@vitest-evals/harness-pi-ai 0.9.0-beta.0 → 0.9.0-beta.2
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.
- package/README.md +41 -7
- package/dist/index.d.mts +6 -3
- package/dist/index.d.ts +6 -3
- package/dist/index.js +56 -5
- package/dist/index.js.map +1 -1
- package/dist/index.mjs +56 -5
- package/dist/index.mjs.map +1 -1
- package/package.json +2 -2
package/README.md
CHANGED
|
@@ -11,13 +11,36 @@ npm install -D vitest-evals @vitest-evals/harness-pi-ai
|
|
|
11
11
|
## Usage
|
|
12
12
|
|
|
13
13
|
```ts
|
|
14
|
+
import { expect } from "vitest";
|
|
14
15
|
import { piAiHarness } from "@vitest-evals/harness-pi-ai";
|
|
16
|
+
import { describeEval, toolCalls } from "vitest-evals";
|
|
15
17
|
|
|
16
18
|
const harness = piAiHarness({
|
|
17
19
|
createAgent: () => createRefundAgent(),
|
|
20
|
+
toolReplay: {
|
|
21
|
+
lookupInvoice: true,
|
|
22
|
+
},
|
|
23
|
+
prompt: sharedJudgePrompt,
|
|
24
|
+
});
|
|
25
|
+
|
|
26
|
+
describeEval("refund agent", { harness }, (it) => {
|
|
27
|
+
it("approves a refundable invoice", async ({ run }) => {
|
|
28
|
+
const result = await run("Refund invoice inv_123");
|
|
29
|
+
|
|
30
|
+
expect(result.output).toMatchObject({
|
|
31
|
+
status: "approved",
|
|
32
|
+
});
|
|
33
|
+
expect(toolCalls(result.session).map((call) => call.name)).toEqual([
|
|
34
|
+
"lookupInvoice",
|
|
35
|
+
"createRefund",
|
|
36
|
+
]);
|
|
37
|
+
});
|
|
18
38
|
});
|
|
19
39
|
```
|
|
20
40
|
|
|
41
|
+
`prompt` gives rubric or factuality judges the same provider/model setup
|
|
42
|
+
through `JudgeContext.harness.prompt`.
|
|
43
|
+
|
|
21
44
|
If the agent already exposes its own tools, the adapter will infer them from
|
|
22
45
|
the agent by default. If your existing Pi Mono agent already exposes its own
|
|
23
46
|
entrypoint, wire that up directly and let the harness provide the runtime
|
|
@@ -26,6 +49,7 @@ seam:
|
|
|
26
49
|
```ts
|
|
27
50
|
const harness = piAiHarness({
|
|
28
51
|
createAgent: () => createRefundAgent(),
|
|
52
|
+
prompt: sharedJudgePrompt,
|
|
29
53
|
run: ({ agent, input, runtime }) => agent.execute(input, runtime),
|
|
30
54
|
});
|
|
31
55
|
```
|
|
@@ -44,6 +68,7 @@ override:
|
|
|
44
68
|
```ts
|
|
45
69
|
const harness = piAiHarness({
|
|
46
70
|
createAgent: () => createRefundAgent(),
|
|
71
|
+
prompt: sharedJudgePrompt,
|
|
47
72
|
tools: hiddenAgentTools,
|
|
48
73
|
});
|
|
49
74
|
```
|
|
@@ -54,6 +79,7 @@ normalization hooks still exist under `normalize`:
|
|
|
54
79
|
```ts
|
|
55
80
|
const harness = piAiHarness({
|
|
56
81
|
createAgent: () => createWrappedRefundAgent(),
|
|
82
|
+
prompt: sharedJudgePrompt,
|
|
57
83
|
run: ({ agent, input, runtime }) => agent.run(input, runtime),
|
|
58
84
|
normalize: {
|
|
59
85
|
output: ({ result }) => result.customDecision,
|
|
@@ -64,9 +90,10 @@ const harness = piAiHarness({
|
|
|
64
90
|
The adapter provides:
|
|
65
91
|
|
|
66
92
|
- a runtime/tool injection seam for an existing agent
|
|
93
|
+
- a required prompt seam for LLM-backed judges
|
|
67
94
|
- normalized session capture from emitted events and wrapped tool calls
|
|
68
95
|
- usage/output inference for common `pi-ai`-style result objects
|
|
69
|
-
- opt-in tool replay/recording
|
|
96
|
+
- opt-in tool replay/recording from harness-level `toolReplay`
|
|
70
97
|
|
|
71
98
|
See the workspace demo in `apps/demo-pi`.
|
|
72
99
|
|
|
@@ -87,17 +114,24 @@ export default defineConfig({
|
|
|
87
114
|
});
|
|
88
115
|
```
|
|
89
116
|
|
|
90
|
-
Then opt individual tools into recording/replay:
|
|
117
|
+
Then opt individual tools into recording/replay from the harness:
|
|
91
118
|
|
|
92
119
|
```ts
|
|
93
|
-
const
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
120
|
+
const harness = piAiHarness({
|
|
121
|
+
createAgent: () => createRefundAgent(),
|
|
122
|
+
toolReplay: {
|
|
123
|
+
lookupInvoice: true,
|
|
97
124
|
},
|
|
98
|
-
|
|
125
|
+
prompt: sharedJudgePrompt,
|
|
126
|
+
});
|
|
99
127
|
```
|
|
100
128
|
|
|
129
|
+
When an agent exposes both a native Pi tool and a runtime tool with the same
|
|
130
|
+
name, a native tool call records in its own cassette namespace. Runtime calls of
|
|
131
|
+
that same name are treated as implementation details while the native tool is
|
|
132
|
+
executing, so delegated runtime calls do not create duplicate trace entries or
|
|
133
|
+
overwrite the native recording.
|
|
134
|
+
|
|
101
135
|
Supported modes:
|
|
102
136
|
|
|
103
137
|
- `off`: never read or write recordings
|
package/dist/index.d.mts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { NormalizedMessage, JsonValue, HarnessMetadata, HarnessContext, NormalizedSession, UsageSummary, TimingSummary, HarnessRun, Harness } from 'vitest-evals/harness';
|
|
1
|
+
import { NormalizedMessage, JsonValue, HarnessMetadata, HarnessContext, NormalizedSession, UsageSummary, TimingSummary, HarnessPrompt, HarnessRun, Harness } from 'vitest-evals/harness';
|
|
2
2
|
import { ReplayMode, ToolRecording, ToolReplayConfig } from 'vitest-evals/replay';
|
|
3
3
|
|
|
4
4
|
type MaybePromise<T> = T | Promise<T>;
|
|
@@ -20,9 +20,10 @@ interface PiAiToolContext<TInput = string, TMetadata extends HarnessMetadata = H
|
|
|
20
20
|
}
|
|
21
21
|
type PiAiToolRecording<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue> = ToolRecording<TArgs, TResult>;
|
|
22
22
|
type PiAiToolReplayConfig<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;
|
|
23
|
+
type PiAiToolReplayPolicy<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = boolean | PiAiToolReplayConfig<Record<string, JsonValue>, JsonValue, TInput, TMetadata>;
|
|
24
|
+
type PiAiToolReplayPolicies<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = Record<string, PiAiToolReplayPolicy<TInput, TMetadata>>;
|
|
23
25
|
interface PiAiToolDefinition<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> {
|
|
24
26
|
description?: string;
|
|
25
|
-
replay?: boolean | PiAiToolReplayConfig<TArgs, TResult, TInput, TMetadata>;
|
|
26
27
|
execute: (args: TArgs, context: PiAiToolContext<TInput, TMetadata>) => MaybePromise<TResult>;
|
|
27
28
|
}
|
|
28
29
|
type PiAiToolset<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = AnyPiAiToolset<TInput, TMetadata>;
|
|
@@ -47,7 +48,9 @@ interface PiAiHarnessResultArgs<TAgent, TInput, TMetadata extends HarnessMetadat
|
|
|
47
48
|
interface PiAiHarnessBaseOptions<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>> {
|
|
48
49
|
agent?: TAgent;
|
|
49
50
|
createAgent?: () => MaybePromise<TAgent>;
|
|
51
|
+
toolReplay?: PiAiToolReplayPolicies<TInput, TMetadata>;
|
|
50
52
|
normalize?: PiAiHarnessNormalizeOptions<TAgent, TInput, TMetadata, TResult, TTools>;
|
|
53
|
+
prompt: HarnessPrompt;
|
|
51
54
|
name?: string;
|
|
52
55
|
}
|
|
53
56
|
interface PiAiHarnessWithToolsOptions<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {
|
|
@@ -70,4 +73,4 @@ interface PiAiHarnessNormalizeOptions<TAgent, TInput = string, TMetadata extends
|
|
|
70
73
|
declare function piAiHarness<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>>(options: PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>): Harness<TInput, TMetadata>;
|
|
71
74
|
declare function piAiHarness<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown>(options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>): Harness<TInput, TMetadata>;
|
|
72
75
|
|
|
73
|
-
export { type PiAiEventSink, type PiAiHarnessInferredToolsOptions, type PiAiHarnessNormalizeOptions, type PiAiHarnessOptions, type PiAiHarnessResultArgs, type PiAiHarnessRunArgs, type PiAiHarnessWithToolsOptions, type PiAiReplayMode, type PiAiRuntime, type PiAiToolContext, type PiAiToolDefinition, type PiAiToolRecording, type PiAiToolReplayConfig, type PiAiToolset, piAiHarness };
|
|
76
|
+
export { type PiAiEventSink, type PiAiHarnessInferredToolsOptions, type PiAiHarnessNormalizeOptions, type PiAiHarnessOptions, type PiAiHarnessResultArgs, type PiAiHarnessRunArgs, type PiAiHarnessWithToolsOptions, type PiAiReplayMode, type PiAiRuntime, type PiAiToolContext, type PiAiToolDefinition, type PiAiToolRecording, type PiAiToolReplayConfig, type PiAiToolReplayPolicies, type PiAiToolReplayPolicy, type PiAiToolset, piAiHarness };
|
package/dist/index.d.ts
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { NormalizedMessage, JsonValue, HarnessMetadata, HarnessContext, NormalizedSession, UsageSummary, TimingSummary, HarnessRun, Harness } from 'vitest-evals/harness';
|
|
1
|
+
import { NormalizedMessage, JsonValue, HarnessMetadata, HarnessContext, NormalizedSession, UsageSummary, TimingSummary, HarnessPrompt, HarnessRun, Harness } from 'vitest-evals/harness';
|
|
2
2
|
import { ReplayMode, ToolRecording, ToolReplayConfig } from 'vitest-evals/replay';
|
|
3
3
|
|
|
4
4
|
type MaybePromise<T> = T | Promise<T>;
|
|
@@ -20,9 +20,10 @@ interface PiAiToolContext<TInput = string, TMetadata extends HarnessMetadata = H
|
|
|
20
20
|
}
|
|
21
21
|
type PiAiToolRecording<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue> = ToolRecording<TArgs, TResult>;
|
|
22
22
|
type PiAiToolReplayConfig<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;
|
|
23
|
+
type PiAiToolReplayPolicy<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = boolean | PiAiToolReplayConfig<Record<string, JsonValue>, JsonValue, TInput, TMetadata>;
|
|
24
|
+
type PiAiToolReplayPolicies<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = Record<string, PiAiToolReplayPolicy<TInput, TMetadata>>;
|
|
23
25
|
interface PiAiToolDefinition<TArgs extends Record<string, JsonValue> = Record<string, JsonValue>, TResult extends JsonValue = JsonValue, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> {
|
|
24
26
|
description?: string;
|
|
25
|
-
replay?: boolean | PiAiToolReplayConfig<TArgs, TResult, TInput, TMetadata>;
|
|
26
27
|
execute: (args: TArgs, context: PiAiToolContext<TInput, TMetadata>) => MaybePromise<TResult>;
|
|
27
28
|
}
|
|
28
29
|
type PiAiToolset<TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata> = AnyPiAiToolset<TInput, TMetadata>;
|
|
@@ -47,7 +48,9 @@ interface PiAiHarnessResultArgs<TAgent, TInput, TMetadata extends HarnessMetadat
|
|
|
47
48
|
interface PiAiHarnessBaseOptions<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>> {
|
|
48
49
|
agent?: TAgent;
|
|
49
50
|
createAgent?: () => MaybePromise<TAgent>;
|
|
51
|
+
toolReplay?: PiAiToolReplayPolicies<TInput, TMetadata>;
|
|
50
52
|
normalize?: PiAiHarnessNormalizeOptions<TAgent, TInput, TMetadata, TResult, TTools>;
|
|
53
|
+
prompt: HarnessPrompt;
|
|
51
54
|
name?: string;
|
|
52
55
|
}
|
|
53
56
|
interface PiAiHarnessWithToolsOptions<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {
|
|
@@ -70,4 +73,4 @@ interface PiAiHarnessNormalizeOptions<TAgent, TInput = string, TMetadata extends
|
|
|
70
73
|
declare function piAiHarness<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown, TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<TInput, TMetadata>>(options: PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>): Harness<TInput, TMetadata>;
|
|
71
74
|
declare function piAiHarness<TAgent, TInput = string, TMetadata extends HarnessMetadata = HarnessMetadata, TResult = unknown>(options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>): Harness<TInput, TMetadata>;
|
|
72
75
|
|
|
73
|
-
export { type PiAiEventSink, type PiAiHarnessInferredToolsOptions, type PiAiHarnessNormalizeOptions, type PiAiHarnessOptions, type PiAiHarnessResultArgs, type PiAiHarnessRunArgs, type PiAiHarnessWithToolsOptions, type PiAiReplayMode, type PiAiRuntime, type PiAiToolContext, type PiAiToolDefinition, type PiAiToolRecording, type PiAiToolReplayConfig, type PiAiToolset, piAiHarness };
|
|
76
|
+
export { type PiAiEventSink, type PiAiHarnessInferredToolsOptions, type PiAiHarnessNormalizeOptions, type PiAiHarnessOptions, type PiAiHarnessResultArgs, type PiAiHarnessRunArgs, type PiAiHarnessWithToolsOptions, type PiAiReplayMode, type PiAiRuntime, type PiAiToolContext, type PiAiToolDefinition, type PiAiToolRecording, type PiAiToolReplayConfig, type PiAiToolReplayPolicies, type PiAiToolReplayPolicy, type PiAiToolset, piAiHarness };
|
package/dist/index.js
CHANGED
|
@@ -29,6 +29,7 @@ var ORIGINAL_NATIVE_EXECUTE = Symbol("vitest-evals.originalNativeExecute");
|
|
|
29
29
|
function piAiHarness(options) {
|
|
30
30
|
return {
|
|
31
31
|
name: options.name ?? "pi-ai",
|
|
32
|
+
prompt: options.prompt,
|
|
32
33
|
run: async (input, context) => {
|
|
33
34
|
const agent = await resolveAgent(options);
|
|
34
35
|
const messages = [
|
|
@@ -64,10 +65,13 @@ function piAiHarness(options) {
|
|
|
64
65
|
};
|
|
65
66
|
}
|
|
66
67
|
async function executePiHarnessRun(options, agent, input, context, messages, runtimeTools, nativeToolsets) {
|
|
68
|
+
const executionState = createPiToolExecutionState();
|
|
67
69
|
const runtime = createRuntime({
|
|
68
70
|
input,
|
|
69
71
|
context,
|
|
70
72
|
tools: runtimeTools,
|
|
73
|
+
toolReplay: options.toolReplay,
|
|
74
|
+
executionState,
|
|
71
75
|
messages
|
|
72
76
|
});
|
|
73
77
|
try {
|
|
@@ -78,7 +82,9 @@ async function executePiHarnessRun(options, agent, input, context, messages, run
|
|
|
78
82
|
input,
|
|
79
83
|
context,
|
|
80
84
|
messages,
|
|
81
|
-
toolCalls: runtime.toolCalls
|
|
85
|
+
toolCalls: runtime.toolCalls,
|
|
86
|
+
toolReplay: options.toolReplay,
|
|
87
|
+
executionState
|
|
82
88
|
},
|
|
83
89
|
() => runAgent(options, {
|
|
84
90
|
agent,
|
|
@@ -256,12 +262,16 @@ async function withInstrumentedAgentTools(agent, toolsets, args, callback) {
|
|
|
256
262
|
signal: args.context.signal,
|
|
257
263
|
setArtifact: args.context.setArtifact
|
|
258
264
|
};
|
|
265
|
+
const leaveNativeTool = enterNativeToolExecution(
|
|
266
|
+
args.executionState,
|
|
267
|
+
tool.name
|
|
268
|
+
);
|
|
259
269
|
try {
|
|
260
270
|
const execution = await executeNativeToolWithReplay({
|
|
261
271
|
toolName: tool.name,
|
|
262
272
|
toolCallId,
|
|
263
273
|
execute: originalExecute,
|
|
264
|
-
replay: tool.
|
|
274
|
+
replay: args.toolReplay?.[tool.name],
|
|
265
275
|
args: rawArgs,
|
|
266
276
|
context: toolContext
|
|
267
277
|
});
|
|
@@ -305,6 +315,8 @@ async function withInstrumentedAgentTools(agent, toolsets, args, callback) {
|
|
|
305
315
|
toolCalls: [call]
|
|
306
316
|
});
|
|
307
317
|
throw error;
|
|
318
|
+
} finally {
|
|
319
|
+
leaveNativeTool();
|
|
308
320
|
}
|
|
309
321
|
};
|
|
310
322
|
instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;
|
|
@@ -381,6 +393,28 @@ function getNativeToolExecuteOrigin(execute) {
|
|
|
381
393
|
const nativeExecute = execute;
|
|
382
394
|
return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;
|
|
383
395
|
}
|
|
396
|
+
function createPiToolExecutionState() {
|
|
397
|
+
return {
|
|
398
|
+
activeNativeToolNames: /* @__PURE__ */ new Map()
|
|
399
|
+
};
|
|
400
|
+
}
|
|
401
|
+
function enterNativeToolExecution(state, toolName) {
|
|
402
|
+
state.activeNativeToolNames.set(
|
|
403
|
+
toolName,
|
|
404
|
+
(state.activeNativeToolNames.get(toolName) ?? 0) + 1
|
|
405
|
+
);
|
|
406
|
+
return () => {
|
|
407
|
+
const nextCount = (state.activeNativeToolNames.get(toolName) ?? 1) - 1;
|
|
408
|
+
if (nextCount <= 0) {
|
|
409
|
+
state.activeNativeToolNames.delete(toolName);
|
|
410
|
+
return;
|
|
411
|
+
}
|
|
412
|
+
state.activeNativeToolNames.set(toolName, nextCount);
|
|
413
|
+
};
|
|
414
|
+
}
|
|
415
|
+
function hasActiveNativeToolExecution(state, toolName) {
|
|
416
|
+
return (state.activeNativeToolNames.get(toolName) ?? 0) > 0;
|
|
417
|
+
}
|
|
384
418
|
async function executeNativeToolWithReplay({
|
|
385
419
|
toolName,
|
|
386
420
|
toolCallId,
|
|
@@ -392,7 +426,7 @@ async function executeNativeToolWithReplay({
|
|
|
392
426
|
let didExecute = false;
|
|
393
427
|
let liveResult;
|
|
394
428
|
const execution = await (0, import_replay.executeWithReplay)({
|
|
395
|
-
toolName,
|
|
429
|
+
toolName: createNativeReplayToolName(toolName),
|
|
396
430
|
args,
|
|
397
431
|
context,
|
|
398
432
|
execute: async (toolArgs) => {
|
|
@@ -414,10 +448,15 @@ async function executeNativeToolWithReplay({
|
|
|
414
448
|
replay: execution.replay
|
|
415
449
|
};
|
|
416
450
|
}
|
|
451
|
+
function createNativeReplayToolName(toolName) {
|
|
452
|
+
return `${toolName}.native`;
|
|
453
|
+
}
|
|
417
454
|
function createRuntime({
|
|
418
455
|
input,
|
|
419
456
|
context,
|
|
420
457
|
tools,
|
|
458
|
+
toolReplay,
|
|
459
|
+
executionState,
|
|
421
460
|
messages
|
|
422
461
|
}) {
|
|
423
462
|
const toolCalls = [];
|
|
@@ -462,6 +501,10 @@ function createRuntime({
|
|
|
462
501
|
toolName,
|
|
463
502
|
async (args) => {
|
|
464
503
|
const startedAt = /* @__PURE__ */ new Date();
|
|
504
|
+
const isNativeImplementationCall = hasActiveNativeToolExecution(
|
|
505
|
+
executionState,
|
|
506
|
+
toolName
|
|
507
|
+
);
|
|
465
508
|
const toolContext = {
|
|
466
509
|
input,
|
|
467
510
|
metadata: context.metadata,
|
|
@@ -472,10 +515,14 @@ function createRuntime({
|
|
|
472
515
|
const execution = await executeToolWithReplay({
|
|
473
516
|
toolName,
|
|
474
517
|
tool,
|
|
518
|
+
replay: isNativeImplementationCall ? void 0 : toolReplay?.[toolName],
|
|
475
519
|
args,
|
|
476
520
|
context: toolContext
|
|
477
521
|
});
|
|
478
522
|
const finishedAt = /* @__PURE__ */ new Date();
|
|
523
|
+
if (isNativeImplementationCall) {
|
|
524
|
+
return execution.result;
|
|
525
|
+
}
|
|
479
526
|
const call = {
|
|
480
527
|
name: toolName,
|
|
481
528
|
arguments: args,
|
|
@@ -500,6 +547,9 @@ function createRuntime({
|
|
|
500
547
|
return execution.result;
|
|
501
548
|
} catch (error) {
|
|
502
549
|
const finishedAt = /* @__PURE__ */ new Date();
|
|
550
|
+
if (isNativeImplementationCall) {
|
|
551
|
+
throw error;
|
|
552
|
+
}
|
|
503
553
|
const call = {
|
|
504
554
|
name: toolName,
|
|
505
555
|
arguments: args,
|
|
@@ -652,6 +702,7 @@ function resolveErrors(result) {
|
|
|
652
702
|
async function executeToolWithReplay({
|
|
653
703
|
toolName,
|
|
654
704
|
tool,
|
|
705
|
+
replay,
|
|
655
706
|
args,
|
|
656
707
|
context
|
|
657
708
|
}) {
|
|
@@ -659,8 +710,8 @@ async function executeToolWithReplay({
|
|
|
659
710
|
toolName,
|
|
660
711
|
args,
|
|
661
712
|
context,
|
|
662
|
-
execute: tool.execute,
|
|
663
|
-
replay
|
|
713
|
+
execute: (toolArgs, toolContext) => tool.execute(toolArgs, toolContext),
|
|
714
|
+
replay
|
|
664
715
|
});
|
|
665
716
|
}
|
|
666
717
|
// Annotate the CommonJS export names for ESM import in node:
|
package/dist/index.js.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import type {\n Harness,\n HarnessContext,\n HarnessMetadata,\n HarnessRun,\n JsonValue,\n NormalizedMessage,\n NormalizedSession,\n TimingSummary,\n ToolCallRecord,\n UsageSummary,\n} from \"vitest-evals/harness\";\nimport {\n attachHarnessRunToError,\n isHarnessRun,\n isNormalizedSession,\n normalizeContent,\n resolveHarnessRunErrors,\n serializeError,\n toJsonValue,\n} from \"vitest-evals/harness\";\nimport {\n executeWithReplay,\n getReplayMetadataFromError,\n normalizeReplayMetadata,\n} from \"vitest-evals/replay\";\nimport type {\n ReplayMode,\n ToolRecording,\n ToolReplayConfig,\n} from \"vitest-evals/replay\";\n\ntype MaybePromise<T> = T | Promise<T>;\ntype AnyPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolDefinition<any, any, TInput, TMetadata>>;\ntype InferredPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<\n string,\n PiAiToolDefinition<Record<string, JsonValue>, JsonValue, TInput, TMetadata>\n>;\n\ntype PiAgentToolLike<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n name: string;\n replay?:\n | boolean\n | PiAiToolReplayConfig<\n Record<string, JsonValue>,\n JsonValue,\n TInput,\n TMetadata\n >;\n execute: (toolCallId: string, args: Record<string, JsonValue>) => unknown;\n};\n\nconst ORIGINAL_NATIVE_EXECUTE = Symbol(\"vitest-evals.originalNativeExecute\");\n\ntype NativeToolExecute<\n TInput,\n TMetadata extends HarnessMetadata,\n> = PiAgentToolLike<TInput, TMetadata>[\"execute\"] & {\n [ORIGINAL_NATIVE_EXECUTE]?: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n};\n\nexport type PiAiReplayMode = ReplayMode;\n\nexport interface PiAiEventSink {\n message: (message: NormalizedMessage) => void;\n system: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n user: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n assistant: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n tool: (\n name: string,\n content: JsonValue,\n metadata?: Record<string, JsonValue>,\n ) => void;\n}\n\nexport interface PiAiToolContext<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n input: TInput;\n metadata: HarnessContext<TMetadata>[\"metadata\"];\n signal?: AbortSignal;\n setArtifact: HarnessContext<TMetadata>[\"setArtifact\"];\n}\n\nexport type PiAiToolRecording<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n> = ToolRecording<TArgs, TResult>;\n\nexport type PiAiToolReplayConfig<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;\n\nexport interface PiAiToolDefinition<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n description?: string;\n replay?: boolean | PiAiToolReplayConfig<TArgs, TResult, TInput, TMetadata>;\n execute: (\n args: TArgs,\n context: PiAiToolContext<TInput, TMetadata>,\n ) => MaybePromise<TResult>;\n}\n\nexport type PiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = AnyPiAiToolset<TInput, TMetadata>;\n\ntype ToolArgs<TTool> = TTool extends PiAiToolDefinition<\n infer TArgs,\n infer _TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TArgs\n : never;\n\ntype ToolResult<TTool> = TTool extends PiAiToolDefinition<\n infer _TArgs,\n infer TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TResult\n : never;\n\nexport type PiAiRuntime<\n TTools extends PiAiToolset<TInput, TMetadata>,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n tools: {\n [K in keyof TTools]: (\n args: ToolArgs<TTools[K]>,\n ) => Promise<ToolResult<TTools[K]>>;\n };\n events: PiAiEventSink;\n signal?: AbortSignal;\n};\n\nexport interface PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> {\n agent: TAgent;\n input: TInput;\n context: HarnessContext<TMetadata>;\n runtime: PiAiRuntime<TTools, TInput, TMetadata>;\n}\n\nexport interface PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> extends PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools> {\n result: TResult;\n}\n\ninterface PiAiHarnessBaseOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n agent?: TAgent;\n createAgent?: () => MaybePromise<TAgent>;\n normalize?: PiAiHarnessNormalizeOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n name?: string;\n}\n\nexport interface PiAiHarnessWithToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {\n tools: TTools;\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport interface PiAiHarnessInferredToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n> extends PiAiHarnessBaseOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n InferredPiAiToolset<TInput, TMetadata>\n > {\n tools?: undefined;\n run?: (\n args: PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata,\n InferredPiAiToolset<TInput, TMetadata>\n >,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport type PiAiHarnessOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> =\n | PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>\n | PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>;\n\ntype PiAiHarnessRunOptions<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> & {\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\ntype PiAiRunnableAgent<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = {\n run: (\n input: TInput,\n runtime: PiAiRuntime<TTools, TInput, TMetadata>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\nexport interface PiAiHarnessNormalizeOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n session?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<NormalizedSession>;\n output?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<JsonValue | undefined>;\n usage?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<UsageSummary>;\n timings?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<TimingSummary | undefined>;\n errors?