@tangle-network/agent-runtime 0.94.6 → 0.94.8
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/dist/agent.d.ts +1 -1
- package/dist/agent.js +4 -4
- package/dist/{agentic-generator-CL1q0H91.d.ts → agentic-generator-1g7GDba7.d.ts} +73 -6
- package/dist/analyst-loop.d.ts +1 -1
- package/dist/analyst-loop.js +2 -2
- package/dist/{chunk-GKZ6DFDN.js → chunk-7HH22XN4.js} +105 -27
- package/dist/chunk-7HH22XN4.js.map +1 -0
- package/dist/{chunk-JMT5VYY3.js → chunk-7NCP74SL.js} +174 -48
- package/dist/chunk-7NCP74SL.js.map +1 -0
- package/dist/{chunk-WQA5Y6PD.js → chunk-EP6RVHMX.js} +3 -3
- package/dist/chunk-EP6RVHMX.js.map +1 -0
- package/dist/{chunk-LWNMCOEA.js → chunk-FRXDOBFP.js} +3 -3
- package/dist/chunk-GHXW56PO.js +526 -0
- package/dist/chunk-GHXW56PO.js.map +1 -0
- package/dist/{chunk-63TIZHMI.js → chunk-KYPVUEJ4.js} +4 -4
- package/dist/{chunk-FVJ7M3DA.js → chunk-M22PL4GH.js} +1 -1
- package/dist/chunk-M22PL4GH.js.map +1 -0
- package/dist/{chunk-4VMRCEI6.js → chunk-MSUWXCHD.js} +2 -2
- package/dist/{chunk-ZQZX77MM.js → chunk-QDSOD7RC.js} +2 -2
- package/dist/{chunk-RNWDH5Z3.js → chunk-TNUBPPZA.js} +2 -2
- package/dist/{completion-gate-CfX2thvL.d.ts → completion-gate-tzwyyD-E.d.ts} +2 -2
- package/dist/{coordination-pOGZuYS7.d.ts → coordination-o0TzS7Ms.d.ts} +5 -5
- package/dist/environment-provider.d.ts +2 -2
- package/dist/{improve-Ciuvu2jh.d.ts → improve-DZeatcs7.d.ts} +19 -8
- package/dist/index.d.ts +20 -17
- package/dist/index.js +37 -20
- package/dist/index.js.map +1 -1
- package/dist/intelligence.d.ts +6 -6
- package/dist/intelligence.js +41 -17
- package/dist/intelligence.js.map +1 -1
- package/dist/knowledge.d.ts +7 -7
- package/dist/knowledge.js +4 -4
- package/dist/lifecycle.d.ts +3 -3
- package/dist/lifecycle.js +2 -2
- package/dist/{local-harness-CePtFrwu.d.ts → local-harness-ZLMkKdZc.d.ts} +1 -1
- package/dist/{loop-runner-bin-B4PwjVaF.d.ts → loop-runner-bin-iTSxSmpv.d.ts} +2 -2
- package/dist/loop-runner-bin.d.ts +6 -6
- package/dist/loop-runner-bin.js +6 -6
- package/dist/loops.d.ts +36 -23
- package/dist/loops.js +4 -4
- package/dist/mcp/bin.js +3 -3
- package/dist/mcp/index.d.ts +10 -10
- package/dist/mcp/index.js +5 -5
- package/dist/{mcp-serve-verifier-D7hCZIkT.d.ts → mcp-serve-verifier-BO4MEsR3.d.ts} +1 -1
- package/dist/{openai-tools-BMOu1t3A.d.ts → openai-tools-0wACz8f8.d.ts} +1 -1
- package/dist/profiles.d.ts +1 -1
- package/dist/{sanitize-Dr3SliRC.d.ts → sanitize-Dcbjp0tU.d.ts} +1 -1
- package/dist/{supervise-Y3w878fN.d.ts → supervise-CRwaJIkd.d.ts} +4 -4
- package/dist/{types-Kjr5W6OF.d.ts → types-CKmyZ9TP.d.ts} +4 -7
- package/dist/{types-DHLXdLVN.d.ts → types-CogNCaw7.d.ts} +1 -1
- package/dist/{worktree-fanout-DmBvR-_O.d.ts → worktree-fanout-BJy24LYe.d.ts} +4 -4
- package/package.json +4 -4
- package/dist/chunk-4WWSICFY.js +0 -210
- package/dist/chunk-4WWSICFY.js.map +0 -1
- package/dist/chunk-FVJ7M3DA.js.map +0 -1
- package/dist/chunk-GKZ6DFDN.js.map +0 -1
- package/dist/chunk-JMT5VYY3.js.map +0 -1
- package/dist/chunk-WQA5Y6PD.js.map +0 -1
- /package/dist/{chunk-LWNMCOEA.js.map → chunk-FRXDOBFP.js.map} +0 -0
- /package/dist/{chunk-63TIZHMI.js.map → chunk-KYPVUEJ4.js.map} +0 -0
- /package/dist/{chunk-4VMRCEI6.js.map → chunk-MSUWXCHD.js.map} +0 -0
- /package/dist/{chunk-ZQZX77MM.js.map → chunk-QDSOD7RC.js.map} +0 -0
- /package/dist/{chunk-RNWDH5Z3.js.map → chunk-TNUBPPZA.js.map} +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import {
|
|
2
2
|
agenticGenerator
|
|
3
|
-
} from "./chunk-
|
|
3
|
+
} from "./chunk-GHXW56PO.js";
|
|
4
4
|
import {
|
|
5
5
|
assertModelAllowed
|
|
6
6
|
} from "./chunk-ISTDY47H.js";
|
|
@@ -9,9 +9,13 @@ import {
|
|
|
9
9
|
} from "./chunk-YEJR7IXO.js";
|
|
10
10
|
|
|
11
11
|
// src/improvement/improvement-driver.ts
|
|
12
|
+
import { spawnSync } from "child_process";
|
|
13
|
+
import {
|
|
14
|
+
verifyCodeSurface
|
|
15
|
+
} from "@tangle-network/agent-eval/campaign";
|
|
12
16
|
function improvementDriver(opts) {
|
|
13
17
|
const baseRef = opts.baseRef ?? "main";
|
|
14
|
-
const
|
|
18
|
+
const owned = /* @__PURE__ */ new Map();
|
|
15
19
|
return {
|
|
16
20
|
kind: `improvement:${opts.generator.kind}`,
|
|
17
21
|
async propose(ctx) {
|
|
@@ -20,32 +24,56 @@ function improvementDriver(opts) {
|
|
|
20
24
|
return [];
|
|
21
25
|
}
|
|
22
26
|
const surfaces = [];
|
|
27
|
+
const incumbent = verifiedCodeIncumbent(ctx.currentSurface);
|
|
28
|
+
const proposalBaseRef = incumbent?.baseCommit ?? baseRef;
|
|
23
29
|
for (let i = 0; i < ctx.populationSize; i++) {
|
|
24
30
|
if (ctx.signal.aborted) break;
|
|
25
31
|
const wt = await opts.worktree.create({
|
|
26
|
-
baseRef,
|
|
32
|
+
baseRef: proposalBaseRef,
|
|
27
33
|
label: `${opts.generator.kind}-gen${ctx.generation}-cand${i}`
|
|
28
34
|
});
|
|
35
|
+
owned.set(wt.path, wt);
|
|
29
36
|
try {
|
|
37
|
+
if (incumbent) advanceToIncumbent(wt, incumbent);
|
|
30
38
|
const { applied, summary } = await opts.generator.generate({
|
|
31
39
|
worktreePath: wt.path,
|
|
32
40
|
report: ctx.report,
|
|
33
41
|
findings,
|
|
34
42
|
dataset: ctx.dataset,
|
|
35
43
|
maxShots: ctx.maxImprovementShots ?? 1,
|
|
36
|
-
signal: ctx.signal
|
|
44
|
+
signal: ctx.signal,
|
|
45
|
+
generation: ctx.generation,
|
|
46
|
+
candidateIndex: i,
|
|
47
|
+
...ctx.costLedger ? { costLedger: ctx.costLedger } : {},
|
|
48
|
+
...ctx.costPhase ? { costPhase: ctx.costPhase } : {}
|
|
37
49
|
});
|
|
38
50
|
if (!applied) {
|
|
39
51
|
await opts.worktree.discard(wt);
|
|
52
|
+
owned.delete(wt.path);
|
|
40
53
|
continue;
|
|
41
54
|
}
|
|
42
55
|
const surface = await opts.worktree.finalize(wt, summary);
|
|
43
56
|
surfaces.push(surface);
|
|
44
|
-
|
|
57
|
+
owned.delete(wt.path);
|
|
58
|
+
owned.set(surface.worktreeRef, wt);
|
|
45
59
|
} catch (err) {
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
60
|
+
const cleanupErrors = [];
|
|
61
|
+
for (let attempt = 0; attempt < 2; attempt += 1) {
|
|
62
|
+
try {
|
|
63
|
+
await opts.worktree.discard(wt);
|
|
64
|
+
owned.delete(wt.path);
|
|
65
|
+
break;
|
|
66
|
+
} catch (cause) {
|
|
67
|
+
cleanupErrors.push(cause);
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
if (cleanupErrors.length === 0) throw err;
|
|
71
|
+
const failure = err instanceof Error ? err.message : String(err);
|
|
72
|
+
const cleanupSucceeded = !owned.has(wt.path);
|
|
73
|
+
throw new AggregateError(
|
|
74
|
+
[err, ...cleanupErrors],
|
|
75
|
+
cleanupSucceeded ? `improvementDriver: ${failure}; candidate cleanup retry succeeded` : `improvementDriver: ${failure}; candidate worktree could not be cleaned`
|
|
76
|
+
);
|
|
49
77
|
}
|
|
50
78
|
}
|
|
51
79
|
return surfaces;
|
|
@@ -53,11 +81,11 @@ function improvementDriver(opts) {
|
|
|
53
81
|
async cleanup(retainWorktreeRefs = []) {
|
|
54
82
|
const retained = new Set(retainWorktreeRefs);
|
|
55
83
|
const errors = [];
|
|
56
|
-
for (const [worktreeRef, worktree] of
|
|
84
|
+
for (const [worktreeRef, worktree] of owned) {
|
|
57
85
|
if (retained.has(worktreeRef)) continue;
|
|
58
86
|
try {
|
|
59
87
|
await opts.worktree.discard(worktree);
|
|
60
|
-
|
|
88
|
+
owned.delete(worktreeRef);
|
|
61
89
|
} catch (cause) {
|
|
62
90
|
errors.push(cause);
|
|
63
91
|
}
|
|
@@ -68,6 +96,38 @@ function improvementDriver(opts) {
|
|
|
68
96
|
}
|
|
69
97
|
};
|
|
70
98
|
}
|
|
99
|
+
function verifiedCodeIncumbent(surface) {
|
|
100
|
+
if (typeof surface !== "object" || surface.kind !== "code") return void 0;
|
|
101
|
+
verifyCodeSurface(surface);
|
|
102
|
+
return surface;
|
|
103
|
+
}
|
|
104
|
+
function advanceToIncumbent(worktree, incumbent) {
|
|
105
|
+
if (worktree.baseCommit !== incumbent.baseCommit || worktree.baseTree !== incumbent.baseTree) {
|
|
106
|
+
throw new Error("improvementDriver: candidate worktree does not match incumbent base identity");
|
|
107
|
+
}
|
|
108
|
+
if (worktree.baseCommit === incumbent.candidateCommit) return;
|
|
109
|
+
const merge = spawnSync("git", ["merge", "--ff-only", incumbent.candidateCommit], {
|
|
110
|
+
cwd: worktree.path,
|
|
111
|
+
encoding: "utf8"
|
|
112
|
+
});
|
|
113
|
+
if (merge.error) {
|
|
114
|
+
throw new Error(
|
|
115
|
+
`improvementDriver: failed to start candidate from incumbent: ${merge.error.message}`
|
|
116
|
+
);
|
|
117
|
+
}
|
|
118
|
+
if (merge.status !== 0) {
|
|
119
|
+
throw new Error(
|
|
120
|
+
`improvementDriver: could not fast-forward candidate to incumbent ${incumbent.candidateCommit}: ${merge.stderr.trim()}`
|
|
121
|
+
);
|
|
122
|
+
}
|
|
123
|
+
const head = spawnSync("git", ["rev-parse", "--verify", "HEAD"], {
|
|
124
|
+
cwd: worktree.path,
|
|
125
|
+
encoding: "utf8"
|
|
126
|
+
});
|
|
127
|
+
if (head.error || head.status !== 0 || head.stdout.trim() !== incumbent.candidateCommit) {
|
|
128
|
+
throw new Error("improvementDriver: candidate worktree did not reach the incumbent commit");
|
|
129
|
+
}
|
|
130
|
+
}
|
|
71
131
|
