@statelyai/agent 2.0.0-alpha.32 → 2.0.0-alpha.33
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +12 -0
- package/dist/testing.cjs +493 -3
- package/dist/testing.d.cts +386 -2
- package/dist/testing.d.mts +386 -2
- package/dist/testing.mjs +489 -4
- package/dist/validate.cjs +1 -1
- package/dist/validate.mjs +1 -1
- package/dist/{verify-D_llRCTr.cjs → verify-CS_WLRXn.cjs} +17 -1
- package/dist/{verify-CoCvR1eM.mjs → verify-DgB8P-Wx.mjs} +12 -2
- package/package.json +1 -1
- package/skills/add-evals/SKILL.md +8 -2
- package/skills/improve-machine/SKILL.md +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,17 @@
|
|
|
1
1
|
# @statelyai/agent
|
|
2
2
|
|
|
3
|
+
## 2.0.0-alpha.33
|
|
4
|
+
|
|
5
|
+
### Minor Changes
|
|
6
|
+
|
|
7
|
+
- [#181](https://github.com/statelyai/agent/pull/181) [`407e2a6`](https://github.com/statelyai/agent/commit/407e2a6c03e463d4b8d92997cf59e72b30301511) Thanks [@davidkpiano](https://github.com/davidkpiano)! - **Four eval helpers in `@statelyai/agent/testing`.**
|
|
8
|
+
|
|
9
|
+
- `traceCoverage(machine, traces, { reachable? })` reports which states and handlers no trace exercised. Coverage is per handler: a state and an event it handles, with the targets it declares and the ones runs reached. `always` handlers are listed as unobservable, because traces record one step per macrostep.
|
|
10
|
+
- `repeatRuns({ scenarios, k, run, passed })` runs each scenario `k` times and reports passes, pass@k, pass^k, distinct paths, and the state where the runs diverged. `report.passHat(j)` estimates the suite's pass^j. `summarizeRepeats` builds the same report from runs collected elsewhere.
|
|
11
|
+
- `judgeAgreement(rows, { positive?, groupBy? })` compares a judge's labels with confirmed human labels: accuracy, per-label recall, the confusion matrix, disagreements, the true positive and true negative rates, and a correction for the judge's bias.
|
|
12
|
+
- `firstDivergence(actual, reference)` finds the first step where a run departs from a reference trace or an expected trajectory, and names the state it was in.
|
|
13
|
+
- `foldTraces` takes a `blame` option and adds a `blamed` column: failed runs counted once, against the state their failure started in. Rows now always include `blamed`, which is 0 without `blame`.
|
|
14
|
+
|
|
3
15
|
## 2.0.0-alpha.32
|
|
4
16
|
|
|
5
17
|
### Patch Changes
|
package/dist/testing.cjs
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
|
|
2
2
|
const require_event_log = require("./event-log-CKWeC7H-.cjs");
|
|
3
3
|
const require_run_agent = require("./run-agent-Cg0xd3K-.cjs");
|
|
4
|
-
const require_verify = require("./verify-
|
|
4
|
+
const require_verify = require("./verify-CS_WLRXn.cjs");
|
|
5
5
|
//#region src/trajectory.ts
|
|
6
6
|
/**
|
|
7
7
|
* Classifies one item. Event-shaped things (log entries, events) become events;
|
|
@@ -381,12 +381,47 @@ function statePaths(value, prefix = "") {
|
|
|
381
381
|
if (entries.length === 0) return prefix ? [prefix.slice(0, -1)] : [];
|
|
382
382
|
return entries.flatMap(([key, child]) => statePaths(child, `${prefix}${key}.`));
|
|
383
383
|
}
|
|
384
|
+
/**
|
|
385
|
+
* A trace's `machine.transition` events as steps, skipping `@xstate.init`.
|
|
386
|
+
* A first-attempt decision's `request.start` names the decision on the step
|
|
387
|
+
* that follows it. Reads live and serialized traces alike.
