4 +503,58 @@ describe("v3 cua handler screenshot behavior", () => {
|
||||
// the CUA client takes a single screenshot after all actions itself.
|
||||
expect(screenshotSpy).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it("still returns provider screenshots when screenshot evidence callbacks fail", async () => {
|
||||
const screenshotBase64 = Buffer.from("fake-image").toString("base64");
|
||||
const onEvidence = vi.fn(async (event: { type: string }) => {
|
||||
if (event.type === "screenshot") {
|
||||
throw new Error("recorder failed");
|
||||
}
|
||||
});
|
||||
|
||||
fakeCuaClient.executeImpl = vi.fn(async () => {
|
||||
await expect(fakeCuaClient.screenshotProvider?.()).resolves.toBe(
|
||||
screenshotBase64,
|
||||
);
|
||||
return {
|
||||
success: true,
|
||||
message: "ok",
|
||||
actions: [],
|
||||
completed: true,
|
||||
};
|
||||
});
|
||||
|
||||
const handler = new V3CuaAgentHandler(
|
||||
{
|
||||
context: {
|
||||
awaitActivePage: async () => page,
|
||||
},
|
||||
isCaptchaAutoSolveEnabled: false,
|
||||
isAdvancedStealth: false,
|
||||
configuredViewport: { width: 1288, height: 711 },
|
||||
isAgentReplayActive: () => false,
|
||||
updateMetrics: vi.fn(),
|
||||
} as never,
|
||||
logger,
|
||||
{
|
||||
modelName: "openai/gpt-5.4",
|
||||
clientOptions: { waitBetweenActions: 1 },
|
||||
} as never,
|
||||
);
|
||||
|
||||
await handler.execute({
|
||||
instruction: "describe the page",
|
||||
highlightCursor: false,
|
||||
callbacks: { onEvidence },
|
||||
});
|
||||
|
||||
expect(onEvidence).toHaveBeenCalledWith(
|
||||
expect.objectContaining({ type: "screenshot" }),
|
||||
);
|
||||
expect(
|
||||
logs.some((line) =>
|
||||
line.message.includes("CUA screenshot evidence callback failed"),
|
||||
),
|
||||
).toBe(true);
|
||||
});
|
||||
});
|
||||
|
||||
55
packages/core/tests/unit/cua-evidence-step-tracker.test.ts
Normal file
55
packages/core/tests/unit/cua-evidence-step-tracker.test.ts
Normal file
@ -0,0 +1,55 @@
|
||||
import { describe, expect, it } from "vitest";
|
||||
|
||||
import { CuaEvidenceStepTracker } from "../../lib/v3/agent/utils/cuaEvidenceStepTracker.js";
|
||||
|
||||
describe("CuaEvidenceStepTracker", () => {
|
||||
it("pairs a fresh provider screenshot with the next action", () => {
|
||||
const tracker = new CuaEvidenceStepTracker();
|
||||
const screenshot = Buffer.from("screen");
|
||||
|
||||
const event = tracker.recordScreenshot(screenshot, "https://example.com");
|
||||
const paired = tracker.pairAction();
|
||||
|
||||
expect(event).toMatchObject({
|
||||
type: "screenshot",
|
||||
stepIndex: 0,
|
||||
evidenceRole: "agent",
|
||||
url: "https://example.com",
|
||||
});
|
||||
expect(paired).toEqual({ stepIndex: 0 });
|
||||
});
|
||||
|
||||
it("allocates an action step without screenshot evidence", () => {
|
||||
const tracker = new CuaEvidenceStepTracker();
|
||||
|
||||
expect(tracker.pairAction()).toEqual({ stepIndex: 0 });
|
||||
});
|
||||
|
||||
it("replays the latest consumed screenshot for later actions", () => {
|
||||
const tracker = new CuaEvidenceStepTracker();
|
||||
const screenshot = Buffer.from("screen");
|
||||
|
||||
tracker.recordScreenshot(screenshot, "https://example.com/start");
|
||||
tracker.pairAction();
|
||||
const paired = tracker.pairAction();
|
||||
|
||||
expect(paired.stepIndex).toBe(1);
|
||||
expect(paired.replayScreenshot).toMatchObject({
|
||||
type: "screenshot",
|
||||
stepIndex: 1,
|
||||
evidenceRole: "agent",
|
||||
url: "https://example.com/start",
|
||||
});
|
||||
expect(paired.replayScreenshot?.screenshot).toEqual(screenshot);
|
||||
});
|
||||
|
||||
it("resets step allocation and pending screenshot state", () => {
|
||||
const tracker = new CuaEvidenceStepTracker();
|
||||
|
||||
tracker.recordScreenshot(Buffer.from("screen"), "https://example.com");
|
||||
tracker.reset();
|
||||
|
||||
expect(tracker.pairAction()).toEqual({ stepIndex: 0 });
|
||||
expect(tracker.latestScreenshotUrl).toBeUndefined();
|
||||
});
|
||||
});
|
||||
@ -32,6 +32,8 @@ const publicApiShape = {
|
||||
V3: Stagehand.V3,
|
||||
V3Evaluator: Stagehand.V3Evaluator,
|
||||
V3FunctionName: Stagehand.V3FunctionName,
|
||||
buildAgentEvidenceFromStepFinished:
|
||||
Stagehand.buildAgentEvidenceFromStepFinished,
|
||||
connectToMCPServer: Stagehand.connectToMCPServer,
|
||||
default: StagehandDefaultExport,
|
||||
defaultExtractSchema: Stagehand.defaultExtractSchema,
|
||||
@ -44,18 +46,22 @@ const publicApiShape = {
|
||||