function resolveFindings(ctx) {
|
|
72
132
|
const report = ctx.report;
|
|
73
133
|
if (report && typeof report === "object" && "findings" in report) {
|
|
@@ -342,48 +402,88 @@ function memoryGenerationDistiller(staticFindings) {
|
|
|
342
402
|
return fresh === staticFindings ? staticFindings : [...staticFindings, ...fresh];
|
|
343
403
|
};
|
|
344
404
|
}
|
|
405
|
+
async function rethrowAfterCleanup(cause, cleanup, message) {
|
|
406
|
+
const cleanupErrors = [];
|
|
407
|
+
for (let attempt = 0; attempt < 2; attempt += 1) {
|
|
408
|
+
try {
|
|
409
|
+
await cleanup();
|
|
410
|
+
} catch (cleanupCause) {
|
|
411
|
+
cleanupErrors.push(cleanupCause);
|
|
412
|
+
continue;
|
|
413
|
+
}
|
|
414
|
+
if (cleanupErrors.length === 0) throw cause;
|
|
415
|
+
throw new AggregateError([cause, ...cleanupErrors], `${message}; the cleanup retry succeeded`);
|
|
416
|
+
}
|
|
417
|
+
throw new AggregateError([cause, ...cleanupErrors], message);
|
|
418
|
+
}
|
|
419
|
+
async function discardPreparedBaseline(worktree, baselineWorktree, cause) {
|
|
420
|
+
return rethrowAfterCleanup(
|
|
421
|
+
cause,
|
|
422
|
+
() => worktree.discard(baselineWorktree),
|
|
423
|
+
"improve(): code preparation failed and its baseline worktree could not be cleaned"
|
|
424
|
+
);
|
|
425
|
+
}
|
|
345
426
|
function isCodeSurface(surface) {
|
|
346
427
|
return typeof surface === "object" && surface !== null && surface.kind === "code";
|
|
347
428
|
}
|
|
348
|
-
async function prepareCodeRun(code
|
|
429
|
+
async function prepareCodeRun(code) {
|
|
349
430
|
const baseRef = code.baseRef ?? "main";
|
|
350
|
-
const worktree = gitWorktreeAdapter({
|
|
431
|
+
const worktree = code.worktree ?? gitWorktreeAdapter({
|
|
351
432
|
repoRoot: code.repoRoot,
|
|
352
433
|
...code.worktreeDir ? { worktreeDir: code.worktreeDir } : {}
|
|
353
434
|
});
|
|
354
435
|
const baselineWorktree = await worktree.create({ baseRef, label: "incumbent-baseline" });
|
|
355
|
-
|
|
356
|
-
|
|
357
|
-
|
|
436
|
+
try {
|
|
437
|
+
const baseline = await worktree.finalize(baselineWorktree, "Incumbent code checkout");
|
|
438
|
+
let baselineDiscarded = false;
|
|
358
439
|
const generator = code.generator ?? agenticGenerator({
|
|
359
440
|
...code.harness ? { harness: code.harness } : {},
|
|
360
441
|
...code.verify ? { verify: code.verify } : {},
|
|
361
442
|
...code.timeoutMs ? { timeoutMs: code.timeoutMs } : {}
|
|
362
443
|
});
|
|
363
|
-
managed = improvementDriver({ worktree, generator, baseRef });
|
|
364
|
-
|
|
365
|
-
|
|
366
|
-
|
|
367
|
-
|
|
368
|
-
|
|
369
|
-
|
|
370
|
-
|
|
371
|
-
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
|
|
376
|
-
|
|
377
|
-
|
|
378
|
-
|
|
379
|
-
|
|
380
|
-
|
|
381
|
-
|
|
382
|
-
|
|
383
|
-
|
|
384
|
-
|
|
385
|
-
|
|
444
|
+
const managed = improvementDriver({ worktree, generator, baseRef });
|
|
445
|
+
return {
|
|
446
|
+
baseline,
|
|
447
|
+
proposer: managed,
|
|
448
|
+
async cleanup(retainedWinner) {
|
|
449
|
+
const errors = [];
|
|
450
|
+
const retainedWorktreeRef = isCodeSurface(retainedWinner) ? retainedWinner.worktreeRef : void 0;
|
|
451
|
+
try {
|
|
452
|
+
await managed?.cleanup(retainedWorktreeRef ? [retainedWorktreeRef] : []);
|
|
453
|
+
} catch (cause) {
|
|
454
|
+
errors.push(cause);
|
|
455
|
+
}
|
|
456
|
+
if (!baselineDiscarded && retainedWorktreeRef !== baseline.worktreeRef) {
|
|
457
|
+
try {
|
|
458
|
+
await worktree.discard(baselineWorktree);
|
|
459
|
+
baselineDiscarded = true;
|
|
460
|
+
} catch (cause) {
|
|
461
|
+
errors.push(cause);
|
|
462
|
+
}
|
|
463
|
+
}
|
|
464
|
+
if (errors.length > 0) {
|
|
465
|
+
throw new AggregateError(errors, "improve(): failed to clean code improvement worktrees");
|
|
466
|
+
}
|
|
386
467
|
}
|
|
468
|
+
};
|
|
469
|
+
} catch (cause) {
|
|
470
|
+
return discardPreparedBaseline(worktree, baselineWorktree, cause);
|
|
471
|
+
}
|
|
472
|
+
}
|
|
473
|
+
function idempotentDispose(dispose) {
|
|
474
|
+
let disposed = false;
|
|
475
|
+
let inFlight;
|
|
476
|
+
return async () => {
|
|
477
|
+
if (disposed) return;
|
|
478
|
+
if (inFlight) return inFlight;
|
|
479
|
+
inFlight = (async () => {
|
|
480
|
+
await dispose();
|
|
481
|
+
disposed = true;
|
|
482
|
+
})();
|
|
483
|
+
try {
|
|
484
|
+
await inFlight;
|
|
485
|
+
} finally {
|
|
486
|
+
inFlight = void 0;
|
|
387
487
|
}
|
|
388
488
|
};
|
|
389
489
|
}
|
|
@@ -465,6 +565,11 @@ async function improve(profile, findings, opts) {
|
|
|
465
565
|
if (surface === "memory" && !memory) {
|
|
466
566
|
throw new ConfigError("improve(): surface 'memory' requires opts.memory.document");
|
|
467
567
|
}
|
|
568
|
+
if (surface === "code" && generator) {
|
|
569
|
+
throw new ConfigError(
|
|
570
|
+
"improve(): surface 'code' forbids opts.generator because an external SurfaceProposer cannot transfer checkout ownership; pass opts.code.generator instead"
|
|
571
|
+
);
|
|
572
|
+
}
|
|
468
573
|
const usesReflectionModel = !generator && (surface === "prompt" || surface === "skills");
|
|
469
574
|
if (usesReflectionModel) {
|
|
470
575
|
assertModelAllowed(sharedOptions.llm?.model ?? defaultReflectionModel, allowedModels);
|
|
@@ -476,7 +581,7 @@ async function improve(profile, findings, opts) {
|
|
|
476
581
|
"improve(): surface 'code' requires opts.code.repoRoot so the incumbent can run from an isolated checkout"
|
|
477
582
|
);
|
|
478
583
|
}
|
|
479
|
-
preparedCode = await prepareCodeRun(code
|
|
584
|
+
preparedCode = await prepareCodeRun(code);
|
|
480
585
|
}
|
|
481
586
|
const proposer = preparedCode?.proposer ?? generator ?? defaultGeneratorFor(surface, sharedOptions.llm);
|
|
482
587
|
if (!proposer) {
|
|
@@ -500,19 +605,33 @@ async function improve(profile, findings, opts) {
|
|
|
500
605
|
});
|
|
501
606
|
} catch (cause) {
|
|
502
607
|
if (!preparedCode) throw cause;
|
|
608
|
+
return rethrowAfterCleanup(
|
|
609
|
+
cause,
|
|
610
|
+
() => preparedCode.cleanup(),
|
|
611
|
+
"improve(): code improvement failed and its worktrees could not be cleaned"
|
|
612
|
+
);
|
|
613
|
+
}
|
|
614
|
+
const shipped = raw.gateDecision === "ship";
|
|
615
|
+
const winnerSurface = raw.winner.surface;
|
|
616
|
+
if (preparedCode) {
|
|
503
617
|
try {
|
|
504
|
-
await preparedCode.cleanup();
|
|
618
|
+
await preparedCode.cleanup(winnerSurface);
|
|
505
619
|
} catch (cleanupCause) {
|
|
620
|
+
try {
|
|
621
|
+
await preparedCode.cleanup();
|
|
622
|
+
} catch (finalCleanupCause) {
|
|
623
|
+
throw new AggregateError(
|
|
624
|
+
[cleanupCause, finalCleanupCause],
|
|
625
|
+
"improve(): code result cleanup failed, including the final all-worktree retry"
|
|
626
|
+
);
|
|
627
|
+
}
|
|
506
628
|
throw new AggregateError(
|
|
507
|
-
[
|
|
508
|
-
"improve(): code
|
|
629
|
+
[cleanupCause],
|
|
630
|
+
"improve(): code result cleanup failed; the final all-worktree retry succeeded"
|
|
509
631
|
);
|
|
510
632
|
}
|
|
511
|
-
throw cause;
|
|
512
633
|
}
|
|
513
|
-
const
|
|
514
|
-
const winnerSurface = raw.winner.surface;
|
|
515
|
-
await preparedCode?.cleanup(shipped ? winnerSurface : void 0);
|
|
634
|
+
const dispose = idempotentDispose(async () => preparedCode?.cleanup());
|
|
516
635
|
const externalDocument = surface === "skills" && skills ? skills : surface === "memory" && memory ? memory : void 0;
|
|
517
636
|
if (shipped && externalDocument) {
|
|
518
637
|
if (typeof winnerSurface !== "string") {
|
|
@@ -523,7 +642,14 @@ async function improve(profile, findings, opts) {
|
|
|
523
642
|
await externalDocument.writeBack?.(winnerSurface);
|
|
524
643
|
}
|
|
525
644
|
const nextProfile = shipped && !externalDocument ? applyImprovementWinnerToProfile(profile, surface, winnerSurface) : profile;
|
|
526
|
-
return {
|
|
645
|
+
return {
|
|
646
|
+
profile: nextProfile,
|
|
647
|
+
shipped,
|
|
648
|
+
lift: raw.lift,
|
|
649
|
+
gateDecision: raw.gateDecision,
|
|
650
|
+
raw,
|
|
651
|
+
dispose
|
|
652
|
+
};
|
|
527
653
|
}
|
|
528
654
|
|
|
529
655
|
export {
|
|
@@ -532,4 +658,4 @@ export {
|
|
|
532
658
|
applyImprovementWinnerToProfile,
|
|
533
659
|
improve
|
|
534
660
|
};
|
|
535
|
-
//# sourceMappingURL=chunk-
|
|
661
|
+