|
|
388
|
+
* @internal
|
|
389
|
+
*/
|
|
390
|
+
function traceSteps(trace) {
|
|
391
|
+
const steps = [];
|
|
392
|
+
let decision;
|
|
393
|
+
for (const event of trace) {
|
|
394
|
+
if (event.type === "request.start" && isFirstDecision(event.request)) {
|
|
395
|
+
const name = event.request.name;
|
|
396
|
+
decision = typeof name === "string" ? name : void 0;
|
|
397
|
+
continue;
|
|
398
|
+
}
|
|
399
|
+
if (event.type !== "machine.transition") continue;
|
|
400
|
+
const type = require_event_log.isRecord(event.event) ? event.event.type : void 0;
|
|
401
|
+
if (type === "@xstate.init") continue;
|
|
402
|
+
const fromValue = valueOf(event.from);
|
|
403
|
+
const toValue = valueOf(event.snapshot);
|
|
404
|
+
steps.push({
|
|
405
|
+
from: statePaths(fromValue),
|
|
406
|
+
fromValue,
|
|
407
|
+
event: typeof type === "string" ? type : "",
|
|
408
|
+
payload: event.event,
|
|
409
|
+
to: statePaths(toValue),
|
|
410
|
+
toValue,
|
|
411
|
+
ignored: event.ignored === true,
|
|
412
|
+
...decision === void 0 ? {} : { decision }
|
|
413
|
+
});
|
|
414
|
+
decision = void 0;
|
|
415
|
+
}
|
|
416
|
+
return steps;
|
|
417
|
+
}
|
|
384
418
|
//#endregion
|
|
385
419
|
//#region src/fold-traces.ts
|
|
386
420
|
/**
|
|
387
421
|
* Folds many runs' traces into per-state stats: how many runs visited each
|
|
388
422
|
* state, how many of those failed, how many options the guards withheld
|
|
389
|
-
* at decisions made there,
|
|
423
|
+
* at decisions made there, how many events it ignored, and how many failures
|
|
424
|
+
* `blame` traced back to it. Pure; reads only:
|
|
390
425
|
*
|
|
391
426
|
* - `machine.transition`: the run visited `from.value` and `snapshot.value`;
|
|
392
427
|
* with `ignored`, each of those states ignored one event.
|
|
@@ -399,6 +434,10 @@ function statePaths(value, prefix = "") {
|
|
|
399
434
|
* `{ type: 'request.start', request: result.request }` onto the trace before
|
|
400
435
|
* sending its event. Parallel states count once per active leaf.
|
|
401
436
|
*
|
|
437
|
+
* `failures` counts a failed run against every state it visited, so states
|
|
438
|
+
* after a bad decision rank as high as the decision. Sort by `blamed` to see
|
|
439
|
+
* where failures start.
|
|
440
|
+
*
|
|
402
441
|
* @example
|
|
403
442
|
* ```ts no-check
|
|
404
443
|
* const stats = foldTraces(traces, {
|
|
@@ -416,7 +455,8 @@ function foldTraces(traces, options) {
|
|
|
416
455
|
visits: 0,
|
|
417
456
|
failures: 0,
|
|
418
457
|
withheld: 0,
|
|
419
|
-
ignored: 0
|
|
458
|
+
ignored: 0,
|
|
459
|
+
blamed: 0
|
|
420
460
|
};
|
|
421
461
|
stats.set(state, entry);
|
|
422
462
|
}
|
|
@@ -445,10 +485,249 @@ function foldTraces(traces, options) {
|
|
|
445
485
|
entry.visits++;
|
|
446
486
|
if (failed) entry.failures++;
|
|
447
487
|
}
|
|
488
|
+
const blamed = failed ? options.blame?.(trace) : void 0;
|
|
489
|
+
if (blamed !== void 0) statsFor(blamed).blamed++;
|
|
448
490
|
}
|
|
449
491
|
return [...stats.values()];
|
|
450
492
|
}
|
|
451
493
|
//#endregion
|
|
494
|
+
//#region src/trace-coverage.ts
|
|
495
|
+
/** Mirrors xstate@6.0.0-alpha.69 `matchesEventDescriptor` (src/utils.ts), which the package does not export. */
|
|
496
|
+
function matchesDescriptor(eventType, descriptor) {
|
|
497
|
+
if (descriptor === eventType || descriptor === "*") return true;
|
|
498
|
+
if (!descriptor.endsWith(".*")) return false;
|
|
499
|
+
const tokens = descriptor.split(".");
|
|
500
|
+
const eventTokens = eventType.split(".");
|
|
501
|
+
for (let index = 0; index < tokens.length; index++) {
|
|
502
|
+
if (tokens[index] === "*") return index === tokens.length - 1;
|
|
503
|
+
if (tokens[index] !== eventTokens[index]) return false;
|
|
504
|
+
}
|
|
505
|
+
return true;
|
|
506
|
+
}
|
|
507
|
+
/** Whether a definition's `matches` pattern (e.g. an invoke's `actorId`) accepts the traced event. */
|
|
508
|
+
function matchesPattern(definition, event) {
|
|
509
|
+
const pattern = definition.matches;
|
|
510
|
+
if (!require_event_log.isRecord(pattern) || !require_event_log.isRecord(event)) return true;
|
|
511
|
+
return Object.entries(pattern).every(([key, value]) => Object.is(event[key], value));
|
|
512
|
+
}
|
|
513
|
+
/**
|
|
514
|
+
* The descriptor XState would try first at `node` for `event`, guards aside:
|
|
515
|
+
* the exact key, then partial and wildcard keys longest first, as in
|
|
516
|
+
* xstate@6.0.0-alpha.69 `getCandidates` (src/stateUtils.ts).