jsonSchemaToZod: Stagehand.jsonSchemaToZod,
|
||||
loadApiKeyFromEnv: Stagehand.loadApiKeyFromEnv,
|
||||
loadTrajectoryFromDisk: Stagehand.loadTrajectoryFromDisk,
|
||||
mergeAgentEvidence: Stagehand.mergeAgentEvidence,
|
||||
localBrowserLaunchOptionsSchema: Stagehand.localBrowserLaunchOptionsSchema,
|
||||
modelToAgentProviderMap: Stagehand.modelToAgentProviderMap,
|
||||
nextResultFilename: Stagehand.nextResultFilename,
|
||||
normalizeRubric: Stagehand.normalizeRubric,
|
||||
pageTextSchema: Stagehand.pageTextSchema,
|
||||
providerEnvVarMap: Stagehand.providerEnvVarMap,
|
||||
redactInlineImagePayloads: Stagehand.redactInlineImagePayloads,
|
||||
shouldPersistTrajectory: Stagehand.shouldPersistTrajectory,
|
||||
toGeminiSchema: Stagehand.toGeminiSchema,
|
||||
toJsonSchema: Stagehand.toJsonSchema,
|
||||
tool: Stagehand.tool,
|
||||
transformSchema: Stagehand.transformSchema,
|
||||
trimTrailingTextNode: Stagehand.trimTrailingTextNode,
|
||||
validateZodSchema: Stagehand.validateZodSchema,
|
||||
writeTrajectoryDir: Stagehand.writeTrajectoryDir,
|
||||
...publicErrorTypes,
|
||||
} as const;
|
||||
|
||||
|
||||
58
packages/core/tests/unit/tool-output-evidence.test.ts
Normal file
58
packages/core/tests/unit/tool-output-evidence.test.ts
Normal file
@ -0,0 +1,58 @@
|
||||
import { describe, expect, it } from "vitest";
|
||||
|
||||
import { inferToolOutput } from "../../lib/v3/agent/utils/toolOutputEvidence.js";
|
||||
|
||||
describe("inferToolOutput", () => {
|
||||
it("preserves raw results while normalizing top-level failure status", () => {
|
||||
const result = { success: false };
|
||||
|
||||
expect(inferToolOutput(result)).toEqual({
|
||||
ok: false,
|
||||
result,
|
||||
error: undefined,
|
||||
});
|
||||
});
|
||||
|
||||
it("normalizes one-level AI SDK output wrappers", () => {
|
||||
const result = {
|
||||
toolCallId: "call-1",
|
||||
output: { success: false, error: { message: "not found" } },
|
||||
};
|
||||
|
||||
expect(inferToolOutput(result)).toEqual({
|
||||
ok: false,
|
||||
result,
|
||||
error: '{"message":"not found"}',
|
||||
});
|
||||
});
|
||||
|
||||
it("handles isError and non-string errors", () => {
|
||||
const result = { isError: true, error: new Error("bad input") };
|
||||
|
||||
expect(inferToolOutput(result)).toEqual({
|
||||
ok: false,
|
||||
result,
|
||||
error: "bad input",
|
||||
});
|
||||
});
|
||||
|
||||
it("normalizes non-json error values", () => {
|
||||
const result = { error: Symbol("bad input") };
|
||||
|
||||
expect(inferToolOutput(result)).toEqual({
|
||||
ok: false,
|
||||
result,
|
||||
error: "Symbol(bad input)",
|
||||
});
|
||||
});
|
||||
|
||||
it("does not recursively treat page data as tool status", () => {
|
||||
const result = { data: { success: false, error: "page field" } };
|
||||
|
||||
expect(inferToolOutput(result)).toEqual({
|
||||
ok: true,
|
||||
result,
|
||||
error: undefined,
|
||||
});
|
||||
});
|
||||
});
|
||||
@ -82,6 +82,39 @@ describe("V3Evaluator verifier facade", () => {
|
||||
expect(result.perCriterion).toBeUndefined();
|
||||
});
|
||||
|
||||
it("passes final observation screenshots to the legacy verifier adapter", async () => {
|
||||
const taskSpec: TaskSpec = {
|
||||
id: "final-observation",
|
||||
instruction: "Complete the task",
|
||||
};
|
||||
const finalScreenshot = Buffer.from("final screenshot");
|
||||
const trajectory = {
|
||||
...makeTrajectory(taskSpec),
|
||||
finalObservation: {
|
||||
url: "https://example.com/done",
|
||||
screenshot: finalScreenshot,
|
||||
},
|
||||
};
|
||||
const ask = vi.fn().mockResolvedValue({
|
||||
evaluation: "YES",
|
||||
reasoning: "The final screenshot shows completion.",
|
||||
});
|
||||
const evaluator = new V3Evaluator({} as V3, {
|
||||
backend: "legacy",
|
||||
});
|
||||
Object.defineProperty(evaluator, "legacyEvaluator", {
|
||||
value: { ask },
|
||||
});
|
||||
|
||||
await evaluator.verify(trajectory);
|
||||
|
||||
expect(ask).toHaveBeenCalledWith(
|
||||
expect.objectContaining({
|
||||
screenshot: [finalScreenshot],
|
||||
}),
|
||||
);
|
||||
});
|
||||
|
||||
it("keeps legacy tool output detail until the overall reasoning budget is reached", async () => {
|
||||
const taskSpec: TaskSpec = {
|
||||
id: "reasoning-budget",
|
||||
|
||||
@ -0,0 +1,20 @@
|
||||
import { describe, expect, it } from "vitest";
|
||||
|
||||
import { buildAgentEvidenceFromStepFinished } from "../../lib/v3/verifier/evidenceNormalization.js";
|
||||
|
||||
describe("buildAgentEvidenceFromStepFinished", () => {
|
||||
it("captures primitive tool results as text evidence", () => {
|
||||
const evidence = buildAgentEvidenceFromStepFinished({
|
||||
type: "step_finished",
|
||||
stepIndex: 0,
|
||||
actionName: "check",
|
||||
actionArgs: {},
|
||||
reasoning: "",
|
||||
toolOutput: { ok: true, result: false },
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(1).toISOString(),
|
||||
});
|
||||
|
||||
expect(evidence.modalities).toEqual([{ type: "text", content: "false" }]);
|
||||
});
|
||||
});
|
||||
@ -1,4 +1,4 @@
|
||||
import { mkdtemp, writeFile } from "node:fs/promises";
|
||||
import { mkdir, mkdtemp, readFile, writeFile } from "node:fs/promises";
|
||||
import { tmpdir } from "node:os";
|
||||
import path from "node:path";
|
||||
|
||||
@ -8,6 +8,7 @@ import {
|
||||
loadTrajectoryFromDisk,
|
||||
nextResultFilename,
|
||||
normalizeRubric,
|
||||
writeTrajectoryDir,
|
||||
} from "../../lib/v3/verifier/trajectory.js";
|
||||
|
||||
describe("verifier trajectory utilities", () => {
|
||||
@ -63,8 +64,67 @@ describe("verifier trajectory utilities", () => {
|
||||
it("loads trajectory screenshots and image modalities from disk", async () => {
|
||||
const dir = await mkdtemp(path.join(tmpdir(), "stagehand-verifier-"));
|
||||
const screenshot = Buffer.from("probe screenshot");
|
||||
const finalScreenshot = Buffer.from("final screenshot");
|
||||
const agentImage = Buffer.from("agent image");
|
||||
await writeFile(path.join(dir, "screenshot_1.png"), screenshot);
|
||||
await writeFile(path.join(dir, "final.png"), finalScreenshot);
|
||||
await mkdir(path.join(dir, "screenshots", "agent"), { recursive: true });
|
||||
await writeFile(
|
||||
path.join(dir, "screenshots", "agent", "1.png"),
|
||||
agentImage,
|
||||
);
|
||||
await writeFile(
|
||||
path.join(dir, "trajectory.json"),
|
||||
JSON.stringify({
|
||||
task: { id: "task", instruction: "Do the task" },
|
||||
status: "complete",
|
||||
usage: { input_tokens: 0, output_tokens: 0 },
|
||||
timing: {
|
||||
startedAt: new Date(0).toISOString(),
|
||||
endedAt: new Date(0).toISOString(),
|
||||
},
|
||||
steps: [
|
||||
{
|
||||
index: 0,
|
||||
actionName: "act",
|
||||
actionArgs: {},
|
||||
reasoning: "",
|
||||
agentEvidence: {
|
||||
modalities: [
|
||||
{
|
||||
type: "image",
|
||||
mediaType: "image/png",
|
||||
imagePath: "screenshots/agent/1.png",
|
||||
},
|
||||
],
|
||||
},
|
||||
probeEvidence: { screenshotPath: "screenshot_1.png" },
|
||||
toolOutput: { ok: true, result: null },
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
},
|
||||
],
|
||||
finalObservation: {
|
||||
url: "https://example.com/done",
|
||||
screenshotPath: "final.png",
|
||||
},
|
||||
}),
|
||||
);
|
||||
|
||||
const trajectory = await loadTrajectoryFromDisk(dir);
|
||||
const modality = trajectory.steps[0].agentEvidence.modalities[0];
|
||||
|
||||
expect(trajectory.steps[0].probeEvidence.screenshot).toEqual(screenshot);
|
||||
expect(trajectory.finalObservation?.screenshot).toEqual(finalScreenshot);
|
||||
expect(modality.type).toBe("image");
|
||||
if (modality.type === "image") {
|
||||
expect(modality.bytes).toEqual(agentImage);
|
||||
}
|
||||
});
|
||||
|
||||
it("loads legacy base64 image modalities from disk", async () => {
|
||||
const dir = await mkdtemp(path.join(tmpdir(), "stagehand-verifier-"));
|
||||
const agentImage = Buffer.from("legacy agent image");
|
||||
await writeFile(
|
||||
path.join(dir, "trajectory.json"),
|
||||
JSON.stringify({
|
||||
@ -90,7 +150,7 @@ describe("verifier trajectory utilities", () => {
|
||||
},
|
||||
],
|
||||
},
|
||||
probeEvidence: { screenshotPath: "screenshot_1.png" },
|
||||
probeEvidence: {},
|
||||
toolOutput: { ok: true, result: null },
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
@ -102,13 +162,83 @@ describe("verifier trajectory utilities", () => {
|
||||
const trajectory = await loadTrajectoryFromDisk(dir);
|
||||
const modality = trajectory.steps[0].agentEvidence.modalities[0];
|
||||
|
||||
expect(trajectory.steps[0].probeEvidence.screenshot).toEqual(screenshot);
|
||||
expect(modality.type).toBe("image");
|
||||
if (modality.type === "image") {
|
||||
expect(modality.bytes).toEqual(agentImage);
|
||||
}
|
||||
});
|
||||
|
||||
it("redacts inline screenshot payloads when writing trajectories", async () => {
|
||||
const dir = await mkdtemp(path.join(tmpdir(), "stagehand-verifier-"));
|
||||
const inlineScreenshot =
|
||||
Buffer.from("inline screenshot").toString("base64");