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<Array<Record<string, JsonValue>>>;\n}\n\ntype InferredToolSurfaces<TInput, TMetadata extends HarnessMetadata> = {\n runtimeTools?: InferredPiAiToolset<TInput, TMetadata>;\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>;\n};\n\n/** Adapts a Pi agent runtime into a normalized vitest-evals harness. */\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n>(\n options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): Harness<TInput, TMetadata> {\n return {\n name: options.name ?? \"pi-ai\",\n run: async (input, context) => {\n const agent = await resolveAgent(options);\n const messages: NormalizedMessage[] = [\n {\n role: \"user\",\n content: normalizeContent(input),\n },\n ];\n const inferredTools = resolveInferredToolSurfaces<TInput, TMetadata>(\n agent,\n );\n\n if (hasExplicitToolset(options)) {\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n options.tools,\n inferredTools.nativeToolsets,\n );\n }\n\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n inferredTools.runtimeTools,\n inferredTools.nativeToolsets,\n );\n },\n };\n}\n\nasync function executePiHarnessRun<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n agent: TAgent,\n input: TInput,\n context: HarnessContext<TMetadata>,\n messages: NormalizedMessage[],\n runtimeTools: TTools | undefined,\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n): Promise<HarnessRun> {\n const runtime = createRuntime({\n input,\n context,\n tools: runtimeTools,\n messages,\n });\n\n try {\n const result = await withInstrumentedAgentTools(\n agent,\n nativeToolsets,\n {\n input,\n context,\n messages,\n toolCalls: runtime.toolCalls,\n },\n () =>\n runAgent(options, {\n agent,\n input,\n context,\n runtime,\n }),\n );\n\n if (isHarnessRun(result) && !hasResultOverrides(options)) {\n if (Object.keys(context.artifacts).length > 0 && !result.artifacts) {\n result.artifacts = context.artifacts;\n }\n return result;\n }\n\n const normalizeResult = result as TResult;\n const resultArgs = {\n agent,\n input,\n context,\n runtime,\n result: normalizeResult,\n } satisfies PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n\n const output = options.normalize?.output\n ? await options.normalize.output(resultArgs)\n : resolveOutput(normalizeResult);\n const usage = options.normalize?.usage\n ? await options.normalize.usage(resultArgs)\n : resolveUsage(normalizeResult, runtime.toolCalls.length);\n const session = options.normalize?.session\n ? await options.normalize.session(resultArgs)\n : resolveSession(normalizeResult, messages, output, usage);\n\n return {\n session,\n output,\n usage,\n timings: options.normalize?.timings\n ? await options.normalize.timings(resultArgs)\n : undefined,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: options.normalize?.errors\n ? await options.normalize.errors(resultArgs)\n : resolveErrors(normalizeResult),\n };\n } catch (error) {\n const usage = resolveUsage(undefined, runtime.toolCalls.length);\n const run = {\n session: resolveSession(undefined, messages, undefined, usage),\n output: undefined,\n usage,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: [serializeError(error)],\n } satisfies HarnessRun;\n\n throw attachHarnessRunToError(error, run);\n }\n}\n\nasync function resolveAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n if (options.agent !== undefined) {\n return options.agent;\n }\n\n if (options.createAgent) {\n return options.createAgent();\n }\n\n throw new Error(\n \"piAiHarness requires either an agent instance or a createAgent() function.\",\n );\n}\n\nfunction hasResultOverrides<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n return Boolean(\n options.normalize?.output ??\n options.normalize?.session ??\n options.normalize?.usage ??\n options.normalize?.timings ??\n options.normalize?.errors,\n );\n}\n\nfunction resolveInferredToolSurfaces<TInput, TMetadata extends HarnessMetadata>(\n agent: unknown,\n): InferredToolSurfaces<TInput, TMetadata> {\n let runtimeTools: InferredPiAiToolset<TInput, TMetadata> | undefined;\n const nativeToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nextRuntimeTools = getRuntimeToolset<TInput, TMetadata>(candidate);\n if (runtimeTools === undefined && nextRuntimeTools !== undefined) {\n runtimeTools = nextRuntimeTools;\n }\n\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n nativeToolsets.push(nativeTools);\n }\n }\n\n return {\n ...(runtimeTools ? { runtimeTools } : {}),\n ...(nativeToolsets.length > 0 ? { nativeToolsets } : {}),\n };\n}\n\nfunction getAgentToolCandidates(agent: unknown): object[] {\n const roots = getAgentRoots(agent);\n const candidates: object[] = [];\n const seen = new Set<object>();\n\n for (const root of roots) {\n addUniqueObject(candidates, seen, root);\n addUniqueObject(candidates, seen, getObjectProperty(root, \"state\"));\n addUniqueObject(candidates, seen, getObjectProperty(root, \"initialState\"));\n }\n\n return candidates;\n}\n\nfunction getAgentRoots(agent: unknown): object[] {\n return [asObject(agent)]\n .concat(asObject(getObjectProperty(agent, \"agent\")))\n .filter((value): value is object => value !== undefined);\n}\n\nfunction addUniqueObject(\n candidates: object[],\n seen: Set<object>,\n value: unknown,\n) {\n if (!value || typeof value !== \"object\" || seen.has(value)) {\n return;\n }\n\n seen.add(value);\n candidates.push(value);\n}\n\nfunction asObject(value: unknown): object | undefined {\n return value && typeof value === \"object\" ? value : undefined;\n}\n\nfunction getObjectProperty(value: unknown, key: string): unknown {\n return value && typeof value === \"object\"\n ? (value as Record<string, unknown>)[key]\n : undefined;\n}\n\nfunction getRuntimeToolset<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): InferredPiAiToolset<TInput, TMetadata> | undefined {\n const candidate =\n getObjectProperty(value, \"tools\") ?? getObjectProperty(value, \"toolset\");\n\n return isPiAiToolset(candidate)\n ? (candidate as InferredPiAiToolset<TInput, TMetadata>)\n : undefined;\n}\n\nfunction getNativeToolArray<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): PiAgentToolLike<TInput, TMetadata>[] | undefined {\n const candidate = getObjectProperty(value, \"tools\");\n if (isAgentToolArray(candidate)) {\n return candidate as PiAgentToolLike<TInput, TMetadata>[];\n }\n\n return undefined;\n}\n\nasync function runAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n): Promise<TResult | HarnessRun> {\n if (options.run) {\n return options.run(args);\n }\n\n if (hasPiAiRunMethod<TInput, TMetadata, TResult, TTools>(args.agent)) {\n return args.agent.run(args.input, args.runtime);\n }\n\n throw new Error(\n \"piAiHarness requires a run() function unless the provided agent exposes run(input, runtime).\",\n );\n}\n\nfunction hasExplicitToolset<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): options is PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n> {\n return options.tools !== undefined;\n}\n\nfunction hasPiAiRunMethod<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n agent: unknown,\n): agent is PiAiRunnableAgent<TInput, TMetadata, TResult, TTools> {\n if (!agent || typeof agent !== \"object\") {\n return false;\n }\n\n return (\n \"run\" in agent && typeof (agent as { run?: unknown }).run === \"function\"\n );\n}\n\nfunction isPiAiToolset(value: unknown): value is PiAiToolset {\n if (!value || typeof value !== \"object\" || Array.isArray(value)) {\n return false;\n }\n\n const tools = Object.values(value);\n return (\n tools.length > 0 &&\n tools.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nfunction isAgentToolArray(value: unknown): value is PiAgentToolLike[] {\n return (\n Array.isArray(value) &&\n value.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"name\" in tool &&\n typeof (tool as { name?: unknown }).name === \"string\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nasync function withInstrumentedAgentTools<\n TResult,\n TInput,\n TMetadata extends HarnessMetadata,\n>(\n agent: unknown,\n toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> | undefined,\n args: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n messages: NormalizedMessage[];\n toolCalls: ToolCallRecord[];\n },\n callback: () => Promise<TResult>,\n) {\n if (!toolsets || toolsets.length === 0) {\n return callback();\n }\n\n const originalExecutions = new Map<\n PiAgentToolLike<TInput, TMetadata>,\n PiAgentToolLike<TInput, TMetadata>[\"execute\"]\n >();\n const originalResets = new Map<ResettableAgent, ResettableAgent[\"reset\"]>();\n\n const patchTool = (tool: PiAgentToolLike<TInput, TMetadata>) => {\n if (originalExecutions.has(tool)) {\n return;\n }\n\n const originalExecute = getNativeToolExecuteOrigin(tool.execute);\n originalExecutions.set(tool, originalExecute);\n const instrumentedExecute: NativeToolExecute<TInput, TMetadata> = async (\n toolCallId: string,\n rawArgs: Record<string, JsonValue>,\n ) => {\n const startedAt = new Date();\n const toolContext = {\n input: args.input,\n metadata: args.context.metadata,\n signal: args.context.signal,\n setArtifact: args.context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeNativeToolWithReplay({\n toolName: tool.name,\n toolCallId,\n execute: originalExecute,\n replay: tool.replay,\n args: rawArgs,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n result: execution.normalizedResult,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n args.messages.push({\n role: \"tool\",\n content: execution.normalizedResult,\n metadata: {\n name: tool.name,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(getReplayMetadataFromError(error)),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n };\n instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;\n tool.execute = instrumentedExecute;\n };\n\n const patchToolsets = (\n nextToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n ) => {\n for (const toolset of nextToolsets) {\n for (const tool of toolset) {\n patchTool(tool);\n }\n }\n };\n\n patchToolsets(toolsets);\n\n for (const target of getAgentResetTargets(agent)) {\n const originalReset = target.reset;\n originalResets.set(target, originalReset);\n target.reset = function patchedReset(this: ResettableAgent, ...resetArgs) {\n const resetResult = originalReset.apply(this, resetArgs);\n\n if (isPromiseLike(resetResult)) {\n return resetResult.finally(() => {\n patchToolsets(\n resolveInferredNativeToolsets<TInput, TMetadata>(agent),\n );\n });\n }\n\n patchToolsets(resolveInferredNativeToolsets<TInput, TMetadata>(agent));\n return resetResult;\n };\n }\n\n try {\n return await callback();\n } finally {\n for (const [target, originalReset] of originalResets) {\n target.reset = originalReset;\n }\n\n for (const [tool, originalExecute] of originalExecutions) {\n tool.execute = originalExecute;\n }\n }\n}\n\ntype ResettableAgent = {\n reset: (...args: unknown[]) => unknown;\n};\n\nfunction getAgentResetTargets(agent: unknown): ResettableAgent[] {\n return getAgentRoots(agent).filter(isResettableAgent);\n}\n\nfunction isResettableAgent(value: object): value is ResettableAgent {\n return \"reset\" in value && typeof value.reset === \"function\";\n}\n\nfunction resolveInferredNativeToolsets<\n TInput,\n TMetadata extends HarnessMetadata,\n>(agent: unknown): Array<PiAgentToolLike<TInput, TMetadata>[]> {\n const toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n toolsets.push(nativeTools);\n }\n }\n\n return toolsets;\n}\n\nfunction isPromiseLike(value: unknown): value is Promise<unknown> {\n return Boolean(\n value && typeof (value as { then?: unknown }).then === \"function\",\n );\n}\n\nfunction serializeToolCallError(\n error: unknown,\n): NonNullable<ToolCallRecord[\"error\"]> {\n const serialized = serializeError(error);\n const { message, type, ...details } = serialized;\n\n return {\n ...details,\n message: typeof message === \"string\" ? message : String(message),\n ...(typeof type === \"string\" ? { type } : {}),\n };\n}\n\nfunction getNativeToolExecuteOrigin<TInput, TMetadata extends HarnessMetadata>(\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"],\n) {\n const nativeExecute = execute as NativeToolExecute<TInput, TMetadata>;\n return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;\n}\n\nasync function executeNativeToolWithReplay<\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n toolCallId,\n execute,\n replay,\n args,\n context,\n}: {\n toolName: string;\n toolCallId: string;\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n replay: PiAgentToolLike<TInput, TMetadata>[\"replay\"];\n args: Record<string, JsonValue>;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n let didExecute = false;\n let liveResult: unknown;\n\n const execution = await executeWithReplay({\n toolName,\n args,\n context,\n execute: async (toolArgs) => {\n didExecute = true;\n liveResult = await execute(toolCallId, toolArgs);\n return createNativeToolReplayEnvelope(liveResult);\n },\n replay,\n });\n\n if (didExecute) {\n return {\n result: liveResult,\n normalizedResult: normalizeReplayToolResult(liveResult),\n replay: execution.replay,\n };\n }\n\n return {\n ...resolveNativeToolReplayResult(execution.result),\n replay: execution.replay,\n };\n}\n\nfunction createRuntime<\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>({\n input,\n context,\n tools,\n messages,\n}: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n tools: TTools | undefined;\n messages: NormalizedMessage[];\n}): PiAiRuntime<TTools, TInput, TMetadata> & {\n toolCalls: ToolCallRecord[];\n} {\n const toolCalls: ToolCallRecord[] = [];\n const eventSink: PiAiEventSink = {\n message: (message) => {\n messages.push(message);\n },\n system: (content, metadata) => {\n messages.push({\n role: \"system\",\n content,\n metadata,\n });\n },\n user: (content, metadata) => {\n messages.push({\n role: \"user\",\n content,\n metadata,\n });\n },\n assistant: (content, metadata) => {\n messages.push({\n role: \"assistant\",\n content,\n metadata,\n });\n },\n tool: (name, content, metadata) => {\n messages.push({\n role: \"tool\",\n content,\n metadata: {\n name,\n ...(metadata ?? {}),\n },\n });\n },\n };\n\n const runtimeTools = Object.fromEntries(\n Object.entries(tools ?? {}).map(([toolName, tool]) => [\n toolName,\n async (args: Record<string, JsonValue>) => {\n const startedAt = new Date();\n const toolContext = {\n input,\n metadata: context.metadata,\n signal: context.signal,\n setArtifact: context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeToolWithReplay({\n toolName,\n tool,\n args,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: toolName,\n arguments: args,\n result: execution.result,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n messages.push({\n role: \"tool\",\n content: execution.result,\n metadata: {\n name: toolName,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: toolName,\n arguments: args,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(\n getReplayMetadataFromError(error),\n ),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n },\n ]),\n ) as unknown as PiAiRuntime<TTools, TInput, TMetadata>[\"tools\"];\n\n return {\n tools: runtimeTools,\n events: eventSink,\n signal: context.signal,\n toolCalls,\n };\n}\n\nfunction resolveOutput(result: unknown): JsonValue | undefined {\n if (!result || typeof result !== \"object\") {\n return toJsonValue(result);\n }\n\n const candidates = [\n \"output\",\n \"decision\",\n \"result\",\n \"final\",\n ] satisfies string[];\n\n for (const key of candidates) {\n const value = (result as Record<string, unknown>)[key];\n const normalized = toJsonValue(value);\n if (normalized !== undefined) {\n return normalized;\n }\n }\n\n return undefined;\n}\n\nfunction normalizeToolResult(result: unknown): JsonValue | undefined {\n const details =\n result && typeof result === \"object\"\n ? toJsonValue((result as { details?: unknown }).details)\n : undefined;\n if (details !== undefined) {\n return details;\n }\n\n return (\n toJsonValue(result) ?? (result === undefined ? undefined : String(result))\n );\n}\n\nfunction normalizeReplayToolResult(result: unknown): JsonValue {\n return normalizeToolResult(result) ?? null;\n}\n\ntype NativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 2;\n };\n agentResult: JsonValue;\n normalizedResult: JsonValue;\n};\n\ntype LegacyNativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 1;\n };\n normalizedResult: JsonValue;\n agentResult?: JsonValue;\n};\n\nfunction createNativeToolReplayEnvelope(\n result: unknown,\n): NativeToolReplayEnvelope {\n const normalizedResult = normalizeReplayToolResult(result);\n\n return {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\",\n version: 2,\n },\n agentResult: toJsonValue(result) ?? normalizedResult,\n normalizedResult,\n };\n}\n\nfunction resolveNativeToolReplayResult(result: JsonValue) {\n if (isNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n if (isLegacyNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult ?? result.normalizedResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n return {\n result,\n normalizedResult: normalizeReplayToolResult(result),\n };\n}\n\nfunction isNativeToolReplayEnvelope(\n value: unknown,\n): value is NativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"agentResult\" in value &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isNativeToolReplayMarker(\n value: unknown,\n): value is NativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 2,\n );\n}\n\nfunction isLegacyNativeToolReplayEnvelope(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isLegacyNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isLegacyNativeToolReplayMarker(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 1,\n );\n}\n\nfunction resolveUsage(result: unknown, toolCallCount: number): UsageSummary {\n if (!result || typeof result !== \"object\") {\n return toolCallCount > 0 ? { toolCalls: toolCallCount } : {};\n }\n\n const usageValue =\n (result as Record<string, unknown>).usage ??\n (result as Record<string, unknown>).metrics;\n\n const usage =\n usageValue && typeof usageValue === \"object\"\n ? ({ ...(usageValue as Record<string, unknown>) } as UsageSummary)\n : {};\n\n if (usage.toolCalls === undefined && toolCallCount > 0) {\n usage.toolCalls = toolCallCount;\n }\n\n return usage;\n}\n\nfunction resolveSession(\n result: unknown,\n messages: NormalizedMessage[],\n output: JsonValue | undefined,\n usage: UsageSummary,\n): NormalizedSession {\n if (\n isNormalizedSession(\n (result as Record<string, unknown> | undefined)?.session,\n )\n ) {\n return (result as { session: NormalizedSession }).session;\n }\n\n if (\n isNormalizedSession((result as Record<string, unknown> | undefined)?.trace)\n ) {\n return (result as { trace: NormalizedSession }).trace;\n }\n\n const sessionMessages = [...messages];\n if (\n output !== undefined &&\n !sessionMessages.some(\n (message) =>\n message.role === \"assistant\" && message.content !== undefined,\n )\n ) {\n sessionMessages.push({\n role: \"assistant\",\n content: output,\n });\n }\n\n return {\n messages: sessionMessages,\n outputText: typeof output === \"string\" ? output : undefined,\n provider:\n ((result as Record<string, unknown> | undefined)?.provider as\n | string\n | undefined) ?? usage.provider,\n model:\n ((result as Record<string, unknown> | undefined)?.model as\n | string\n | undefined) ?? usage.model,\n };\n}\n\nfunction resolveErrors(result: unknown): Array<Record<string, JsonValue>> {\n return resolveHarnessRunErrors(result);\n}\n\nasync function executeToolWithReplay<\n TArgs extends Record<string, JsonValue>,\n TResult extends JsonValue,\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n tool,\n args,\n context,\n}: {\n toolName: string;\n tool: PiAiToolDefinition<TArgs, TResult, TInput, TMetadata>;\n args: TArgs;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n return executeWithReplay({\n toolName,\n args,\n context,\n execute: tool.execute,\n replay: tool.replay,\n });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAYA,qBAQO;AACP,oBAIO;AAoCP,IAAM,0BAA0B,OAAO,oCAAoC;AAmRpE,SAAS,YAUd,SAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,QAAQ,MAAM,aAAa,OAAO;AACxC,YAAM,WAAgC;AAAA,QACpC;AAAA,UACE,MAAM;AAAA,UACN,aAAS,iCAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AACA,YAAM,gBAAgB;AAAA,QACpB;AAAA,MACF;AAEA,UAAI,mBAAmB,OAAO,GAAG;AAC/B,eAAO;AAAA,UACL;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,cAAc;AAAA,QAChB;AAAA,MACF;AAEA,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,cAAc;AAAA,QACd,cAAc;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AACF;AAEA,eAAe,oBAOb,SACA,OACA,OACA,SACA,UACA,cACA,gBACqB;AACrB,QAAM,UAAU,cAAc;AAAA,IAC5B;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP;AAAA,EACF,CAAC;AAED,MAAI;AACF,UAAM,SAAS,MAAM;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW,QAAQ;AAAA,MACrB;AAAA,MACA,MACE,SAAS,SAAS;AAAA,QAChB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACL;AAEA,YAAI,6BAAa,MAAM,KAAK,CAAC,mBAAmB,OAAO,GAAG;AACxD,UAAI,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KAAK,CAAC,OAAO,WAAW;AAClE,eAAO,YAAY,QAAQ;AAAA,MAC7B;AACA,aAAO;AAAA,IACT;AAEA,UAAM,kBAAkB;AACxB,UAAM,aAAa;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,IACV;AAQA,UAAM,SAAS,QAAQ,WAAW,SAC9B,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AACjC,UAAM,QAAQ,QAAQ,WAAW,QAC7B,MAAM,QAAQ,UAAU,MAAM,UAAU,IACxC,aAAa,iBAAiB,QAAQ,UAAU,MAAM;AAC1D,UAAM,UAAU,QAAQ,WAAW,UAC/B,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C,eAAe,iBAAiB,UAAU,QAAQ,KAAK;AAE3D,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,QAAQ,WAAW,UACxB,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C;AAAA,MACJ,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,QAAQ,WAAW,SACvB,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AAAA,IACnC;AAAA,EACF,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,QAAW,QAAQ,UAAU,MAAM;AAC9D,UAAM,MAAM;AAAA,MACV,SAAS,eAAe,QAAW,UAAU,QAAW,KAAK;AAAA,MAC7D,QAAQ;AAAA,MACR;AAAA,MACA,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,KAAC,+BAAe,KAAK,CAAC;AAAA,IAChC;AAEA,cAAM,wCAAwB,OAAO,GAAG;AAAA,EAC1C;AACF;AAEA,eAAe,aAMb,SAAyE;AACzE,MAAI,QAAQ,UAAU,QAAW;AAC/B,WAAO,QAAQ;AAAA,EACjB;AAEA,MAAI,QAAQ,aAAa;AACvB,WAAO,QAAQ,YAAY;AAAA,EAC7B;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAMP,SAA4E;AAC5E,SAAO;AAAA,IACL,QAAQ,WAAW,UACjB,QAAQ,WAAW,WACnB,QAAQ,WAAW,SACnB,QAAQ,WAAW,WACnB,QAAQ,WAAW;AAAA,EACvB;AACF;AAEA,SAAS,4BACP,OACyC;AACzC,MAAI;AACJ,QAAM,iBAA8D,CAAC;AACrE,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,mBAAmB,kBAAqC,SAAS;AACvE,QAAI,iBAAiB,UAAa,qBAAqB,QAAW;AAChE,qBAAe;AAAA,IACjB;AAEA,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,qBAAe,KAAK,WAAW;AAAA,IACjC;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;AAAA,IACvC,GAAI,eAAe,SAAS,IAAI,EAAE,eAAe,IAAI,CAAC;AAAA,EACxD;AACF;AAEA,SAAS,uBAAuB,OAA0B;AACxD,QAAM,QAAQ,cAAc,KAAK;AACjC,QAAM,aAAuB,CAAC;AAC9B,QAAM,OAAO,oBAAI,IAAY;AAE7B,aAAW,QAAQ,OAAO;AACxB,oBAAgB,YAAY,MAAM,IAAI;AACtC,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,OAAO,CAAC;AAClE,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,cAAc,CAAC;AAAA,EAC3E;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA0B;AAC/C,SAAO,CAAC,SAAS,KAAK,CAAC,EACpB,OAAO,SAAS,kBAAkB,OAAO,OAAO,CAAC,CAAC,EAClD,OAAO,CAAC,UAA2B,UAAU,MAAS;AAC3D;AAEA,SAAS,gBACP,YACA,MACA,OACA;AACA,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,KAAK,IAAI,KAAK,GAAG;AAC1D;AAAA,EACF;AAEA,OAAK,IAAI,KAAK;AACd,aAAW,KAAK,KAAK;AACvB;AAEA,SAAS,SAAS,OAAoC;AACpD,SAAO,SAAS,OAAO,UAAU,WAAW,QAAQ;AACtD;AAEA,SAAS,kBAAkB,OAAgB,KAAsB;AAC/D,SAAO,SAAS,OAAO,UAAU,WAC5B,MAAkC,GAAG,IACtC;AACN;AAEA,SAAS,kBACP,OACoD;AACpD,QAAM,YACJ,kBAAkB,OAAO,OAAO,KAAK,kBAAkB,OAAO,SAAS;AAEzE,SAAO,cAAc,SAAS,IACzB,YACD;AACN;AAEA,SAAS,mBACP,OACkD;AAClD,QAAM,YAAY,kBAAkB,OAAO,OAAO;AAClD,MAAI,iBAAiB,SAAS,GAAG;AAC/B,WAAO;AAAA,EACT;AAEA,SAAO;AACT;AAEA,eAAe,SAOb,SACA,MAC+B;AAC/B,MAAI,QAAQ,KAAK;AACf,WAAO,QAAQ,IAAI,IAAI;AAAA,EACzB;AAEA,MAAI,iBAAqD,KAAK,KAAK,GAAG;AACpE,WAAO,KAAK,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO;AAAA,EAChD;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAOP,SAOA;AACA,SAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,iBAMP,OACgE;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,SACE,SAAS,SAAS,OAAQ,MAA4B,QAAQ;AAElE;AAEA,SAAS,cAAc,OAAsC;AAC3D,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AAC/D,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,OAAO,OAAO,KAAK;AACjC,SACE,MAAM,SAAS,KACf,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,SAAS,iBAAiB,OAA4C;AACpE,SACE,MAAM,QAAQ,KAAK,KACnB,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,UAAU,QACV,OAAQ,KAA4B,SAAS,YAC7C,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,eAAe,2BAKb,OACA,UACA,MAMA,UACA;AACA,MAAI,CAAC,YAAY,SAAS,WAAW,GAAG;AACtC,WAAO,SAAS;AAAA,EAClB;AAEA,QAAM,qBAAqB,oBAAI,IAG7B;AACF,QAAM,iBAAiB,oBAAI,IAA+C;AAE1E,QAAM,YAAY,CAAC,SAA6C;AAC9D,QAAI,mBAAmB,IAAI,IAAI,GAAG;AAChC;AAAA,IACF;AAEA,UAAM,kBAAkB,2BAA2B,KAAK,OAAO;AAC/D,uBAAmB,IAAI,MAAM,eAAe;AAC5C,UAAM,sBAA4D,OAChE,YACA,YACG;AACH,YAAM,YAAY,oBAAI,KAAK;AAC3B,YAAM,cAAc;AAAA,QAClB,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK,QAAQ;AAAA,QACvB,QAAQ,KAAK,QAAQ;AAAA,QACrB,aAAa,KAAK,QAAQ;AAAA,MAC5B;AAEA,UAAI;AACF,cAAM,YAAY,MAAM,4BAA4B;AAAA,UAClD,UAAU,KAAK;AAAA,UACf;AAAA,UACA,SAAS;AAAA,UACT,QAAQ,KAAK;AAAA,UACb,MAAM;AAAA,UACN,SAAS;AAAA,QACX,CAAC;AACD,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,UAAU;AAAA,UAClB,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,cAAU,uCAAwB,UAAU,MAAM;AAAA,QACpD;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,SAAS,UAAU;AAAA,UACnB,UAAU;AAAA,YACR,MAAM,KAAK;AAAA,UACb;AAAA,QACF,CAAC;AACD,eAAO,UAAU;AAAA,MACnB,SAAS,OAAO;AACd,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,OAAO,uBAAuB,KAAK;AAAA,UACnC,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,cAAU,2CAAwB,0CAA2B,KAAK,CAAC;AAAA,QACrE;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,cAAM;AAAA,MACR;AAAA,IACF;AACA,wBAAoB,uBAAuB,IAAI;AAC/C,SAAK,UAAU;AAAA,EACjB;AAEA,QAAM,gBAAgB,CACpB,iBACG;AACH,eAAW,WAAW,cAAc;AAClC,iBAAW,QAAQ,SAAS;AAC1B,kBAAU,IAAI;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,gBAAc,QAAQ;AAEtB,aAAW,UAAU,qBAAqB,KAAK,GAAG;AAChD,UAAM,gBAAgB,OAAO;AAC7B,mBAAe,IAAI,QAAQ,aAAa;AACxC,WAAO,QAAQ,SAAS,gBAAuC,WAAW;AACxE,YAAM,cAAc,cAAc,MAAM,MAAM,SAAS;AAEvD,UAAI,cAAc,WAAW,GAAG;AAC9B,eAAO,YAAY,QAAQ,MAAM;AAC/B;AAAA,YACE,8BAAiD,KAAK;AAAA,UACxD;AAAA,QACF,CAAC;AAAA,MACH;AAEA,oBAAc,8BAAiD,KAAK,CAAC;AACrE,aAAO;AAAA,IACT;AAAA,EACF;AAEA,MAAI;AACF,WAAO,MAAM,SAAS;AAAA,EACxB,UAAE;AACA,eAAW,CAAC,QAAQ,aAAa,KAAK,gBAAgB;AACpD,aAAO,QAAQ;AAAA,IACjB;AAEA,eAAW,CAAC,MAAM,eAAe,KAAK,oBAAoB;AACxD,WAAK,UAAU;AAAA,IACjB;AAAA,EACF;AACF;AAMA,SAAS,qBAAqB,OAAmC;AAC/D,SAAO,cAAc,KAAK,EAAE,OAAO,iBAAiB;AACtD;AAEA,SAAS,kBAAkB,OAAyC;AAClE,SAAO,WAAW,SAAS,OAAO,MAAM,UAAU;AACpD;AAEA,SAAS,8BAGP,OAA6D;AAC7D,QAAM,WAAwD,CAAC;AAC/D,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,eAAS,KAAK,WAAW;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA2C;AAChE,SAAO;AAAA,IACL,SAAS,OAAQ,MAA6B,SAAS;AAAA,EACzD;AACF;AAEA,SAAS,uBACP,OACsC;AACtC,QAAM,iBAAa,+BAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAEA,SAAS,2BACP,SACA;AACA,QAAM,gBAAgB;AACtB,SAAO,cAAc,uBAAuB,KAAK;AACnD;AAEA,eAAe,4BAGb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOG;AACD,MAAI,aAAa;AACjB,MAAI;AAEJ,QAAM,YAAY,UAAM,iCAAkB;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,OAAO,aAAa;AAC3B,mBAAa;AACb,mBAAa,MAAM,QAAQ,YAAY,QAAQ;AAC/C,aAAO,+BAA+B,UAAU;AAAA,IAClD;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI,YAAY;AACd,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,kBAAkB,0BAA0B,UAAU;AAAA,MACtD,QAAQ,UAAU;AAAA,IACpB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAG,8BAA8B,UAAU,MAAM;AAAA,IACjD,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,cAIP;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOE;AACA,QAAM,YAA8B,CAAC;AACrC,QAAM,YAA2B;AAAA,IAC/B,SAAS,CAAC,YAAY;AACpB,eAAS,KAAK,OAAO;AAAA,IACvB;AAAA,IACA,QAAQ,CAAC,SAAS,aAAa;AAC7B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,SAAS,aAAa;AAC3B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,WAAW,CAAC,SAAS,aAAa;AAChC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,MAAM,SAAS,aAAa;AACjC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA,UAAU;AAAA,UACR;AAAA,UACA,GAAI,YAAY,CAAC;AAAA,QACnB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,OAAO;AAAA,IAC1B,OAAO,QAAQ,SAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,UAAU,IAAI,MAAM;AAAA,MACpD;AAAA,MACA,OAAO,SAAoC;AACzC,cAAM,YAAY,oBAAI,KAAK;AAC3B,cAAM,cAAc;AAAA,UAClB;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB,QAAQ,QAAQ;AAAA,UAChB,aAAa,QAAQ;AAAA,QACvB;AAEA,YAAI;AACF,gBAAM,YAAY,MAAM,sBAAsB;AAAA,YAC5C;AAAA,YACA;AAAA,YACA;AAAA,YACA,SAAS;AAAA,UACX,CAAC;AACD,gBAAM,aAAa,oBAAI,KAAK;AAC5B,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,UAAU;AAAA,YAClB,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,cAAU,uCAAwB,UAAU,MAAM;AAAA,UACpD;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,SAAS,UAAU;AAAA,YACnB,UAAU;AAAA,cACR,MAAM;AAAA,YACR;AAAA,UACF,CAAC;AACD,iBAAO,UAAU;AAAA,QACnB,SAAS,OAAO;AACd,gBAAM,aAAa,oBAAI,KAAK;AAC5B,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,OAAO,uBAAuB,KAAK;AAAA,YACnC,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,cAAU;AAAA,kBACR,0CAA2B,KAAK;AAAA,YAClC;AAAA,UACF;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,QAAQ,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;AAEA,SAAS,cAAc,QAAwC;AAC7D,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,eAAO,4BAAY,MAAM;AAAA,EAC3B;AAEA,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAS,OAAmC,GAAG;AACrD,UAAM,iBAAa,4BAAY,KAAK;AACpC,QAAI,eAAe,QAAW;AAC5B,aAAO;AAAA,IACT;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,oBAAoB,QAAwC;AACnE,QAAM,UACJ,UAAU,OAAO,WAAW,eACxB,4BAAa,OAAiC,OAAO,IACrD;AACN,MAAI,YAAY,QAAW;AACzB,WAAO;AAAA,EACT;AAEA,aACE,4BAAY,MAAM,MAAM,WAAW,SAAY,SAAY,OAAO,MAAM;AAE5E;AAEA,SAAS,0BAA0B,QAA4B;AAC7D,SAAO,oBAAoB,MAAM,KAAK;AACxC;AAoBA,SAAS,+BACP,QAC0B;AAC1B,QAAM,mBAAmB,0BAA0B,MAAM;AAEzD,SAAO;AAAA,IACL,eAAe;AAAA,MACb,MAAM;AAAA,MACN,SAAS;AAAA,IACX;AAAA,IACA,iBAAa,4BAAY,MAAM,KAAK;AAAA,IACpC;AAAA,EACF;AACF;AAEA,SAAS,8BAA8B,QAAmB;AACxD,MAAI,2BAA2B,MAAM,GAAG;AACtC,WAAO;AAAA,MACL,QAAQ,OAAO;AAAA,MACf,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,MAAI,iCAAiC,MAAM,GAAG;AAC5C,WAAO;AAAA,MACL,QAAQ,OAAO,eAAe,OAAO;AAAA,MACrC,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,0BAA0B,MAAM;AAAA,EACpD;AACF;AAEA,SAAS,2BACP,OACmC;AACnC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,iBAAiB,SACjB,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,yBACP,OACoD;AACpD,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,iCACP,OACyC;AACzC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,+BACP,OAC0D;AAC1D,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,aAAa,QAAiB,eAAqC;AAC1E,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,gBAAgB,IAAI,EAAE,WAAW,cAAc,IAAI,CAAC;AAAA,EAC7D;AAEA,QAAM,aACH,OAAmC,SACnC,OAAmC;AAEtC,QAAM,QACJ,cAAc,OAAO,eAAe,WAC/B,EAAE,GAAI,WAAuC,IAC9C,CAAC;AAEP,MAAI,MAAM,cAAc,UAAa,gBAAgB,GAAG;AACtD,UAAM,YAAY;AAAA,EACpB;AAEA,SAAO;AACT;AAEA,SAAS,eACP,QACA,UACA,QACA,OACmB;AACnB,UACE;AAAA,IACG,QAAgD;AAAA,EACnD,GACA;AACA,WAAQ,OAA0C;AAAA,EACpD;AAEA,UACE,oCAAqB,QAAgD,KAAK,GAC1E;AACA,WAAQ,OAAwC;AAAA,EAClD;AAEA,QAAM,kBAAkB,CAAC,GAAG,QAAQ;AACpC,MACE,WAAW,UACX,CAAC,gBAAgB;AAAA,IACf,CAAC,YACC,QAAQ,SAAS,eAAe,QAAQ,YAAY;AAAA,EACxD,GACA;AACA,oBAAgB,KAAK;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,UAAU;AAAA,IACV,YAAY,OAAO,WAAW,WAAW,SAAS;AAAA,IAClD,UACI,QAAgD,YAEhC,MAAM;AAAA,IAC1B,OACI,QAAgD,SAEhC,MAAM;AAAA,EAC5B;AACF;AAEA,SAAS,cAAc,QAAmD;AACxE,aAAO,wCAAwB,MAAM;AACvC;AAEA,eAAe,sBAKb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAKG;AACD,aAAO,iCAAkB;AAAA,IACvB;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,KAAK;AAAA,IACd,QAAQ,KAAK;AAAA,EACf,CAAC;AACH;","names":[]}
|
|
1
|
+
{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import type {\n Harness,\n HarnessContext,\n HarnessMetadata,\n HarnessPrompt,\n HarnessRun,\n JsonValue,\n NormalizedMessage,\n NormalizedSession,\n TimingSummary,\n ToolCallRecord,\n UsageSummary,\n} from \"vitest-evals/harness\";\nimport {\n attachHarnessRunToError,\n isHarnessRun,\n isNormalizedSession,\n normalizeContent,\n resolveHarnessRunErrors,\n serializeError,\n toJsonValue,\n} from \"vitest-evals/harness\";\nimport {\n executeWithReplay,\n getReplayMetadataFromError,\n normalizeReplayMetadata,\n} from \"vitest-evals/replay\";\nimport type {\n ReplayMode,\n ToolRecording,\n ToolReplayConfig,\n} from \"vitest-evals/replay\";\n\ntype MaybePromise<T> = T | Promise<T>;\ntype AnyPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolDefinition<any, any, TInput, TMetadata>>;\ntype InferredPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<\n string,\n PiAiToolDefinition<Record<string, JsonValue>, JsonValue, TInput, TMetadata>\n>;\n\ntype PiAgentToolLike<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n name: string;\n execute: (toolCallId: string, args: Record<string, JsonValue>) => unknown;\n};\n\nconst ORIGINAL_NATIVE_EXECUTE = Symbol(\"vitest-evals.originalNativeExecute\");\n\ntype NativeToolExecute<\n TInput,\n TMetadata extends HarnessMetadata,\n> = PiAgentToolLike<TInput, TMetadata>[\"execute\"] & {\n [ORIGINAL_NATIVE_EXECUTE]?: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n};\n\nexport type PiAiReplayMode = ReplayMode;\n\nexport interface PiAiEventSink {\n message: (message: NormalizedMessage) => void;\n system: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n user: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n assistant: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n tool: (\n name: string,\n content: JsonValue,\n metadata?: Record<string, JsonValue>,\n ) => void;\n}\n\nexport interface PiAiToolContext<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n input: TInput;\n metadata: HarnessContext<TMetadata>[\"metadata\"];\n signal?: AbortSignal;\n setArtifact: HarnessContext<TMetadata>[\"setArtifact\"];\n}\n\nexport type PiAiToolRecording<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n> = ToolRecording<TArgs, TResult>;\n\nexport type PiAiToolReplayConfig<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;\n\nexport type PiAiToolReplayPolicy<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> =\n | boolean\n | PiAiToolReplayConfig<\n Record<string, JsonValue>,\n JsonValue,\n TInput,\n TMetadata\n >;\n\nexport type PiAiToolReplayPolicies<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolReplayPolicy<TInput, TMetadata>>;\n\nexport interface PiAiToolDefinition<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n description?: string;\n execute: (\n args: TArgs,\n context: PiAiToolContext<TInput, TMetadata>,\n ) => MaybePromise<TResult>;\n}\n\nexport type PiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = AnyPiAiToolset<TInput, TMetadata>;\n\ntype ToolArgs<TTool> = TTool extends PiAiToolDefinition<\n infer TArgs,\n infer _TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TArgs\n : never;\n\ntype ToolResult<TTool> = TTool extends PiAiToolDefinition<\n infer _TArgs,\n infer TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TResult\n : never;\n\nexport type PiAiRuntime<\n TTools extends PiAiToolset<TInput, TMetadata>,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n tools: {\n [K in keyof TTools]: (\n args: ToolArgs<TTools[K]>,\n ) => Promise<ToolResult<TTools[K]>>;\n };\n events: PiAiEventSink;\n signal?: AbortSignal;\n};\n\nexport interface PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> {\n agent: TAgent;\n input: TInput;\n context: HarnessContext<TMetadata>;\n runtime: PiAiRuntime<TTools, TInput, TMetadata>;\n}\n\nexport interface PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> extends PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools> {\n result: TResult;\n}\n\ninterface PiAiHarnessBaseOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n agent?: TAgent;\n createAgent?: () => MaybePromise<TAgent>;\n toolReplay?: PiAiToolReplayPolicies<TInput, TMetadata>;\n normalize?: PiAiHarnessNormalizeOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n prompt: HarnessPrompt;\n name?: string;\n}\n\nexport interface PiAiHarnessWithToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {\n tools: TTools;\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport interface PiAiHarnessInferredToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n> extends PiAiHarnessBaseOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n InferredPiAiToolset<TInput, TMetadata>\n > {\n tools?: undefined;\n run?: (\n args: PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata,\n InferredPiAiToolset<TInput, TMetadata>\n >,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport type PiAiHarnessOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> =\n | PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>\n | PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>;\n\ntype PiAiHarnessRunOptions<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> & {\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\ntype PiAiRunnableAgent<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = {\n run: (\n input: TInput,\n runtime: PiAiRuntime<TTools, TInput, TMetadata>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\nexport interface PiAiHarnessNormalizeOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n session?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<NormalizedSession>;\n output?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<JsonValue | undefined>;\n usage?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<UsageSummary>;\n timings?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<TimingSummary | undefined>;\n errors?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<Array<Record<string, JsonValue>>>;\n}\n\ntype InferredToolSurfaces<TInput, TMetadata extends HarnessMetadata> = {\n runtimeTools?: InferredPiAiToolset<TInput, TMetadata>;\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>;\n};\n\ntype PiToolExecutionState = {\n activeNativeToolNames: Map<string, number>;\n};\n\n/** Adapts a Pi agent runtime into a normalized vitest-evals harness. */\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n>(\n options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): Harness<TInput, TMetadata> {\n return {\n name: options.name ?? \"pi-ai\",\n prompt: options.prompt,\n run: async (input, context) => {\n const agent = await resolveAgent(options);\n const messages: NormalizedMessage[] = [\n {\n role: \"user\",\n content: normalizeContent(input),\n },\n ];\n const inferredTools = resolveInferredToolSurfaces<TInput, TMetadata>(\n agent,\n );\n\n if (hasExplicitToolset(options)) {\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n options.tools,\n inferredTools.nativeToolsets,\n );\n }\n\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n inferredTools.runtimeTools,\n inferredTools.nativeToolsets,\n );\n },\n };\n}\n\nasync function executePiHarnessRun<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n agent: TAgent,\n input: TInput,\n context: HarnessContext<TMetadata>,\n messages: NormalizedMessage[],\n runtimeTools: TTools | undefined,\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n): Promise<HarnessRun> {\n const executionState = createPiToolExecutionState();\n const runtime = createRuntime({\n input,\n context,\n tools: runtimeTools,\n toolReplay: options.toolReplay,\n executionState,\n messages,\n });\n\n try {\n const result = await withInstrumentedAgentTools(\n agent,\n nativeToolsets,\n {\n input,\n context,\n messages,\n toolCalls: runtime.toolCalls,\n toolReplay: options.toolReplay,\n executionState,\n },\n () =>\n runAgent(options, {\n agent,\n input,\n context,\n runtime,\n }),\n );\n\n if (isHarnessRun(result) && !hasResultOverrides(options)) {\n if (Object.keys(context.artifacts).length > 0 && !result.artifacts) {\n result.artifacts = context.artifacts;\n }\n return result;\n }\n\n const normalizeResult = result as TResult;\n const resultArgs = {\n agent,\n input,\n context,\n runtime,\n result: normalizeResult,\n } satisfies PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n\n const output = options.normalize?.output\n ? await options.normalize.output(resultArgs)\n : resolveOutput(normalizeResult);\n const usage = options.normalize?.usage\n ? await options.normalize.usage(resultArgs)\n : resolveUsage(normalizeResult, runtime.toolCalls.length);\n const session = options.normalize?.session\n ? await options.normalize.session(resultArgs)\n : resolveSession(normalizeResult, messages, output, usage);\n\n return {\n session,\n output,\n usage,\n timings: options.normalize?.timings\n ? await options.normalize.timings(resultArgs)\n : undefined,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: options.normalize?.errors\n ? await options.normalize.errors(resultArgs)\n : resolveErrors(normalizeResult),\n };\n } catch (error) {\n const usage = resolveUsage(undefined, runtime.toolCalls.length);\n const run = {\n session: resolveSession(undefined, messages, undefined, usage),\n output: undefined,\n usage,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: [serializeError(error)],\n } satisfies HarnessRun;\n\n throw attachHarnessRunToError(error, run);\n }\n}\n\nasync function resolveAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n if (options.agent !== undefined) {\n return options.agent;\n }\n\n if (options.createAgent) {\n return options.createAgent();\n }\n\n throw new Error(\n \"piAiHarness requires either an agent instance or a createAgent() function.\",\n );\n}\n\nfunction hasResultOverrides<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n return Boolean(\n options.normalize?.output ??\n options.normalize?.session ??\n options.normalize?.usage ??\n options.normalize?.timings ??\n options.normalize?.errors,\n );\n}\n\nfunction resolveInferredToolSurfaces<TInput, TMetadata extends HarnessMetadata>(\n agent: unknown,\n): InferredToolSurfaces<TInput, TMetadata> {\n let runtimeTools: InferredPiAiToolset<TInput, TMetadata> | undefined;\n const nativeToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nextRuntimeTools = getRuntimeToolset<TInput, TMetadata>(candidate);\n if (runtimeTools === undefined && nextRuntimeTools !== undefined) {\n runtimeTools = nextRuntimeTools;\n }\n\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n nativeToolsets.push(nativeTools);\n }\n }\n\n return {\n ...(runtimeTools ? { runtimeTools } : {}),\n ...(nativeToolsets.length > 0 ? { nativeToolsets } : {}),\n };\n}\n\nfunction getAgentToolCandidates(agent: unknown): object[] {\n const roots = getAgentRoots(agent);\n const candidates: object[] = [];\n const seen = new Set<object>();\n\n for (const root of roots) {\n addUniqueObject(candidates, seen, root);\n addUniqueObject(candidates, seen, getObjectProperty(root, \"state\"));\n addUniqueObject(candidates, seen, getObjectProperty(root, \"initialState\"));\n }\n\n return candidates;\n}\n\nfunction getAgentRoots(agent: unknown): object[] {\n return [asObject(agent)]\n .concat(asObject(getObjectProperty(agent, \"agent\")))\n .filter((value): value is object => value !== undefined);\n}\n\nfunction addUniqueObject(\n candidates: object[],\n seen: Set<object>,\n value: unknown,\n) {\n if (!value || typeof value !== \"object\" || seen.has(value)) {\n return;\n }\n\n seen.add(value);\n candidates.push(value);\n}\n\nfunction asObject(value: unknown): object | undefined {\n return value && typeof value === \"object\" ? value : undefined;\n}\n\nfunction getObjectProperty(value: unknown, key: string): unknown {\n return value && typeof value === \"object\"\n ? (value as Record<string, unknown>)[key]\n : undefined;\n}\n\nfunction getRuntimeToolset<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): InferredPiAiToolset<TInput, TMetadata> | undefined {\n const candidate =\n getObjectProperty(value, \"tools\") ?? getObjectProperty(value, \"toolset\");\n\n return isPiAiToolset(candidate)\n ? (candidate as InferredPiAiToolset<TInput, TMetadata>)\n : undefined;\n}\n\nfunction getNativeToolArray<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): PiAgentToolLike<TInput, TMetadata>[] | undefined {\n const candidate = getObjectProperty(value, \"tools\");\n if (isAgentToolArray(candidate)) {\n return candidate as PiAgentToolLike<TInput, TMetadata>[];\n }\n\n return undefined;\n}\n\nasync function runAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n): Promise<TResult | HarnessRun> {\n if (options.run) {\n return options.run(args);\n }\n\n if (hasPiAiRunMethod<TInput, TMetadata, TResult, TTools>(args.agent)) {\n return args.agent.run(args.input, args.runtime);\n }\n\n throw new Error(\n \"piAiHarness requires a run() function unless the provided agent exposes run(input, runtime).\",\n );\n}\n\nfunction hasExplicitToolset<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): options is PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n> {\n return options.tools !== undefined;\n}\n\nfunction hasPiAiRunMethod<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n agent: unknown,\n): agent is PiAiRunnableAgent<TInput, TMetadata, TResult, TTools> {\n if (!agent || typeof agent !== \"object\") {\n return false;\n }\n\n return (\n \"run\" in agent && typeof (agent as { run?: unknown }).run === \"function\"\n );\n}\n\nfunction isPiAiToolset(value: unknown): value is PiAiToolset {\n if (!value || typeof value !== \"object\" || Array.isArray(value)) {\n return false;\n }\n\n const tools = Object.values(value);\n return (\n tools.length > 0 &&\n tools.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nfunction isAgentToolArray(value: unknown): value is PiAgentToolLike[] {\n return (\n Array.isArray(value) &&\n value.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"name\" in tool &&\n typeof (tool as { name?: unknown }).name === \"string\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nasync function withInstrumentedAgentTools<\n TResult,\n TInput,\n TMetadata extends HarnessMetadata,\n>(\n agent: unknown,\n toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> | undefined,\n args: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n messages: NormalizedMessage[];\n toolCalls: ToolCallRecord[];\n toolReplay: PiAiToolReplayPolicies<TInput, TMetadata> | undefined;\n executionState: PiToolExecutionState;\n },\n callback: () => Promise<TResult>,\n) {\n if (!toolsets || toolsets.length === 0) {\n return callback();\n }\n\n const originalExecutions = new Map<\n PiAgentToolLike<TInput, TMetadata>,\n PiAgentToolLike<TInput, TMetadata>[\"execute\"]\n >();\n const originalResets = new Map<ResettableAgent, ResettableAgent[\"reset\"]>();\n\n const patchTool = (tool: PiAgentToolLike<TInput, TMetadata>) => {\n if (originalExecutions.has(tool)) {\n return;\n }\n\n const originalExecute = getNativeToolExecuteOrigin(tool.execute);\n originalExecutions.set(tool, originalExecute);\n const instrumentedExecute: NativeToolExecute<TInput, TMetadata> = async (\n toolCallId: string,\n rawArgs: Record<string, JsonValue>,\n ) => {\n const startedAt = new Date();\n const toolContext = {\n input: args.input,\n metadata: args.context.metadata,\n signal: args.context.signal,\n setArtifact: args.context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n const leaveNativeTool = enterNativeToolExecution(\n args.executionState,\n tool.name,\n );\n\n try {\n const execution = await executeNativeToolWithReplay({\n toolName: tool.name,\n toolCallId,\n execute: originalExecute,\n replay: args.toolReplay?.[tool.name],\n args: rawArgs,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n result: execution.normalizedResult,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n args.messages.push({\n role: \"tool\",\n content: execution.normalizedResult,\n metadata: {\n name: tool.name,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(getReplayMetadataFromError(error)),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n } finally {\n leaveNativeTool();\n }\n };\n instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;\n tool.execute = instrumentedExecute;\n };\n\n const patchToolsets = (\n nextToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n ) => {\n for (const toolset of nextToolsets) {\n for (const tool of toolset) {\n patchTool(tool);\n }\n }\n };\n\n patchToolsets(toolsets);\n\n for (const target of getAgentResetTargets(agent)) {\n const originalReset = target.reset;\n originalResets.set(target, originalReset);\n target.reset = function patchedReset(this: ResettableAgent, ...resetArgs) {\n const resetResult = originalReset.apply(this, resetArgs);\n\n if (isPromiseLike(resetResult)) {\n return resetResult.finally(() => {\n patchToolsets(\n resolveInferredNativeToolsets<TInput, TMetadata>(agent),\n );\n });\n }\n\n patchToolsets(resolveInferredNativeToolsets<TInput, TMetadata>(agent));\n return resetResult;\n };\n }\n\n try {\n return await callback();\n } finally {\n for (const [target, originalReset] of originalResets) {\n target.reset = originalReset;\n }\n\n for (const [tool, originalExecute] of originalExecutions) {\n tool.execute = originalExecute;\n }\n }\n}\n\ntype ResettableAgent = {\n reset: (...args: unknown[]) => unknown;\n};\n\nfunction getAgentResetTargets(agent: unknown): ResettableAgent[] {\n return getAgentRoots(agent).filter(isResettableAgent);\n}\n\nfunction isResettableAgent(value: object): value is ResettableAgent {\n return \"reset\" in value && typeof value.reset === \"function\";\n}\n\nfunction resolveInferredNativeToolsets<\n TInput,\n TMetadata extends HarnessMetadata,\n>(agent: unknown): Array<PiAgentToolLike<TInput, TMetadata>[]> {\n const toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n toolsets.push(nativeTools);\n }\n }\n\n return toolsets;\n}\n\nfunction isPromiseLike(value: unknown): value is Promise<unknown> {\n return Boolean(\n value && typeof (value as { then?: unknown }).then === \"function\",\n );\n}\n\nfunction serializeToolCallError(\n error: unknown,\n): NonNullable<ToolCallRecord[\"error\"]> {\n const serialized = serializeError(error);\n const { message, type, ...details } = serialized;\n\n return {\n ...details,\n message: typeof message === \"string\" ? message : String(message),\n ...