|
|
517
|
+
*/
|
|
518
|
+
function candidateAt(node, type, event) {
|
|
519
|
+
const keys = [...node.transitions.keys()].filter((key) => key !== type && matchesDescriptor(type, key)).sort((a, b) => b.length - a.length);
|
|
520
|
+
if (node.transitions.has(type)) keys.unshift(type);
|
|
521
|
+
return keys.find((key) => node.transitions.get(key).some((d) => matchesPattern(d, event)));
|
|
522
|
+
}
|
|
523
|
+
const pathOf = (node) => node.path.join(".");
|
|
524
|
+
const ratioOf = (visited, total) => total === 0 ? 1 : visited / total;
|
|
525
|
+
/**
|
|
526
|
+
* Measures how much of a machine a set of runs exercised: which atomic and
|
|
527
|
+
* final states they visited, and which handlers they took. A handler is a
|
|
528
|
+
* `(declaring state, event descriptor)` pair from `on`, invoke
|
|
529
|
+
* `onDone`/`onError`, and `after`, with the targets it declares and the
|
|
530
|
+
* leaves the runs landed in after it. Pure; reads only `machine.transition`
|
|
531
|
+
* events, live (`AgentTraceEvent`) or serialized (`serializeTraceEvent`).
|
|
532
|
+
*
|
|
533
|
+
* A step credits one handler per active leaf: the nearest node, from the leaf
|
|
534
|
+
* up, whose exact, partial (`a.*`), or wildcard (`*`) descriptor matches the
|
|
535
|
+
* event, preferred in XState's order. Guards are not replayed, so a step
|
|
536
|
+
* whose nearest handler was rejected and fell through to an ancestor is
|
|
537
|
+
* credited to the nearest one. XState keys invoke and `after` handlers by
|
|
538
|
+
* bare type (`xstate.done.actor`), one per state however many invokes; a
|
|
539
|
+
* definition's `matches` (an invoke's `actorId`) must accept the traced event
|
|
540
|
+
* for its state to be credited. Ignored steps credit no handler. `always`
|
|
541
|
+
* handlers are listed as `unobservable`, never as unvisited.
|
|
542
|
+
*
|
|
543
|
+
* @example
|
|
544
|
+
* ```ts no-check
|
|
545
|
+
* const report = traceCoverage(machine, traces, { reachable: await explorePaths(machine) });
|
|
546
|
+
* expect(report.unvisitedHandlers).toEqual([]);
|
|
547
|
+
* ```
|
|
548
|
+
*/
|
|
549
|
+
function traceCoverage(machine, traces, options = {}) {
|
|
550
|
+
const nodes = /* @__PURE__ */ new Map();
|
|
551
|
+
const states = [];
|
|
552
|
+
const handlers = [];
|
|
553
|
+
const handlerAt = /* @__PURE__ */ new Map();
|
|
554
|
+
const unobservable = [];
|
|
555
|
+
const walk = (node) => {
|
|
556
|
+
const state = pathOf(node);
|
|
557
|
+
nodes.set(state, node);
|
|
558
|
+
if (node.type === "atomic" || node.type === "final") states.push({
|
|
559
|
+
state,
|
|
560
|
+
runs: 0
|
|
561
|
+
});
|
|
562
|
+
const byEvent = /* @__PURE__ */ new Map();
|
|
563
|
+
for (const [event, definitions] of node.transitions) {
|
|
564
|
+
const dynamic = definitions.some((d) => typeof d.to === "function");
|
|
565
|
+
const declared = [...new Set(definitions.flatMap((d) => (d.target ?? []).map(pathOf)))];
|
|
566
|
+
const handler = {
|
|
567
|
+
state,
|
|
568
|
+
event,
|
|
569
|
+
runs: 0,
|
|
570
|
+
targets: {
|
|
571
|
+
declared: dynamic ? "dynamic" : declared,
|
|
572
|
+
seen: [],
|
|
573
|
+
unreached: []
|
|
574
|
+
}
|
|
575
|
+
};
|
|
576
|
+
handlers.push(handler);
|
|
577
|
+
byEvent.set(event, handler);
|
|
578
|
+
}
|
|
579
|
+
handlerAt.set(node, byEvent);
|
|
580
|
+
if (node.always?.length) unobservable.push({
|
|
581
|
+
state,
|
|
582
|
+
event: "always"
|
|
583
|
+
});
|
|
584
|
+
for (const child of Object.values(node.states)) walk(child);
|
|
585
|
+
};
|
|
586
|
+
walk(machine.root);
|
|
587
|
+
const stateRuns = new Map(states.map((row) => [row.state, row]));
|
|
588
|
+
const seen = /* @__PURE__ */ new Map();
|
|
589
|
+
for (const trace of traces) {
|
|
590
|
+
const steps = traceSteps(trace);
|
|
591
|
+
const visited = new Set(steps[0]?.from);