|
||||
|
||||
await writeTrajectoryDir(dir, {
|
||||
task: { id: "task", instruction: "Do the task" },
|
||||
status: "complete",
|
||||
usage: { input_tokens: 0, output_tokens: 0 },
|
||||
timing: {
|
||||
startedAt: new Date(0).toISOString(),
|
||||
endedAt: new Date(0).toISOString(),
|
||||
},
|
||||
finalObservation: {
|
||||
url: "https://example.com/done",
|
||||
screenshot: Buffer.from("final screenshot"),
|
||||
},
|
||||
steps: [
|
||||
{
|
||||
index: 0,
|
||||
actionName: "click",
|
||||
actionArgs: {},
|
||||
reasoning: "",
|
||||
agentEvidence: {
|
||||
modalities: [
|
||||
{
|
||||
type: "json",
|
||||
content: {
|
||||
output: {
|
||||
success: true,
|
||||
screenshotBase64: inlineScreenshot,
|
||||
},
|
||||
},
|
||||
},
|
||||
],
|
||||
},
|
||||
probeEvidence: {},
|
||||
toolOutput: {
|
||||
ok: true,
|
||||
result: {
|
||||
output: {
|
||||
success: true,
|
||||
screenshotBase64: inlineScreenshot,
|
||||
},
|
||||
},
|
||||
},
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
},
|
||||
],
|
||||
});
|
||||
|
||||
const raw = await readFile(path.join(dir, "trajectory.json"), "utf8");
|
||||
const trajectory = JSON.parse(raw);
|
||||
|
||||
expect(raw).not.toContain(inlineScreenshot);
|
||||
expect(
|
||||
trajectory.steps[0].agentEvidence.modalities[0].content.output
|
||||
.screenshotBase64,
|
||||
).toBe("[redacted inline image payload]");
|
||||
expect(trajectory.steps[0].toolOutput.result.output.screenshotBase64).toBe(
|
||||
"[redacted inline image payload]",
|
||||
);
|
||||
expect(trajectory.finalObservation.screenshotPath).toBe(
|
||||
"screenshots/probe/final.png",
|
||||
);
|
||||
await expect(
|
||||
readFile(path.join(dir, "screenshots", "probe", "final.png")),
|
||||
).resolves.toEqual(Buffer.from("final screenshot"));
|
||||
});
|
||||
|
||||
it("rejects screenshot paths outside the trajectory directory", async () => {
|
||||
const dir = await mkdtemp(path.join(tmpdir(), "stagehand-verifier-"));
|
||||
await writeFile(
|
||||
|
||||
305
packages/evals/framework/trajectoryRecorder.ts
Normal file
305
packages/evals/framework/trajectoryRecorder.ts
Normal file
@ -0,0 +1,305 @@
|
||||
import fs from "node:fs/promises";
|
||||
import path from "node:path";
|
||||
import {
|
||||
buildAgentEvidenceFromStepFinished,
|
||||
mergeAgentEvidence,
|
||||
redactInlineImagePayloads,
|
||||
shouldPersistTrajectory,
|
||||
writeTrajectoryDir,
|
||||
} from "@browserbasehq/stagehand";
|
||||
import type {
|
||||
AgentEvidence,
|
||||
AgentEvidenceEvent,
|
||||
AgentFinalAnswerEvent,
|
||||
AgentScreenshotEvidenceEvent,
|
||||
AgentStepFinishedEvent,
|
||||
AgentStepObservedEvent,
|
||||
ProbeEvidence,
|
||||
TaskSpec,
|
||||
Trajectory,
|
||||
TrajectoryStatus,
|
||||
TrajectoryStep,
|
||||
TrajectoryUsage,
|
||||
EvaluationResult,
|
||||
} from "@browserbasehq/stagehand";
|
||||
|
||||
interface PartialStep {
|
||||
index: number;
|
||||
actionName: string;
|
||||
actionArgs: Record<string, unknown>;
|
||||
reasoning: string;
|
||||
agentEvidence: AgentEvidence;
|
||||
probeEvidence: ProbeEvidence;
|
||||
toolOutput: { ok: boolean; result: unknown; error?: string };
|
||||
startedAt: string;
|
||||
finishedAt: string;
|
||||
}
|
||||
|
||||
export interface TrajectoryRecorderOptions {
|
||||
taskSpec: TaskSpec;
|
||||
/**
|
||||
* Root directory under which trajectory dirs are written. Each task run
|
||||
* gets a subdirectory named by runId/task.id.
|
||||
* Defaults to `<cwd>/.trajectories`.
|
||||
*/
|
||||
outputRoot?: string;
|
||||
/** Run identifier (e.g., ISO timestamp + env). Defaults to a fresh timestamp. */
|
||||
runId?: string;
|
||||
/**
|
||||
* Override the env-gated persistence default. `true` always persists,
|
||||
* `false` never does, `undefined` defers to VERIFIER_PERSIST_TRAJECTORIES.
|
||||
*/
|
||||
persist?: boolean;
|
||||
}
|
||||
|
||||
export interface TrajectoryFinishOptions {
|
||||
status: TrajectoryStatus;
|
||||
finalAnswer?: string;
|
||||
usage?: Partial<TrajectoryUsage>;
|
||||
}
|
||||
|
||||
const ZERO_USAGE: TrajectoryUsage = {
|
||||
input_tokens: 0,
|
||||
output_tokens: 0,
|
||||
};
|
||||
|
||||
export class TrajectoryRecorder {
|
||||
private readonly taskSpec: TaskSpec;
|
||||
private readonly runId: string;
|
||||
private readonly outputDir: string;
|
||||
private readonly persistEnabled: boolean;
|
||||
|
||||
// Events can arrive out-of-order across step indices; same-step events all
|
||||
// fire in one microtask.