(typeof type === \"string\" ? { type } : {}),\n };\n}\n\nfunction getNativeToolExecuteOrigin<TInput, TMetadata extends HarnessMetadata>(\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"],\n) {\n const nativeExecute = execute as NativeToolExecute<TInput, TMetadata>;\n return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;\n}\n\nfunction createPiToolExecutionState(): PiToolExecutionState {\n return {\n activeNativeToolNames: new Map(),\n };\n}\n\nfunction enterNativeToolExecution(\n state: PiToolExecutionState,\n toolName: string,\n) {\n state.activeNativeToolNames.set(\n toolName,\n (state.activeNativeToolNames.get(toolName) ?? 0) + 1,\n );\n\n return () => {\n const nextCount = (state.activeNativeToolNames.get(toolName) ?? 1) - 1;\n if (nextCount <= 0) {\n state.activeNativeToolNames.delete(toolName);\n return;\n }\n\n state.activeNativeToolNames.set(toolName, nextCount);\n };\n}\n\nfunction hasActiveNativeToolExecution(\n state: PiToolExecutionState,\n toolName: string,\n) {\n return (state.activeNativeToolNames.get(toolName) ?? 0) > 0;\n}\n\nasync function executeNativeToolWithReplay<\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n toolCallId,\n execute,\n replay,\n args,\n context,\n}: {\n toolName: string;\n toolCallId: string;\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n replay: PiAiToolReplayPolicy<TInput, TMetadata> | undefined;\n args: Record<string, JsonValue>;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n let didExecute = false;\n let liveResult: unknown;\n\n const execution = await executeWithReplay({\n toolName: createNativeReplayToolName(toolName),\n args,\n context,\n execute: async (toolArgs) => {\n didExecute = true;\n liveResult = await execute(toolCallId, toolArgs);\n return createNativeToolReplayEnvelope(liveResult);\n },\n replay,\n });\n\n if (didExecute) {\n return {\n result: liveResult,\n normalizedResult: normalizeReplayToolResult(liveResult),\n replay: execution.replay,\n };\n }\n\n return {\n ...resolveNativeToolReplayResult(execution.result),\n replay: execution.replay,\n };\n}\n\nfunction createNativeReplayToolName(toolName: string) {\n return `${toolName}.native`;\n}\n\nfunction createRuntime<\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>({\n input,\n context,\n tools,\n toolReplay,\n executionState,\n messages,\n}: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n tools: TTools | undefined;\n toolReplay: PiAiToolReplayPolicies<TInput, TMetadata> | undefined;\n executionState: PiToolExecutionState;\n messages: NormalizedMessage[];\n}): PiAiRuntime<TTools, TInput, TMetadata> & {\n toolCalls: ToolCallRecord[];\n} {\n const toolCalls: ToolCallRecord[] = [];\n const eventSink: PiAiEventSink = {\n message: (message) => {\n messages.push(message);\n },\n system: (content, metadata) => {\n messages.push({\n role: \"system\",\n content,\n metadata,\n });\n },\n user: (content, metadata) => {\n messages.push({\n role: \"user\",\n content,\n metadata,\n });\n },\n assistant: (content, metadata) => {\n messages.push({\n role: \"assistant\",\n content,\n metadata,\n });\n },\n tool: (name, content, metadata) => {\n messages.push({\n role: \"tool\",\n content,\n metadata: {\n name,\n ...(metadata ?? {}),\n },\n });\n },\n };\n\n const runtimeTools = Object.fromEntries(\n Object.entries(tools ?? {}).map(([toolName, tool]) => [\n toolName,\n async (args: Record<string, JsonValue>) => {\n const startedAt = new Date();\n const isNativeImplementationCall = hasActiveNativeToolExecution(\n executionState,\n toolName,\n );\n const toolContext = {\n input,\n metadata: context.metadata,\n signal: context.signal,\n setArtifact: context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeToolWithReplay({\n toolName,\n tool,\n replay: isNativeImplementationCall\n ? undefined\n : toolReplay?.[toolName],\n args,\n context: toolContext,\n });\n const finishedAt = new Date();\n\n if (isNativeImplementationCall) {\n return execution.result;\n }\n\n const call = {\n name: toolName,\n arguments: args,\n result: execution.result,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n messages.push({\n role: \"tool\",\n content: execution.result,\n metadata: {\n name: toolName,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n if (isNativeImplementationCall) {\n throw error;\n }\n\n const call = {\n name: toolName,\n arguments: args,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(\n getReplayMetadataFromError(error),\n ),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n },\n ]),\n ) as unknown as PiAiRuntime<TTools, TInput, TMetadata>[\"tools\"];\n\n return {\n tools: runtimeTools,\n events: eventSink,\n signal: context.signal,\n toolCalls,\n };\n}\n\nfunction resolveOutput(result: unknown): JsonValue | undefined {\n if (!result || typeof result !== \"object\") {\n return toJsonValue(result);\n }\n\n const candidates = [\n \"output\",\n \"decision\",\n \"result\",\n \"final\",\n ] satisfies string[];\n\n for (const key of candidates) {\n const value = (result as Record<string, unknown>)[key];\n const normalized = toJsonValue(value);\n if (normalized !== undefined) {\n return normalized;\n }\n }\n\n return undefined;\n}\n\nfunction normalizeToolResult(result: unknown): JsonValue | undefined {\n const details =\n result && typeof result === \"object\"\n ? toJsonValue((result as { details?: unknown }).details)\n : undefined;\n if (details !== undefined) {\n return details;\n }\n\n return (\n toJsonValue(result) ?? (result === undefined ? undefined : String(result))\n );\n}\n\nfunction normalizeReplayToolResult(result: unknown): JsonValue {\n return normalizeToolResult(result) ?? null;\n}\n\ntype NativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 2;\n };\n agentResult: JsonValue;\n normalizedResult: JsonValue;\n};\n\ntype LegacyNativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 1;\n };\n normalizedResult: JsonValue;\n agentResult?: JsonValue;\n};\n\nfunction createNativeToolReplayEnvelope(\n result: unknown,\n): NativeToolReplayEnvelope {\n const normalizedResult = normalizeReplayToolResult(result);\n\n return {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\",\n version: 2,\n },\n agentResult: toJsonValue(result) ?? normalizedResult,\n normalizedResult,\n };\n}\n\nfunction resolveNativeToolReplayResult(result: JsonValue) {\n if (isNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n if (isLegacyNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult ?? result.normalizedResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n return {\n result,\n normalizedResult: normalizeReplayToolResult(result),\n };\n}\n\nfunction isNativeToolReplayEnvelope(\n value: unknown,\n): value is NativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"agentResult\" in value &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isNativeToolReplayMarker(\n value: unknown,\n): value is NativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 2,\n );\n}\n\nfunction isLegacyNativeToolReplayEnvelope(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isLegacyNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isLegacyNativeToolReplayMarker(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 1,\n );\n}\n\nfunction resolveUsage(result: unknown, toolCallCount: number): UsageSummary {\n if (!result || typeof result !== \"object\") {\n return toolCallCount > 0 ? { toolCalls: toolCallCount } : {};\n }\n\n const usageValue =\n (result as Record<string, unknown>).usage ??\n (result as Record<string, unknown>).metrics;\n\n const usage =\n usageValue && typeof usageValue === \"object\"\n ? ({ ...(usageValue as Record<string, unknown>) } as UsageSummary)\n : {};\n\n if (usage.toolCalls === undefined && toolCallCount > 0) {\n usage.toolCalls = toolCallCount;\n }\n\n return usage;\n}\n\nfunction resolveSession(\n result: unknown,\n messages: NormalizedMessage[],\n output: JsonValue | undefined,\n usage: UsageSummary,\n): NormalizedSession {\n if (\n isNormalizedSession(\n (result as Record<string, unknown> | undefined)?.session,\n )\n ) {\n return (result as { session: NormalizedSession }).session;\n }\n\n if (\n isNormalizedSession((result as Record<string, unknown> | undefined)?.trace)\n ) {\n return (result as { trace: NormalizedSession }).trace;\n }\n\n const sessionMessages = [...messages];\n if (\n output !== undefined &&\n !sessionMessages.some(\n (message) =>\n message.role === \"assistant\" && message.content !== undefined,\n )\n ) {\n sessionMessages.push({\n role: \"assistant\",\n content: output,\n });\n }\n\n return {\n messages: sessionMessages,\n outputText: typeof output === \"string\" ? output : undefined,\n provider:\n ((result as Record<string, unknown> | undefined)?.provider as\n | string\n | undefined) ?? usage.provider,\n model:\n ((result as Record<string, unknown> | undefined)?.model as\n | string\n | undefined) ?? usage.model,\n };\n}\n\nfunction resolveErrors(result: unknown): Array<Record<string, JsonValue>> {\n return resolveHarnessRunErrors(result);\n}\n\nasync function executeToolWithReplay<\n TArgs extends Record<string, JsonValue>,\n TResult extends JsonValue,\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n tool,\n replay,\n args,\n context,\n}: {\n toolName: string;\n tool: PiAiToolDefinition<TArgs, TResult, TInput, TMetadata>;\n replay: PiAiToolReplayPolicy<TInput, TMetadata> | undefined;\n args: TArgs;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n return executeWithReplay<\n Record<string, JsonValue>,\n JsonValue,\n PiAiToolContext<TInput, TMetadata>\n >({\n toolName,\n args,\n context,\n execute: (toolArgs, toolContext) =>\n tool.execute(toolArgs as TArgs, toolContext),\n replay,\n });\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAaA,qBAQO;AACP,oBAIO;AA4BP,IAAM,0BAA0B,OAAO,oCAAoC;AAySpE,SAAS,YAUd,SAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,QAAQ,QAAQ;AAAA,IAChB,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,QAAQ,MAAM,aAAa,OAAO;AACxC,YAAM,WAAgC;AAAA,QACpC;AAAA,UACE,MAAM;AAAA,UACN,aAAS,iCAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AACA,YAAM,gBAAgB;AAAA,QACpB;AAAA,MACF;AAEA,UAAI,mBAAmB,OAAO,GAAG;AAC/B,eAAO;AAAA,UACL;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,cAAc;AAAA,QAChB;AAAA,MACF;AAEA,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,cAAc;AAAA,QACd,cAAc;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AACF;AAEA,eAAe,oBAOb,SACA,OACA,OACA,SACA,UACA,cACA,gBACqB;AACrB,QAAM,iBAAiB,2BAA2B;AAClD,QAAM,UAAU,cAAc;AAAA,IAC5B;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,YAAY,QAAQ;AAAA,IACpB;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI;AACF,UAAM,SAAS,MAAM;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW,QAAQ;AAAA,QACnB,YAAY,QAAQ;AAAA,QACpB;AAAA,MACF;AAAA,MACA,MACE,SAAS,SAAS;AAAA,QAChB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACL;AAEA,YAAI,6BAAa,MAAM,KAAK,CAAC,mBAAmB,OAAO,GAAG;AACxD,UAAI,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KAAK,CAAC,OAAO,WAAW;AAClE,eAAO,YAAY,QAAQ;AAAA,MAC7B;AACA,aAAO;AAAA,IACT;AAEA,UAAM,kBAAkB;AACxB,UAAM,aAAa;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,IACV;AAQA,UAAM,SAAS,QAAQ,WAAW,SAC9B,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AACjC,UAAM,QAAQ,QAAQ,WAAW,QAC7B,MAAM,QAAQ,UAAU,MAAM,UAAU,IACxC,aAAa,iBAAiB,QAAQ,UAAU,MAAM;AAC1D,UAAM,UAAU,QAAQ,WAAW,UAC/B,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C,eAAe,iBAAiB,UAAU,QAAQ,KAAK;AAE3D,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,QAAQ,WAAW,UACxB,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C;AAAA,MACJ,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,QAAQ,WAAW,SACvB,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AAAA,IACnC;AAAA,EACF,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,QAAW,QAAQ,UAAU,MAAM;AAC9D,UAAM,MAAM;AAAA,MACV,SAAS,eAAe,QAAW,UAAU,QAAW,KAAK;AAAA,MAC7D,QAAQ;AAAA,MACR;AAAA,MACA,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,KAAC,+BAAe,KAAK,CAAC;AAAA,IAChC;AAEA,cAAM,wCAAwB,OAAO,GAAG;AAAA,EAC1C;AACF;AAEA,eAAe,aAMb,SAAyE;AACzE,MAAI,QAAQ,UAAU,QAAW;AAC/B,WAAO,QAAQ;AAAA,EACjB;AAEA,MAAI,QAAQ,aAAa;AACvB,WAAO,QAAQ,YAAY;AAAA,EAC7B;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAMP,SAA4E;AAC5E,SAAO;AAAA,IACL,QAAQ,WAAW,UACjB,QAAQ,WAAW,WACnB,QAAQ,WAAW,SACnB,QAAQ,WAAW,WACnB,QAAQ,WAAW;AAAA,EACvB;AACF;AAEA,SAAS,4BACP,OACyC;AACzC,MAAI;AACJ,QAAM,iBAA8D,CAAC;AACrE,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,mBAAmB,kBAAqC,SAAS;AACvE,QAAI,iBAAiB,UAAa,qBAAqB,QAAW;AAChE,qBAAe;AAAA,IACjB;AAEA,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,qBAAe,KAAK,WAAW;AAAA,IACjC;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;AAAA,IACvC,GAAI,eAAe,SAAS,IAAI,EAAE,eAAe,IAAI,CAAC;AAAA,EACxD;AACF;AAEA,SAAS,uBAAuB,OAA0B;AACxD,QAAM,QAAQ,cAAc,KAAK;AACjC,QAAM,aAAuB,CAAC;AAC9B,QAAM,OAAO,oBAAI,IAAY;AAE7B,aAAW,QAAQ,OAAO;AACxB,oBAAgB,YAAY,MAAM,IAAI;AACtC,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,OAAO,CAAC;AAClE,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,cAAc,CAAC;AAAA,EAC3E;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA0B;AAC/C,SAAO,CAAC,SAAS,KAAK,CAAC,EACpB,OAAO,SAAS,kBAAkB,OAAO,OAAO,CAAC,CAAC,EAClD,OAAO,CAAC,UAA2B,UAAU,MAAS;AAC3D;AAEA,SAAS,gBACP,YACA,MACA,OACA;AACA,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,KAAK,IAAI,KAAK,GAAG;AAC1D;AAAA,EACF;AAEA,OAAK,IAAI,KAAK;AACd,aAAW,KAAK,KAAK;AACvB;AAEA,SAAS,SAAS,OAAoC;AACpD,SAAO,SAAS,OAAO,UAAU,WAAW,QAAQ;AACtD;AAEA,SAAS,kBAAkB,OAAgB,KAAsB;AAC/D,SAAO,SAAS,OAAO,UAAU,WAC5B,MAAkC,GAAG,IACtC;AACN;AAEA,SAAS,kBACP,OACoD;AACpD,QAAM,YACJ,kBAAkB,OAAO,OAAO,KAAK,kBAAkB,OAAO,SAAS;AAEzE,SAAO,cAAc,SAAS,IACzB,YACD;AACN;AAEA,SAAS,mBACP,OACkD;AAClD,QAAM,YAAY,kBAAkB,OAAO,OAAO;AAClD,MAAI,iBAAiB,SAAS,GAAG;AAC/B,WAAO;AAAA,EACT;AAEA,SAAO;AACT;AAEA,eAAe,SAOb,SACA,MAC+B;AAC/B,MAAI,QAAQ,KAAK;AACf,WAAO,QAAQ,IAAI,IAAI;AAAA,EACzB;AAEA,MAAI,iBAAqD,KAAK,KAAK,GAAG;AACpE,WAAO,KAAK,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO;AAAA,EAChD;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAOP,SAOA;AACA,SAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,iBAMP,OACgE;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,SACE,SAAS,SAAS,OAAQ,MAA4B,QAAQ;AAElE;AAEA,SAAS,cAAc,OAAsC;AAC3D,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AAC/D,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,OAAO,OAAO,KAAK;AACjC,SACE,MAAM,SAAS,KACf,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,SAAS,iBAAiB,OAA4C;AACpE,SACE,MAAM,QAAQ,KAAK,KACnB,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,UAAU,QACV,OAAQ,KAA4B,SAAS,YAC7C,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,eAAe,2BAKb,OACA,UACA,MAQA,UACA;AACA,MAAI,CAAC,YAAY,SAAS,WAAW,GAAG;AACtC,WAAO,SAAS;AAAA,EAClB;AAEA,QAAM,qBAAqB,oBAAI,IAG7B;AACF,QAAM,iBAAiB,oBAAI,IAA+C;AAE1E,QAAM,YAAY,CAAC,SAA6C;AAC9D,QAAI,mBAAmB,IAAI,IAAI,GAAG;AAChC;AAAA,IACF;AAEA,UAAM,kBAAkB,2BAA2B,KAAK,OAAO;AAC/D,uBAAmB,IAAI,MAAM,eAAe;AAC5C,UAAM,sBAA4D,OAChE,YACA,YACG;AACH,YAAM,YAAY,oBAAI,KAAK;AAC3B,YAAM,cAAc;AAAA,QAClB,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK,QAAQ;AAAA,QACvB,QAAQ,KAAK,QAAQ;AAAA,QACrB,aAAa,KAAK,QAAQ;AAAA,MAC5B;AACA,YAAM,kBAAkB;AAAA,QACtB,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AAEA,UAAI;AACF,cAAM,YAAY,MAAM,4BAA4B;AAAA,UAClD,UAAU,KAAK;AAAA,UACf;AAAA,UACA,SAAS;AAAA,UACT,QAAQ,KAAK,aAAa,KAAK,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,SAAS;AAAA,QACX,CAAC;AACD,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,UAAU;AAAA,UAClB,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,cAAU,uCAAwB,UAAU,MAAM;AAAA,QACpD;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,SAAS,UAAU;AAAA,UACnB,UAAU;AAAA,YACR,MAAM,KAAK;AAAA,UACb;AAAA,QACF,CAAC;AACD,eAAO,UAAU;AAAA,MACnB,SAAS,OAAO;AACd,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,OAAO,uBAAuB,KAAK;AAAA,UACnC,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,cAAU,2CAAwB,0CAA2B,KAAK,CAAC;AAAA,QACrE;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,cAAM;AAAA,MACR,UAAE;AACA,wBAAgB;AAAA,MAClB;AAAA,IACF;AACA,wBAAoB,uBAAuB,IAAI;AAC/C,SAAK,UAAU;AAAA,EACjB;AAEA,QAAM,gBAAgB,CACpB,iBACG;AACH,eAAW,WAAW,cAAc;AAClC,iBAAW,QAAQ,SAAS;AAC1B,kBAAU,IAAI;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,gBAAc,QAAQ;AAEtB,aAAW,UAAU,qBAAqB,KAAK,GAAG;AAChD,UAAM,gBAAgB,OAAO;AAC7B,mBAAe,IAAI,QAAQ,aAAa;AACxC,WAAO,QAAQ,SAAS,gBAAuC,WAAW;AACxE,YAAM,cAAc,cAAc,MAAM,MAAM,SAAS;AAEvD,UAAI,cAAc,WAAW,GAAG;AAC9B,eAAO,YAAY,QAAQ,MAAM;AAC/B;AAAA,YACE,8BAAiD,KAAK;AAAA,UACxD;AAAA,QACF,CAAC;AAAA,MACH;AAEA,oBAAc,8BAAiD,KAAK,CAAC;AACrE,aAAO;AAAA,IACT;AAAA,EACF;AAEA,MAAI;AACF,WAAO,MAAM,SAAS;AAAA,EACxB,UAAE;AACA,eAAW,CAAC,QAAQ,aAAa,KAAK,gBAAgB;AACpD,aAAO,QAAQ;AAAA,IACjB;AAEA,eAAW,CAAC,MAAM,eAAe,KAAK,oBAAoB;AACxD,WAAK,UAAU;AAAA,IACjB;AAAA,EACF;AACF;AAMA,SAAS,qBAAqB,OAAmC;AAC/D,SAAO,cAAc,KAAK,EAAE,OAAO,iBAAiB;AACtD;AAEA,SAAS,kBAAkB,OAAyC;AAClE,SAAO,WAAW,SAAS,OAAO,MAAM,UAAU;AACpD;AAEA,SAAS,8BAGP,OAA6D;AAC7D,QAAM,WAAwD,CAAC;AAC/D,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,eAAS,KAAK,WAAW;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA2C;AAChE,SAAO;AAAA,IACL,SAAS,OAAQ,MAA6B,SAAS;AAAA,EACzD;AACF;AAEA,SAAS,uBACP,OACsC;AACtC,QAAM,iBAAa,+BAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAEA,SAAS,2BACP,SACA;AACA,QAAM,gBAAgB;AACtB,SAAO,cAAc,uBAAuB,KAAK;AACnD;AAEA,SAAS,6BAAmD;AAC1D,SAAO;AAAA,IACL,uBAAuB,oBAAI,IAAI;AAAA,EACjC;AACF;AAEA,SAAS,yBACP,OACA,UACA;AACA,QAAM,sBAAsB;AAAA,IAC1B;AAAA,KACC,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AAAA,EACrD;AAEA,SAAO,MAAM;AACX,UAAM,aAAa,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AACrE,QAAI,aAAa,GAAG;AAClB,YAAM,sBAAsB,OAAO,QAAQ;AAC3C;AAAA,IACF;AAEA,UAAM,sBAAsB,IAAI,UAAU,SAAS;AAAA,EACrD;AACF;AAEA,SAAS,6BACP,OACA,UACA;AACA,UAAQ,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AAC5D;AAEA,eAAe,4BAGb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOG;AACD,MAAI,aAAa;AACjB,MAAI;AAEJ,QAAM,YAAY,UAAM,iCAAkB;AAAA,IACxC,UAAU,2BAA2B,QAAQ;AAAA,IAC7C;AAAA,IACA;AAAA,IACA,SAAS,OAAO,aAAa;AAC3B,mBAAa;AACb,mBAAa,MAAM,QAAQ,YAAY,QAAQ;AAC/C,aAAO,+BAA+B,UAAU;AAAA,IAClD;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI,YAAY;AACd,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,kBAAkB,0BAA0B,UAAU;AAAA,MACtD,QAAQ,UAAU;AAAA,IACpB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAG,8BAA8B,UAAU,MAAM;AAAA,IACjD,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,2BAA2B,UAAkB;AACpD,SAAO,GAAG,QAAQ;AACpB;AAEA,SAAS,cAIP;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GASE;AACA,QAAM,YAA8B,CAAC;AACrC,QAAM,YAA2B;AAAA,IAC/B,SAAS,CAAC,YAAY;AACpB,eAAS,KAAK,OAAO;AAAA,IACvB;AAAA,IACA,QAAQ,CAAC,SAAS,aAAa;AAC7B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,SAAS,aAAa;AAC3B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,WAAW,CAAC,SAAS,aAAa;AAChC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,MAAM,SAAS,aAAa;AACjC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA,UAAU;AAAA,UACR;AAAA,UACA,GAAI,YAAY,CAAC;AAAA,QACnB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,OAAO;AAAA,IAC1B,OAAO,QAAQ,SAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,UAAU,IAAI,MAAM;AAAA,MACpD;AAAA,MACA,OAAO,SAAoC;AACzC,cAAM,YAAY,oBAAI,KAAK;AAC3B,cAAM,6BAA6B;AAAA,UACjC;AAAA,UACA;AAAA,QACF;AACA,cAAM,cAAc;AAAA,UAClB;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB,QAAQ,QAAQ;AAAA,UAChB,aAAa,QAAQ;AAAA,QACvB;AAEA,YAAI;AACF,gBAAM,YAAY,MAAM,sBAAsB;AAAA,YAC5C;AAAA,YACA;AAAA,YACA,QAAQ,6BACJ,SACA,aAAa,QAAQ;AAAA,YACzB;AAAA,YACA,SAAS;AAAA,UACX,CAAC;AACD,gBAAM,aAAa,oBAAI,KAAK;AAE5B,cAAI,4BAA4B;AAC9B,mBAAO,UAAU;AAAA,UACnB;AAEA,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,UAAU;AAAA,YAClB,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,cAAU,uCAAwB,UAAU,MAAM;AAAA,UACpD;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,SAAS,UAAU;AAAA,YACnB,UAAU;AAAA,cACR,MAAM;AAAA,YACR;AAAA,UACF,CAAC;AACD,iBAAO,UAAU;AAAA,QACnB,SAAS,OAAO;AACd,gBAAM,aAAa,oBAAI,KAAK;AAC5B,cAAI,4BAA4B;AAC9B,kBAAM;AAAA,UACR;AAEA,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,OAAO,uBAAuB,KAAK;AAAA,YACnC,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,cAAU;AAAA,kBACR,0CAA2B,KAAK;AAAA,YAClC;AAAA,UACF;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,QAAQ,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;AAEA,SAAS,cAAc,QAAwC;AAC7D,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,eAAO,4BAAY,MAAM;AAAA,EAC3B;AAEA,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAS,OAAmC,GAAG;AACrD,UAAM,iBAAa,4BAAY,KAAK;AACpC,QAAI,eAAe,QAAW;AAC5B,aAAO;AAAA,IACT;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,oBAAoB,QAAwC;AACnE,QAAM,UACJ,UAAU,OAAO,WAAW,eACxB,4BAAa,OAAiC,OAAO,IACrD;AACN,MAAI,YAAY,QAAW;AACzB,WAAO;AAAA,EACT;AAEA,aACE,4BAAY,MAAM,MAAM,WAAW,SAAY,SAAY,OAAO,MAAM;AAE5E;AAEA,SAAS,0BAA0B,QAA4B;AAC7D,SAAO,oBAAoB,MAAM,KAAK;AACxC;AAoBA,SAAS,+BACP,QAC0B;AAC1B,QAAM,mBAAmB,0BAA0B,MAAM;AAEzD,SAAO;AAAA,IACL,eAAe;AAAA,MACb,MAAM;AAAA,MACN,SAAS;AAAA,IACX;AAAA,IACA,iBAAa,4BAAY,MAAM,KAAK;AAAA,IACpC;AAAA,EACF;AACF;AAEA,SAAS,8BAA8B,QAAmB;AACxD,MAAI,2BAA2B,MAAM,GAAG;AACtC,WAAO;AAAA,MACL,QAAQ,OAAO;AAAA,MACf,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,MAAI,iCAAiC,MAAM,GAAG;AAC5C,WAAO;AAAA,MACL,QAAQ,OAAO,eAAe,OAAO;AAAA,MACrC,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,0BAA0B,MAAM;AAAA,EACpD;AACF;AAEA,SAAS,2BACP,OACmC;AACnC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,iBAAiB,SACjB,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,yBACP,OACoD;AACpD,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,iCACP,OACyC;AACzC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,+BACP,OAC0D;AAC1D,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,aAAa,QAAiB,eAAqC;AAC1E,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,gBAAgB,IAAI,EAAE,WAAW,cAAc,IAAI,CAAC;AAAA,EAC7D;AAEA,QAAM,aACH,OAAmC,SACnC,OAAmC;AAEtC,QAAM,QACJ,cAAc,OAAO,eAAe,WAC/B,EAAE,GAAI,WAAuC,IAC9C,CAAC;AAEP,MAAI,MAAM,cAAc,UAAa,gBAAgB,GAAG;AACtD,UAAM,YAAY;AAAA,EACpB;AAEA,SAAO;AACT;AAEA,SAAS,eACP,QACA,UACA,QACA,OACmB;AACnB,UACE;AAAA,IACG,QAAgD;AAAA,EACnD,GACA;AACA,WAAQ,OAA0C;AAAA,EACpD;AAEA,UACE,oCAAqB,QAAgD,KAAK,GAC1E;AACA,WAAQ,OAAwC;AAAA,EAClD;AAEA,QAAM,kBAAkB,CAAC,GAAG,QAAQ;AACpC,MACE,WAAW,UACX,CAAC,gBAAgB;AAAA,IACf,CAAC,YACC,QAAQ,SAAS,eAAe,QAAQ,YAAY;AAAA,EACxD,GACA;AACA,oBAAgB,KAAK;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,UAAU;AAAA,IACV,YAAY,OAAO,WAAW,WAAW,SAAS;AAAA,IAClD,UACI,QAAgD,YAEhC,MAAM;AAAA,IAC1B,OACI,QAAgD,SAEhC,MAAM;AAAA,EAC5B;AACF;AAEA,SAAS,cAAc,QAAmD;AACxE,aAAO,wCAAwB,MAAM;AACvC;AAEA,eAAe,sBAKb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAMG;AACD,aAAO,iCAIL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,CAAC,UAAU,gBAClB,KAAK,QAAQ,UAAmB,WAAW;AAAA,IAC7C;AAAA,EACF,CAAC;AACH;","names":[]}
|
package/dist/index.mjs
CHANGED
|
@@ -17,6 +17,7 @@ var ORIGINAL_NATIVE_EXECUTE = Symbol("vitest-evals.originalNativeExecute");
|
|
|
17
17
|
function piAiHarness(options) {
|
|
18
18
|
return {
|
|
19
19
|
name: options.name ?? "pi-ai",
|
|
20
|
+
prompt: options.prompt,
|
|
20
21
|
run: async (input, context) => {
|
|
21
22
|
const agent = await resolveAgent(options);
|
|
22
23
|
const messages = [
|
|
@@ -52,10 +53,13 @@ function piAiHarness(options) {
|
|
|
52
53
|
};
|
|
53
54
|
}
|
|
54
55
|
async function executePiHarnessRun(options, agent, input, context, messages, runtimeTools, nativeToolsets) {
|
|
56
|
+
const executionState = createPiToolExecutionState();
|
|
55
57
|
const runtime = createRuntime({
|
|
56
58
|
input,
|
|
57
59
|
context,
|
|
58
60
|
tools: runtimeTools,
|
|
61
|
+
toolReplay: options.toolReplay,
|
|
62
|
+
executionState,
|
|
59
63
|
messages
|
|
60
64
|
});
|
|
61
65
|
try {
|
|
@@ -66,7 +70,9 @@ async function executePiHarnessRun(options, agent, input, context, messages, run
|
|
|
66
70
|
input,
|
|
67
71
|
context,
|
|
68
72
|
messages,
|
|
69
|
-
toolCalls: runtime.toolCalls
|
|
73
|
+
toolCalls: runtime.toolCalls,
|
|
74
|
+
toolReplay: options.toolReplay,
|
|
75
|
+
executionState
|
|
70
76
|
},
|
|
71
77
|
() => runAgent(options, {
|
|
72
78
|
agent,
|
|
@@ -244,12 +250,16 @@ async function withInstrumentedAgentTools(agent, toolsets, args, callback) {
|
|
|
244
250
|
signal: args.context.signal,
|
|
245
251
|
setArtifact: args.context.setArtifact
|
|
246
252
|
};
|
|
253
|
+
const leaveNativeTool = enterNativeToolExecution(
|
|
254
|
+
args.executionState,
|
|
255
|
+
tool.name
|
|
256
|
+
);
|
|
247
257
|
try {
|
|
248
258
|
const execution = await executeNativeToolWithReplay({
|
|
249
259
|
toolName: tool.name,
|
|
250
260
|
toolCallId,
|
|
251
261
|
execute: originalExecute,
|
|
252
|
-
replay: tool.