|
|
592
|
+
for (const event of trace) {
|
|
593
|
+
if (event.type !== "machine.transition" || !require_event_log.isRecord(event.event)) continue;
|
|
594
|
+
if (event.event.type !== "@xstate.init" || !require_event_log.isRecord(event.snapshot)) continue;
|
|
595
|
+
for (const leaf of statePaths(event.snapshot.value)) visited.add(leaf);
|
|
596
|
+
}
|
|
597
|
+
const credited = /* @__PURE__ */ new Set();
|
|
598
|
+
for (const step of steps) {
|
|
599
|
+
for (const leaf of step.to) visited.add(leaf);
|
|
600
|
+
if (step.ignored) continue;
|
|
601
|
+
const hits = /* @__PURE__ */ new Set();
|
|
602
|
+
for (const leaf of step.from) for (let node = nodes.get(leaf); node; node = node.parent) {
|
|
603
|
+
if (node.type === "final") continue;
|
|
604
|
+
const event = candidateAt(node, step.event, step.payload);
|
|
605
|
+
if (event === void 0) continue;
|
|
606
|
+
hits.add(handlerAt.get(node).get(event));
|
|
607
|
+
break;
|
|
608
|
+
}
|
|
609
|
+
for (const handler of hits) {
|
|
610
|
+
credited.add(handler);
|
|
611
|
+
let leaves = seen.get(handler);
|
|
612
|
+
if (!leaves) seen.set(handler, leaves = /* @__PURE__ */ new Set());
|
|
613
|
+
for (const leaf of step.to) leaves.add(leaf);
|
|
614
|
+
}
|
|
615
|
+
}
|
|
616
|
+
for (const state of visited) {
|
|
617
|
+
const row = stateRuns.get(state);
|
|
618
|
+
if (row) row.runs++;
|
|
619
|
+
}
|
|
620
|
+
for (const handler of credited) handler.runs++;
|
|
621
|
+
}
|
|
622
|
+
for (const handler of handlers) {
|
|
623
|
+
const leaves = [...seen.get(handler) ?? []];
|
|
624
|
+
handler.targets.seen = leaves;
|
|
625
|
+
const { declared } = handler.targets;
|
|
626
|
+
if (declared === "dynamic") continue;
|
|
627
|
+
handler.targets.unreached = declared.filter((target) => !leaves.some((leaf) => leaf === target || leaf.startsWith(`${target}.`)));
|
|
628
|
+
}
|
|
629
|
+
const reachableComplete = options.reachable ? require_verify.isExplorationComplete(options.reachable) : void 0;
|
|
630
|
+
let unreachable;
|
|
631
|
+
if (options.reachable && reachableComplete) {
|
|
632
|
+
const reached = new Set(options.reachable.reachedStates.flatMap((value) => statePaths(value)));
|
|
633
|
+
unreachable = states.filter((row) => !reached.has(row.state) && row.runs === 0).map((row) => row.state);
|
|
634
|
+
}
|
|
635
|
+
const counted = unreachable ? states.filter((row) => !unreachable.includes(row.state)) : states;
|
|
636
|
+
const unvisitedStates = counted.filter((row) => row.runs === 0).map((row) => row.state);
|
|
637
|
+
const unvisitedHandlers = handlers.filter((handler) => handler.runs === 0).map(({ state, event }) => ({
|
|
638
|
+
state,
|
|
639
|
+
event
|
|
640
|
+
}));
|
|
641
|
+
return {
|
|
642
|
+
states,
|
|
643
|
+
handlers,
|
|
644
|
+
unvisitedStates,
|
|
645
|
+
unvisitedHandlers,
|
|
646
|
+
unobservable,
|
|
647
|
+
...unreachable ? { unreachable } : {},
|
|
648
|
+
...reachableComplete === void 0 ? {} : { reachableComplete },
|
|
649
|
+
ratio: {
|
|
650
|
+
states: ratioOf(counted.length - unvisitedStates.length, counted.length),
|
|
651
|
+
handlers: ratioOf(handlers.length - unvisitedHandlers.length, handlers.length)
|
|
652
|
+
}
|
|
653
|
+
};
|
|
654
|
+
}
|
|
655
|
+
//#endregion
|
|
656
|
+
//#region src/first-divergence.ts
|
|
657
|
+
const defaultKey = (step) => `${step.from.join(",")}|${step.event}|${step.to.join(",")}`;
|
|
658
|
+
const isTrace = (items) => items.length === 0 || items.some((item) => require_event_log.isRecord(item) && (item.type === "machine.transition" || item.type === "run.start" || item.type === "run.end"));
|
|
659
|
+
/**
|
|
660
|
+
* Finds the first step where a run departs from a reference, so a failure is
|
|
661
|
+
* blamed on the state it started in rather than on every state the run went
|
|
662
|
+
* through afterwards.