|
||||
private readonly partialSteps = new Map<number, Partial<PartialStep>>();
|
||||
private finalAnswerEvent?: AgentFinalAnswerEvent;
|
||||
private finalObservation?: ProbeEvidence;
|
||||
private startedAt = "";
|
||||
private endedAt = "";
|
||||
|
||||
private onScreenshot(e: AgentScreenshotEvidenceEvent): void {
|
||||
const partial = this.ensurePartial(e.stepIndex);
|
||||
|
||||
// Default to probe when the emit site doesn't tag a role: matches
|
||||
// v3AgentHandler's post-step screenshot. For CUA the pre-action shot is
|
||||
// NOT a probe — emitCuaActionStep fills that role post-action.
|
||||
const role = e.evidenceRole ?? "probe";
|
||||
|
||||
if (role === "probe") {
|
||||
const probe: ProbeEvidence = { ...(partial.probeEvidence ?? {}) };
|
||||
probe.screenshot = e.screenshot;
|
||||
probe.url = e.url;
|
||||
partial.probeEvidence = probe;
|
||||
} else if (!partial.probeEvidence?.url) {
|
||||
// Capture URL even for tier-1-only events; a later post-action URL
|
||||
// can still overwrite it.
|
||||
partial.probeEvidence = {
|
||||
...(partial.probeEvidence ?? {}),
|
||||
url: e.url,
|
||||
};
|
||||
}
|
||||
|
||||
if (role === "agent") {
|
||||
partial.agentEvidence = mergeAgentEvidence(partial.agentEvidence, {
|
||||
modalities: [
|
||||
{ type: "image", bytes: e.screenshot, mediaType: "image/png" },
|
||||
],
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
private onStepFinished(e: AgentStepFinishedEvent): void {
|
||||
const partial = this.ensurePartial(e.stepIndex);
|
||||
partial.actionName = e.actionName;
|
||||
partial.actionArgs = e.actionArgs;
|
||||
partial.reasoning = e.reasoning;
|
||||
partial.toolOutput = {
|
||||
...e.toolOutput,
|
||||
result: redactInlineImagePayloads(e.toolOutput.result, e.actionName),
|
||||
};
|
||||
partial.startedAt = e.startedAt ?? e.finishedAt;
|
||||
partial.finishedAt = e.finishedAt;
|
||||
partial.agentEvidence = mergeAgentEvidence(
|
||||
partial.agentEvidence,
|
||||
buildAgentEvidenceFromStepFinished(e),
|
||||
);
|
||||
}
|
||||
|
||||
private onStepObserved(e: AgentStepObservedEvent): void {
|
||||
const partial = this.ensurePartial(e.stepIndex);
|
||||
const probe: ProbeEvidence = { ...(partial.probeEvidence ?? {}) };
|
||||
probe.url = e.url;
|
||||
if (e.ariaTree !== undefined) probe.ariaTree = e.ariaTree;
|
||||
if (e.scroll !== undefined) probe.scroll = e.scroll;
|
||||
partial.probeEvidence = probe;
|
||||
}
|
||||
|
||||
private onFinalAnswer(e: AgentFinalAnswerEvent): void {
|
||||
this.finalAnswerEvent = e;
|
||||
if (e.observation) {
|
||||
this.finalObservation = {
|
||||
url: e.observation.url,
|
||||
...(e.observation.screenshot
|
||||
? { screenshot: e.observation.screenshot }
|
||||
: {}),
|
||||
...(e.observation.ariaTree !== undefined
|
||||
? { ariaTree: e.observation.ariaTree }
|
||||
: {}),
|
||||
...(e.observation.scroll !== undefined
|
||||
? { scroll: e.observation.scroll }
|
||||
: {}),
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
constructor(opts: TrajectoryRecorderOptions) {
|
||||
this.taskSpec = opts.taskSpec;
|
||||
this.runId = opts.runId ?? new Date().toISOString().replace(/[:.]/g, "-");
|
||||
const root = opts.outputRoot ?? path.join(process.cwd(), ".trajectories");
|
||||
this.outputDir = path.join(root, this.runId, opts.taskSpec.id);
|
||||
this.persistEnabled = shouldPersistTrajectory(opts.persist);
|
||||
}
|
||||
|
||||
/** Mark the beginning of collection. Call once before agent.execute(). */
|
||||
start(): void {
|
||||
if (this.startedAt) return;
|
||||
this.startedAt = new Date().toISOString();
|
||||
}
|
||||
|
||||
/** Ingest an evidence callback event from agent.execute(). */
|
||||
record(event: AgentEvidenceEvent): void {
|
||||
if (!this.startedAt) this.start();
|
||||
switch (event.type) {
|
||||
case "screenshot":
|
||||
this.onScreenshot(event);
|
||||
break;
|
||||
case "step_finished":
|
||||
this.onStepFinished(event);
|
||||
break;
|
||||
case "step_observed":
|
||||
this.onStepObserved(event);
|
||||
break;
|
||||
case "final_answer":
|
||||
this.onFinalAnswer(event);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Detach listeners, assemble the Trajectory, and (if persistence is on)
|
||||
* write the on-disk layout. Idempotent.