//# sourceMappingURL=chunk-7NCP74SL.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"sources":["../src/improvement/improvement-driver.ts","../src/improvement/raw-trace-distiller.ts","../src/improvement/improve.ts"],"sourcesContent":["/**\n *\n * `improvementDriver` — the ONE reflective/agentic improvement proposer for\n * agent-eval's improvement loop. It implements `SurfaceProposer` and owns\n * the candidate lifecycle (worktree create → generate → finalize/discard,\n * × populationSize); it delegates the only thing that genuinely varies — HOW\n * a candidate change is produced — to a pluggable `CandidateGenerator`.\n *\n * There is no separate \"analyst driver\" vs \"autoresearch driver\": those are\n * the SAME driver at two settings of a dial.\n * - cheap reflective path → `reflectiveGenerator` (shots=1, no sandbox;\n * applies pre-drafted patches)\n * - full agentic path → `agenticGenerator` (shots=N, multi-shot\n * verify-in-session loop; an agent reads code +\n * report, edits, and re-tries on verifier failure)\n * Both emit changes into a worktree the driver finalizes into a\n * `CodeSurface{ worktreeRef }` the loop measures on the holdout. See\n * agent-eval's `docs/design/self-improvement-engine.md`.\n *\n * @experimental\n */\n\nimport { spawnSync } from 'node:child_process'\nimport type { AnalystFinding, CostLedger } from '@tangle-network/agent-eval'\nimport {\n type CodeSurface,\n type LabeledScenarioStore,\n type ProposeContext,\n type SurfaceProposer,\n verifyCodeSurface,\n type Worktree,\n type WorktreeAdapter,\n} from '@tangle-network/agent-eval/campaign'\n\n/** The byte-producing seam — the ONE thing that differs between the cheap\n * reflective path and the full agentic path. A generator makes (uncommitted)\n * changes inside `worktreePath`; the driver commits them via the worktree\n * adapter's `finalize`. */\nexport interface CandidateGenerator {\n kind: string\n /** Whether this generator can produce a candidate from an EMPTY findings set\n * and no phase-2 report — i.e. it draws its change signal from the repo and\n * the raw-trace filesystem context on disk, not only from pre-summarized\n * findings. An agentic coder (`agenticGenerator`) sets this: the seed repo +\n * raw traces ARE the signal, so it must still run the full `populationSize`\n * when the distiller yielded nothing (this is the meta-harness contract — the\n * agent diagnoses from the raw traces itself). A patch-applier\n * (`reflectiveGenerator`) leaves it unset — with no findings there is no\n * patch to draft, so the driver short-circuits rather than spin up worktrees\n * for a guaranteed no-op. Default `false`. */\n proposesWithoutFindings?: boolean\n generate(args: {\n /** The candidate worktree — a clean checkout of the current incumbent. */\n worktreePath: string\n /** Phase-2 research report (analyst findings + diff), opaque. */\n report: unknown\n /** Findings resolved from the report or the loop context. */\n findings: AnalystFinding[]\n /** Handle to all captured data, to ground the change. */\n dataset?: LabeledScenarioStore\n /** DEPTH: max iterations the generator may take (agentic uses this; the\n * reflective generator ignores it). */\n maxShots: number\n signal: AbortSignal\n /** Improvement-loop coordinates. Present when called through improvementDriver. */\n generation?: number\n candidateIndex?: number\n /** Shared run-wide paid-call account supplied by agent-eval 0.117+. */\n costLedger?: CostLedger\n /** Receipt attribution phase supplied alongside `costLedger`. */\n costPhase?: string\n }): Promise<{ applied: boolean; summary: string }>\n}\n\nexport interface ImprovementDriverOptions {\n worktree: WorktreeAdapter\n generator: CandidateGenerator\n /** Root ref for first-generation/direct callers. Default `main`.\n * Later code generations retain the incumbent's original root. */\n baseRef?: string\n}\n\nexport interface ManagedImprovementDriver extends SurfaceProposer<AnalystFinding> {\n /** Remove every owned candidate except explicitly retained finalized winners. */\n cleanup(retainWorktreeRefs?: readonly string[]): Promise<void>\n}\n\n/** The one reflective/agentic improvement proposer (`SurfaceProposer`): owns the candidate worktree lifecycle and delegates HOW a change is produced to a pluggable `CandidateGenerator`. */\nexport function improvementDriver(opts: ImprovementDriverOptions): ManagedImprovementDriver {\n const baseRef = opts.baseRef ?? 'main'\n const owned = new Map<string, Worktree>()\n\n return {\n kind: `improvement:${opts.generator.kind}`,\n async propose(ctx: ProposeContext<AnalystFinding>) {\n const findings = resolveFindings(ctx)\n // No findings AND no report AND a generator that can only act on findings\n // (the reflective patch-applier) — propose nothing rather than spin up\n // worktrees for a guaranteed no-op. An agentic coder draws its signal from\n // the repo + raw traces on disk, so it opts in via `proposesWithoutFindings`\n // and still runs the full populationSize even on an empty findings set —\n // otherwise the FIRST generation (whose seed findings are empty and whose\n // rawTraceDistiller has not run yet) would always generate ZERO candidates.\n if (\n findings.length === 0 &&\n ctx.report === undefined &&\n !opts.generator.proposesWithoutFindings\n ) {\n return []\n }\n\n const surfaces: CodeSurface[] = []\n const incumbent = verifiedCodeIncumbent(ctx.currentSurface)\n const proposalBaseRef = incumbent?.baseCommit ?? baseRef\n for (let i = 0; i < ctx.populationSize; i++) {\n if (ctx.signal.aborted) break\n const wt = await opts.worktree.create({\n baseRef: proposalBaseRef,\n label: `${opts.generator.kind}-gen${ctx.generation}-cand${i}`,\n })\n owned.set(wt.path, wt)\n // Once a worktree exists it MUST be accounted for: finalized into a\n // surface, or discarded. A throw from generate()/finalize() must not\n // leak the worktree + branch — discard best-effort, then rethrow loud.\n try {\n if (incumbent) advanceToIncumbent(wt, incumbent)\n const { applied, summary } = await opts.generator.generate({\n worktreePath: wt.path,\n report: ctx.report,\n findings,\n dataset: ctx.dataset,\n maxShots: ctx.maxImprovementShots ?? 1,\n signal: ctx.signal,\n generation: ctx.generation,\n candidateIndex: i,\n ...(ctx.costLedger ? { costLedger: ctx.costLedger } : {}),\n ...(ctx.costPhase ? { costPhase: ctx.costPhase } : {}),\n })\n if (!applied) {\n await opts.worktree.discard(wt)\n owned.delete(wt.path)\n continue\n }\n const surface = await opts.worktree.finalize(wt, summary)\n surfaces.push(surface)\n owned.delete(wt.path)\n owned.set(surface.worktreeRef, wt)\n } catch (err) {\n const cleanupErrors: unknown[] = []\n for (let attempt = 0; attempt < 2; attempt += 1) {\n try {\n await opts.worktree.discard(wt)\n owned.delete(wt.path)\n break\n } catch (cause) {\n cleanupErrors.push(cause)\n }\n }\n if (cleanupErrors.length === 0) throw err\n const failure = err instanceof Error ? err.message : String(err)\n const cleanupSucceeded = !owned.has(wt.path)\n throw new AggregateError(\n [err, ...cleanupErrors],\n cleanupSucceeded\n ? `improvementDriver: ${failure}; candidate cleanup retry succeeded`\n : `improvementDriver: ${failure}; candidate worktree could not be cleaned`,\n )\n }\n }\n return surfaces\n },\n async cleanup(retainWorktreeRefs = []) {\n const retained = new Set(retainWorktreeRefs)\n const errors: unknown[] = []\n for (const [worktreeRef, worktree] of owned) {\n if (retained.has(worktreeRef)) continue\n try {\n await opts.worktree.discard(worktree)\n owned.delete(worktreeRef)\n } catch (cause) {\n errors.push(cause)\n }\n }\n if (errors.length > 0) {\n throw new AggregateError(errors, 'improvementDriver: failed to discard candidate worktrees')\n }\n },\n }\n}\n\n/** A code incumbent must still match the immutable identity that was measured. */\nfunction verifiedCodeIncumbent(surface: ProposeContext['currentSurface']): CodeSurface | undefined {\n if (typeof surface !== 'object' || surface.kind !== 'code') return undefined\n verifyCodeSurface(surface)\n return surface\n}\n\n/** Start at the root commit recorded by the incumbent, then fast-forward the\n * fresh branch to the incumbent commit. The worktree stays clean for the\n * generator while `finalize()` still emits one cumulative root-to-candidate\n * patch that can be applied or rolled back independently of prior branches. */\nfunction advanceToIncumbent(worktree: Worktree, incumbent: CodeSurface): void {\n if (worktree.baseCommit !== incumbent.baseCommit || worktree.baseTree !== incumbent.baseTree) {\n throw new Error('improvementDriver: candidate worktree does not match incumbent base identity')\n }\n if (worktree.baseCommit === incumbent.candidateCommit) return\n\n const merge = spawnSync('git', ['merge', '--ff-only', incumbent.candidateCommit], {\n cwd: worktree.path,\n encoding: 'utf8',\n })\n if (merge.error) {\n throw new Error(\n `improvementDriver: failed to start candidate from incumbent: ${merge.error.message}`,\n )\n }\n if (merge.status !== 0) {\n throw new Error(\n `improvementDriver: could not fast-forward candidate to incumbent ${incumbent.candidateCommit}: ${merge.stderr.trim()}`,\n )\n }\n\n const head = spawnSync('git', ['rev-parse', '--verify', 'HEAD'], {\n cwd: worktree.path,\n encoding: 'utf8',\n })\n if (head.error || head.status !== 0 || head.stdout.trim() !