|
|
262
|
+
replay: args.toolReplay?.[tool.name],
|
|
253
263
|
args: rawArgs,
|
|
254
264
|
context: toolContext
|
|
255
265
|
});
|
|
@@ -293,6 +303,8 @@ async function withInstrumentedAgentTools(agent, toolsets, args, callback) {
|
|
|
293
303
|
toolCalls: [call]
|
|
294
304
|
});
|
|
295
305
|
throw error;
|
|
306
|
+
} finally {
|
|
307
|
+
leaveNativeTool();
|
|
296
308
|
}
|
|
297
309
|
};
|
|
298
310
|
instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;
|
|
@@ -369,6 +381,28 @@ function getNativeToolExecuteOrigin(execute) {
|
|
|
369
381
|
const nativeExecute = execute;
|
|
370
382
|
return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;
|
|
371
383
|
}
|
|
384
|
+
function createPiToolExecutionState() {
|
|
385
|
+
return {
|
|
386
|
+
activeNativeToolNames: /* @__PURE__ */ new Map()
|
|
387
|
+
};
|
|
388
|
+
}
|
|
389
|
+
function enterNativeToolExecution(state, toolName) {
|
|
390
|
+
state.activeNativeToolNames.set(
|
|
391
|
+
toolName,
|
|
392
|
+
(state.activeNativeToolNames.get(toolName) ?? 0) + 1
|
|
393
|
+
);
|
|
394
|
+
return () => {
|
|
395
|
+
const nextCount = (state.activeNativeToolNames.get(toolName) ?? 1) - 1;
|
|
396
|
+
if (nextCount <= 0) {
|
|
397
|
+
state.activeNativeToolNames.delete(toolName);
|
|
398
|
+
return;
|
|
399
|
+
}
|
|
400
|
+
state.activeNativeToolNames.set(toolName, nextCount);
|
|
401
|
+
};
|
|
402
|
+
}
|
|
403
|
+
function hasActiveNativeToolExecution(state, toolName) {
|
|
404
|
+
return (state.activeNativeToolNames.get(toolName) ?? 0) > 0;
|
|
405
|
+
}
|
|
372
406
|
async function executeNativeToolWithReplay({
|
|
373
407
|
toolName,
|
|
374
408
|
toolCallId,
|
|
@@ -380,7 +414,7 @@ async function executeNativeToolWithReplay({
|
|
|
380
414
|
let didExecute = false;
|
|
381
415
|
let liveResult;
|
|
382
416
|
const execution = await executeWithReplay({
|
|
383
|
-
toolName,
|
|
417
|
+
toolName: createNativeReplayToolName(toolName),
|
|
384
418
|
args,
|
|
385
419
|
context,
|
|
386
420
|
execute: async (toolArgs) => {
|
|
@@ -402,10 +436,15 @@ async function executeNativeToolWithReplay({
|
|
|
402
436
|
replay: execution.replay
|
|
403
437
|
};
|
|
404
438
|
}
|
|
439
|
+
function createNativeReplayToolName(toolName) {
|
|
440
|
+
return `${toolName}.native`;
|
|
441
|
+
}
|
|
405
442
|
function createRuntime({
|
|
406
443
|
input,
|
|
407
444
|
context,
|
|
408
445
|
tools,
|
|
446
|
+
toolReplay,
|
|
447
|
+
executionState,
|
|
409
448
|
messages
|
|
410
449
|
}) {
|
|
411
450
|
const toolCalls = [];
|
|
@@ -450,6 +489,10 @@ function createRuntime({
|
|
|
450
489
|
toolName,
|
|
451
490
|
async (args) => {
|
|
452
491
|
const startedAt = /* @__PURE__ */ new Date();
|
|
492
|
+
const isNativeImplementationCall = hasActiveNativeToolExecution(
|
|
493
|
+
executionState,
|
|
494
|
+
toolName
|
|
495
|
+
);
|
|
453
496
|
const toolContext = {
|
|
454
497
|
input,
|
|
455
498
|
metadata: context.metadata,
|
|
@@ -460,10 +503,14 @@ function createRuntime({
|
|
|
460
503
|
const execution = await executeToolWithReplay({
|
|
461
504
|
toolName,
|
|
462
505
|
tool,
|
|
506
|
+
replay: isNativeImplementationCall ? void 0 : toolReplay?.[toolName],
|
|
463
507
|
args,
|
|
464
508
|
context: toolContext
|
|
465
509
|
});
|
|
466
510
|
const finishedAt = /* @__PURE__ */ new Date();
|
|
511
|
+
if (isNativeImplementationCall) {
|
|
512
|
+
return execution.result;
|
|
513
|
+
}
|
|
467
514
|
const call = {
|
|
468
515
|
name: toolName,
|
|
469
516
|
arguments: args,
|
|
@@ -488,6 +535,9 @@ function createRuntime({
|
|
|
488
535
|
return execution.result;
|
|
489
536
|
} catch (error) {
|
|
490
537
|
const finishedAt = /* @__PURE__ */ new Date();
|
|
538
|
+
if (isNativeImplementationCall) {
|
|
539
|
+
throw error;
|
|
540
|
+
}
|
|
491
541
|
const call = {
|
|
492
542
|
name: toolName,
|
|
493
543
|
arguments: args,
|
|
@@ -640,6 +690,7 @@ function resolveErrors(result) {
|
|
|
640
690
|
async function executeToolWithReplay({
|
|
641
691
|
toolName,
|
|
642
692
|
tool,
|
|
693
|
+
replay,
|
|
643
694
|
args,
|
|
644
695
|
context
|
|
645
696
|
}) {
|
|
@@ -647,8 +698,8 @@ async function executeToolWithReplay({
|
|
|
647
698
|
toolName,
|
|
648
699
|
args,
|
|
649
700
|
context,
|
|
650
|
-
execute: tool.execute,
|
|
651
|
-
replay
|
|
701
|
+
execute: (toolArgs, toolContext) => tool.execute(toolArgs, toolContext),
|
|
702
|
+
replay
|
|
652
703
|
});
|
|
653
704
|
}
|
|
654
705
|
export {
|
package/dist/index.mjs.map
CHANGED
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import type {\n Harness,\n HarnessContext,\n HarnessMetadata,\n HarnessRun,\n JsonValue,\n NormalizedMessage,\n NormalizedSession,\n TimingSummary,\n ToolCallRecord,\n UsageSummary,\n} from \"vitest-evals/harness\";\nimport {\n attachHarnessRunToError,\n isHarnessRun,\n isNormalizedSession,\n normalizeContent,\n resolveHarnessRunErrors,\n serializeError,\n toJsonValue,\n} from \"vitest-evals/harness\";\nimport {\n executeWithReplay,\n getReplayMetadataFromError,\n normalizeReplayMetadata,\n} from \"vitest-evals/replay\";\nimport type {\n ReplayMode,\n ToolRecording,\n ToolReplayConfig,\n} from \"vitest-evals/replay\";\n\ntype MaybePromise<T> = T | Promise<T>;\ntype AnyPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolDefinition<any, any, TInput, TMetadata>>;\ntype InferredPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<\n string,\n PiAiToolDefinition<Record<string, JsonValue>, JsonValue, TInput, TMetadata>\n>;\n\ntype PiAgentToolLike<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n name: string;\n replay?:\n | boolean\n | PiAiToolReplayConfig<\n Record<string, JsonValue>,\n JsonValue,\n TInput,\n TMetadata\n >;\n execute: (toolCallId: string, args: Record<string, JsonValue>) => unknown;\n};\n\nconst ORIGINAL_NATIVE_EXECUTE = Symbol(\"vitest-evals.originalNativeExecute\");\n\ntype NativeToolExecute<\n TInput,\n TMetadata extends HarnessMetadata,\n> = PiAgentToolLike<TInput, TMetadata>[\"execute\"] & {\n [ORIGINAL_NATIVE_EXECUTE]?: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n};\n\nexport type PiAiReplayMode = ReplayMode;\n\nexport interface PiAiEventSink {\n message: (message: NormalizedMessage) => void;\n system: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n user: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n assistant: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n tool: (\n name: string,\n content: JsonValue,\n metadata?: Record<string, JsonValue>,\n ) => void;\n}\n\nexport interface PiAiToolContext<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n input: TInput;\n metadata: HarnessContext<TMetadata>[\"metadata\"];\n signal?: AbortSignal;\n setArtifact: HarnessContext<TMetadata>[\"setArtifact\"];\n}\n\nexport type PiAiToolRecording<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n> = ToolRecording<TArgs, TResult>;\n\nexport type PiAiToolReplayConfig<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;\n\nexport interface PiAiToolDefinition<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n description?: string;\n replay?: boolean | PiAiToolReplayConfig<TArgs, TResult, TInput, TMetadata>;\n execute: (\n args: TArgs,\n context: PiAiToolContext<TInput, TMetadata>,\n ) => MaybePromise<TResult>;\n}\n\nexport type PiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = AnyPiAiToolset<TInput, TMetadata>;\n\ntype ToolArgs<TTool> = TTool extends PiAiToolDefinition<\n infer TArgs,\n infer _TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TArgs\n : never;\n\ntype ToolResult<TTool> = TTool extends PiAiToolDefinition<\n infer _TArgs,\n infer TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TResult\n : never;\n\nexport type PiAiRuntime<\n TTools extends PiAiToolset<TInput, TMetadata>,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n tools: {\n [K in keyof TTools]: (\n args: ToolArgs<TTools[K]>,\n ) => Promise<ToolResult<TTools[K]>>;\n };\n events: PiAiEventSink;\n signal?: AbortSignal;\n};\n\nexport interface PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> {\n agent: TAgent;\n input: TInput;\n context: HarnessContext<TMetadata>;\n runtime: PiAiRuntime<TTools, TInput, TMetadata>;\n}\n\nexport interface PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> extends PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools> {\n result: TResult;\n}\n\ninterface PiAiHarnessBaseOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n agent?: TAgent;\n createAgent?: () => MaybePromise<TAgent>;\n normalize?: PiAiHarnessNormalizeOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n name?: string;\n}\n\nexport interface PiAiHarnessWithToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {\n tools: TTools;\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport interface PiAiHarnessInferredToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n> extends PiAiHarnessBaseOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n InferredPiAiToolset<TInput, TMetadata>\n > {\n tools?: undefined;\n run?: (\n args: PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata,\n InferredPiAiToolset<TInput, TMetadata>\n >,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport type PiAiHarnessOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> =\n | PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>\n | PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>;\n\ntype PiAiHarnessRunOptions<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> & {\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\ntype PiAiRunnableAgent<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = {\n run: (\n input: TInput,\n runtime: PiAiRuntime<TTools, TInput, TMetadata>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\nexport interface PiAiHarnessNormalizeOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n session?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<NormalizedSession>;\n output?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<JsonValue | undefined>;\n usage?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<UsageSummary>;\n timings?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<TimingSummary | undefined>;\n errors?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<Array<Record<string, JsonValue>>>;\n}\n\ntype InferredToolSurfaces<TInput, TMetadata extends HarnessMetadata> = {\n runtimeTools?: InferredPiAiToolset<TInput, TMetadata>;\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>;\n};\n\n/** Adapts a Pi agent runtime into a normalized vitest-evals harness. */\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n>(\n options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): Harness<TInput, TMetadata> {\n return {\n name: options.name ?? \"pi-ai\",\n run: async (input, context) => {\n const agent = await resolveAgent(options);\n const messages: NormalizedMessage[] = [\n {\n role: \"user\",\n content: normalizeContent(input),\n },\n ];\n const inferredTools = resolveInferredToolSurfaces<TInput, TMetadata>(\n agent,\n );\n\n if (hasExplicitToolset(options)) {\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n options.tools,\n inferredTools.nativeToolsets,\n );\n }\n\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n inferredTools.runtimeTools,\n inferredTools.nativeToolsets,\n );\n },\n };\n}\n\nasync function executePiHarnessRun<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n agent: TAgent,\n input: TInput,\n context: HarnessContext<TMetadata>,\n messages: NormalizedMessage[],\n runtimeTools: TTools | undefined,\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n): Promise<HarnessRun> {\n const runtime = createRuntime({\n input,\n context,\n tools: runtimeTools,\n messages,\n });\n\n try {\n const result = await withInstrumentedAgentTools(\n agent,\n nativeToolsets,\n {\n input,\n context,\n messages,\n toolCalls: runtime.toolCalls,\n },\n () =>\n runAgent(options, {\n agent,\n input,\n context,\n runtime,\n }),\n );\n\n if (isHarnessRun(result) && !hasResultOverrides(options)) {\n if (Object.keys(context.artifacts).length > 0 && !result.artifacts) {\n result.artifacts = context.artifacts;\n }\n return result;\n }\n\n const normalizeResult = result as TResult;\n const resultArgs = {\n agent,\n input,\n context,\n runtime,\n result: normalizeResult,\n } satisfies PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n\n const output = options.normalize?.output\n ? await options.normalize.output(resultArgs)\n : resolveOutput(normalizeResult);\n const usage = options.normalize?.usage\n ? await options.normalize.usage(resultArgs)\n : resolveUsage(normalizeResult, runtime.toolCalls.length);\n const session = options.normalize?.session\n ? await options.normalize.session(resultArgs)\n : resolveSession(normalizeResult, messages, output, usage);\n\n return {\n session,\n output,\n usage,\n timings: options.normalize?.timings\n ? await options.normalize.timings(resultArgs)\n : undefined,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: options.normalize?.errors\n ? await options.normalize.errors(resultArgs)\n : resolveErrors(normalizeResult),\n };\n } catch (error) {\n const usage = resolveUsage(undefined, runtime.toolCalls.length);\n const run = {\n session: resolveSession(undefined, messages, undefined, usage),\n output: undefined,\n usage,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: [serializeError(error)],\n } satisfies HarnessRun;\n\n throw attachHarnessRunToError(error, run);\n }\n}\n\nasync function resolveAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n if (options.agent !== undefined) {\n return options.agent;\n }\n\n if (options.createAgent) {\n return options.createAgent();\n }\n\n throw new Error(\n \"piAiHarness requires either an agent instance or a createAgent() function.\",\n );\n}\n\nfunction hasResultOverrides<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n return Boolean(\n options.normalize?.output ??\n options.normalize?.session ??\n options.normalize?.usage ??\n options.normalize?.timings ??\n options.normalize?.errors,\n );\n}\n\nfunction resolveInferredToolSurfaces<TInput, TMetadata extends HarnessMetadata>(\n agent: unknown,\n): InferredToolSurfaces<TInput, TMetadata> {\n let runtimeTools: InferredPiAiToolset<TInput, TMetadata> | undefined;\n const nativeToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nextRuntimeTools = getRuntimeToolset<TInput, TMetadata>(candidate);\n if (runtimeTools === undefined && nextRuntimeTools !== undefined) {\n runtimeTools = nextRuntimeTools;\n }\n\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n nativeToolsets.push(nativeTools);\n }\n }\n\n return {\n ...(runtimeTools ? { runtimeTools } : {}),\n ...(nativeToolsets.length > 0 ? { nativeToolsets } : {}),\n };\n}\n\nfunction getAgentToolCandidates(agent: unknown): object[] {\n const roots = getAgentRoots(agent);\n const candidates: object[] = [];\n const seen = new Set<object>();\n\n for (const root of roots) {\n addUniqueObject(candidates, seen, root);\n addUniqueObject(candidates, seen, getObjectProperty(root, \"state\"));\n addUniqueObject(candidates, seen, getObjectProperty(root, \"initialState\"));\n }\n\n return candidates;\n}\n\nfunction getAgentRoots(agent: unknown): object[] {\n return [asObject(agent)]\n .concat(asObject(getObjectProperty(agent, \"agent\")))\n .filter((value): value is object => value !== undefined);\n}\n\nfunction addUniqueObject(\n candidates: object[],\n seen: Set<object>,\n value: unknown,\n) {\n if (!value || typeof value !== \"object\" || seen.has(value)) {\n return;\n }\n\n seen.add(value);\n candidates.push(value);\n}\n\nfunction asObject(value: unknown): object | undefined {\n return value && typeof value === \"object\" ? value : undefined;\n}\n\nfunction getObjectProperty(value: unknown, key: string): unknown {\n return value && typeof value === \"object\"\n ? (value as Record<string, unknown>)[key]\n : undefined;\n}\n\nfunction getRuntimeToolset<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): InferredPiAiToolset<TInput, TMetadata> | undefined {\n const candidate =\n getObjectProperty(value, \"tools\") ?? getObjectProperty(value, \"toolset\");\n\n return isPiAiToolset(candidate)\n ? (candidate as InferredPiAiToolset<TInput, TMetadata>)\n : undefined;\n}\n\nfunction getNativeToolArray<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): PiAgentToolLike<TInput, TMetadata>[] | undefined {\n const candidate = getObjectProperty(value, \"tools\");\n if (isAgentToolArray(candidate)) {\n return candidate as PiAgentToolLike<TInput, TMetadata>[];\n }\n\n return undefined;\n}\n\nasync function runAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n): Promise<TResult | HarnessRun> {\n if (options.run) {\n return options.run(args);\n }\n\n if (hasPiAiRunMethod<TInput, TMetadata, TResult, TTools>(args.agent)) {\n return args.agent.run(args.input, args.runtime);\n }\n\n throw new Error(\n \"piAiHarness requires a run() function unless the provided agent exposes run(input, runtime).\",\n );\n}\n\nfunction hasExplicitToolset<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): options is PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n> {\n return options.tools !== undefined;\n}\n\nfunction hasPiAiRunMethod<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n agent: unknown,\n): agent is PiAiRunnableAgent<TInput, TMetadata, TResult, TTools> {\n if (!agent || typeof agent !== \"object\") {\n return false;\n }\n\n return (\n \"run\" in agent && typeof (agent as { run?: unknown }).run === \"function\"\n );\n}\n\nfunction isPiAiToolset(value: unknown): value is PiAiToolset {\n if (!value || typeof value !== \"object\" || Array.isArray(value)) {\n return false;\n }\n\n const tools = Object.values(value);\n return (\n tools.length > 0 &&\n tools.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nfunction isAgentToolArray(value: unknown): value is PiAgentToolLike[] {\n return (\n Array.isArray(value) &&\n value.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"name\" in tool &&\n typeof (tool as { name?: unknown }).name === \"string\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nasync function withInstrumentedAgentTools<\n TResult,\n TInput,\n TMetadata extends HarnessMetadata,\n>(\n agent: unknown,\n toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> | undefined,\n args: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n messages: NormalizedMessage[];\n toolCalls: ToolCallRecord[];\n },\n callback: () => Promise<TResult>,\n) {\n if (!toolsets || toolsets.length === 0) {\n return callback();\n }\n\n const originalExecutions = new Map<\n PiAgentToolLike<TInput, TMetadata>,\n PiAgentToolLike<TInput, TMetadata>[\"execute\"]\n >();\n const originalResets = new Map<ResettableAgent, ResettableAgent[\"reset\"]>();\n\n const patchTool = (tool: PiAgentToolLike<TInput, TMetadata>) => {\n if (originalExecutions.has(tool)) {\n return;\n }\n\n const originalExecute = getNativeToolExecuteOrigin(tool.execute);\n originalExecutions.set(tool, originalExecute);\n const instrumentedExecute: NativeToolExecute<TInput, TMetadata> = async (\n toolCallId: string,\n rawArgs: Record<string, JsonValue>,\n ) => {\n const startedAt = new Date();\n const toolContext = {\n input: args.input,\n metadata: args.context.metadata,\n signal: args.context.signal,\n setArtifact: args.context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeNativeToolWithReplay({\n toolName: tool.name,\n toolCallId,\n execute: originalExecute,\n replay: tool.replay,\n args: rawArgs,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n result: execution.normalizedResult,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n args.messages.push({\n role: \"tool\",\n content: execution.normalizedResult,\n metadata: {\n name: tool.name,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(getReplayMetadataFromError(error)),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n };\n instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;\n tool.execute = instrumentedExecute;\n };\n\n const patchToolsets = (\n nextToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n ) => {\n for (const toolset of nextToolsets) {\n for (const tool of toolset) {\n patchTool(tool);\n }\n }\n };\n\n patchToolsets(toolsets);\n\n for (const target of getAgentResetTargets(agent)) {\n const originalReset = target.reset;\n originalResets.set(target, originalReset);\n target.reset = function patchedReset(this: ResettableAgent, ...resetArgs) {\n const resetResult = originalReset.apply(this, resetArgs);\n\n if (isPromiseLike(resetResult)) {\n return resetResult.finally(() => {\n patchToolsets(\n resolveInferredNativeToolsets<TInput, TMetadata>(agent),\n );\n });\n }\n\n patchToolsets(resolveInferredNativeToolsets<TInput, TMetadata>(agent));\n return resetResult;\n };\n }\n\n try {\n return await callback();\n } finally {\n for (const [target, originalReset] of originalResets) {\n target.reset = originalReset;\n }\n\n for (const [tool, originalExecute] of originalExecutions) {\n tool.execute = originalExecute;\n }\n }\n}\n\ntype ResettableAgent = {\n reset: (...args: unknown[]) => unknown;\n};\n\nfunction getAgentResetTargets(agent: unknown): ResettableAgent[] {\n return getAgentRoots(agent).filter(isResettableAgent);\n}\n\nfunction isResettableAgent(value: object): value is ResettableAgent {\n return \"reset\" in value && typeof value.reset === \"function\";\n}\n\nfunction resolveInferredNativeToolsets<\n TInput,\n TMetadata extends HarnessMetadata,\n>(agent: unknown): Array<PiAgentToolLike<TInput, TMetadata>[]> {\n const toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n toolsets.push(nativeTools);\n }\n }\n\n return toolsets;\n}\n\nfunction isPromiseLike(value: unknown): value is Promise<unknown> {\n return Boolean(\n value && typeof (value as { then?: unknown }).then === \"function\",\n );\n}\n\nfunction serializeToolCallError(\n error: unknown,\n): NonNullable<ToolCallRecord[\"error\"]> {\n const serialized = serializeError(error);\n const { message, type, ...details } = serialized;\n\n return {\n ...details,\n message: typeof message === \"string\" ? message : String(message),\n ...