|
||||
*/
|
||||
async finish(opts: TrajectoryFinishOptions): Promise<Trajectory> {
|
||||
if (!this.startedAt) this.start();
|
||||
this.endedAt = new Date().toISOString();
|
||||
|
||||
const steps = this.assembleSteps();
|
||||
const trajectory: Trajectory = {
|
||||
task: this.taskSpec,
|
||||
steps,
|
||||
finalAnswer: opts.finalAnswer ?? this.finalAnswerEvent?.message,
|
||||
...(this.finalObservation
|
||||
? { finalObservation: this.finalObservation }
|
||||
: {}),
|
||||
status: opts.status,
|
||||
usage: { ...ZERO_USAGE, ...(opts.usage ?? {}) },
|
||||
timing: { startedAt: this.startedAt, endedAt: this.endedAt },
|
||||
};
|
||||
|
||||
if (this.persistEnabled) {
|
||||
await writeTrajectoryDir(this.outputDir, trajectory);
|
||||
}
|
||||
|
||||
return trajectory;
|
||||
}
|
||||
|
||||
/** Throw away in-memory state without writing to disk. Used on early abort. */
|
||||
cancel(): void {
|
||||
this.partialSteps.clear();
|
||||
this.finalAnswerEvent = undefined;
|
||||
this.finalObservation = undefined;
|
||||
}
|
||||
|
||||
/** Where the trajectory dir lives (whether or not it was persisted). */
|
||||
get directory(): string {
|
||||
return this.outputDir;
|
||||
}
|
||||
|
||||
/** Whether this recorder wrote the trajectory directory on finish(). */
|
||||
get persisted(): boolean {
|
||||
return this.persistEnabled;
|
||||
}
|
||||
|
||||
/**
|
||||
* Persist evaluator result next to the trajectory. No-op when trajectory
|
||||
* persistence is disabled.
|
||||
*/
|
||||
async persistResult(
|
||||
result: EvaluationResult,
|
||||
filename = "result.json",
|
||||
): Promise<void> {
|
||||
if (!this.persistEnabled) return;
|
||||
|
||||
const scoresDir = path.join(this.outputDir, "scores");
|
||||
await fs.mkdir(scoresDir, { recursive: true });
|
||||
await fs.writeFile(
|
||||
path.join(scoresDir, filename),
|
||||
JSON.stringify(result, null, 2),
|
||||
);
|
||||
|
||||
const taskDataPath = path.join(this.outputDir, "task_data.json");
|
||||
let taskData: Record<string, unknown>;
|
||||
try {
|
||||
taskData = JSON.parse(await fs.readFile(taskDataPath, "utf8")) as Record<
|
||||
string,
|
||||
unknown
|
||||
>;
|
||||
} catch {
|
||||
taskData = { task: this.taskSpec };
|
||||
}
|
||||
await fs.writeFile(
|
||||
taskDataPath,
|
||||
JSON.stringify({ ...taskData, result }, null, 2),
|
||||
);
|
||||
}
|
||||
|
||||
private ensurePartial(stepIndex: number): Partial<PartialStep> {
|
||||
let p = this.partialSteps.get(stepIndex);
|
||||
if (!p) {
|
||||
p = { index: stepIndex };
|
||||
this.partialSteps.set(stepIndex, p);
|
||||
}
|
||||
return p;
|
||||
}
|
||||
|
||||
private assembleSteps(): TrajectoryStep[] {
|
||||
const out: TrajectoryStep[] = [];
|
||||
const indices = [...this.partialSteps.keys()].sort((a, b) => a - b);
|
||||
for (const i of indices) {
|
||||
const p = this.partialSteps.get(i)!;
|
||||
if (
|
||||
p.actionName === undefined ||
|
||||
p.toolOutput === undefined ||
|
||||
p.startedAt === undefined ||
|
||||
p.finishedAt === undefined
|
||||
) {
|
||||
// Provider-only screenshot refreshes are transport evidence for the
|
||||
// next CUA action. If no action arrives for this index, there is no
|
||||
// completed trajectory step to persist.
|
||||
continue;
|
||||
}
|
||||
out.push({
|
||||
index: i,
|
||||
actionName: p.actionName,
|
||||
actionArgs: p.actionArgs ?? {},
|
||||
reasoning: p.reasoning ?? "",
|
||||
agentEvidence: p.agentEvidence ?? { modalities: [] },
|
||||
probeEvidence: p.probeEvidence ?? {},
|
||||
toolOutput: p.toolOutput,
|
||||
startedAt: p.startedAt,
|
||||
finishedAt: p.finishedAt,
|
||||
});
|
||||
}
|
||||
return out;
|
||||
}
|
||||
}
|
||||
327
packages/evals/tests/framework/trajectoryRecorder.test.ts
Normal file
327
packages/evals/tests/framework/trajectoryRecorder.test.ts
Normal file
@ -0,0 +1,327 @@
|
||||
import fs from "node:fs/promises";
|
||||
import os from "node:os";
|
||||
import path from "node:path";
|
||||
|
||||
import { afterEach, describe, expect, it } from "vitest";
|
||||
import type { TaskSpec } from "@browserbasehq/stagehand";
|
||||
|
||||
import { TrajectoryRecorder } from "../../framework/trajectoryRecorder.js";
|
||||
|
||||
const tempDirs: string[] = [];
|
||||
|
||||
afterEach(async () => {
|
||||
while (tempDirs.length > 0) {
|
||||
const dir = tempDirs.pop();
|
||||
if (dir) await fs.rm(dir, { recursive: true, force: true });
|
||||
}
|
||||
});
|
||||
|
||||
function makeTempDir(): Promise<string> {
|
||||
return fs
|
||||
.mkdtemp(path.join(os.tmpdir(), "trajectory-recorder-"))
|
||||
.then((dir) => {
|
||||
tempDirs.push(dir);
|
||||
return dir;
|
||||
});
|
||||
}
|
||||
|
||||
function makeTaskSpec(): TaskSpec {
|
||||
return {
|
||||
id: "recorder-task",