== incumbent.candidateCommit) {\n throw new Error('improvementDriver: candidate worktree did not reach the incumbent commit')\n }\n}\n\n/** Phase-2 report carries `findings` when present; else fall back to the\n * loop's `ctx.findings`. The report is opaque to the substrate, so probe it\n * structurally. */\nfunction resolveFindings(ctx: ProposeContext<AnalystFinding>): AnalystFinding[] {\n const report = ctx.report\n if (report && typeof report === 'object' && 'findings' in report) {\n const f = (report as { findings: unknown }).findings\n if (Array.isArray(f) && f.length > 0) return f as AnalystFinding[]\n }\n return ctx.findings\n}\n","/**\n *\n * `rawTraceDistiller` — the meta-harness `analyzeGeneration` producer.\n *\n * The default `generationFailureDistiller` (in `improve.ts`) COMPRESSES each\n * generation's failing cells into ~400-char structured findings before the next\n * proposal round. That is the ACE-style recipe: a small summary is the proposer's\n * whole view of what went wrong. This producer does the opposite — the\n * meta-harness recipe (yoonholee.com/meta-harness): it does NOT summarize. It\n * points the coding-agent proposer at the generation's RAW run traces already on\n * disk under `runDir` — the durable per-cell `spans.jsonl` event logs,\n * `cached-result.json` scores, and any artifacts the substrate persisted — and\n * instructs the agent to `grep`/`cat`/`ls` them to diagnose the failures itself\n * (up to the harness's full context, ~millions of tokens, vs a ~400-char digest).\n *\n * It emits `AnalystFinding[]` so it drops into the SAME `opts.analyzeGeneration`\n * slot the default distiller uses, and renders through the same\n * `agenticGenerator` prompt path (`claim` + `recommended_action`). The findings\n * carry ABSOLUTE paths — the coding harness runs with `cwd` = a candidate\n * worktree, so a relative `runDir` would be uncattable from there.\n *\n * Runtime layout it reads (written by agent-eval's optimization loop):\n *\n * <runDir>/gen-<N>/ ← the generation dir (input.runDir)\n * candidate-<i>/ ← one candidate campaign (campaign.runDir)\n * <sanitized cellId>/ ← one scenario×rep cell\n * spans.jsonl ← the raw trace (event/span log)\n * cached-result.json ← the cell's score + artifact ref\n * <artifacts…> ← whatever the dispatch wrote\n *\n * @experimental\n */\n\nimport { type Dirent, existsSync, readdirSync } from 'node:fs'\nimport { basename, join, resolve } from 'node:path'\nimport { type AnalystFinding, makeFinding } from '@tangle-network/agent-eval'\nimport type { Scenario, SelfImproveOptions } from '@tangle-network/agent-eval/contract'\n\nconst ANALYST_ID = 'raw-trace-distiller'\n/** A cell counts as \"failing\" below this mean composite (matches the default\n * distiller's near-perfect threshold) or when it recorded an `error`. */\nconst PASS_THRESHOLD = 0.999\n\nexport interface RawTraceDistillerOptions {\n /** Anchor the emitted paths at this run root instead of the generation `runDir`\n * the loop passes in. Normally unset — each call points at that generation's\n * own directory (`input.runDir`). Pass an absolute path when you construct the\n * producer ahead of the loop and want a fixed anchor (e.g. a test fixture). */\n runDir?: string\n /** Max candidates to surface trace paths for, worst-scoring first. Default 12. */\n maxCandidates?: number\n /** Max failing cells to enumerate per candidate before collapsing the rest into\n * an \"ls the candidate dir\" pointer. Default 8. */\n maxCellsPerCandidate?: number\n /** Max concrete file paths to list per cell (the agent can always `ls` the dir\n * for the rest). Default 24. */\n maxFilesPerCell?: number\n /** Findings to fall back to when the generation had NO failing cells, so a\n * clean round never wipes the proposer's steering context. Mirrors the default\n * distiller's static-seed fallback. Default: a single instruction finding. */\n fallbackFindings?: unknown[]\n}\n\ninterface CellTrace {\n scenarioId: string\n composite: number\n error?: string\n cellDir: string\n files: string[]\n truncatedFiles: boolean\n}\n\n/**\n * Build an `analyzeGeneration` producer that feeds the proposer RAW-TRACE\n * FILESYSTEM CONTEXT — paths into the prior generation's real run traces plus a\n * grep/cat-to-diagnose instruction — instead of a pre-summarized digest.\n *\n * Drop-in for `opts.analyzeGeneration` on `improve()` / `selfImprove()`:\n *\n * await improve(profile, seedFindings, {\n * surface: 'code',\n * code: { repoRoot },\n * runDir: '/abs/run', // MUST be a real path — the traces live here\n * analyzeGeneration: rawTraceDistiller(),\n * scenarios, judge, agent,\n * })\n */\nexport function rawTraceDistiller<TScenario extends Scenario = Scenario, TArtifact = unknown>(\n options: RawTraceDistillerOptions = {},\n): NonNullable<SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration']> {\n const maxCandidates = options.maxCandidates ?? 12\n const maxCellsPerCandidate = options.maxCellsPerCandidate ?? 8\n const maxFilesPerCell = options.maxFilesPerCell ?? 24\n\n return async (input) => {\n const genRoot = absoluteRunDir(options.runDir ?? input.runDir)\n const durable = isDurable(genRoot)\n\n // Rank candidates worst-first; the worst failures are the highest-signal\n // context for the next edit. Stable sort keeps equal-composite order.\n const ranked = [...input.candidates]\n .map((c) => ({\n surfaceHash: c.surfaceHash,\n composite: c.composite,\n campaignDir: absoluteRunDir(c.campaign.runDir),\n cells: failingCells(c.campaign, maxCellsPerCandidate, maxFilesPerCell),\n }))\n .sort((a, b) => a.composite - b.composite)\n .slice(0, maxCandidates)\n\n const totalFailingCells = ranked.reduce((n, c) => n + c.cells.length, 0)\n\n // A clean generation: keep the proposer's steering context rather than\n // wiping it (parity with the default distiller's static-seed fallback). An\n // EMPTY fallback array means there is no STATIC seed to preserve — it must NOT\n // wipe the context to nothing. Fall through to the default raw-trace\n // instruction so the meta-harness discipline stays live (the agent still\n // inspects the on-disk traces next round, and the finding's `raw-trace-context`\n // area keeps the agenticGenerator's diagnosis-evidence gate armed). A bare\n // `??` would return the empty array and silently disable both.\n if (totalFailingCells === 0) {\n if (options.fallbackFindings && options.fallbackFindings.length > 0) {\n return options.fallbackFindings\n }\n return [\n makeFinding({\n analyst_id: ANALYST_ID,\n severity: 'info',\n area: 'raw-trace-context',\n confidence: 1,\n claim: `Generation ${input.generation} had no failing cells. The full raw run traces are on disk under ${genRoot}.`,\n recommended_action: `To keep improving, grep/cat the raw traces under ${genRoot} (per-cell spans.jsonl + cached-result.json) to find the weakest passing runs, then make a targeted harness-code edit.`,\n evidence_refs: [{ kind: 'artifact', uri: genRoot }],\n metadata: { generation: input.generation, runDir: genRoot, failingCells: 0 },\n }),\n ]\n }\n\n const findings: AnalystFinding[] = []\n\n // 1. The meta-harness instruction: diagnose from the RAW traces, not a digest.\n findings.push(\n makeFinding({\n analyst_id: ANALYST_ID,\n severity: 'high',\n area: 'raw-trace-context',\n confidence: 1,\n claim: `Generation ${input.generation} produced ${totalFailingCells} failing/low-scoring cell(s) across ${ranked.length} candidate(s). Their FULL RAW run traces are on disk under ${genRoot} — the actual event logs (spans.jsonl), scores (cached-result.json), and artifacts, not a summary.${\n durable\n ? ''\n : ' (WARNING: this run root does not exist on disk — it looks like an in-memory run; pass a real runDir to improve() to get raw-trace context.)'\n }`,\n recommended_action: `Do NOT rely on a pre-summarized finding. Before editing, DIAGNOSE from the raw traces: run \\`grep\\`/\\`cat\\`/\\`ls\\` over the trace files and directories named in the following findings to see exactly what each failing run did and why it scored low, then make the smallest harness-code edit that fixes the dominant failure. Start with \\`grep -rIn \"error\" ${genRoot}\\` then \\`cat\\` the spans.jsonl of the worst cell.`,\n evidence_refs: [{ kind: 'artifact', uri: genRoot }],\n metadata: {\n generation: input.generation,\n runDir: genRoot,\n failingCells: totalFailingCells,\n candidates: ranked.length,\n },\n }),\n )\n\n // 2. One finding per failing candidate: its campaign dir + the concrete raw\n // trace files to grep/cat.\n for (const cand of ranked) {\n if (cand.cells.length === 0) continue\n const scenarioList = cand.cells.map((c) => c.scenarioId).join(', ')\n const fileLines = cand.cells\n .map((c) => {\n const header = ` cell ${c.scenarioId} (composite ${c.composite.toFixed(3)}${\n c.error ? `, error: ${truncate(c.error, 160)}` : ''\n }) — dir ${c.cellDir}`\n const files = c.files.map((f) => ` - ${f}`).join('\\n')\n const more = c.truncatedFiles ? `\\n - …(ls ${c.cellDir} for the rest)` : ''\n return c.files.length > 0 ? `${header}\\n${files}${more}` : header\n })\n .join('\\n')\n\n findings.push(\n makeFinding({\n analyst_id: ANALYST_ID,\n severity: cand.composite < 0.5 ? 