(typeof type === \"string\" ? { type } : {}),\n };\n}\n\nfunction getNativeToolExecuteOrigin<TInput, TMetadata extends HarnessMetadata>(\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"],\n) {\n const nativeExecute = execute as NativeToolExecute<TInput, TMetadata>;\n return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;\n}\n\nasync function executeNativeToolWithReplay<\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n toolCallId,\n execute,\n replay,\n args,\n context,\n}: {\n toolName: string;\n toolCallId: string;\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n replay: PiAgentToolLike<TInput, TMetadata>[\"replay\"];\n args: Record<string, JsonValue>;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n let didExecute = false;\n let liveResult: unknown;\n\n const execution = await executeWithReplay({\n toolName,\n args,\n context,\n execute: async (toolArgs) => {\n didExecute = true;\n liveResult = await execute(toolCallId, toolArgs);\n return createNativeToolReplayEnvelope(liveResult);\n },\n replay,\n });\n\n if (didExecute) {\n return {\n result: liveResult,\n normalizedResult: normalizeReplayToolResult(liveResult),\n replay: execution.replay,\n };\n }\n\n return {\n ...resolveNativeToolReplayResult(execution.result),\n replay: execution.replay,\n };\n}\n\nfunction createRuntime<\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>({\n input,\n context,\n tools,\n messages,\n}: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n tools: TTools | undefined;\n messages: NormalizedMessage[];\n}): PiAiRuntime<TTools, TInput, TMetadata> & {\n toolCalls: ToolCallRecord[];\n} {\n const toolCalls: ToolCallRecord[] = [];\n const eventSink: PiAiEventSink = {\n message: (message) => {\n messages.push(message);\n },\n system: (content, metadata) => {\n messages.push({\n role: \"system\",\n content,\n metadata,\n });\n },\n user: (content, metadata) => {\n messages.push({\n role: \"user\",\n content,\n metadata,\n });\n },\n assistant: (content, metadata) => {\n messages.push({\n role: \"assistant\",\n content,\n metadata,\n });\n },\n tool: (name, content, metadata) => {\n messages.push({\n role: \"tool\",\n content,\n metadata: {\n name,\n ...(metadata ?? {}),\n },\n });\n },\n };\n\n const runtimeTools = Object.fromEntries(\n Object.entries(tools ?? {}).map(([toolName, tool]) => [\n toolName,\n async (args: Record<string, JsonValue>) => {\n const startedAt = new Date();\n const toolContext = {\n input,\n metadata: context.metadata,\n signal: context.signal,\n setArtifact: context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeToolWithReplay({\n toolName,\n tool,\n args,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: toolName,\n arguments: args,\n result: execution.result,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n messages.push({\n role: \"tool\",\n content: execution.result,\n metadata: {\n name: toolName,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: toolName,\n arguments: args,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(\n getReplayMetadataFromError(error),\n ),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n },\n ]),\n ) as unknown as PiAiRuntime<TTools, TInput, TMetadata>[\"tools\"];\n\n return {\n tools: runtimeTools,\n events: eventSink,\n signal: context.signal,\n toolCalls,\n };\n}\n\nfunction resolveOutput(result: unknown): JsonValue | undefined {\n if (!result || typeof result !== \"object\") {\n return toJsonValue(result);\n }\n\n const candidates = [\n \"output\",\n \"decision\",\n \"result\",\n \"final\",\n ] satisfies string[];\n\n for (const key of candidates) {\n const value = (result as Record<string, unknown>)[key];\n const normalized = toJsonValue(value);\n if (normalized !== undefined) {\n return normalized;\n }\n }\n\n return undefined;\n}\n\nfunction normalizeToolResult(result: unknown): JsonValue | undefined {\n const details =\n result && typeof result === \"object\"\n ? toJsonValue((result as { details?: unknown }).details)\n : undefined;\n if (details !== undefined) {\n return details;\n }\n\n return (\n toJsonValue(result) ?? (result === undefined ? undefined : String(result))\n );\n}\n\nfunction normalizeReplayToolResult(result: unknown): JsonValue {\n return normalizeToolResult(result) ?? null;\n}\n\ntype NativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 2;\n };\n agentResult: JsonValue;\n normalizedResult: JsonValue;\n};\n\ntype LegacyNativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 1;\n };\n normalizedResult: JsonValue;\n agentResult?: JsonValue;\n};\n\nfunction createNativeToolReplayEnvelope(\n result: unknown,\n): NativeToolReplayEnvelope {\n const normalizedResult = normalizeReplayToolResult(result);\n\n return {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\",\n version: 2,\n },\n agentResult: toJsonValue(result) ?? normalizedResult,\n normalizedResult,\n };\n}\n\nfunction resolveNativeToolReplayResult(result: JsonValue) {\n if (isNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n if (isLegacyNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult ?? result.normalizedResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n return {\n result,\n normalizedResult: normalizeReplayToolResult(result),\n };\n}\n\nfunction isNativeToolReplayEnvelope(\n value: unknown,\n): value is NativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"agentResult\" in value &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isNativeToolReplayMarker(\n value: unknown,\n): value is NativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 2,\n );\n}\n\nfunction isLegacyNativeToolReplayEnvelope(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isLegacyNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isLegacyNativeToolReplayMarker(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 1,\n );\n}\n\nfunction resolveUsage(result: unknown, toolCallCount: number): UsageSummary {\n if (!result || typeof result !== \"object\") {\n return toolCallCount > 0 ? { toolCalls: toolCallCount } : {};\n }\n\n const usageValue =\n (result as Record<string, unknown>).usage ??\n (result as Record<string, unknown>).metrics;\n\n const usage =\n usageValue && typeof usageValue === \"object\"\n ? ({ ...(usageValue as Record<string, unknown>) } as UsageSummary)\n : {};\n\n if (usage.toolCalls === undefined && toolCallCount > 0) {\n usage.toolCalls = toolCallCount;\n }\n\n return usage;\n}\n\nfunction resolveSession(\n result: unknown,\n messages: NormalizedMessage[],\n output: JsonValue | undefined,\n usage: UsageSummary,\n): NormalizedSession {\n if (\n isNormalizedSession(\n (result as Record<string, unknown> | undefined)?.session,\n )\n ) {\n return (result as { session: NormalizedSession }).session;\n }\n\n if (\n isNormalizedSession((result as Record<string, unknown> | undefined)?.trace)\n ) {\n return (result as { trace: NormalizedSession }).trace;\n }\n\n const sessionMessages = [...messages];\n if (\n output !== undefined &&\n !sessionMessages.some(\n (message) =>\n message.role === \"assistant\" && message.content !== undefined,\n )\n ) {\n sessionMessages.push({\n role: \"assistant\",\n content: output,\n });\n }\n\n return {\n messages: sessionMessages,\n outputText: typeof output === \"string\" ? output : undefined,\n provider:\n ((result as Record<string, unknown> | undefined)?.provider as\n | string\n | undefined) ?? usage.provider,\n model:\n ((result as Record<string, unknown> | undefined)?.model as\n | string\n | undefined) ?? usage.model,\n };\n}\n\nfunction resolveErrors(result: unknown): Array<Record<string, JsonValue>> {\n return resolveHarnessRunErrors(result);\n}\n\nasync function executeToolWithReplay<\n TArgs extends Record<string, JsonValue>,\n TResult extends JsonValue,\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n tool,\n args,\n context,\n}: {\n toolName: string;\n tool: PiAiToolDefinition<TArgs, TResult, TInput, TMetadata>;\n args: TArgs;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n return executeWithReplay({\n toolName,\n args,\n context,\n execute: tool.execute,\n replay: tool.replay,\n });\n}\n"],"mappings":";AAYA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AAoCP,IAAM,0BAA0B,OAAO,oCAAoC;AAmRpE,SAAS,YAUd,SAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,QAAQ,MAAM,aAAa,OAAO;AACxC,YAAM,WAAgC;AAAA,QACpC;AAAA,UACE,MAAM;AAAA,UACN,SAAS,iBAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AACA,YAAM,gBAAgB;AAAA,QACpB;AAAA,MACF;AAEA,UAAI,mBAAmB,OAAO,GAAG;AAC/B,eAAO;AAAA,UACL;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,cAAc;AAAA,QAChB;AAAA,MACF;AAEA,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,cAAc;AAAA,QACd,cAAc;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AACF;AAEA,eAAe,oBAOb,SACA,OACA,OACA,SACA,UACA,cACA,gBACqB;AACrB,QAAM,UAAU,cAAc;AAAA,IAC5B;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP;AAAA,EACF,CAAC;AAED,MAAI;AACF,UAAM,SAAS,MAAM;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW,QAAQ;AAAA,MACrB;AAAA,MACA,MACE,SAAS,SAAS;AAAA,QAChB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACL;AAEA,QAAI,aAAa,MAAM,KAAK,CAAC,mBAAmB,OAAO,GAAG;AACxD,UAAI,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KAAK,CAAC,OAAO,WAAW;AAClE,eAAO,YAAY,QAAQ;AAAA,MAC7B;AACA,aAAO;AAAA,IACT;AAEA,UAAM,kBAAkB;AACxB,UAAM,aAAa;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,IACV;AAQA,UAAM,SAAS,QAAQ,WAAW,SAC9B,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AACjC,UAAM,QAAQ,QAAQ,WAAW,QAC7B,MAAM,QAAQ,UAAU,MAAM,UAAU,IACxC,aAAa,iBAAiB,QAAQ,UAAU,MAAM;AAC1D,UAAM,UAAU,QAAQ,WAAW,UAC/B,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C,eAAe,iBAAiB,UAAU,QAAQ,KAAK;AAE3D,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,QAAQ,WAAW,UACxB,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C;AAAA,MACJ,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,QAAQ,WAAW,SACvB,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AAAA,IACnC;AAAA,EACF,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,QAAW,QAAQ,UAAU,MAAM;AAC9D,UAAM,MAAM;AAAA,MACV,SAAS,eAAe,QAAW,UAAU,QAAW,KAAK;AAAA,MAC7D,QAAQ;AAAA,MACR;AAAA,MACA,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,CAAC,eAAe,KAAK,CAAC;AAAA,IAChC;AAEA,UAAM,wBAAwB,OAAO,GAAG;AAAA,EAC1C;AACF;AAEA,eAAe,aAMb,SAAyE;AACzE,MAAI,QAAQ,UAAU,QAAW;AAC/B,WAAO,QAAQ;AAAA,EACjB;AAEA,MAAI,QAAQ,aAAa;AACvB,WAAO,QAAQ,YAAY;AAAA,EAC7B;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAMP,SAA4E;AAC5E,SAAO;AAAA,IACL,QAAQ,WAAW,UACjB,QAAQ,WAAW,WACnB,QAAQ,WAAW,SACnB,QAAQ,WAAW,WACnB,QAAQ,WAAW;AAAA,EACvB;AACF;AAEA,SAAS,4BACP,OACyC;AACzC,MAAI;AACJ,QAAM,iBAA8D,CAAC;AACrE,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,mBAAmB,kBAAqC,SAAS;AACvE,QAAI,iBAAiB,UAAa,qBAAqB,QAAW;AAChE,qBAAe;AAAA,IACjB;AAEA,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,qBAAe,KAAK,WAAW;AAAA,IACjC;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;AAAA,IACvC,GAAI,eAAe,SAAS,IAAI,EAAE,eAAe,IAAI,CAAC;AAAA,EACxD;AACF;AAEA,SAAS,uBAAuB,OAA0B;AACxD,QAAM,QAAQ,cAAc,KAAK;AACjC,QAAM,aAAuB,CAAC;AAC9B,QAAM,OAAO,oBAAI,IAAY;AAE7B,aAAW,QAAQ,OAAO;AACxB,oBAAgB,YAAY,MAAM,IAAI;AACtC,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,OAAO,CAAC;AAClE,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,cAAc,CAAC;AAAA,EAC3E;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA0B;AAC/C,SAAO,CAAC,SAAS,KAAK,CAAC,EACpB,OAAO,SAAS,kBAAkB,OAAO,OAAO,CAAC,CAAC,EAClD,OAAO,CAAC,UAA2B,UAAU,MAAS;AAC3D;AAEA,SAAS,gBACP,YACA,MACA,OACA;AACA,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,KAAK,IAAI,KAAK,GAAG;AAC1D;AAAA,EACF;AAEA,OAAK,IAAI,KAAK;AACd,aAAW,KAAK,KAAK;AACvB;AAEA,SAAS,SAAS,OAAoC;AACpD,SAAO,SAAS,OAAO,UAAU,WAAW,QAAQ;AACtD;AAEA,SAAS,kBAAkB,OAAgB,KAAsB;AAC/D,SAAO,SAAS,OAAO,UAAU,WAC5B,MAAkC,GAAG,IACtC;AACN;AAEA,SAAS,kBACP,OACoD;AACpD,QAAM,YACJ,kBAAkB,OAAO,OAAO,KAAK,kBAAkB,OAAO,SAAS;AAEzE,SAAO,cAAc,SAAS,IACzB,YACD;AACN;AAEA,SAAS,mBACP,OACkD;AAClD,QAAM,YAAY,kBAAkB,OAAO,OAAO;AAClD,MAAI,iBAAiB,SAAS,GAAG;AAC/B,WAAO;AAAA,EACT;AAEA,SAAO;AACT;AAEA,eAAe,SAOb,SACA,MAC+B;AAC/B,MAAI,QAAQ,KAAK;AACf,WAAO,QAAQ,IAAI,IAAI;AAAA,EACzB;AAEA,MAAI,iBAAqD,KAAK,KAAK,GAAG;AACpE,WAAO,KAAK,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO;AAAA,EAChD;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAOP,SAOA;AACA,SAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,iBAMP,OACgE;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,SACE,SAAS,SAAS,OAAQ,MAA4B,QAAQ;AAElE;AAEA,SAAS,cAAc,OAAsC;AAC3D,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AAC/D,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,OAAO,OAAO,KAAK;AACjC,SACE,MAAM,SAAS,KACf,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,SAAS,iBAAiB,OAA4C;AACpE,SACE,MAAM,QAAQ,KAAK,KACnB,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,UAAU,QACV,OAAQ,KAA4B,SAAS,YAC7C,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,eAAe,2BAKb,OACA,UACA,MAMA,UACA;AACA,MAAI,CAAC,YAAY,SAAS,WAAW,GAAG;AACtC,WAAO,SAAS;AAAA,EAClB;AAEA,QAAM,qBAAqB,oBAAI,IAG7B;AACF,QAAM,iBAAiB,oBAAI,IAA+C;AAE1E,QAAM,YAAY,CAAC,SAA6C;AAC9D,QAAI,mBAAmB,IAAI,IAAI,GAAG;AAChC;AAAA,IACF;AAEA,UAAM,kBAAkB,2BAA2B,KAAK,OAAO;AAC/D,uBAAmB,IAAI,MAAM,eAAe;AAC5C,UAAM,sBAA4D,OAChE,YACA,YACG;AACH,YAAM,YAAY,oBAAI,KAAK;AAC3B,YAAM,cAAc;AAAA,QAClB,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK,QAAQ;AAAA,QACvB,QAAQ,KAAK,QAAQ;AAAA,QACrB,aAAa,KAAK,QAAQ;AAAA,MAC5B;AAEA,UAAI;AACF,cAAM,YAAY,MAAM,4BAA4B;AAAA,UAClD,UAAU,KAAK;AAAA,UACf;AAAA,UACA,SAAS;AAAA,UACT,QAAQ,KAAK;AAAA,UACb,MAAM;AAAA,UACN,SAAS;AAAA,QACX,CAAC;AACD,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,UAAU;AAAA,UAClB,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,UAAU,wBAAwB,UAAU,MAAM;AAAA,QACpD;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,SAAS,UAAU;AAAA,UACnB,UAAU;AAAA,YACR,MAAM,KAAK;AAAA,UACb;AAAA,QACF,CAAC;AACD,eAAO,UAAU;AAAA,MACnB,SAAS,OAAO;AACd,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,OAAO,uBAAuB,KAAK;AAAA,UACnC,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,UAAU,wBAAwB,2BAA2B,KAAK,CAAC;AAAA,QACrE;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,cAAM;AAAA,MACR;AAAA,IACF;AACA,wBAAoB,uBAAuB,IAAI;AAC/C,SAAK,UAAU;AAAA,EACjB;AAEA,QAAM,gBAAgB,CACpB,iBACG;AACH,eAAW,WAAW,cAAc;AAClC,iBAAW,QAAQ,SAAS;AAC1B,kBAAU,IAAI;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,gBAAc,QAAQ;AAEtB,aAAW,UAAU,qBAAqB,KAAK,GAAG;AAChD,UAAM,gBAAgB,OAAO;AAC7B,mBAAe,IAAI,QAAQ,aAAa;AACxC,WAAO,QAAQ,SAAS,gBAAuC,WAAW;AACxE,YAAM,cAAc,cAAc,MAAM,MAAM,SAAS;AAEvD,UAAI,cAAc,WAAW,GAAG;AAC9B,eAAO,YAAY,QAAQ,MAAM;AAC/B;AAAA,YACE,8BAAiD,KAAK;AAAA,UACxD;AAAA,QACF,CAAC;AAAA,MACH;AAEA,oBAAc,8BAAiD,KAAK,CAAC;AACrE,aAAO;AAAA,IACT;AAAA,EACF;AAEA,MAAI;AACF,WAAO,MAAM,SAAS;AAAA,EACxB,UAAE;AACA,eAAW,CAAC,QAAQ,aAAa,KAAK,gBAAgB;AACpD,aAAO,QAAQ;AAAA,IACjB;AAEA,eAAW,CAAC,MAAM,eAAe,KAAK,oBAAoB;AACxD,WAAK,UAAU;AAAA,IACjB;AAAA,EACF;AACF;AAMA,SAAS,qBAAqB,OAAmC;AAC/D,SAAO,cAAc,KAAK,EAAE,OAAO,iBAAiB;AACtD;AAEA,SAAS,kBAAkB,OAAyC;AAClE,SAAO,WAAW,SAAS,OAAO,MAAM,UAAU;AACpD;AAEA,SAAS,8BAGP,OAA6D;AAC7D,QAAM,WAAwD,CAAC;AAC/D,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,eAAS,KAAK,WAAW;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA2C;AAChE,SAAO;AAAA,IACL,SAAS,OAAQ,MAA6B,SAAS;AAAA,EACzD;AACF;AAEA,SAAS,uBACP,OACsC;AACtC,QAAM,aAAa,eAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAEA,SAAS,2BACP,SACA;AACA,QAAM,gBAAgB;AACtB,SAAO,cAAc,uBAAuB,KAAK;AACnD;AAEA,eAAe,4BAGb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOG;AACD,MAAI,aAAa;AACjB,MAAI;AAEJ,QAAM,YAAY,MAAM,kBAAkB;AAAA,IACxC;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,OAAO,aAAa;AAC3B,mBAAa;AACb,mBAAa,MAAM,QAAQ,YAAY,QAAQ;AAC/C,aAAO,+BAA+B,UAAU;AAAA,IAClD;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI,YAAY;AACd,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,kBAAkB,0BAA0B,UAAU;AAAA,MACtD,QAAQ,UAAU;AAAA,IACpB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAG,8BAA8B,UAAU,MAAM;AAAA,IACjD,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,cAIP;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOE;AACA,QAAM,YAA8B,CAAC;AACrC,QAAM,YAA2B;AAAA,IAC/B,SAAS,CAAC,YAAY;AACpB,eAAS,KAAK,OAAO;AAAA,IACvB;AAAA,IACA,QAAQ,CAAC,SAAS,aAAa;AAC7B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,SAAS,aAAa;AAC3B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,WAAW,CAAC,SAAS,aAAa;AAChC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,MAAM,SAAS,aAAa;AACjC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA,UAAU;AAAA,UACR;AAAA,UACA,GAAI,YAAY,CAAC;AAAA,QACnB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,OAAO;AAAA,IAC1B,OAAO,QAAQ,SAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,UAAU,IAAI,MAAM;AAAA,MACpD;AAAA,MACA,OAAO,SAAoC;AACzC,cAAM,YAAY,oBAAI,KAAK;AAC3B,cAAM,cAAc;AAAA,UAClB;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB,QAAQ,QAAQ;AAAA,UAChB,aAAa,QAAQ;AAAA,QACvB;AAEA,YAAI;AACF,gBAAM,YAAY,MAAM,sBAAsB;AAAA,YAC5C;AAAA,YACA;AAAA,YACA;AAAA,YACA,SAAS;AAAA,UACX,CAAC;AACD,gBAAM,aAAa,oBAAI,KAAK;AAC5B,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,UAAU;AAAA,YAClB,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,UAAU,wBAAwB,UAAU,MAAM;AAAA,UACpD;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,SAAS,UAAU;AAAA,YACnB,UAAU;AAAA,cACR,MAAM;AAAA,YACR;AAAA,UACF,CAAC;AACD,iBAAO,UAAU;AAAA,QACnB,SAAS,OAAO;AACd,gBAAM,aAAa,oBAAI,KAAK;AAC5B,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,OAAO,uBAAuB,KAAK;AAAA,YACnC,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,UAAU;AAAA,cACR,2BAA2B,KAAK;AAAA,YAClC;AAAA,UACF;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,QAAQ,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;AAEA,SAAS,cAAc,QAAwC;AAC7D,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,YAAY,MAAM;AAAA,EAC3B;AAEA,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAS,OAAmC,GAAG;AACrD,UAAM,aAAa,YAAY,KAAK;AACpC,QAAI,eAAe,QAAW;AAC5B,aAAO;AAAA,IACT;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,oBAAoB,QAAwC;AACnE,QAAM,UACJ,UAAU,OAAO,WAAW,WACxB,YAAa,OAAiC,OAAO,IACrD;AACN,MAAI,YAAY,QAAW;AACzB,WAAO;AAAA,EACT;AAEA,SACE,YAAY,MAAM,MAAM,WAAW,SAAY,SAAY,OAAO,MAAM;AAE5E;AAEA,SAAS,0BAA0B,QAA4B;AAC7D,SAAO,oBAAoB,MAAM,KAAK;AACxC;AAoBA,SAAS,+BACP,QAC0B;AAC1B,QAAM,mBAAmB,0BAA0B,MAAM;AAEzD,SAAO;AAAA,IACL,eAAe;AAAA,MACb,MAAM;AAAA,MACN,SAAS;AAAA,IACX;AAAA,IACA,aAAa,YAAY,MAAM,KAAK;AAAA,IACpC;AAAA,EACF;AACF;AAEA,SAAS,8BAA8B,QAAmB;AACxD,MAAI,2BAA2B,MAAM,GAAG;AACtC,WAAO;AAAA,MACL,QAAQ,OAAO;AAAA,MACf,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,MAAI,iCAAiC,MAAM,GAAG;AAC5C,WAAO;AAAA,MACL,QAAQ,OAAO,eAAe,OAAO;AAAA,MACrC,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,0BAA0B,MAAM;AAAA,EACpD;AACF;AAEA,SAAS,2BACP,OACmC;AACnC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,iBAAiB,SACjB,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,yBACP,OACoD;AACpD,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,iCACP,OACyC;AACzC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,+BACP,OAC0D;AAC1D,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,aAAa,QAAiB,eAAqC;AAC1E,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,gBAAgB,IAAI,EAAE,WAAW,cAAc,IAAI,CAAC;AAAA,EAC7D;AAEA,QAAM,aACH,OAAmC,SACnC,OAAmC;AAEtC,QAAM,QACJ,cAAc,OAAO,eAAe,WAC/B,EAAE,GAAI,WAAuC,IAC9C,CAAC;AAEP,MAAI,MAAM,cAAc,UAAa,gBAAgB,GAAG;AACtD,UAAM,YAAY;AAAA,EACpB;AAEA,SAAO;AACT;AAEA,SAAS,eACP,QACA,UACA,QACA,OACmB;AACnB,MACE;AAAA,IACG,QAAgD;AAAA,EACnD,GACA;AACA,WAAQ,OAA0C;AAAA,EACpD;AAEA,MACE,oBAAqB,QAAgD,KAAK,GAC1E;AACA,WAAQ,OAAwC;AAAA,EAClD;AAEA,QAAM,kBAAkB,CAAC,GAAG,QAAQ;AACpC,MACE,WAAW,UACX,CAAC,gBAAgB;AAAA,IACf,CAAC,YACC,QAAQ,SAAS,eAAe,QAAQ,YAAY;AAAA,EACxD,GACA;AACA,oBAAgB,KAAK;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,UAAU;AAAA,IACV,YAAY,OAAO,WAAW,WAAW,SAAS;AAAA,IAClD,UACI,QAAgD,YAEhC,MAAM;AAAA,IAC1B,OACI,QAAgD,SAEhC,MAAM;AAAA,EAC5B;AACF;AAEA,SAAS,cAAc,QAAmD;AACxE,SAAO,wBAAwB,MAAM;AACvC;AAEA,eAAe,sBAKb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAKG;AACD,SAAO,kBAAkB;AAAA,IACvB;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,KAAK;AAAA,IACd,QAAQ,KAAK;AAAA,EACf,CAAC;AACH;","names":[]}
|
|
1
|
+