|
||||
instruction: "Compare economy and business fares.",
|
||||
initUrl: "https://example.com",
|
||||
precomputedRubric: {
|
||||
items: [
|
||||
{
|
||||
criterion: "Report fare delta",
|
||||
description: "Report the difference between two fares.",
|
||||
maxPoints: 1,
|
||||
},
|
||||
],
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
describe("TrajectoryRecorder", () => {
|
||||
it("assembles trajectory evidence from callback events", async () => {
|
||||
const recorder = new TrajectoryRecorder({
|
||||
taskSpec: makeTaskSpec(),
|
||||
persist: false,
|
||||
});
|
||||
const screenshot = Buffer.from("screen-1");
|
||||
|
||||
recorder.start();
|
||||
recorder.record({
|
||||
type: "screenshot",
|
||||
stepIndex: 0,
|
||||
screenshot,
|
||||
url: "https://example.com/search",
|
||||
evidenceRole: "agent",
|
||||
});
|
||||
recorder.record({
|
||||
type: "screenshot",
|
||||
stepIndex: 0,
|
||||
screenshot,
|
||||
url: "https://example.com/search",
|
||||
evidenceRole: "probe",
|
||||
});
|
||||
recorder.record({
|
||||
type: "step_finished",
|
||||
stepIndex: 0,
|
||||
actionName: "extract",
|
||||
actionArgs: { instruction: "Read fares" },
|
||||
reasoning: "Read visible fare cells.",
|
||||
toolOutput: {
|
||||
ok: true,
|
||||
result: { economy: "$100", business: "$250" },
|
||||
},
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
});
|
||||
recorder.record({
|
||||
type: "step_observed",
|
||||
stepIndex: 0,
|
||||
url: "https://example.com/search",
|
||||
ariaTree: "RootWebArea\nStaticText: Economy $100",
|
||||
});
|
||||
recorder.record({
|
||||
type: "final_answer",
|
||||
message: "Business is $150 more than economy.",
|
||||
observation: {
|
||||
url: "https://example.com/checkout",
|
||||
screenshot: Buffer.from("final-screen"),
|
||||
ariaTree: "RootWebArea\nStaticText: Complete",
|
||||
},
|
||||
});
|
||||
|
||||
const trajectory = await recorder.finish({
|
||||
status: "complete",
|
||||
usage: { input_tokens: 10, output_tokens: 5 },
|
||||
});
|
||||
|
||||
expect(trajectory.steps).toHaveLength(1);
|
||||
expect(trajectory.steps[0]).toMatchObject({
|
||||
index: 0,
|
||||
actionName: "extract",
|
||||
actionArgs: { instruction: "Read fares" },
|
||||
reasoning: "Read visible fare cells.",
|
||||
toolOutput: {
|
||||
ok: true,
|
||||
result: { economy: "$100", business: "$250" },
|
||||
},
|
||||
probeEvidence: {
|
||||
url: "https://example.com/search",
|
||||
ariaTree: "RootWebArea\nStaticText: Economy $100",
|
||||
},
|
||||
});
|
||||
expect(trajectory.steps[0].probeEvidence.screenshot).toEqual(screenshot);
|
||||
expect(trajectory.steps[0].startedAt).toBe(new Date(0).toISOString());
|
||||
expect(trajectory.steps[0].agentEvidence.modalities).toEqual(
|
||||
expect.arrayContaining([
|
||||
{ type: "image", bytes: screenshot, mediaType: "image/png" },
|
||||
{ type: "text", content: "Read visible fare cells." },
|
||||
{ type: "json", content: { economy: "$100", business: "$250" } },
|
||||
]),
|
||||
);
|
||||
expect(trajectory.finalAnswer).toBe("Business is $150 more than economy.");
|
||||
expect(trajectory.finalObservation).toMatchObject({
|
||||
url: "https://example.com/checkout",
|
||||
ariaTree: "RootWebArea\nStaticText: Complete",
|
||||
});
|
||||
expect(trajectory.finalObservation?.screenshot).toEqual(
|
||||
Buffer.from("final-screen"),
|
||||
);
|
||||
});
|
||||
|
||||
it("persists trajectory files and evaluator results", async () => {
|
||||
const outputRoot = await makeTempDir();
|
||||
const recorder = new TrajectoryRecorder({
|
||||
taskSpec: makeTaskSpec(),
|
||||
outputRoot,
|
||||
runId: "run-1",
|
||||
persist: true,
|
||||
});
|
||||
const screenshot = Buffer.from("screen-1");
|
||||
|
||||
recorder.start();
|
||||
recorder.record({
|
||||
type: "screenshot",
|
||||
stepIndex: 0,
|
||||
screenshot,
|
||||
url: "https://example.com/search",
|
||||
evidenceRole: "agent",
|
||||
});
|
||||
recorder.record({
|
||||
type: "screenshot",
|
||||
stepIndex: 0,
|
||||
screenshot,
|
||||
url: "https://example.com/search",
|
||||
evidenceRole: "probe",
|
||||
});
|
||||
recorder.record({
|
||||
type: "step_finished",
|
||||
stepIndex: 0,
|
||||
actionName: "act",
|
||||
actionArgs: { instruction: "Search fares" },
|
||||
reasoning: "Search for fares.",
|
||||
toolOutput: { ok: true, result: "done" },
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
});
|
||||
recorder.record({
|
||||
type: "step_observed",
|
||||
stepIndex: 0,
|
||||
url: "https://example.com/search",
|
||||
});
|
||||
recorder.record({
|
||||
type: "final_answer",
|
||||
message: "Complete.",
|
||||
observation: {
|
||||
url: "https://example.com/complete",
|
||||
screenshot: Buffer.from("final-screen"),
|
||||
},
|
||||
});