'critical' : 'high',\n area: 'raw-trace-context',\n confidence: 1,\n subject: cand.surfaceHash,\n claim: `Candidate ${cand.surfaceHash} scored composite ${cand.composite.toFixed(3)} with ${cand.cells.length} failing cell(s) [${scenarioList}]. Its raw traces are under ${cand.campaignDir}.`,\n recommended_action: `grep/cat these raw trace files to diagnose WHY this candidate failed before editing:\\n${fileLines}\\nOr scan the whole candidate at once: \\`grep -rIn . ${cand.campaignDir}\\` and \\`ls -R ${cand.campaignDir}\\`.`,\n evidence_refs: [\n { kind: 'artifact', uri: cand.campaignDir },\n ...cand.cells.flatMap((c) =>\n c.files.map((f) => ({ kind: 'artifact' as const, uri: f })),\n ),\n ],\n metadata: {\n surfaceHash: cand.surfaceHash,\n composite: cand.composite,\n campaignDir: cand.campaignDir,\n cells: cand.cells.map((c) => ({\n scenarioId: c.scenarioId,\n composite: c.composite,\n cellDir: c.cellDir,\n files: c.files,\n ...(c.error ? { error: c.error } : {}),\n })),\n },\n }),\n )\n }\n\n return findings\n }\n}\n\n/** The failing cells of a candidate campaign, each with its on-disk trace files.\n * Mirrors the default distiller's per-cell composite (mean of judge composites,\n * 0 when a cell produced no judge score) and its failing predicate. */\nfunction failingCells(\n campaign: {\n runDir: string\n cells: ReadonlyArray<{\n cellId: string\n scenarioId: string\n error?: string\n judgeScores: Record<string, { composite?: number }>\n }>\n artifactsByPath?: Record<string, string>\n },\n maxCells: number,\n maxFiles: number,\n): CellTrace[] {\n const campaignDir = absoluteRunDir(campaign.runDir)\n const durable = isDurable(campaignDir)\n const out: CellTrace[] = []\n for (const cell of campaign.cells) {\n const scores = Object.values(cell.judgeScores ?? {})\n const composite =\n scores.length === 0\n ? 0\n : scores.reduce((sum, s) => sum + (s.composite ?? 0), 0) / scores.length\n if (!cell.error && composite >= PASS_THRESHOLD) continue\n\n const cellDir = join(campaignDir, sanitizeCellId(cell.cellId))\n const artifactPaths = artifactPathsForCell(campaign.artifactsByPath, cell.cellId)\n const discovered = durable ? listTraceFiles(cellDir) : []\n // Canonical anchors the substrate always writes, kept even when a mem:// run\n // never flushed them to disk (so the agent still learns the expected path).\n const canonical = [join(cellDir, 'spans.jsonl'), join(cellDir, 'cached-result.json')]\n const files = dedupeSorted([...discovered, ...artifactPaths, ...canonical])\n\n out.push({\n scenarioId: cell.scenarioId,\n composite: Number(composite.toFixed(3)),\n ...(cell.error ? { error: cell.error } : {}),\n cellDir,\n files: files.slice(0, maxFiles),\n truncatedFiles: files.length > maxFiles,\n })\n if (out.length >= maxCells) break\n }\n return out\n}\n\n/** Absolute paths of artifacts the campaign recorded for a cell. `artifactsByPath`\n * is keyed `${cellId}/${relPath}` → absolute path. */\nfunction artifactPathsForCell(\n artifactsByPath: Record<string, string> | undefined,\n cellId: string,\n): string[] {\n if (!artifactsByPath) return []\n const prefix = `${cellId}/`\n return Object.entries(artifactsByPath)\n .filter(([key]) => key.startsWith(prefix))\n .map(([, absPath]) => resolve(absPath))\n}\n\n/** Real files directly under `dir` and one level of sub-directories (artifacts\n * are sometimes nested). Absolute paths, sorted. `[]` when the dir is absent,\n * stale, unreadable, or contains symlinked dirs — trace context is advisory and\n * the canonical anchors below still tell the proposer where to inspect. */\nfunction listTraceFiles(dir: string): string[] {\n const out: string[] = []\n for (const entry of safeReadDir(dir)) {\n const full = join(dir, entry.name)\n if (entry.isFile()) {\n out.push(full)\n } else if (!entry.isSymbolicLink() && entry.isDirectory()) {\n for (const sub of safeReadDir(full)) {\n if (sub.isFile()) out.push(join(full, sub.name))\n }\n }\n }\n return out\n}\n\nfunction safeReadDir(dir: string): Dirent[] {\n try {\n return readdirSync(dir, { withFileTypes: true })\n } catch {\n return []\n }\n}\n\n/** Substrate cell-dir sanitization — must match agent-eval's\n * `cellId.replace(/[^a-zA-Z0-9_-]/g, '_')` so the computed dir matches disk. */\nfunction sanitizeCellId(cellId: string): string {\n return cellId.replace(/[^a-zA-Z0-9_-]/g, '_')\n}\n\n/** A run root is durable (has real files) when it is not an in-memory sentinel\n * and exists on disk. `mem://` runs keep everything in-process — no traces. */\nfunction isDurable(runDir: string): boolean {\n return !runDir.startsWith('mem://') && existsSync(runDir)\n}\n\n/** Resolve a run dir to absolute (the coding harness runs from a worktree cwd, so\n * relative paths are uncattable there). `mem://` sentinels pass through untouched. */\nfunction absoluteRunDir(runDir: string): string {\n return runDir.startsWith('mem://') ? runDir : resolve(runDir)\n}\n\nfunction dedupeSorted(paths: string[]): string[] {\n return [...new Set(paths)].sort((a, b) => {\n // Group by directory then filename for a stable, readable listing.\n const da = a.slice(0, a.length - basename(a).length)\n const db = b.slice(0, b.length - basename(b).length)\n return da === db ? basename(a).localeCompare(basename(b)) : da.localeCompare(db)\n })\n}\n\nfunction truncate(s: string, n: number): string {\n return s.length <= n ? s : `${s.slice(0, n - 1)}…`\n}\n","/**\n *\n * `improve` — the ONE public, surface-pluggable RSI verb.\n *\n * A thin facade over agent-eval's `selfImprove` (the held-out-gated closed\n * loop). It removes the two things a caller otherwise has to know to drive the\n * loop by hand: WHICH `MutableSurface` of the profile is being optimized, and\n * WHICH `SurfaceProposer` mutates that surface. You name a `surface`; the\n * facade picks the matching default proposer, extracts the baseline surface from\n * the profile, runs `selfImprove`, and (on a ship verdict) writes the promoted\n * winner back into the corresponding profile field.\n *\n * - `surface: 'prompt'` → `gepaProposer` mutates `profile.prompt.systemPrompt`.\n * - `surface: 'skills'` → `skillOptProposer` mutates a skills document string.\n * - `surface: 'memory'` → `memoryCurationProposer` curates a bounded durable\n * lesson document supplied through `opts.memory`.\n * - `surface: 'agent-profile'` → caller-supplied proposer mutates the complete\n * canonical AgentProfile JSON in one candidate.\n * - `surface` ∈ {`tools`, `mcp`, `hooks`, `subagents`, `agent-profile`} → no zero-config default\n * proposer exists (a code/config proposer needs caller-supplied wiring — a\n * worktree repo root, a candidate generator, a serializer). The facade\n * requires an explicit `opts.generator` for these and throws a `ConfigError`\n * otherwise. This is a designed boundary, not a missing default: there is\n * no safe value the facade could invent for those surfaces. Code instead\n * requires `opts.code.repoRoot` and accepts only the runtime-owned\n * `opts.code.generator` path so every isolated checkout can be released.\n *\n * Everything else (`scenarios`, `judge`, `agent`, `budget`, `llm`) passes\n * straight through to `selfImprove`.\n *\n * @experimental\n */\n\nimport { canonicalJson } from '@tangle-network/agent-eval'\nimport {\n gepaProposer,\n gitWorktreeAdapter,\n memoryCurationProposer,\n skillOptProposer,\n type Worktree,\n type WorktreeAdapter,\n} from '@tangle-network/agent-eval/campaign'\nimport {\n type CodeSurface,\n type MutableSurface,\n type Scenario,\n type SelfImproveBudget,\n type SelfImproveLlm,\n type SelfImproveOptions,\n type SelfImproveResult,\n type SurfaceProposer,\n selfImprove,\n} from '@tangle-network/agent-eval/contract'\nimport { type AgentProfile, agentProfileSchema } from '@tangle-network/agent-interface'\nimport { ConfigError } from '../errors'\nimport type { LocalHarness } from '../mcp/local-harness'\nimport { assertModelAllowed } from '../runtime/supervise/model-policy'\nimport { agenticGenerator, type Verifier } from './agentic-generator'\nimport {\n type CandidateGenerator,\n improvementDriver,\n type ManagedImprovementDriver,\n} from './improvement-driver'\nimport { rawTraceDistiller } from './raw-trace-distiller'\n\n/** The executable agent lever `improve` optimizes. Profile fields remain\n * portable AgentProfile coordinates; implementation and orchestration files\n * use the code surface so a winner can be sealed into an exact candidate. */\nexport type ImproveSurface =\n | 'prompt'\n | 'skills'\n | 'tools'\n | 'mcp'\n | 'hooks'\n | 'subagents'\n | 'agent-profile'\n | 'memory'\n | 'code'\n\nexport type ImproveOptions<TScenario extends Scenario, TArtifact> = Omit<\n SelfImproveOptions<TScenario, TArtifact>,\n 'analyzeGeneration' | 'baselineSurface' | 'findings' | 'gate' | 'proposer'\n> & {\n /** Which profile lever to optimize. Default `'prompt'`. Selects the default\n * generator + the baseline-surface extraction shape. */\n surface?: ImproveSurface\n /** The `SurfaceProposer` that mutates a profile surface. When unset, the facade\n * picks the default for prompt, skills, and memory; surfaces\n * with no default REQUIRE this (fail-loud otherwise). Forbidden for code;\n * use `code.generator` so the runtime owns candidate cleanup. */\n generator?: SurfaceProposer\n /** Gate mode. `'holdout'` (default) runs the held-out promotion gate;\n * `'none'` is a baseline-only run (`budget.generations = 0`). */\n gate?: 'holdout' | 'none'\n /** Restrict the run to this subset of models. When set, the reflection model\n * (`llm.model`, or the default when unset) must be a member, or `improve()` throws\n * a `ConfigError` before the generator is built. Unset = unrestricted. */\n allowedModels?: readonly string[]\n /** Per-generation findings producer passthrough (see selfImprove.analyzeGeneration).