|
|
663
|
+
*
|
|
664
|
+
* - Against a reference **trace**, such as a passing run of the same
|
|
665
|
+
* scenario, steps are compared in order by `key`. The first step that
|
|
666
|
+
* differs, or the point where either trace ends early, is the divergence.
|
|
667
|
+
* - Against an expected **trajectory** (`TrajectoryItem[]`, as for
|
|
668
|
+
* {@link matchesTrajectory}), the run is laid out as its initial state, then
|
|
669
|
+
* each step's event and new state. The first expected item it never reaches
|
|
670
|
+
* is the divergence, blamed on the step after the last item it did reach. A
|
|
671
|
+
* run that goes on past the trajectory has not diverged.
|
|
672
|
+
*
|
|
673
|
+
* An empty reference is a trace with no steps, so any step diverges from it.
|
|
674
|
+
* A trace is one run's events, live or serialized. Steps the machine ignored
|
|
675
|
+
* changed nothing and are left out on both sides. Returns `undefined` when the
|
|
676
|
+
* run does not diverge.
|
|
677
|
+
*
|
|
678
|
+
* @example
|
|
679
|
+
* ```ts no-check
|
|
680
|
+
* const stats = foldTraces(traces, {
|
|
681
|
+
* failed,
|
|
682
|
+
* blame: (trace) => firstDivergence(trace, passingRunFor(trace))?.state,
|
|
683
|
+
* });
|
|
684
|
+
* ```
|
|
685
|
+
*/
|
|
686
|
+
function firstDivergence(actual, reference, options = {}) {
|
|
687
|
+
const steps = traceSteps(actual).filter((step) => !step.ignored);
|
|
688
|
+
return isTrace(reference) ? againstTrace(steps, traceSteps(reference).filter((step) => !step.ignored), options.key ?? defaultKey) : againstTrajectory(actual, steps, reference);
|
|
689
|
+
}
|
|
690
|
+
function diverged(steps, index, expected, initial = steps[0]?.fromValue) {
|
|
691
|
+
const step = steps[index];
|
|
692
|
+
return {
|
|
693
|
+
index,
|
|
694
|
+
state: (step ? step.from[0] : steps.at(-1)?.to[0] ?? statePaths(initial)[0]) ?? "",
|
|
695
|
+
...step?.decision === void 0 ? {} : { decision: step.decision },
|
|
696
|
+
...step ? { actual: step } : {},
|
|
697
|
+
...expected === void 0 ? {} : { expected }
|
|
698
|
+
};
|
|
699
|
+
}
|
|
700
|
+
function againstTrace(steps, reference, key) {
|
|
701
|
+
const length = Math.max(steps.length, reference.length);
|
|
702
|
+
for (let index = 0; index < length; index++) {
|
|
703
|
+
const step = steps[index];
|
|
704
|
+
const want = reference[index];
|
|
705
|
+
if (step && want && key(step) === key(want)) continue;
|
|
706
|
+
return diverged(steps, index, want, steps[0]?.fromValue ?? reference[0]?.fromValue);
|
|
707
|
+
}
|
|
708
|
+
}
|
|
709
|
+
function againstTrajectory(trace, steps, expected) {
|
|
710
|
+
const initial = initialValue(trace, steps);
|
|
711
|
+
const timeline = initial === void 0 ? [] : [{ value: initial }];
|
|
712
|
+
const stepAt = initial === void 0 ? [] : [0];
|
|
713
|
+
steps.forEach((step, index) => {
|
|
714
|
+
timeline.push(step.payload);
|
|
715
|
+
stepAt.push(index);
|
|
716
|
+
if (JSON.stringify(step.toValue) !== JSON.stringify(step.fromValue)) {
|
|
717
|
+
timeline.push({ value: step.toValue });
|
|
718
|
+
stepAt.push(index);
|
|
719
|
+
}
|
|
720
|
+
});
|
|
721
|
+
const match = matchesTrajectory(timeline, expected);
|
|
722
|
+
if (!match.firstMiss) return void 0;
|
|
723
|
+
const at = match.firstMiss.searchedFrom;
|
|
724
|
+
return diverged(steps, at < timeline.length ? stepAt[at] ?? steps.length : steps.length, match.firstMiss.expected, initial);
|
|
725
|
+
}
|
|
726
|
+
function initialValue(trace, steps) {
|
|
727
|
+
for (const event of trace) if (event.type === "machine.transition" && require_event_log.isRecord(event.snapshot)) return event.snapshot.value;
|
|
728
|
+
return steps[0]?.fromValue;
|
|
729
|
+
}
|
|
730
|
+
//#endregion
|
|
452
731
|
//#region src/decision-requests.ts
|
|
453
732
|
/**
|
|
454
733
|
* The decisions a run asked for, read from its trace: each decision's first
|
|
@@ -481,6 +760,212 @@ function getDecisionRequests(trace, options = {}) {
|
|
|
481
760
|
return requests;
|
|
482
761
|
}
|
|
483
762
|
//#endregion
|
|
763
|
+
//#region src/repeat-runs.ts
|
|
764
|
+
/**
|
|
765
|
+
* Summarizes repeated runs of each scenario: how often each passed, whether
|
|
766
|
+
* all its runs passed (pass^k), how many distinct state paths they took, and
|
|
767
|
+
* where they first went different ways. Pure. Use it for runs collected
|
|
768
|
+
* elsewhere, such as CI shards or replays; {@link repeatRuns} collects them and
|
|
769
|
+
* calls it.