{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import type {\n Harness,\n HarnessContext,\n HarnessMetadata,\n HarnessPrompt,\n HarnessRun,\n JsonValue,\n NormalizedMessage,\n NormalizedSession,\n TimingSummary,\n ToolCallRecord,\n UsageSummary,\n} from \"vitest-evals/harness\";\nimport {\n attachHarnessRunToError,\n isHarnessRun,\n isNormalizedSession,\n normalizeContent,\n resolveHarnessRunErrors,\n serializeError,\n toJsonValue,\n} from \"vitest-evals/harness\";\nimport {\n executeWithReplay,\n getReplayMetadataFromError,\n normalizeReplayMetadata,\n} from \"vitest-evals/replay\";\nimport type {\n ReplayMode,\n ToolRecording,\n ToolReplayConfig,\n} from \"vitest-evals/replay\";\n\ntype MaybePromise<T> = T | Promise<T>;\ntype AnyPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolDefinition<any, any, TInput, TMetadata>>;\ntype InferredPiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<\n string,\n PiAiToolDefinition<Record<string, JsonValue>, JsonValue, TInput, TMetadata>\n>;\n\ntype PiAgentToolLike<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n name: string;\n execute: (toolCallId: string, args: Record<string, JsonValue>) => unknown;\n};\n\nconst ORIGINAL_NATIVE_EXECUTE = Symbol(\"vitest-evals.originalNativeExecute\");\n\ntype NativeToolExecute<\n TInput,\n TMetadata extends HarnessMetadata,\n> = PiAgentToolLike<TInput, TMetadata>[\"execute\"] & {\n [ORIGINAL_NATIVE_EXECUTE]?: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n};\n\nexport type PiAiReplayMode = ReplayMode;\n\nexport interface PiAiEventSink {\n message: (message: NormalizedMessage) => void;\n system: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n user: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n assistant: (content: JsonValue, metadata?: Record<string, JsonValue>) => void;\n tool: (\n name: string,\n content: JsonValue,\n metadata?: Record<string, JsonValue>,\n ) => void;\n}\n\nexport interface PiAiToolContext<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n input: TInput;\n metadata: HarnessContext<TMetadata>[\"metadata\"];\n signal?: AbortSignal;\n setArtifact: HarnessContext<TMetadata>[\"setArtifact\"];\n}\n\nexport type PiAiToolRecording<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n> = ToolRecording<TArgs, TResult>;\n\nexport type PiAiToolReplayConfig<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = ToolReplayConfig<TArgs, TResult, PiAiToolContext<TInput, TMetadata>>;\n\nexport type PiAiToolReplayPolicy<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> =\n | boolean\n | PiAiToolReplayConfig<\n Record<string, JsonValue>,\n JsonValue,\n TInput,\n TMetadata\n >;\n\nexport type PiAiToolReplayPolicies<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = Record<string, PiAiToolReplayPolicy<TInput, TMetadata>>;\n\nexport interface PiAiToolDefinition<\n TArgs extends Record<string, JsonValue> = Record<string, JsonValue>,\n TResult extends JsonValue = JsonValue,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> {\n description?: string;\n execute: (\n args: TArgs,\n context: PiAiToolContext<TInput, TMetadata>,\n ) => MaybePromise<TResult>;\n}\n\nexport type PiAiToolset<\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = AnyPiAiToolset<TInput, TMetadata>;\n\ntype ToolArgs<TTool> = TTool extends PiAiToolDefinition<\n infer TArgs,\n infer _TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TArgs\n : never;\n\ntype ToolResult<TTool> = TTool extends PiAiToolDefinition<\n infer _TArgs,\n infer TResult,\n infer _TInput,\n infer _TMetadata\n>\n ? TResult\n : never;\n\nexport type PiAiRuntime<\n TTools extends PiAiToolset<TInput, TMetadata>,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n> = {\n tools: {\n [K in keyof TTools]: (\n args: ToolArgs<TTools[K]>,\n ) => Promise<ToolResult<TTools[K]>>;\n };\n events: PiAiEventSink;\n signal?: AbortSignal;\n};\n\nexport interface PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> {\n agent: TAgent;\n input: TInput;\n context: HarnessContext<TMetadata>;\n runtime: PiAiRuntime<TTools, TInput, TMetadata>;\n}\n\nexport interface PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> extends PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools> {\n result: TResult;\n}\n\ninterface PiAiHarnessBaseOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n agent?: TAgent;\n createAgent?: () => MaybePromise<TAgent>;\n toolReplay?: PiAiToolReplayPolicies<TInput, TMetadata>;\n normalize?: PiAiHarnessNormalizeOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n prompt: HarnessPrompt;\n name?: string;\n}\n\nexport interface PiAiHarnessWithToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> extends PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> {\n tools: TTools;\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport interface PiAiHarnessInferredToolsOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n> extends PiAiHarnessBaseOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n InferredPiAiToolset<TInput, TMetadata>\n > {\n tools?: undefined;\n run?: (\n args: PiAiHarnessRunArgs<\n TAgent,\n TInput,\n TMetadata,\n InferredPiAiToolset<TInput, TMetadata>\n >,\n ) => MaybePromise<TResult | HarnessRun>;\n}\n\nexport type PiAiHarnessOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> =\n | PiAiHarnessWithToolsOptions<TAgent, TInput, TMetadata, TResult, TTools>\n | PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>;\n\ntype PiAiHarnessRunOptions<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = PiAiHarnessBaseOptions<TAgent, TInput, TMetadata, TResult, TTools> & {\n run?: (\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\ntype PiAiRunnableAgent<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n> = {\n run: (\n input: TInput,\n runtime: PiAiRuntime<TTools, TInput, TMetadata>,\n ) => MaybePromise<TResult | HarnessRun>;\n};\n\nexport interface PiAiHarnessNormalizeOptions<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n> {\n session?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<NormalizedSession>;\n output?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<JsonValue | undefined>;\n usage?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<UsageSummary>;\n timings?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<TimingSummary | undefined>;\n errors?: (\n args: PiAiHarnessResultArgs<TAgent, TInput, TMetadata, TResult, TTools>,\n ) => MaybePromise<Array<Record<string, JsonValue>>>;\n}\n\ntype InferredToolSurfaces<TInput, TMetadata extends HarnessMetadata> = {\n runtimeTools?: InferredPiAiToolset<TInput, TMetadata>;\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>;\n};\n\ntype PiToolExecutionState = {\n activeNativeToolNames: Map<string, number>;\n};\n\n/** Adapts a Pi agent runtime into a normalized vitest-evals harness. */\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n>(\n options: PiAiHarnessInferredToolsOptions<TAgent, TInput, TMetadata, TResult>,\n): Harness<TInput, TMetadata>;\nexport function piAiHarness<\n TAgent,\n TInput = string,\n TMetadata extends HarnessMetadata = HarnessMetadata,\n TResult = unknown,\n TTools extends PiAiToolset<TInput, TMetadata> = PiAiToolset<\n TInput,\n TMetadata\n >,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): Harness<TInput, TMetadata> {\n return {\n name: options.name ?? \"pi-ai\",\n prompt: options.prompt,\n run: async (input, context) => {\n const agent = await resolveAgent(options);\n const messages: NormalizedMessage[] = [\n {\n role: \"user\",\n content: normalizeContent(input),\n },\n ];\n const inferredTools = resolveInferredToolSurfaces<TInput, TMetadata>(\n agent,\n );\n\n if (hasExplicitToolset(options)) {\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n options.tools,\n inferredTools.nativeToolsets,\n );\n }\n\n return executePiHarnessRun(\n options,\n agent,\n input,\n context,\n messages,\n inferredTools.runtimeTools,\n inferredTools.nativeToolsets,\n );\n },\n };\n}\n\nasync function executePiHarnessRun<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n agent: TAgent,\n input: TInput,\n context: HarnessContext<TMetadata>,\n messages: NormalizedMessage[],\n runtimeTools: TTools | undefined,\n nativeToolsets?: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n): Promise<HarnessRun> {\n const executionState = createPiToolExecutionState();\n const runtime = createRuntime({\n input,\n context,\n tools: runtimeTools,\n toolReplay: options.toolReplay,\n executionState,\n messages,\n });\n\n try {\n const result = await withInstrumentedAgentTools(\n agent,\n nativeToolsets,\n {\n input,\n context,\n messages,\n toolCalls: runtime.toolCalls,\n toolReplay: options.toolReplay,\n executionState,\n },\n () =>\n runAgent(options, {\n agent,\n input,\n context,\n runtime,\n }),\n );\n\n if (isHarnessRun(result) && !hasResultOverrides(options)) {\n if (Object.keys(context.artifacts).length > 0 && !result.artifacts) {\n result.artifacts = context.artifacts;\n }\n return result;\n }\n\n const normalizeResult = result as TResult;\n const resultArgs = {\n agent,\n input,\n context,\n runtime,\n result: normalizeResult,\n } satisfies PiAiHarnessResultArgs<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n >;\n\n const output = options.normalize?.output\n ? await options.normalize.output(resultArgs)\n : resolveOutput(normalizeResult);\n const usage = options.normalize?.usage\n ? await options.normalize.usage(resultArgs)\n : resolveUsage(normalizeResult, runtime.toolCalls.length);\n const session = options.normalize?.session\n ? await options.normalize.session(resultArgs)\n : resolveSession(normalizeResult, messages, output, usage);\n\n return {\n session,\n output,\n usage,\n timings: options.normalize?.timings\n ? await options.normalize.timings(resultArgs)\n : undefined,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: options.normalize?.errors\n ? await options.normalize.errors(resultArgs)\n : resolveErrors(normalizeResult),\n };\n } catch (error) {\n const usage = resolveUsage(undefined, runtime.toolCalls.length);\n const run = {\n session: resolveSession(undefined, messages, undefined, usage),\n output: undefined,\n usage,\n artifacts:\n Object.keys(context.artifacts).length > 0\n ? context.artifacts\n : undefined,\n errors: [serializeError(error)],\n } satisfies HarnessRun;\n\n throw attachHarnessRunToError(error, run);\n }\n}\n\nasync function resolveAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n if (options.agent !== undefined) {\n return options.agent;\n }\n\n if (options.createAgent) {\n return options.createAgent();\n }\n\n throw new Error(\n \"piAiHarness requires either an agent instance or a createAgent() function.\",\n );\n}\n\nfunction hasResultOverrides<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>) {\n return Boolean(\n options.normalize?.output ??\n options.normalize?.session ??\n options.normalize?.usage ??\n options.normalize?.timings ??\n options.normalize?.errors,\n );\n}\n\nfunction resolveInferredToolSurfaces<TInput, TMetadata extends HarnessMetadata>(\n agent: unknown,\n): InferredToolSurfaces<TInput, TMetadata> {\n let runtimeTools: InferredPiAiToolset<TInput, TMetadata> | undefined;\n const nativeToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nextRuntimeTools = getRuntimeToolset<TInput, TMetadata>(candidate);\n if (runtimeTools === undefined && nextRuntimeTools !== undefined) {\n runtimeTools = nextRuntimeTools;\n }\n\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n nativeToolsets.push(nativeTools);\n }\n }\n\n return {\n ...(runtimeTools ? { runtimeTools } : {}),\n ...(nativeToolsets.length > 0 ? { nativeToolsets } : {}),\n };\n}\n\nfunction getAgentToolCandidates(agent: unknown): object[] {\n const roots = getAgentRoots(agent);\n const candidates: object[] = [];\n const seen = new Set<object>();\n\n for (const root of roots) {\n addUniqueObject(candidates, seen, root);\n addUniqueObject(candidates, seen, getObjectProperty(root, \"state\"));\n addUniqueObject(candidates, seen, getObjectProperty(root, \"initialState\"));\n }\n\n return candidates;\n}\n\nfunction getAgentRoots(agent: unknown): object[] {\n return [asObject(agent)]\n .concat(asObject(getObjectProperty(agent, \"agent\")))\n .filter((value): value is object => value !== undefined);\n}\n\nfunction addUniqueObject(\n candidates: object[],\n seen: Set<object>,\n value: unknown,\n) {\n if (!value || typeof value !== \"object\" || seen.has(value)) {\n return;\n }\n\n seen.add(value);\n candidates.push(value);\n}\n\nfunction asObject(value: unknown): object | undefined {\n return value && typeof value === \"object\" ? value : undefined;\n}\n\nfunction getObjectProperty(value: unknown, key: string): unknown {\n return value && typeof value === \"object\"\n ? (value as Record<string, unknown>)[key]\n : undefined;\n}\n\nfunction getRuntimeToolset<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): InferredPiAiToolset<TInput, TMetadata> | undefined {\n const candidate =\n getObjectProperty(value, \"tools\") ?? getObjectProperty(value, \"toolset\");\n\n return isPiAiToolset(candidate)\n ? (candidate as InferredPiAiToolset<TInput, TMetadata>)\n : undefined;\n}\n\nfunction getNativeToolArray<TInput, TMetadata extends HarnessMetadata>(\n value: object,\n): PiAgentToolLike<TInput, TMetadata>[] | undefined {\n const candidate = getObjectProperty(value, \"tools\");\n if (isAgentToolArray(candidate)) {\n return candidate as PiAgentToolLike<TInput, TMetadata>[];\n }\n\n return undefined;\n}\n\nasync function runAgent<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessRunOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n args: PiAiHarnessRunArgs<TAgent, TInput, TMetadata, TTools>,\n): Promise<TResult | HarnessRun> {\n if (options.run) {\n return options.run(args);\n }\n\n if (hasPiAiRunMethod<TInput, TMetadata, TResult, TTools>(args.agent)) {\n return args.agent.run(args.input, args.runtime);\n }\n\n throw new Error(\n \"piAiHarness requires a run() function unless the provided agent exposes run(input, runtime).\",\n );\n}\n\nfunction hasExplicitToolset<\n TAgent,\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n options: PiAiHarnessOptions<TAgent, TInput, TMetadata, TResult, TTools>,\n): options is PiAiHarnessWithToolsOptions<\n TAgent,\n TInput,\n TMetadata,\n TResult,\n TTools\n> {\n return options.tools !== undefined;\n}\n\nfunction hasPiAiRunMethod<\n TInput,\n TMetadata extends HarnessMetadata,\n TResult,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>(\n agent: unknown,\n): agent is PiAiRunnableAgent<TInput, TMetadata, TResult, TTools> {\n if (!agent || typeof agent !== \"object\") {\n return false;\n }\n\n return (\n \"run\" in agent && typeof (agent as { run?: unknown }).run === \"function\"\n );\n}\n\nfunction isPiAiToolset(value: unknown): value is PiAiToolset {\n if (!value || typeof value !== \"object\" || Array.isArray(value)) {\n return false;\n }\n\n const tools = Object.values(value);\n return (\n tools.length > 0 &&\n tools.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nfunction isAgentToolArray(value: unknown): value is PiAgentToolLike[] {\n return (\n Array.isArray(value) &&\n value.every((tool) =>\n Boolean(\n tool &&\n typeof tool === \"object\" &&\n \"name\" in tool &&\n typeof (tool as { name?: unknown }).name === \"string\" &&\n \"execute\" in tool &&\n typeof (tool as { execute?: unknown }).execute === \"function\",\n ),\n )\n );\n}\n\nasync function withInstrumentedAgentTools<\n TResult,\n TInput,\n TMetadata extends HarnessMetadata,\n>(\n agent: unknown,\n toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> | undefined,\n args: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n messages: NormalizedMessage[];\n toolCalls: ToolCallRecord[];\n toolReplay: PiAiToolReplayPolicies<TInput, TMetadata> | undefined;\n executionState: PiToolExecutionState;\n },\n callback: () => Promise<TResult>,\n) {\n if (!toolsets || toolsets.length === 0) {\n return callback();\n }\n\n const originalExecutions = new Map<\n PiAgentToolLike<TInput, TMetadata>,\n PiAgentToolLike<TInput, TMetadata>[\"execute\"]\n >();\n const originalResets = new Map<ResettableAgent, ResettableAgent[\"reset\"]>();\n\n const patchTool = (tool: PiAgentToolLike<TInput, TMetadata>) => {\n if (originalExecutions.has(tool)) {\n return;\n }\n\n const originalExecute = getNativeToolExecuteOrigin(tool.execute);\n originalExecutions.set(tool, originalExecute);\n const instrumentedExecute: NativeToolExecute<TInput, TMetadata> = async (\n toolCallId: string,\n rawArgs: Record<string, JsonValue>,\n ) => {\n const startedAt = new Date();\n const toolContext = {\n input: args.input,\n metadata: args.context.metadata,\n signal: args.context.signal,\n setArtifact: args.context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n const leaveNativeTool = enterNativeToolExecution(\n args.executionState,\n tool.name,\n );\n\n try {\n const execution = await executeNativeToolWithReplay({\n toolName: tool.name,\n toolCallId,\n execute: originalExecute,\n replay: args.toolReplay?.[tool.name],\n args: rawArgs,\n context: toolContext,\n });\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n result: execution.normalizedResult,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n args.messages.push({\n role: \"tool\",\n content: execution.normalizedResult,\n metadata: {\n name: tool.name,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n const call = {\n name: tool.name,\n arguments: rawArgs,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(getReplayMetadataFromError(error)),\n } satisfies ToolCallRecord;\n args.toolCalls.push(call);\n args.messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n } finally {\n leaveNativeTool();\n }\n };\n instrumentedExecute[ORIGINAL_NATIVE_EXECUTE] = originalExecute;\n tool.execute = instrumentedExecute;\n };\n\n const patchToolsets = (\n nextToolsets: Array<PiAgentToolLike<TInput, TMetadata>[]>,\n ) => {\n for (const toolset of nextToolsets) {\n for (const tool of toolset) {\n patchTool(tool);\n }\n }\n };\n\n patchToolsets(toolsets);\n\n for (const target of getAgentResetTargets(agent)) {\n const originalReset = target.reset;\n originalResets.set(target, originalReset);\n target.reset = function patchedReset(this: ResettableAgent, ...resetArgs) {\n const resetResult = originalReset.apply(this, resetArgs);\n\n if (isPromiseLike(resetResult)) {\n return resetResult.finally(() => {\n patchToolsets(\n resolveInferredNativeToolsets<TInput, TMetadata>(agent),\n );\n });\n }\n\n patchToolsets(resolveInferredNativeToolsets<TInput, TMetadata>(agent));\n return resetResult;\n };\n }\n\n try {\n return await callback();\n } finally {\n for (const [target, originalReset] of originalResets) {\n target.reset = originalReset;\n }\n\n for (const [tool, originalExecute] of originalExecutions) {\n tool.execute = originalExecute;\n }\n }\n}\n\ntype ResettableAgent = {\n reset: (...args: unknown[]) => unknown;\n};\n\nfunction getAgentResetTargets(agent: unknown): ResettableAgent[] {\n return getAgentRoots(agent).filter(isResettableAgent);\n}\n\nfunction isResettableAgent(value: object): value is ResettableAgent {\n return \"reset\" in value && typeof value.reset === \"function\";\n}\n\nfunction resolveInferredNativeToolsets<\n TInput,\n TMetadata extends HarnessMetadata,\n>(agent: unknown): Array<PiAgentToolLike<TInput, TMetadata>[]> {\n const toolsets: Array<PiAgentToolLike<TInput, TMetadata>[]> = [];\n const seenToolsets = new Set<PiAgentToolLike<TInput, TMetadata>[]>();\n\n for (const candidate of getAgentToolCandidates(agent)) {\n const nativeTools = getNativeToolArray(candidate);\n if (nativeTools && !seenToolsets.has(nativeTools)) {\n seenToolsets.add(nativeTools);\n toolsets.push(nativeTools);\n }\n }\n\n return toolsets;\n}\n\nfunction isPromiseLike(value: unknown): value is Promise<unknown> {\n return Boolean(\n value && typeof (value as { then?: unknown }).then === \"function\",\n );\n}\n\nfunction serializeToolCallError(\n error: unknown,\n): NonNullable<ToolCallRecord[\"error\"]> {\n const serialized = serializeError(error);\n const { message, type, ...details } = serialized;\n\n return {\n ...details,\n message: typeof message === \"string\" ? message : String(message),\n ...(typeof type === \"string\" ? { type } : {}),\n };\n}\n\nfunction getNativeToolExecuteOrigin<TInput, TMetadata extends HarnessMetadata>(\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"],\n) {\n const nativeExecute = execute as NativeToolExecute<TInput, TMetadata>;\n return nativeExecute[ORIGINAL_NATIVE_EXECUTE] ?? nativeExecute;\n}\n\nfunction createPiToolExecutionState(): PiToolExecutionState {\n return {\n activeNativeToolNames: new Map(),\n };\n}\n\nfunction enterNativeToolExecution(\n state: PiToolExecutionState,\n toolName: string,\n) {\n state.activeNativeToolNames.set(\n toolName,\n (state.activeNativeToolNames.get(toolName) ?? 0) + 1,\n );\n\n return () => {\n const nextCount = (state.activeNativeToolNames.get(toolName) ?? 1) - 1;\n if (nextCount <= 0) {\n state.activeNativeToolNames.delete(toolName);\n return;\n }\n\n state.activeNativeToolNames.set(toolName, nextCount);\n };\n}\n\nfunction hasActiveNativeToolExecution(\n state: PiToolExecutionState,\n toolName: string,\n) {\n return (state.activeNativeToolNames.get(toolName) ?? 0) > 0;\n}\n\nasync function executeNativeToolWithReplay<\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n toolCallId,\n execute,\n replay,\n args,\n context,\n}: {\n toolName: string;\n toolCallId: string;\n execute: PiAgentToolLike<TInput, TMetadata>[\"execute\"];\n replay: PiAiToolReplayPolicy<TInput, TMetadata> | undefined;\n args: Record<string, JsonValue>;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n let didExecute = false;\n let liveResult: unknown;\n\n const execution = await executeWithReplay({\n toolName: createNativeReplayToolName(toolName),\n args,\n context,\n execute: async (toolArgs) => {\n didExecute = true;\n liveResult = await execute(toolCallId, toolArgs);\n return createNativeToolReplayEnvelope(liveResult);\n },\n replay,\n });\n\n if (didExecute) {\n return {\n result: liveResult,\n normalizedResult: normalizeReplayToolResult(liveResult),\n replay: execution.replay,\n };\n }\n\n return {\n ...resolveNativeToolReplayResult(execution.result),\n replay: execution.replay,\n };\n}\n\nfunction createNativeReplayToolName(toolName: string) {\n return `${toolName}.native`;\n}\n\nfunction createRuntime<\n TInput,\n TMetadata extends HarnessMetadata,\n TTools extends PiAiToolset<TInput, TMetadata>,\n>({\n input,\n context,\n tools,\n toolReplay,\n executionState,\n messages,\n}: {\n input: TInput;\n context: HarnessContext<TMetadata>;\n tools: TTools | undefined;\n toolReplay: PiAiToolReplayPolicies<TInput, TMetadata> | undefined;\n executionState: PiToolExecutionState;\n messages: NormalizedMessage[];\n}): PiAiRuntime<TTools, TInput, TMetadata> & {\n toolCalls: ToolCallRecord[];\n} {\n const toolCalls: ToolCallRecord[] = [];\n const eventSink: PiAiEventSink = {\n message: (message) => {\n messages.push(message);\n },\n system: (content, metadata) => {\n messages.push({\n role: \"system\",\n content,\n metadata,\n });\n },\n user: (content, metadata) => {\n messages.push({\n role: \"user\",\n content,\n metadata,\n });\n },\n assistant: (content, metadata) => {\n messages.push({\n role: \"assistant\",\n content,\n metadata,\n });\n },\n tool: (name, content, metadata) => {\n messages.push({\n role: \"tool\",\n content,\n metadata: {\n name,\n ...