\n * DEFAULT: the built-in failure distiller — after each generation it turns the\n * worst-scoring/errored cells into structured findings ({ scenario, composite,\n * notes, error }) for the NEXT proposal round, so the proposer reasons over what\n * actually failed instead of a static seed. Pass your own producer (e.g. a\n * trace-analyst over the runDir's traces) to replace it; pass `null` to disable\n * and keep the static `findings` all the way through. */\n analyzeGeneration?: SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration'] | null\n /** META-HARNESS mode: instead of the ~400-char distilled findings, feed the\n * proposer RAW-TRACE FILESYSTEM CONTEXT — the PATHS into the prior generation's\n * real run traces under `runDir` (per-cell `spans.jsonl` event logs +\n * `cached-result.json` scores + artifacts) plus a `grep`/`cat`-to-diagnose\n * instruction — so the coding agent reads the actual failures itself rather than\n * a pre-summary. Requires a REAL `runDir` (that is where the traces live).\n * Ignored when `analyzeGeneration` is set explicitly (that wins) or is `null`\n * (disabled). Equivalent to `analyzeGeneration: rawTraceDistiller()`; this flag\n * is the one-line enable. Default `false` (the distiller stays the default). */\n rawTraceContext?: boolean\n /** CODE-surface wiring: name `surface: 'code'`, point at a repo, and the\n * facade assembles the whole candidate pipeline — an isolated incumbent plus git worktrees\n * (`gitWorktreeAdapter`) driven by `improvementDriver` with the full agentic\n * generator (a real coding harness edits each candidate worktree; a `verify`\n * hook gates candidates before they are ever measured). Ignored when\n * `opts.generator` is supplied. Required for every code run because a real\n * repository and base ref are necessary to measure the incumbent. */\n code?: ImproveCodeOptions\n /** SKILLS-surface wiring for real skill-DOCUMENT optimization. Without this,\n * `surface: 'skills'` optimizes the profile's skills REFS array (file pointers)\n * — which `skillOptProposer` (a document patcher) cannot meaningfully edit.\n * Provide the document CONTENT to optimize + a `writeBack` to persist the\n * shipped winner (the profile ref points at a file the caller owns). This is\n * what makes skillOpt reachable through improve(). */\n skills?: ImproveSkillsOptions\n /** MEMORY-surface wiring for a curated durable memory document. The default\n * deterministic proposer deduplicates and ranks lessons from findings, then\n * replaces its managed block instead of growing memory without bound. */\n memory?: ImproveMemoryOptions\n /** Custom held-back-exam decision. The string `gate` above controls whether\n * the exam runs; this callback controls how its evidence decides promotion. */\n promotionGate?: SelfImproveOptions<TScenario, TArtifact>['gate']\n}\n\nexport interface ImproveSkillsOptions {\n /** The skill document's current text — the baseline `skillOptProposer` patches. */\n document: string\n /** Persist the shipped winner document (write the file the profile ref points at).\n * Called only on a ship verdict. When omitted, the winner is still returned in\n * `result.raw.winner.surface` for the caller to materialize. */\n writeBack?: (winnerDocument: string) => void | Promise<void>\n}\n\nexport interface ImproveMemoryOptions {\n /** Current durable memory text used as the measured baseline. */\n document: string\n /** Persist the promoted memory document. Never called on hold or error. */\n writeBack?: (winnerDocument: string) => void | Promise<void>\n}\n\nexport interface ImproveCodeOptions {\n /** Repo root candidate worktrees fork from. */\n repoRoot: string\n /** Base ref candidates fork from. Default `main`. */\n baseRef?: string\n /** Directory worktrees are created under. Default `<repoRoot>/.worktrees`. */\n worktreeDir?: string\n /** Git-compatible adapter override, primarily for tests. Candidate advancement\n * still requires normal Git worktree and commit semantics. */\n worktree?: WorktreeAdapter\n /** Coding harness the agentic generator runs in each worktree. Default `claude`. */\n harness?: LocalHarness\n /** Verify a candidate worktree before it becomes a measurable surface; failures\n * feed the next shot (see `agenticGenerator.verify` / `commandVerifier`). */\n verify?: Verifier\n /** Per-shot wall-clock timeout for the harness (ms). */\n timeoutMs?: number\n /** Byte-producer override — the test seam and the escape hatch for custom\n * candidate production. When set, `harness`/`verify`/`timeoutMs` are unused. */\n generator?: CandidateGenerator\n}\n\nexport interface ImproveResult<TScenario extends Scenario, TArtifact> {\n /** The profile after improvement: the winner surface applied back into the\n * matching field when the gate shipped, else the input profile unchanged. */\n profile: AgentProfile\n /** True when `gateDecision === 'ship'`. */\n shipped: boolean\n /** Held-out lift (`winner − baseline` composite). */\n lift: number\n /** The five-valued gate verdict from `selfImprove`. */\n gateDecision: SelfImproveResult<TScenario, TArtifact>['gateDecision']\n /** Full `selfImprove` result for advanced inspection. For code runs,\n * `raw.winner.surface.worktreeRef` remains live after return whether the\n * candidate shipped or held; call `dispose()` after consuming it. */\n raw: SelfImproveResult<TScenario, TArtifact>\n /** Release resources owned by this result. Idempotent; currently disposes\n * the returned code worktree and is a no-op for profile-only surfaces. */\n dispose(): Promise<void>\n}\n\n/** Default model id for the reflective drivers when `llm.model` is unset — a model the Tangle\n * router actually serves (callers should pass their own `llm.model`). */\nconst defaultReflectionModel = 'deepseek-v4-flash'\n\n/** The reflective proposers (`gepaProposer`/`skillOptProposer`) take a full\n * `LlmClientOptions`; `SelfImproveLlm` is the thin user-facing subset. */\nfunction llmClientOptions(llm: SelfImproveLlm | undefined): { baseUrl?: string; apiKey?: string } {\n return { baseUrl: llm?.baseUrl, apiKey: llm?.apiKey }\n}\n\n/** The default proposer for a surface, or `undefined` when the surface has no\n * zero-config default (the caller must supply `opts.generator`). */\nfunction defaultGeneratorFor(\n surface: ImproveSurface,\n llm: SelfImproveLlm | undefined,\n): SurfaceProposer | undefined {\n const model = llm?.model ?? defaultReflectionModel\n switch (surface) {\n case 'prompt':\n return gepaProposer({ llm: llmClientOptions(llm), model, target: 'agent system prompt' })\n case 'skills':\n return skillOptProposer({ llm: llmClientOptions(llm), model, target: 'agent skill document' })\n case 'memory':\n return memoryCurationProposer()\n default:\n return undefined\n }\n}\n\n/** Extract the baseline surface a driver mutates from the profile field that\n * backs `surface`. Prompt, skills, and memory are text surfaces; config\n * surfaces serialize the matching profile record. */\nfunction baselineSurfaceFor(\n profile: AgentProfile,\n surface: ImproveSurface,\n skills?: ImproveSkillsOptions,\n memory?: ImproveMemoryOptions,\n): MutableSurface {\n switch (surface) {\n case 'prompt':\n return profile.prompt?.systemPrompt ?? ''\n case 'skills':\n // With a document supplied, optimize its CONTENT (the real skillOpt path);\n // otherwise fall back to the refs-array surface for back-compat.