|
|
770
|
+
*
|
|
771
|
+
* A scenario that passes 3 of 5 runs and `divergedAt: 'deciding'` points at one
|
|
772
|
+
* decision, not at a flaky agent.
|
|
773
|
+
*
|
|
774
|
+
* @example
|
|
775
|
+
* ```ts no-check
|
|
776
|
+
* const report = summarizeRepeats(shards.map((s) => ({ id: s.id, scenario: s, runs: s.runs })));
|
|
777
|
+
* report.passHat(3); // the chance 3 runs of a scenario all pass
|
|
778
|
+
* ```
|
|
779
|
+
*/
|
|
780
|
+
function summarizeRepeats(groups) {
|
|
781
|
+
const scenarios = groups.map((group) => summarize$1(group));
|
|
782
|
+
const traces = groups.flatMap((group) => group.runs.flatMap((run) => run.trace ? [run.trace] : []));
|
|
783
|
+
const most = Math.max(0, ...groups.map((group) => group.runs.length));
|
|
784
|
+
const passHat = (j) => {
|
|
785
|
+
if (!Number.isInteger(j) || j < 1 || j > most) throw new require_event_log.AgentError("pass-hat-out-of-range", `passHat: j must be a whole number from 1 to ${most}, got ${j}.`);
|
|
786
|
+
const eligible = scenarios.filter((scenario) => scenario.runs.length >= j);
|
|
787
|
+
return eligible.reduce((total, scenario) => total + allPass(scenario.passes, scenario.runs.length, j), 0) / eligible.length;
|
|
788
|
+
};
|
|
789
|
+
return {
|
|
790
|
+
scenarios,
|
|
791
|
+
passHat,
|
|
792
|
+
traces
|
|
793
|
+
};
|
|
794
|
+
}
|
|
795
|
+
/**
|
|
796
|
+
* Runs every scenario `k` times and summarizes the runs with
|
|
797
|
+
* {@link summarizeRepeats}. Runs are sequential, scenario by scenario. A run
|
|
798
|
+
* that throws fails, with its message as `error` and no trace.
|
|
799
|
+
*
|
|
800
|
+
* With a live model, one run per scenario says little: τ-bench found pass^k
|
|
801
|
+
* falls well below the single-run rate. Report pass^k next to it.
|
|
802
|
+
*
|
|
803
|
+
* @example
|
|
804
|
+
* ```ts no-check
|
|
805
|
+
* const report = await repeatRuns({
|
|
806
|
+
* scenarios,
|
|
807
|
+
* k: 5,
|
|
808
|
+
* run: async (scenario) => {
|
|
809
|
+
* const trace: AgentTraceEvent[] = [];
|
|
810
|
+
* const result = await runToQuiescence(
|
|
811
|
+
* createAgentRuntime(machine, { executors, onTrace: (event) => trace.push(event) }),
|
|
812
|
+
* { input: scenario.input },
|
|
813
|
+
* );
|
|
814
|
+
* return { trace, outcome: String(result.snapshot.value) };
|
|
815
|
+
* },
|
|
816
|
+
* passed: (scenario, { outcome }) => outcome === scenario.expected,
|
|
817
|
+
* });
|
|
818
|
+
* report.passHat(5);
|
|
819
|
+
* ```
|
|
820
|
+
*/
|
|
821
|
+
async function repeatRuns(options) {
|
|
822
|
+
if (!Number.isInteger(options.k) || options.k < 1) throw new require_event_log.AgentError("invalid-repeat-count", `repeatRuns: k must be a whole number of 1 or more, got ${options.k}.`);
|
|
823
|
+
const groups = [];
|
|
824
|
+
for (const scenario of options.scenarios) {
|
|
825
|
+
const runs = [];
|
|
826
|
+
for (let attempt = 0; attempt < options.k; attempt++) try {
|
|
827
|
+