|
||||
|
||||
await recorder.finish({ status: "complete" });
|
||||
await recorder.persistResult({
|
||||
outcomeSuccess: true,
|
||||
explanation: "The task was completed.",
|
||||
});
|
||||
|
||||
const taskDir = path.join(outputRoot, "run-1", "recorder-task");
|
||||
await expect(fs.readdir(taskDir)).resolves.toEqual(
|
||||
expect.arrayContaining([
|
||||
"core.log",
|
||||
"scores",
|
||||
"screenshots",
|
||||
"task_data.json",
|
||||
"times.json",
|
||||
"trajectory.json",
|
||||
]),
|
||||
);
|
||||
await expect(
|
||||
fs.readFile(path.join(taskDir, "screenshots", "probe", "1.png")),
|
||||
).resolves.toEqual(screenshot);
|
||||
await expect(
|
||||
fs.readFile(path.join(taskDir, "screenshots", "probe", "final.png")),
|
||||
).resolves.toEqual(Buffer.from("final-screen"));
|
||||
await expect(
|
||||
fs.readFile(path.join(taskDir, "screenshots", "agent", "1.png")),
|
||||
).resolves.toEqual(screenshot);
|
||||
await expect(
|
||||
fs.readFile(path.join(taskDir, "scores", "result.json"), "utf8"),
|
||||
).resolves.toContain('"outcomeSuccess": true');
|
||||
|
||||
const trajectory = JSON.parse(
|
||||
await fs.readFile(path.join(taskDir, "trajectory.json"), "utf8"),
|
||||
);
|
||||
expect(trajectory.steps[0].probeEvidence.screenshotPath).toBe(
|
||||
"screenshots/probe/1.png",
|
||||
);
|
||||
expect(trajectory.finalObservation.screenshotPath).toBe(
|
||||
"screenshots/probe/final.png",
|
||||
);
|
||||
expect(trajectory.steps[0].agentEvidence.modalities).toContainEqual({
|
||||
type: "image",
|
||||
imagePath: "screenshots/agent/1.png",
|
||||
mediaType: "image/png",
|
||||
});
|
||||
|
||||
const taskData = JSON.parse(
|
||||
await fs.readFile(path.join(taskDir, "task_data.json"), "utf8"),
|
||||
);
|
||||
expect(taskData.result).toMatchObject({
|
||||
outcomeSuccess: true,
|
||||
explanation: "The task was completed.",
|
||||
});
|
||||
});
|
||||
|
||||
it("normalizes missing step startedAt to finishedAt", async () => {
|
||||
const recorder = new TrajectoryRecorder({
|
||||
taskSpec: makeTaskSpec(),
|
||||
persist: false,
|
||||
});
|
||||
const finishedAt = new Date(1).toISOString();
|
||||
|
||||
recorder.record({
|
||||
type: "step_finished",
|
||||
stepIndex: 0,
|
||||
actionName: "extract",
|
||||
actionArgs: { instruction: "Read fares" },
|
||||
reasoning: "",
|
||||
toolOutput: { ok: true, result: false },
|
||||
finishedAt,
|
||||
});
|
||||
|
||||
const trajectory = await recorder.finish({ status: "complete" });
|
||||
|
||||
expect(trajectory.steps[0].startedAt).toBe(finishedAt);
|
||||
expect(trajectory.steps[0].finishedAt).toBe(finishedAt);
|
||||
});
|
||||
|
||||
it("lifts inline screenshot payloads into image evidence and redacts JSON", async () => {
|
||||
const inlineScreenshot =
|
||||
Buffer.from("inline screenshot").toString("base64");
|
||||
const recorder = new TrajectoryRecorder({
|
||||
taskSpec: makeTaskSpec(),
|
||||
persist: false,
|
||||
});
|
||||
|
||||
recorder.record({
|
||||
type: "step_finished",
|
||||
stepIndex: 0,
|
||||
actionName: "click",
|
||||
actionArgs: { describe: "Open fare details" },
|
||||
reasoning: "Click the fare details button.",
|
||||
toolOutput: {
|
||||
ok: true,
|
||||
result: {
|
||||
output: {
|
||||
success: true,
|
||||
describe: "Open fare details",
|
||||
screenshotBase64: inlineScreenshot,
|
||||
},
|
||||
},
|
||||
},
|
||||
startedAt: new Date(0).toISOString(),
|
||||
finishedAt: new Date(0).toISOString(),
|
||||
});
|
||||
|
||||
const trajectory = await recorder.finish({ status: "complete" });
|
||||
const step = trajectory.steps[0];
|
||||
const rawTrajectory = JSON.stringify(trajectory);
|
||||
const imageModalities = step.agentEvidence.modalities.filter(
|
||||
(m) => m.type === "image",
|
||||
);
|
||||
const jsonModality = step.agentEvidence.modalities.find(
|
||||
(m) => m.type === "json",
|
||||
);
|
||||
|
||||
expect(rawTrajectory).not.toContain(inlineScreenshot);
|
||||
expect(step.toolOutput.result).toMatchObject({
|
||||
output: {
|
||||
success: true,
|
||||
describe: "Open fare details",
|
||||
screenshotBase64: "[redacted inline image payload]",
|
||||
},
|
||||
});
|
||||
expect(jsonModality).toMatchObject({
|
||||
type: "json",
|
||||
content: {
|
||||
output: {
|
||||
screenshotBase64: "[redacted inline image payload]",
|
||||
},
|
||||
},
|
||||
});
|
||||
expect(imageModalities).toHaveLength(1);
|
||||
expect(imageModalities[0]).toMatchObject({
|
||||
type: "image",
|
||||
mediaType: "image/png",
|
||||
});
|
||||
if (imageModalities[0].type === "image") {
|
||||
expect(imageModalities[0].bytes).toEqual(
|
||||
Buffer.from(inlineScreenshot, "base64"),
|
||||
);
|
||||
}
|
||||
});
|
||||
});
|
||||
Loading…
x
Reference in New Issue
Block a user