\n return skills?.document ?? JSON.stringify(profile.resources?.skills ?? [])\n case 'tools':\n return JSON.stringify(profile.tools ?? {})\n case 'mcp':\n return JSON.stringify(profile.mcp ?? {})\n case 'hooks':\n return JSON.stringify(profile.hooks ?? {})\n case 'subagents':\n return JSON.stringify(profile.subagents ?? {})\n case 'agent-profile':\n return canonicalJson(profile)\n case 'memory':\n if (!memory) {\n throw new ConfigError(\"improve(): surface 'memory' requires opts.memory.document\")\n }\n return memory.document\n case 'code':\n throw new ConfigError(\n 'improve(): code requires the isolated baseline created from opts.code.repoRoot',\n )\n }\n}\n\n/** The default `analyzeGeneration`: distill each generation's failing cells into\n * findings for the next proposal round. Deliberately dependency-free — judge notes\n * and errors are already the domain's own diagnosis (executable gates put their\n * reasons there); a trace-analyst can replace this wholesale via\n * `opts.analyzeGeneration`. Falls back to the static seed findings when the\n * generation had no failures, so a clean round never wipes the seed context. */\nfunction generationFailureDistiller<TScenario extends Scenario, TArtifact>(\n staticFindings: unknown[],\n): NonNullable<SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration']> {\n const CAP = 12\n return async (input) => {\n const failures: Array<{\n scenario: string\n composite: number\n notes: string\n claim?: string\n error?: string\n }> = []\n for (const candidate of input.candidates) {\n for (const rawCell of candidate.campaign.cells) {\n const cell = rawCell as unknown as Record<string, unknown>\n const scenario = String(cell.scenarioId ?? 'unknown')\n const error = typeof cell.error === 'string' ? cell.error : undefined\n const judgeScores =\n cell.judgeScores && typeof cell.judgeScores === 'object'\n ? Object.values(\n cell.judgeScores as Record<string, { composite?: number; notes?: string }>,\n )\n : []\n const composite =\n judgeScores.length === 0\n ? 0\n : judgeScores.reduce((sum, j) => sum + (j.composite ?? 0), 0) / judgeScores.length\n if (!error && composite >= 0.999) continue\n const notes = judgeScores\n .map((j) => j.notes)\n .filter((n): n is string => typeof n === 'string' && n.length > 0)\n .join('; ')\n .slice(0, 400)\n const claim = notes || (error ? `Scenario ${scenario} failed: ${error.slice(0, 200)}` : '')\n failures.push({\n scenario,\n composite: Number(composite.toFixed(3)),\n notes,\n ...(claim ? { claim } : {}),\n ...(error ? { error: error.slice(0, 200) } : {}),\n })\n }\n }\n if (failures.length === 0) return staticFindings\n failures.sort((a, b) => a.composite - b.composite)\n return failures.slice(0, CAP)\n }\n}\n\n/** Memory accumulates durable lessons, so keep the caller's seed findings while\n * adding fresh judge failures. Curator proposers consume `claim`; the generic\n * distiller retains the richer diagnostic fields for reflective proposers. */\nfunction memoryGenerationDistiller<TScenario extends Scenario, TArtifact>(\n staticFindings: unknown[],\n): NonNullable<SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration']> {\n const distillFailures = generationFailureDistiller<TScenario, TArtifact>(staticFindings)\n return async (input) => {\n const fresh = await distillFailures(input)\n return fresh === staticFindings ? staticFindings : [...staticFindings, ...fresh]\n }\n}\n\ninterface PreparedCodeRun {\n baseline: CodeSurface\n proposer: SurfaceProposer\n cleanup(retainedWinner?: MutableSurface): Promise<void>\n}\n\n/** Preserve the primary failure while making two best-effort cleanup attempts.\n * A failed first attempt is retained in the thrown AggregateError even when the\n * retry succeeds, so callers can diagnose degraded cleanup without losing the\n * error that caused cleanup to run. */\nasync function rethrowAfterCleanup(\n cause: unknown,\n cleanup: () => Promise<void>,\n message: string,\n): Promise<never> {\n const cleanupErrors: unknown[] = []\n for (let attempt = 0; attempt < 2; attempt += 1) {\n try {\n await cleanup()\n } catch (cleanupCause) {\n cleanupErrors.push(cleanupCause)\n continue\n }\n if (cleanupErrors.length === 0) throw cause\n throw new AggregateError([cause, ...cleanupErrors], `${message}; the cleanup retry succeeded`)\n }\n throw new AggregateError([cause, ...cleanupErrors], message)\n}\n\nasync function discardPreparedBaseline(\n worktree: WorktreeAdapter,\n baselineWorktree: Worktree,\n cause: unknown,\n): Promise<never> {\n return rethrowAfterCleanup(\n cause,\n () => worktree.discard(baselineWorktree),\n 'improve(): code preparation failed and its baseline worktree could not be cleaned',\n )\n}\n\nfunction isCodeSurface(surface: MutableSurface | undefined): surface is CodeSurface {\n return typeof surface === 'object' && surface !== null && surface.kind === 'code'\n}\n\n/** Create a clean incumbent checkout and the candidate producer for a code run. */\nasync function prepareCodeRun(code: ImproveCodeOptions): Promise<PreparedCodeRun> {\n const baseRef = code.baseRef ?? 'main'\n const worktree =\n code.worktree ??\n gitWorktreeAdapter({\n repoRoot: code.repoRoot,\n ...(code.worktreeDir ? { worktreeDir: code.worktreeDir } : {}),\n })\n const baselineWorktree = await worktree.create({ baseRef, label: 'incumbent-baseline' })\n try {\n const baseline = await worktree.finalize(baselineWorktree, 'Incumbent code checkout')\n let baselineDiscarded = false\n const generator =\n code.generator ??\n agenticGenerator({\n ...(code.harness ? { harness: code.harness } : {}),\n ...(code.verify ? { verify: code.verify } : {}),\n ...(code.timeoutMs ? { timeoutMs: code.timeoutMs } : {}),\n })\n const managed: ManagedImprovementDriver = improvementDriver({ worktree, generator, baseRef })\n\n return {\n baseline,\n proposer: managed,\n async cleanup(retainedWinner) {\n const errors: unknown[] = []\n const retainedWorktreeRef = isCodeSurface(retainedWinner)\n ? retainedWinner.worktreeRef\n : undefined\n try {\n await managed?.cleanup(retainedWorktreeRef ? [retainedWorktreeRef] : [])\n } catch (cause) {\n errors.push(cause)\n }\n if (!baselineDiscarded && retainedWorktreeRef !== baseline.worktreeRef) {\n try {\n await worktree.discard(baselineWorktree)\n baselineDiscarded = true\n } catch (cause) {\n errors.push(cause)\n }\n }\n if (errors.length > 0) {\n throw new AggregateError(errors, 'improve(): failed to clean code improvement worktrees')\n }\n },\n }\n } catch (cause) {\n return discardPreparedBaseline(worktree, baselineWorktree, cause)\n }\n}\n\nfunction idempotentDispose(dispose: () => Promise<void>): () => Promise<void> {\n let disposed = false\n let inFlight: Promise<void> | undefined\n return async () => {\n if (disposed) return\n if (inFlight) return inFlight\n inFlight = (async () => {\n await dispose()\n disposed = true\n })()\n try {\n await inFlight\n } finally {\n inFlight = undefined\n }\n }\n}\n\n/** Parse a JSON winner surface (`tools`/`mcp`/`hooks`/`subagents`/`agent-profile`) with a typed,\n * contextual error. A malformed generator output must fail loud here, not throw\n * a raw `SyntaxError` to the caller after a ship verdict. */\nfunction parseWinnerJson<T>(winner: string, surface: ImproveSurface): T {\n try {\n return JSON.parse(winner) as T\n } catch (cause) {\n throw new ConfigError(\n `improve(): the shipped '${surface}' winner is not valid JSON, so it cannot be applied back to the profile: ${\n (cause as Error).message\n }`,\n )\n }\n}\n\n/** Apply a promoted winner surface back into the profile field for `surface`.\n * Returns a shallow copy; never mutates the input profile. */\nexport function applyImprovementWinnerToProfile(\n profile: AgentProfile,\n surface: ImproveSurface,\n winner: MutableSurface,\n): AgentProfile {\n // Only string surfaces map cleanly back onto a profile field. A `CodeSurface`\n // winner (the `code` lever) is a worktree ref, not a profile value — the\n // caller materializes it from `raw.winner.surface`; the returned profile is\n // unchanged for that lever.\n if (typeof winner !== 'string') return profile\n let candidate: AgentProfile\n switch (surface) {\n case 'prompt':\n candidate = { ...profile, prompt: { ...profile.prompt, systemPrompt: winner } }\n break\n case 'skills':\n candidate = {\n ...profile,\n resources: { ...profile.resources, skills: parseWinnerJson(winner, surface) },\n }\n break\n case 'tools':\n candidate = { ...profile, tools: parseWinnerJson(winner, surface) }\n break\n case 'mcp':\n candidate = { ...profile, mcp: parseWinnerJson(winner, surface) }\n break\n case 'hooks':\n candidate = { ...profile, hooks: parseWinnerJson(winner, surface) }\n break\n case 'subagents':\n candidate = { ...profile, subagents: parseWinnerJson(winner, surface) }\n break\n case 'agent-profile':\n candidate = parseWinnerJson(winner, surface)\n break\n case 'memory':\n return profile\n case 'code':\n return profile\n }\n const parsed = agentProfileSchema.safeParse(candidate)\n if (!parsed.success) {\n throw new ConfigError(\n `improve(): the shipped '${surface}' winner does not produce a valid AgentProfile: ${parsed.error.message}`,\n )\n }\n return parsed.data\n}\n\n/**\n * Run the held-out-gated self-improvement loop on ONE profile surface.