const result = await options.run(scenario, attempt);
|
|
828
|
+
runs.push({
|
|
829
|
+
passed: options.passed(scenario, result),
|
|
830
|
+
trace: result.trace,
|
|
831
|
+
...result.outcome === void 0 ? {} : { outcome: result.outcome }
|
|
832
|
+
});
|
|
833
|
+
} catch (error) {
|
|
834
|
+
runs.push({
|
|
835
|
+
passed: false,
|
|
836
|
+
error: error instanceof Error ? error.message : String(error)
|
|
837
|
+
});
|
|
838
|
+
}
|
|
839
|
+
groups.push({
|
|
840
|
+
id: scenario.id,
|
|
841
|
+
scenario,
|
|
842
|
+
runs
|
|
843
|
+
});
|
|
844
|
+
}
|
|
845
|
+
return summarizeRepeats(groups);
|
|
846
|
+
}
|
|
847
|
+
function summarize$1(group) {
|
|
848
|
+
const passes = group.runs.filter((run) => run.passed).length;
|
|
849
|
+
const traces = group.runs.flatMap((run) => run.trace ? [run.trace] : []);
|
|
850
|
+
const paths = new Set(traces.map((trace) => traceSteps(trace).filter((step) => !step.ignored).map((step) => `${step.from.join(",")}|${step.event}|${step.to.join(",")}`).join(" ")));
|
|
851
|
+
const reference = traces.find((trace) => trace.some((event) => event.type === "machine.transition"));
|
|
852
|
+
let earliest;
|
|
853
|
+
for (const trace of reference ? traces : []) {
|
|
854
|
+
if (trace === reference) continue;
|
|
855
|
+
const divergence = firstDivergence(trace, reference);
|
|
856
|
+
if (divergence && (!earliest || divergence.index < earliest.index)) earliest = divergence;
|
|
857
|
+
}
|
|
858
|
+
return {
|
|
859
|
+
...group,
|
|
860
|
+
passes,
|
|
861
|
+
passAny: passes > 0,
|
|
862
|
+
passAll: group.runs.length > 0 && passes === group.runs.length,
|
|
863
|
+
distinctPaths: paths.size,
|
|
864
|
+
...earliest ? { divergedAt: earliest.state } : {}
|
|
865
|
+
};
|
|
866
|
+
}
|
|
867
|
+
/** C(passes, j) / C(runs, j), as a product so large counts don't overflow. */
|
|
868
|
+
function allPass(passes, runs, j) {
|
|
869
|
+
let chance = 1;
|
|
870
|
+
for (let i = 0; i < j; i++) chance *= Math.max(0, passes - i) / (runs - i);
|
|
871
|
+
return chance;
|
|
872
|
+
}
|
|
873
|
+
//#endregion
|
|
874
|
+
//#region src/judge-agreement.ts
|
|
875
|
+
/**
|
|
876
|
+
* Measures how well a judge agrees with people on the same cases: accuracy,
|
|
877
|
+
* per-label recall, the confusion matrix, and the rows that disagree. With
|
|
878
|
+
* `positive`, also the TPR and TNR, the recommended view for binary judges
|
|
879
|
+
* (Husain/Shankar), and a correction for the judge's bias on unlabeled rows.
|
|
880
|
+
* Pure.
|
|
881
|
+
*
|
|
882
|
+
* Build rows from confirmed labels only; draft labels gate nothing. Group by
|
|
883
|
+
* decision state or request name so each narrow judge is aligned on its own.
|
|
884
|
+
* A row with no group key counts in the totals and in no group. Run it on a
|
|
885
|
+
* dev split and report the held-out split separately.
|
|
886
|
+
*
|
|
887
|
+
* Throws `AgentError` with code `'empty-agreement-rows'` when `rows` is empty.