(metadata ?? {}),\n },\n });\n },\n };\n\n const runtimeTools = Object.fromEntries(\n Object.entries(tools ?? {}).map(([toolName, tool]) => [\n toolName,\n async (args: Record<string, JsonValue>) => {\n const startedAt = new Date();\n const isNativeImplementationCall = hasActiveNativeToolExecution(\n executionState,\n toolName,\n );\n const toolContext = {\n input,\n metadata: context.metadata,\n signal: context.signal,\n setArtifact: context.setArtifact,\n } satisfies PiAiToolContext<TInput, TMetadata>;\n\n try {\n const execution = await executeToolWithReplay({\n toolName,\n tool,\n replay: isNativeImplementationCall\n ? undefined\n : toolReplay?.[toolName],\n args,\n context: toolContext,\n });\n const finishedAt = new Date();\n\n if (isNativeImplementationCall) {\n return execution.result;\n }\n\n const call = {\n name: toolName,\n arguments: args,\n result: execution.result,\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(execution.replay),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n messages.push({\n role: \"tool\",\n content: execution.result,\n metadata: {\n name: toolName,\n },\n });\n return execution.result;\n } catch (error) {\n const finishedAt = new Date();\n if (isNativeImplementationCall) {\n throw error;\n }\n\n const call = {\n name: toolName,\n arguments: args,\n error: serializeToolCallError(error),\n startedAt: startedAt.toISOString(),\n finishedAt: finishedAt.toISOString(),\n durationMs: finishedAt.getTime() - startedAt.getTime(),\n metadata: normalizeReplayMetadata(\n getReplayMetadataFromError(error),\n ),\n } satisfies ToolCallRecord;\n toolCalls.push(call);\n messages.push({\n role: \"assistant\",\n toolCalls: [call],\n });\n throw error;\n }\n },\n ]),\n ) as unknown as PiAiRuntime<TTools, TInput, TMetadata>[\"tools\"];\n\n return {\n tools: runtimeTools,\n events: eventSink,\n signal: context.signal,\n toolCalls,\n };\n}\n\nfunction resolveOutput(result: unknown): JsonValue | undefined {\n if (!result || typeof result !== \"object\") {\n return toJsonValue(result);\n }\n\n const candidates = [\n \"output\",\n \"decision\",\n \"result\",\n \"final\",\n ] satisfies string[];\n\n for (const key of candidates) {\n const value = (result as Record<string, unknown>)[key];\n const normalized = toJsonValue(value);\n if (normalized !== undefined) {\n return normalized;\n }\n }\n\n return undefined;\n}\n\nfunction normalizeToolResult(result: unknown): JsonValue | undefined {\n const details =\n result && typeof result === \"object\"\n ? toJsonValue((result as { details?: unknown }).details)\n : undefined;\n if (details !== undefined) {\n return details;\n }\n\n return (\n toJsonValue(result) ?? (result === undefined ? undefined : String(result))\n );\n}\n\nfunction normalizeReplayToolResult(result: unknown): JsonValue {\n return normalizeToolResult(result) ?? null;\n}\n\ntype NativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 2;\n };\n agentResult: JsonValue;\n normalizedResult: JsonValue;\n};\n\ntype LegacyNativeToolReplayEnvelope = {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\";\n version: 1;\n };\n normalizedResult: JsonValue;\n agentResult?: JsonValue;\n};\n\nfunction createNativeToolReplayEnvelope(\n result: unknown,\n): NativeToolReplayEnvelope {\n const normalizedResult = normalizeReplayToolResult(result);\n\n return {\n __vitestEvals: {\n kind: \"pi-ai-native-tool-result\",\n version: 2,\n },\n agentResult: toJsonValue(result) ?? normalizedResult,\n normalizedResult,\n };\n}\n\nfunction resolveNativeToolReplayResult(result: JsonValue) {\n if (isNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n if (isLegacyNativeToolReplayEnvelope(result)) {\n return {\n result: result.agentResult ?? result.normalizedResult,\n normalizedResult: result.normalizedResult,\n };\n }\n\n return {\n result,\n normalizedResult: normalizeReplayToolResult(result),\n };\n}\n\nfunction isNativeToolReplayEnvelope(\n value: unknown,\n): value is NativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"agentResult\" in value &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isNativeToolReplayMarker(\n value: unknown,\n): value is NativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 2,\n );\n}\n\nfunction isLegacyNativeToolReplayEnvelope(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"__vitestEvals\" in value &&\n isLegacyNativeToolReplayMarker(\n (value as { __vitestEvals?: unknown }).__vitestEvals,\n ) &&\n \"normalizedResult\" in value,\n );\n}\n\nfunction isLegacyNativeToolReplayMarker(\n value: unknown,\n): value is LegacyNativeToolReplayEnvelope[\"__vitestEvals\"] {\n return Boolean(\n value &&\n typeof value === \"object\" &&\n \"kind\" in value &&\n (value as { kind?: unknown }).kind === \"pi-ai-native-tool-result\" &&\n \"version\" in value &&\n (value as { version?: unknown }).version === 1,\n );\n}\n\nfunction resolveUsage(result: unknown, toolCallCount: number): UsageSummary {\n if (!result || typeof result !== \"object\") {\n return toolCallCount > 0 ? { toolCalls: toolCallCount } : {};\n }\n\n const usageValue =\n (result as Record<string, unknown>).usage ??\n (result as Record<string, unknown>).metrics;\n\n const usage =\n usageValue && typeof usageValue === \"object\"\n ? ({ ...(usageValue as Record<string, unknown>) } as UsageSummary)\n : {};\n\n if (usage.toolCalls === undefined && toolCallCount > 0) {\n usage.toolCalls = toolCallCount;\n }\n\n return usage;\n}\n\nfunction resolveSession(\n result: unknown,\n messages: NormalizedMessage[],\n output: JsonValue | undefined,\n usage: UsageSummary,\n): NormalizedSession {\n if (\n isNormalizedSession(\n (result as Record<string, unknown> | undefined)?.session,\n )\n ) {\n return (result as { session: NormalizedSession }).session;\n }\n\n if (\n isNormalizedSession((result as Record<string, unknown> | undefined)?.trace)\n ) {\n return (result as { trace: NormalizedSession }).trace;\n }\n\n const sessionMessages = [...messages];\n if (\n output !== undefined &&\n !sessionMessages.some(\n (message) =>\n message.role === \"assistant\" && message.content !== undefined,\n )\n ) {\n sessionMessages.push({\n role: \"assistant\",\n content: output,\n });\n }\n\n return {\n messages: sessionMessages,\n outputText: typeof output === \"string\" ? output : undefined,\n provider:\n ((result as Record<string, unknown> | undefined)?.provider as\n | string\n | undefined) ?? usage.provider,\n model:\n ((result as Record<string, unknown> | undefined)?.model as\n | string\n | undefined) ?? usage.model,\n };\n}\n\nfunction resolveErrors(result: unknown): Array<Record<string, JsonValue>> {\n return resolveHarnessRunErrors(result);\n}\n\nasync function executeToolWithReplay<\n TArgs extends Record<string, JsonValue>,\n TResult extends JsonValue,\n TInput,\n TMetadata extends HarnessMetadata,\n>({\n toolName,\n tool,\n replay,\n args,\n context,\n}: {\n toolName: string;\n tool: PiAiToolDefinition<TArgs, TResult, TInput, TMetadata>;\n replay: PiAiToolReplayPolicy<TInput, TMetadata> | undefined;\n args: TArgs;\n context: PiAiToolContext<TInput, TMetadata>;\n}) {\n return executeWithReplay<\n Record<string, JsonValue>,\n JsonValue,\n PiAiToolContext<TInput, TMetadata>\n >({\n toolName,\n args,\n context,\n execute: (toolArgs, toolContext) =>\n tool.execute(toolArgs as TArgs, toolContext),\n replay,\n });\n}\n"],"mappings":";AAaA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OACK;AACP;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,OACK;AA4BP,IAAM,0BAA0B,OAAO,oCAAoC;AAySpE,SAAS,YAUd,SAC4B;AAC5B,SAAO;AAAA,IACL,MAAM,QAAQ,QAAQ;AAAA,IACtB,QAAQ,QAAQ;AAAA,IAChB,KAAK,OAAO,OAAO,YAAY;AAC7B,YAAM,QAAQ,MAAM,aAAa,OAAO;AACxC,YAAM,WAAgC;AAAA,QACpC;AAAA,UACE,MAAM;AAAA,UACN,SAAS,iBAAiB,KAAK;AAAA,QACjC;AAAA,MACF;AACA,YAAM,gBAAgB;AAAA,QACpB;AAAA,MACF;AAEA,UAAI,mBAAmB,OAAO,GAAG;AAC/B,eAAO;AAAA,UACL;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,cAAc;AAAA,QAChB;AAAA,MACF;AAEA,aAAO;AAAA,QACL;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,QACA,cAAc;AAAA,QACd,cAAc;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AACF;AAEA,eAAe,oBAOb,SACA,OACA,OACA,SACA,UACA,cACA,gBACqB;AACrB,QAAM,iBAAiB,2BAA2B;AAClD,QAAM,UAAU,cAAc;AAAA,IAC5B;AAAA,IACA;AAAA,IACA,OAAO;AAAA,IACP,YAAY,QAAQ;AAAA,IACpB;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI;AACF,UAAM,SAAS,MAAM;AAAA,MACnB;AAAA,MACA;AAAA,MACA;AAAA,QACE;AAAA,QACA;AAAA,QACA;AAAA,QACA,WAAW,QAAQ;AAAA,QACnB,YAAY,QAAQ;AAAA,QACpB;AAAA,MACF;AAAA,MACA,MACE,SAAS,SAAS;AAAA,QAChB;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACL;AAEA,QAAI,aAAa,MAAM,KAAK,CAAC,mBAAmB,OAAO,GAAG;AACxD,UAAI,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,KAAK,CAAC,OAAO,WAAW;AAClE,eAAO,YAAY,QAAQ;AAAA,MAC7B;AACA,aAAO;AAAA,IACT;AAEA,UAAM,kBAAkB;AACxB,UAAM,aAAa;AAAA,MACjB;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,QAAQ;AAAA,IACV;AAQA,UAAM,SAAS,QAAQ,WAAW,SAC9B,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AACjC,UAAM,QAAQ,QAAQ,WAAW,QAC7B,MAAM,QAAQ,UAAU,MAAM,UAAU,IACxC,aAAa,iBAAiB,QAAQ,UAAU,MAAM;AAC1D,UAAM,UAAU,QAAQ,WAAW,UAC/B,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C,eAAe,iBAAiB,UAAU,QAAQ,KAAK;AAE3D,WAAO;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,QAAQ,WAAW,UACxB,MAAM,QAAQ,UAAU,QAAQ,UAAU,IAC1C;AAAA,MACJ,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,QAAQ,WAAW,SACvB,MAAM,QAAQ,UAAU,OAAO,UAAU,IACzC,cAAc,eAAe;AAAA,IACnC;AAAA,EACF,SAAS,OAAO;AACd,UAAM,QAAQ,aAAa,QAAW,QAAQ,UAAU,MAAM;AAC9D,UAAM,MAAM;AAAA,MACV,SAAS,eAAe,QAAW,UAAU,QAAW,KAAK;AAAA,MAC7D,QAAQ;AAAA,MACR;AAAA,MACA,WACE,OAAO,KAAK,QAAQ,SAAS,EAAE,SAAS,IACpC,QAAQ,YACR;AAAA,MACN,QAAQ,CAAC,eAAe,KAAK,CAAC;AAAA,IAChC;AAEA,UAAM,wBAAwB,OAAO,GAAG;AAAA,EAC1C;AACF;AAEA,eAAe,aAMb,SAAyE;AACzE,MAAI,QAAQ,UAAU,QAAW;AAC/B,WAAO,QAAQ;AAAA,EACjB;AAEA,MAAI,QAAQ,aAAa;AACvB,WAAO,QAAQ,YAAY;AAAA,EAC7B;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAMP,SAA4E;AAC5E,SAAO;AAAA,IACL,QAAQ,WAAW,UACjB,QAAQ,WAAW,WACnB,QAAQ,WAAW,SACnB,QAAQ,WAAW,WACnB,QAAQ,WAAW;AAAA,EACvB;AACF;AAEA,SAAS,4BACP,OACyC;AACzC,MAAI;AACJ,QAAM,iBAA8D,CAAC;AACrE,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,mBAAmB,kBAAqC,SAAS;AACvE,QAAI,iBAAiB,UAAa,qBAAqB,QAAW;AAChE,qBAAe;AAAA,IACjB;AAEA,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,qBAAe,KAAK,WAAW;AAAA,IACjC;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;AAAA,IACvC,GAAI,eAAe,SAAS,IAAI,EAAE,eAAe,IAAI,CAAC;AAAA,EACxD;AACF;AAEA,SAAS,uBAAuB,OAA0B;AACxD,QAAM,QAAQ,cAAc,KAAK;AACjC,QAAM,aAAuB,CAAC;AAC9B,QAAM,OAAO,oBAAI,IAAY;AAE7B,aAAW,QAAQ,OAAO;AACxB,oBAAgB,YAAY,MAAM,IAAI;AACtC,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,OAAO,CAAC;AAClE,oBAAgB,YAAY,MAAM,kBAAkB,MAAM,cAAc,CAAC;AAAA,EAC3E;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA0B;AAC/C,SAAO,CAAC,SAAS,KAAK,CAAC,EACpB,OAAO,SAAS,kBAAkB,OAAO,OAAO,CAAC,CAAC,EAClD,OAAO,CAAC,UAA2B,UAAU,MAAS;AAC3D;AAEA,SAAS,gBACP,YACA,MACA,OACA;AACA,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,KAAK,IAAI,KAAK,GAAG;AAC1D;AAAA,EACF;AAEA,OAAK,IAAI,KAAK;AACd,aAAW,KAAK,KAAK;AACvB;AAEA,SAAS,SAAS,OAAoC;AACpD,SAAO,SAAS,OAAO,UAAU,WAAW,QAAQ;AACtD;AAEA,SAAS,kBAAkB,OAAgB,KAAsB;AAC/D,SAAO,SAAS,OAAO,UAAU,WAC5B,MAAkC,GAAG,IACtC;AACN;AAEA,SAAS,kBACP,OACoD;AACpD,QAAM,YACJ,kBAAkB,OAAO,OAAO,KAAK,kBAAkB,OAAO,SAAS;AAEzE,SAAO,cAAc,SAAS,IACzB,YACD;AACN;AAEA,SAAS,mBACP,OACkD;AAClD,QAAM,YAAY,kBAAkB,OAAO,OAAO;AAClD,MAAI,iBAAiB,SAAS,GAAG;AAC/B,WAAO;AAAA,EACT;AAEA,SAAO;AACT;AAEA,eAAe,SAOb,SACA,MAC+B;AAC/B,MAAI,QAAQ,KAAK;AACf,WAAO,QAAQ,IAAI,IAAI;AAAA,EACzB;AAEA,MAAI,iBAAqD,KAAK,KAAK,GAAG;AACpE,WAAO,KAAK,MAAM,IAAI,KAAK,OAAO,KAAK,OAAO;AAAA,EAChD;AAEA,QAAM,IAAI;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,mBAOP,SAOA;AACA,SAAO,QAAQ,UAAU;AAC3B;AAEA,SAAS,iBAMP,OACgE;AAChE,MAAI,CAAC,SAAS,OAAO,UAAU,UAAU;AACvC,WAAO;AAAA,EACT;AAEA,SACE,SAAS,SAAS,OAAQ,MAA4B,QAAQ;AAElE;AAEA,SAAS,cAAc,OAAsC;AAC3D,MAAI,CAAC,SAAS,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AAC/D,WAAO;AAAA,EACT;AAEA,QAAM,QAAQ,OAAO,OAAO,KAAK;AACjC,SACE,MAAM,SAAS,KACf,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,SAAS,iBAAiB,OAA4C;AACpE,SACE,MAAM,QAAQ,KAAK,KACnB,MAAM;AAAA,IAAM,CAAC,SACX;AAAA,MACE,QACE,OAAO,SAAS,YAChB,UAAU,QACV,OAAQ,KAA4B,SAAS,YAC7C,aAAa,QACb,OAAQ,KAA+B,YAAY;AAAA,IACvD;AAAA,EACF;AAEJ;AAEA,eAAe,2BAKb,OACA,UACA,MAQA,UACA;AACA,MAAI,CAAC,YAAY,SAAS,WAAW,GAAG;AACtC,WAAO,SAAS;AAAA,EAClB;AAEA,QAAM,qBAAqB,oBAAI,IAG7B;AACF,QAAM,iBAAiB,oBAAI,IAA+C;AAE1E,QAAM,YAAY,CAAC,SAA6C;AAC9D,QAAI,mBAAmB,IAAI,IAAI,GAAG;AAChC;AAAA,IACF;AAEA,UAAM,kBAAkB,2BAA2B,KAAK,OAAO;AAC/D,uBAAmB,IAAI,MAAM,eAAe;AAC5C,UAAM,sBAA4D,OAChE,YACA,YACG;AACH,YAAM,YAAY,oBAAI,KAAK;AAC3B,YAAM,cAAc;AAAA,QAClB,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK,QAAQ;AAAA,QACvB,QAAQ,KAAK,QAAQ;AAAA,QACrB,aAAa,KAAK,QAAQ;AAAA,MAC5B;AACA,YAAM,kBAAkB;AAAA,QACtB,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AAEA,UAAI;AACF,cAAM,YAAY,MAAM,4BAA4B;AAAA,UAClD,UAAU,KAAK;AAAA,UACf;AAAA,UACA,SAAS;AAAA,UACT,QAAQ,KAAK,aAAa,KAAK,IAAI;AAAA,UACnC,MAAM;AAAA,UACN,SAAS;AAAA,QACX,CAAC;AACD,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,QAAQ,UAAU;AAAA,UAClB,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,UAAU,wBAAwB,UAAU,MAAM;AAAA,QACpD;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,SAAS,UAAU;AAAA,UACnB,UAAU;AAAA,YACR,MAAM,KAAK;AAAA,UACb;AAAA,QACF,CAAC;AACD,eAAO,UAAU;AAAA,MACnB,SAAS,OAAO;AACd,cAAM,aAAa,oBAAI,KAAK;AAC5B,cAAM,OAAO;AAAA,UACX,MAAM,KAAK;AAAA,UACX,WAAW;AAAA,UACX,OAAO,uBAAuB,KAAK;AAAA,UACnC,WAAW,UAAU,YAAY;AAAA,UACjC,YAAY,WAAW,YAAY;AAAA,UACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,UACrD,UAAU,wBAAwB,2BAA2B,KAAK,CAAC;AAAA,QACrE;AACA,aAAK,UAAU,KAAK,IAAI;AACxB,aAAK,SAAS,KAAK;AAAA,UACjB,MAAM;AAAA,UACN,WAAW,CAAC,IAAI;AAAA,QAClB,CAAC;AACD,cAAM;AAAA,MACR,UAAE;AACA,wBAAgB;AAAA,MAClB;AAAA,IACF;AACA,wBAAoB,uBAAuB,IAAI;AAC/C,SAAK,UAAU;AAAA,EACjB;AAEA,QAAM,gBAAgB,CACpB,iBACG;AACH,eAAW,WAAW,cAAc;AAClC,iBAAW,QAAQ,SAAS;AAC1B,kBAAU,IAAI;AAAA,MAChB;AAAA,IACF;AAAA,EACF;AAEA,gBAAc,QAAQ;AAEtB,aAAW,UAAU,qBAAqB,KAAK,GAAG;AAChD,UAAM,gBAAgB,OAAO;AAC7B,mBAAe,IAAI,QAAQ,aAAa;AACxC,WAAO,QAAQ,SAAS,gBAAuC,WAAW;AACxE,YAAM,cAAc,cAAc,MAAM,MAAM,SAAS;AAEvD,UAAI,cAAc,WAAW,GAAG;AAC9B,eAAO,YAAY,QAAQ,MAAM;AAC/B;AAAA,YACE,8BAAiD,KAAK;AAAA,UACxD;AAAA,QACF,CAAC;AAAA,MACH;AAEA,oBAAc,8BAAiD,KAAK,CAAC;AACrE,aAAO;AAAA,IACT;AAAA,EACF;AAEA,MAAI;AACF,WAAO,MAAM,SAAS;AAAA,EACxB,UAAE;AACA,eAAW,CAAC,QAAQ,aAAa,KAAK,gBAAgB;AACpD,aAAO,QAAQ;AAAA,IACjB;AAEA,eAAW,CAAC,MAAM,eAAe,KAAK,oBAAoB;AACxD,WAAK,UAAU;AAAA,IACjB;AAAA,EACF;AACF;AAMA,SAAS,qBAAqB,OAAmC;AAC/D,SAAO,cAAc,KAAK,EAAE,OAAO,iBAAiB;AACtD;AAEA,SAAS,kBAAkB,OAAyC;AAClE,SAAO,WAAW,SAAS,OAAO,MAAM,UAAU;AACpD;AAEA,SAAS,8BAGP,OAA6D;AAC7D,QAAM,WAAwD,CAAC;AAC/D,QAAM,eAAe,oBAAI,IAA0C;AAEnE,aAAW,aAAa,uBAAuB,KAAK,GAAG;AACrD,UAAM,cAAc,mBAAmB,SAAS;AAChD,QAAI,eAAe,CAAC,aAAa,IAAI,WAAW,GAAG;AACjD,mBAAa,IAAI,WAAW;AAC5B,eAAS,KAAK,WAAW;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,cAAc,OAA2C;AAChE,SAAO;AAAA,IACL,SAAS,OAAQ,MAA6B,SAAS;AAAA,EACzD;AACF;AAEA,SAAS,uBACP,OACsC;AACtC,QAAM,aAAa,eAAe,KAAK;AACvC,QAAM,EAAE,SAAS,MAAM,GAAG,QAAQ,IAAI;AAEtC,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS,OAAO,YAAY,WAAW,UAAU,OAAO,OAAO;AAAA,IAC/D,GAAI,OAAO,SAAS,WAAW,EAAE,KAAK,IAAI,CAAC;AAAA,EAC7C;AACF;AAEA,SAAS,2BACP,SACA;AACA,QAAM,gBAAgB;AACtB,SAAO,cAAc,uBAAuB,KAAK;AACnD;AAEA,SAAS,6BAAmD;AAC1D,SAAO;AAAA,IACL,uBAAuB,oBAAI,IAAI;AAAA,EACjC;AACF;AAEA,SAAS,yBACP,OACA,UACA;AACA,QAAM,sBAAsB;AAAA,IAC1B;AAAA,KACC,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AAAA,EACrD;AAEA,SAAO,MAAM;AACX,UAAM,aAAa,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AACrE,QAAI,aAAa,GAAG;AAClB,YAAM,sBAAsB,OAAO,QAAQ;AAC3C;AAAA,IACF;AAEA,UAAM,sBAAsB,IAAI,UAAU,SAAS;AAAA,EACrD;AACF;AAEA,SAAS,6BACP,OACA,UACA;AACA,UAAQ,MAAM,sBAAsB,IAAI,QAAQ,KAAK,KAAK;AAC5D;AAEA,eAAe,4BAGb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAOG;AACD,MAAI,aAAa;AACjB,MAAI;AAEJ,QAAM,YAAY,MAAM,kBAAkB;AAAA,IACxC,UAAU,2BAA2B,QAAQ;AAAA,IAC7C;AAAA,IACA;AAAA,IACA,SAAS,OAAO,aAAa;AAC3B,mBAAa;AACb,mBAAa,MAAM,QAAQ,YAAY,QAAQ;AAC/C,aAAO,+BAA+B,UAAU;AAAA,IAClD;AAAA,IACA;AAAA,EACF,CAAC;AAED,MAAI,YAAY;AACd,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,kBAAkB,0BAA0B,UAAU;AAAA,MACtD,QAAQ,UAAU;AAAA,IACpB;AAAA,EACF;AAEA,SAAO;AAAA,IACL,GAAG,8BAA8B,UAAU,MAAM;AAAA,IACjD,QAAQ,UAAU;AAAA,EACpB;AACF;AAEA,SAAS,2BAA2B,UAAkB;AACpD,SAAO,GAAG,QAAQ;AACpB;AAEA,SAAS,cAIP;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GASE;AACA,QAAM,YAA8B,CAAC;AACrC,QAAM,YAA2B;AAAA,IAC/B,SAAS,CAAC,YAAY;AACpB,eAAS,KAAK,OAAO;AAAA,IACvB;AAAA,IACA,QAAQ,CAAC,SAAS,aAAa;AAC7B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,SAAS,aAAa;AAC3B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,WAAW,CAAC,SAAS,aAAa;AAChC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA;AAAA,MACF,CAAC;AAAA,IACH;AAAA,IACA,MAAM,CAAC,MAAM,SAAS,aAAa;AACjC,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN;AAAA,QACA,UAAU;AAAA,UACR;AAAA,UACA,GAAI,YAAY,CAAC;AAAA,QACnB;AAAA,MACF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,eAAe,OAAO;AAAA,IAC1B,OAAO,QAAQ,SAAS,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC,UAAU,IAAI,MAAM;AAAA,MACpD;AAAA,MACA,OAAO,SAAoC;AACzC,cAAM,YAAY,oBAAI,KAAK;AAC3B,cAAM,6BAA6B;AAAA,UACjC;AAAA,UACA;AAAA,QACF;AACA,cAAM,cAAc;AAAA,UAClB;AAAA,UACA,UAAU,QAAQ;AAAA,UAClB,QAAQ,QAAQ;AAAA,UAChB,aAAa,QAAQ;AAAA,QACvB;AAEA,YAAI;AACF,gBAAM,YAAY,MAAM,sBAAsB;AAAA,YAC5C;AAAA,YACA;AAAA,YACA,QAAQ,6BACJ,SACA,aAAa,QAAQ;AAAA,YACzB;AAAA,YACA,SAAS;AAAA,UACX,CAAC;AACD,gBAAM,aAAa,oBAAI,KAAK;AAE5B,cAAI,4BAA4B;AAC9B,mBAAO,UAAU;AAAA,UACnB;AAEA,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,QAAQ,UAAU;AAAA,YAClB,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,UAAU,wBAAwB,UAAU,MAAM;AAAA,UACpD;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,SAAS,UAAU;AAAA,YACnB,UAAU;AAAA,cACR,MAAM;AAAA,YACR;AAAA,UACF,CAAC;AACD,iBAAO,UAAU;AAAA,QACnB,SAAS,OAAO;AACd,gBAAM,aAAa,oBAAI,KAAK;AAC5B,cAAI,4BAA4B;AAC9B,kBAAM;AAAA,UACR;AAEA,gBAAM,OAAO;AAAA,YACX,MAAM;AAAA,YACN,WAAW;AAAA,YACX,OAAO,uBAAuB,KAAK;AAAA,YACnC,WAAW,UAAU,YAAY;AAAA,YACjC,YAAY,WAAW,YAAY;AAAA,YACnC,YAAY,WAAW,QAAQ,IAAI,UAAU,QAAQ;AAAA,YACrD,UAAU;AAAA,cACR,2BAA2B,KAAK;AAAA,YAClC;AAAA,UACF;AACA,oBAAU,KAAK,IAAI;AACnB,mBAAS,KAAK;AAAA,YACZ,MAAM;AAAA,YACN,WAAW,CAAC,IAAI;AAAA,UAClB,CAAC;AACD,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,QAAQ,QAAQ;AAAA,IAChB;AAAA,EACF;AACF;AAEA,SAAS,cAAc,QAAwC;AAC7D,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,YAAY,MAAM;AAAA,EAC3B;AAEA,QAAM,aAAa;AAAA,IACjB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,OAAO,YAAY;AAC5B,UAAM,QAAS,OAAmC,GAAG;AACrD,UAAM,aAAa,YAAY,KAAK;AACpC,QAAI,eAAe,QAAW;AAC5B,aAAO;AAAA,IACT;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,oBAAoB,QAAwC;AACnE,QAAM,UACJ,UAAU,OAAO,WAAW,WACxB,YAAa,OAAiC,OAAO,IACrD;AACN,MAAI,YAAY,QAAW;AACzB,WAAO;AAAA,EACT;AAEA,SACE,YAAY,MAAM,MAAM,WAAW,SAAY,SAAY,OAAO,MAAM;AAE5E;AAEA,SAAS,0BAA0B,QAA4B;AAC7D,SAAO,oBAAoB,MAAM,KAAK;AACxC;AAoBA,SAAS,+BACP,QAC0B;AAC1B,QAAM,mBAAmB,0BAA0B,MAAM;AAEzD,SAAO;AAAA,IACL,eAAe;AAAA,MACb,MAAM;AAAA,MACN,SAAS;AAAA,IACX;AAAA,IACA,aAAa,YAAY,MAAM,KAAK;AAAA,IACpC;AAAA,EACF;AACF;AAEA,SAAS,8BAA8B,QAAmB;AACxD,MAAI,2BAA2B,MAAM,GAAG;AACtC,WAAO;AAAA,MACL,QAAQ,OAAO;AAAA,MACf,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,MAAI,iCAAiC,MAAM,GAAG;AAC5C,WAAO;AAAA,MACL,QAAQ,OAAO,eAAe,OAAO;AAAA,MACrC,kBAAkB,OAAO;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,kBAAkB,0BAA0B,MAAM;AAAA,EACpD;AACF;AAEA,SAAS,2BACP,OACmC;AACnC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,iBAAiB,SACjB,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,yBACP,OACoD;AACpD,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,iCACP,OACyC;AACzC,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,mBAAmB,SACnB;AAAA,MACG,MAAsC;AAAA,IACzC,KACA,sBAAsB;AAAA,EAC1B;AACF;AAEA,SAAS,+BACP,OAC0D;AAC1D,SAAO;AAAA,IACL,SACE,OAAO,UAAU,YACjB,UAAU,SACT,MAA6B,SAAS,8BACvC,aAAa,SACZ,MAAgC,YAAY;AAAA,EACjD;AACF;AAEA,SAAS,aAAa,QAAiB,eAAqC;AAC1E,MAAI,CAAC,UAAU,OAAO,WAAW,UAAU;AACzC,WAAO,gBAAgB,IAAI,EAAE,WAAW,cAAc,IAAI,CAAC;AAAA,EAC7D;AAEA,QAAM,aACH,OAAmC,SACnC,OAAmC;AAEtC,QAAM,QACJ,cAAc,OAAO,eAAe,WAC/B,EAAE,GAAI,WAAuC,IAC9C,CAAC;AAEP,MAAI,MAAM,cAAc,UAAa,gBAAgB,GAAG;AACtD,UAAM,YAAY;AAAA,EACpB;AAEA,SAAO;AACT;AAEA,SAAS,eACP,QACA,UACA,QACA,OACmB;AACnB,MACE;AAAA,IACG,QAAgD;AAAA,EACnD,GACA;AACA,WAAQ,OAA0C;AAAA,EACpD;AAEA,MACE,oBAAqB,QAAgD,KAAK,GAC1E;AACA,WAAQ,OAAwC;AAAA,EAClD;AAEA,QAAM,kBAAkB,CAAC,GAAG,QAAQ;AACpC,MACE,WAAW,UACX,CAAC,gBAAgB;AAAA,IACf,CAAC,YACC,QAAQ,SAAS,eAAe,QAAQ,YAAY;AAAA,EACxD,GACA;AACA,oBAAgB,KAAK;AAAA,MACnB,MAAM;AAAA,MACN,SAAS;AAAA,IACX,CAAC;AAAA,EACH;AAEA,SAAO;AAAA,IACL,UAAU;AAAA,IACV,YAAY,OAAO,WAAW,WAAW,SAAS;AAAA,IAClD,UACI,QAAgD,YAEhC,MAAM;AAAA,IAC1B,OACI,QAAgD,SAEhC,MAAM;AAAA,EAC5B;AACF;AAEA,SAAS,cAAc,QAAmD;AACxE,SAAO,wBAAwB,MAAM;AACvC;AAEA,eAAe,sBAKb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF,GAMG;AACD,SAAO,kBAIL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,SAAS,CAAC,UAAU,gBAClB,KAAK,QAAQ,UAAmB,WAAW;AAAA,IAC7C;AAAA,EACF,CAAC;AACH;","names":[]}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@vitest-evals/harness-pi-ai",
|
|
3
|
-
"version": "0.9.0-beta.
|
|
3
|
+
"version": "0.9.0-beta.2",
|
|
4
4
|
"sideEffects": false,
|
|
5
5
|
"types": "./dist/index.d.ts",
|
|
6
6
|
"main": "./dist/index.js",
|
|
@@ -27,7 +27,7 @@
|
|
|
27
27
|
"devDependencies": {
|
|
28
28
|
"@mariozechner/pi-agent-core": "^0.67.68",
|
|
29
29
|
"@mariozechner/pi-ai": "^0.67.68",
|
|
30
|
-
"vitest-evals": "0.9.0-beta.
|
|
30
|
+
"vitest-evals": "0.9.0-beta.2"
|
|
31
31
|
},
|
|
32
32
|
"scripts": {
|
|
33
33
|
"build": "tsup --config ./tsup.config.ts"
|