|
|
888
|
+
*
|
|
889
|
+
* @example
|
|
890
|
+
* ```ts no-check
|
|
891
|
+
* const report = judgeAgreement(
|
|
892
|
+
* labeled.map((s) => ({ id: s.id, human: s.label, judge: verdicts[s.id], group: s.request })),
|
|
893
|
+
* { positive: 'fail' },
|
|
894
|
+
* );
|
|
895
|
+
* report.binary; // { tpr: 0.9, tnr: 0.8 }
|
|
896
|
+
* report.correct?.(0.3); // estimated true failure rate in production
|
|
897
|
+
* ```
|
|
898
|
+
*/
|
|
899
|
+
function judgeAgreement(rows, options = {}) {
|
|
900
|
+
if (rows.length === 0) throw new require_event_log.AgentError("empty-agreement-rows", "judgeAgreement: rows must contain at least one row.");
|
|
901
|
+
const report = summarize(rows, options.positive);
|
|
902
|
+
const groupOf = options.groupBy ?? ((row) => row.group);
|
|
903
|
+
if (options.groupBy || rows.some((row) => row.group !== void 0)) {
|
|
904
|
+
const grouped = /* @__PURE__ */ new Map();
|
|
905
|
+
for (const row of rows) {
|
|
906
|
+
const key = groupOf(row);
|
|
907
|
+
if (key === void 0) continue;
|
|
908
|
+
const members = grouped.get(key);
|
|
909
|
+
if (members) members.push(row);
|
|
910
|
+
else grouped.set(key, [row]);
|
|
911
|
+
}
|
|
912
|
+
const groups = Object.create(null);
|
|
913
|
+
for (const [key, members] of grouped) groups[key] = summarize(members, options.positive);
|
|
914
|
+
report.groups = groups;
|
|
915
|
+
}
|
|
916
|
+
return report;
|
|
917
|
+
}
|
|
918
|
+
function summarize(rows, positive) {
|
|
919
|
+
const perLabel = Object.create(null);
|
|
920
|
+
const confusion = Object.create(null);
|
|
921
|
+
const disagreements = [];
|
|
922
|
+
let agreed = 0;
|
|
923
|
+
for (const { id, human, judge } of rows) {
|
|
924
|
+
const label = perLabel[human] ??= {
|
|
925
|
+
human: 0,
|
|
926
|
+
agreed: 0
|
|
927
|
+
};
|
|
928
|
+
const row = confusion[human] ??= Object.create(null);
|
|
929
|
+
label.human++;
|
|
930
|
+
row[judge] = (row[judge] ?? 0) + 1;
|
|
931
|
+
if (human === judge) {
|
|
932
|
+
label.agreed++;
|
|
933
|
+
agreed++;
|
|
934
|
+
} else disagreements.push({
|
|
935
|
+
id,
|
|
936
|
+
human,
|
|
937
|
+
judge
|
|
938
|
+
});
|
|
939
|
+
}
|
|
940
|
+
const summary = {
|
|
941
|
+
total: rows.length,
|
|
942
|
+
accuracy: agreed / rows.length,
|
|
943
|
+
perLabel,
|
|
944
|
+
confusion,
|
|
945
|
+
disagreements
|
|
946
|
+
};
|
|
947
|
+
if (positive === void 0) return summary;
|
|
948
|
+
let positives = 0;
|
|
949
|
+
let truePositives = 0;
|
|
950
|
+
let negatives = 0;
|
|
951
|
+
let trueNegatives = 0;
|
|
952
|
+
for (const { human, judge } of rows) if (human === positive) {
|
|
953
|
+
positives++;
|
|
954
|
+
if (judge === positive) truePositives++;
|
|
955
|
+
} else {
|
|
956
|
+
negatives++;
|
|
957
|
+
if (judge !== positive) trueNegatives++;
|
|
958
|
+
}
|
|
959
|
+
const tpr = positives === 0 ? void 0 : truePositives / positives;
|
|
960
|
+
const tnr = negatives === 0 ? void 0 : trueNegatives / negatives;
|
|
961
|
+
summary.binary = {
|
|
962
|
+
tpr,
|
|
963
|
+
tnr
|
|
964
|
+
};
|
|
965
|
+
if (tpr !== void 0 && tnr !== void 0 && tpr + tnr > 1) summary.correct = (p) => Math.min(1, Math.max(0, (p + tnr - 1) / (tpr + tnr - 1)));
|
|
966
|
+
return summary;
|
|
967
|
+
}
|
|
968
|
+
//#endregion
|
|
484
969
|
//#region src/compare-runs.ts
|
|
485
970
|
/**
|
|
486
971
|
* Runs every scenario through a baseline and a candidate implementation and
|
|
@@ -592,9 +1077,14 @@ exports.assertAgentMachine = require_verify.assertAgentMachine;
|
|
|
592
1077
|
exports.canReach = require_verify.canReach;
|
|
593
1078
|
exports.compareRuns = compareRuns;
|
|
594
1079
|
exports.explorePaths = require_verify.explorePaths;
|
|
1080
|
+
exports.firstDivergence = firstDivergence;
|
|
595
1081
|
exports.foldTraces = foldTraces;
|
|
596
1082
|
exports.getDecisionRequests = getDecisionRequests;
|
|
1083
|
+
exports.judgeAgreement = judgeAgreement;
|
|
597
1084
|
exports.lintAgentMachine = require_verify.lintAgentMachine;
|
|
598
1085
|
exports.matchesTrajectory = matchesTrajectory;
|
|
1086
|
+
exports.repeatRuns = repeatRuns;
|
|
599
1087
|
exports.runSeam = runSeam;
|
|
600
1088
|
exports.simulateAgent = require_verify.simulateAgent;
|
|
1089
|
+
exports.summarizeRepeats = summarizeRepeats;
|
|
1090
|
+
exports.traceCoverage = traceCoverage;
|