\n *\n * @example Optimize the system prompt, default holdout gate:\n *\n * const out = await improve(profile, findings, {\n * surface: 'prompt',\n * scenarios,\n * judge,\n * agent: (surface, scenario, ctx) => runAgent(surface, scenario, ctx.signal),\n * })\n * if (out.shipped) deploy(out.profile)\n */\nexport async function improve<TScenario extends Scenario, TArtifact>(\n profile: AgentProfile,\n findings: unknown[],\n opts: ImproveOptions<TScenario, TArtifact>,\n): Promise<ImproveResult<TScenario, TArtifact>> {\n const {\n surface = 'prompt',\n gate = 'holdout',\n generator,\n allowedModels,\n rawTraceContext,\n code,\n skills,\n memory,\n promotionGate,\n analyzeGeneration,\n ...sharedOptions\n } = opts\n\n const parsedProfile = agentProfileSchema.safeParse(profile)\n if (!parsedProfile.success) {\n throw new ConfigError(\n `improve(): input is not a valid AgentProfile: ${parsedProfile.error.message}`,\n )\n }\n if (surface === 'skills' && !generator && !skills) {\n throw new ConfigError(\n 'improve(): the default skills optimizer requires opts.skills.document; pass the skill text or an explicit generator that understands resource refs',\n )\n }\n if (surface === 'memory' && !memory) {\n throw new ConfigError(\"improve(): surface 'memory' requires opts.memory.document\")\n }\n if (surface === 'code' && generator) {\n throw new ConfigError(\n \"improve(): surface 'code' forbids opts.generator because an external SurfaceProposer cannot transfer checkout ownership; pass opts.code.generator instead\",\n )\n }\n const usesReflectionModel = !generator && (surface === 'prompt' || surface === 'skills')\n if (usesReflectionModel) {\n assertModelAllowed(sharedOptions.llm?.model ?? defaultReflectionModel, allowedModels)\n }\n\n let preparedCode: PreparedCodeRun | undefined\n if (surface === 'code') {\n if (!code) {\n throw new ConfigError(\n \"improve(): surface 'code' requires opts.code.repoRoot so the incumbent can run from an isolated checkout\",\n )\n }\n preparedCode = await prepareCodeRun(code)\n }\n const proposer =\n preparedCode?.proposer ?? generator ?? defaultGeneratorFor(surface, sharedOptions.llm)\n if (!proposer) {\n throw new ConfigError(\n `improve(): surface '${surface}' has no default generator — pass opts.generator (a SurfaceProposer) explicitly`,\n )\n }\n\n const budget: SelfImproveBudget =\n gate === 'none' ? { ...sharedOptions.budget, generations: 0 } : { ...sharedOptions.budget }\n\n let raw: SelfImproveResult<TScenario, TArtifact>\n try {\n raw = await selfImprove<TScenario, TArtifact>({\n ...sharedOptions,\n baselineSurface:\n preparedCode?.baseline ?? baselineSurfaceFor(profile, surface, skills, memory),\n proposer,\n budget,\n findings,\n ...(promotionGate !== undefined ? { gate: promotionGate } : {}),\n ...(analyzeGeneration === null\n ? {}\n : {\n analyzeGeneration:\n analyzeGeneration ??\n (rawTraceContext\n ? rawTraceDistiller<TScenario, TArtifact>({ fallbackFindings: findings })\n : surface === 'memory'\n ? memoryGenerationDistiller<TScenario, TArtifact>(findings)\n : generationFailureDistiller<TScenario, TArtifact>(findings)),\n }),\n })\n } catch (cause) {\n if (!preparedCode) throw cause\n return rethrowAfterCleanup(\n cause,\n () => preparedCode.cleanup(),\n 'improve(): code improvement failed and its worktrees could not be cleaned',\n )\n }\n\n const shipped = raw.gateDecision === 'ship'\n const winnerSurface = raw.winner.surface\n if (preparedCode) {\n try {\n await preparedCode.cleanup(winnerSurface)\n } catch (cleanupCause) {\n try {\n await preparedCode.cleanup()\n } catch (finalCleanupCause) {\n throw new AggregateError(\n [cleanupCause, finalCleanupCause],\n 'improve(): code result cleanup failed, including the final all-worktree retry',\n )\n }\n throw new AggregateError(\n [cleanupCause],\n 'improve(): code result cleanup failed; the final all-worktree retry succeeded',\n )\n }\n }\n const dispose = idempotentDispose(async () => preparedCode?.cleanup())\n // When a skill DOCUMENT was optimized, the winner is document text — persist it\n // via writeBack (the profile ref points at the caller's file, unchanged) rather\n // than parsing it as a refs array. Otherwise use the standard field write-back.\n const externalDocument =\n surface === 'skills' && skills ? skills : surface === 'memory' && memory ? memory : undefined\n if (shipped && externalDocument) {\n if (typeof winnerSurface !== 'string') {\n throw new ConfigError(\n `improve(): the shipped '${surface}' winner must be text before it can be persisted`,\n )\n }\n await externalDocument.writeBack?.(winnerSurface)\n }\n const nextProfile =\n shipped && !externalDocument\n ? applyImprovementWinnerToProfile(profile, surface, winnerSurface)\n : profile\n\n return {\n profile: nextProfile,\n shipped,\n lift: raw.lift,\n gateDecision: raw.gateDecision,\n raw,\n dispose,\n }\n}\n"],"mappings":";;;;;;;;;;;AAsBA,SAAS,iBAAiB;AAE1B;AAAA,EAKE;AAAA,OAGK;AAwDA,SAAS,kBAAkB,MAA0D;AAC1F,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,QAAQ,oBAAI,IAAsB;AAExC,SAAO;AAAA,IACL,MAAM,eAAe,KAAK,UAAU,IAAI;AAAA,IACxC,MAAM,QAAQ,KAAqC;AACjD,YAAM,WAAW,gBAAgB,GAAG;AAQpC,UACE,SAAS,WAAW,KACpB,IAAI,WAAW,UACf,CAAC,KAAK,UAAU,yBAChB;AACA,eAAO,CAAC;AAAA,MACV;AAEA,YAAM,WAA0B,CAAC;AACjC,YAAM,YAAY,sBAAsB,IAAI,cAAc;AAC1D,YAAM,kBAAkB,WAAW,cAAc;AACjD,eAAS,IAAI,GAAG,IAAI,IAAI,gBAAgB,KAAK;AAC3C,YAAI,IAAI,OAAO,QAAS;AACxB,cAAM,KAAK,MAAM,KAAK,SAAS,OAAO;AAAA,UACpC,SAAS;AAAA,UACT,OAAO,GAAG,KAAK,UAAU,IAAI,OAAO,IAAI,UAAU,QAAQ,CAAC;AAAA,QAC7D,CAAC;AACD,cAAM,IAAI,GAAG,MAAM,EAAE;AAIrB,YAAI;AACF,cAAI,UAAW,oBAAmB,IAAI,SAAS;AAC/C,gBAAM,EAAE,SAAS,QAAQ,IAAI,MAAM,KAAK,UAAU,SAAS;AAAA,YACzD,cAAc,GAAG;AAAA,YACjB,QAAQ,IAAI;AAAA,YACZ;AAAA,YACA,SAAS,IAAI;AAAA,YACb,UAAU,IAAI,uBAAuB;AAAA,YACrC,QAAQ,IAAI;AAAA,YACZ,YAAY,IAAI;AAAA,YAChB,gBAAgB;AAAA,YAChB,GAAI,IAAI,aAAa,EAAE,YAAY,IAAI,WAAW,IAAI,CAAC;AAAA,YACvD,GAAI,IAAI,YAAY,EAAE,WAAW,IAAI,UAAU,IAAI,CAAC;AAAA,UACtD,CAAC;AACD,cAAI,CAAC,SAAS;AACZ,kBAAM,KAAK,SAAS,QAAQ,EAAE;AAC9B,kBAAM,OAAO,GAAG,IAAI;AACpB;AAAA,UACF;AACA,gBAAM,UAAU,MAAM,KAAK,SAAS,SAAS,IAAI,OAAO;AACxD,mBAAS,KAAK,OAAO;AACrB,gBAAM,OAAO,GAAG,IAAI;AACpB,gBAAM,IAAI,QAAQ,aAAa,EAAE;AAAA,QACnC,SAAS,KAAK;AACZ,gBAAM,gBAA2B,CAAC;AAClC,mBAAS,UAAU,GAAG,UAAU,GAAG,WAAW,GAAG;AAC/C,gBAAI;AACF,oBAAM,KAAK,SAAS,QAAQ,EAAE;AAC9B,oBAAM,OAAO,GAAG,IAAI;AACpB;AAAA,YACF,SAAS,OAAO;AACd,4BAAc,KAAK,KAAK;AAAA,YAC1B;AAAA,UACF;AACA,cAAI,cAAc,WAAW,EAAG,OAAM;AACtC,gBAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,gBAAM,mBAAmB,CAAC,MAAM,IAAI,GAAG,IAAI;AAC3C,gBAAM,IAAI;AAAA,YACR,CAAC,KAAK,GAAG,aAAa;AAAA,YACtB,mBACI,sBAAsB,OAAO,wCAC7B,sBAAsB,OAAO;AAAA,UACnC;AAAA,QACF;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,IACA,MAAM,QAAQ,qBAAqB,CAAC,GAAG;AACrC,YAAM,WAAW,IAAI,IAAI,kBAAkB;AAC3C,YAAM,SAAoB,CAAC;AAC3B,iBAAW,CAAC,aAAa,QAAQ,KAAK,OAAO;AAC3C,YAAI,SAAS,IAAI,WAAW,EAAG;AAC/B,YAAI;AACF,gBAAM,KAAK,SAAS,QAAQ,QAAQ;AACpC,gBAAM,OAAO,WAAW;AAAA,QAC1B,SAAS,OAAO;AACd,iBAAO,KAAK,KAAK;AAAA,QACnB;AAAA,MACF;AACA,UAAI,OAAO,SAAS,GAAG;AACrB,cAAM,IAAI,eAAe,QAAQ,0DAA0D;AAAA,MAC7F;AAAA,IACF;AAAA,EACF;AACF;AAGA,SAAS,sBAAsB,SAAoE;AACjG,MAAI,OAAO,YAAY,YAAY,QAAQ,SAAS,OAAQ,QAAO;AACnE,oBAAkB,OAAO;AACzB,SAAO;AACT;AAMA,SAAS,mBAAmB,UAAoB,WAA8B;AAC5E,MAAI,SAAS,eAAe,UAAU,cAAc,SAAS,aAAa,UAAU,UAAU;AAC5F,UAAM,IAAI,MAAM,8EAA8E;AAAA,EAChG;AACA,MAAI,SAAS,eAAe,UAAU,gBAAiB;AAEvD,QAAM,QAAQ,UAAU,OAAO,CAAC,SAAS,aAAa,UAAU,eAAe,GAAG;AAAA,IAChF,KAAK,SAAS;AAAA,IACd,UAAU;AAAA,EACZ,CAAC;AACD,MAAI,MAAM,OAAO;AACf,UAAM,IAAI;AAAA,MACR,gEAAgE,MAAM,MAAM,OAAO;AAAA,IACrF;AAAA,EACF;AACA,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI;AAAA,MACR,oEAAoE,UAAU,eAAe,KAAK,MAAM,OAAO,KAAK,CAAC;AAAA,IACvH;AAAA,EACF;AAEA,QAAM,OAAO,UAAU,OAAO,CAAC,aAAa,YAAY,MAAM,GAAG;AAAA,IAC/D,KAAK,SAAS;AAAA,IACd,UAAU;AAAA,EACZ,CAAC;AACD,MAAI,KAAK,SAAS,KAAK,WAAW,KAAK,KAAK,OAAO,KAAK,MAAM,UAAU,iBAAiB;AACvF,UAAM,IAAI,MAAM,0EAA0E;AAAA,EAC5F;AACF;AAKA,SAAS,gBAAgB,KAAuD;AAC9E,QAAM,SAAS,IAAI;AACnB,MAAI,UAAU,OAAO,WAAW,YAAY,cAAc,QAAQ;AAChE,UAAM,IAAK,OAAiC;AAC5C,QAAI,MAAM,QAAQ,CAAC,KAAK,EAAE,SAAS,EAAG,QAAO;AAAA,EAC/C;AACA,SAAO,IAAI;AACb;;;AChNA,SAAsB,YAAY,mBAAmB;AACrD,SAAS,UAAU,MAAM,eAAe;AACxC,SAA8B,mBAAmB;AAGjD,IAAM,aAAa;AAGnB,IAAM,iBAAiB;AA8ChB,SAAS,kBACd,UAAoC,CAAC,GACuC;AAC5E,QAAM,gBAAgB,QAAQ,iBAAiB;AAC/C,QAAM,uBAAuB,QAAQ,wBAAwB;AAC7D,QAAM,kBAAkB,QAAQ,mBAAmB;AAEnD,SAAO,OAAO,UAAU;AACtB,UAAM,UAAU,eAAe,QAAQ,UAAU,MAAM,MAAM;AAC7D,UAAM,UAAU,UAAU,OAAO;AAIjC,UAAM,SAAS,CAAC,GAAG,MAAM,UAAU,EAChC,IAAI,CAAC,OAAO;AAAA,MACX,aAAa,EAAE;AAAA,MACf,WAAW,EAAE;AAAA,MACb,aAAa,eAAe,EAAE,SAAS,MAAM;AAAA,MAC7C,OAAO,aAAa,EAAE,UAAU,sBAAsB,eAAe;AAAA,IACvE,EAAE,EACD,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS,EACxC,MAAM,GAAG,aAAa;AAEzB,UAAM,oBAAoB,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,MAAM,QAAQ,CAAC;AAUvE,QAAI,sBAAsB,GAAG;AAC3B,UAAI,QAAQ,oBAAoB,QAAQ,iBAAiB,SAAS,GAAG;AACnE,eAAO,QAAQ;AAAA,MACjB;AACA,aAAO;AAAA,QACL,YAAY;AAAA,UACV,YAAY;AAAA,UACZ,UAAU;AAAA,UACV,MAAM;AAAA,UACN,YAAY;AAAA,UACZ,OAAO,cAAc,MAAM,UAAU,oEAAoE,OAAO;AAAA,UAChH,oBAAoB,oDAAoD,OAAO;AAAA,UAC/E,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,QAAQ,CAAC;AAAA,UAClD,UAAU,EAAE,YAAY,MAAM,YAAY,QAAQ,SAAS,cAAc,EAAE;AAAA,QAC7E,CAAC;AAAA,MACH;AAAA,IACF;AAEA,UAAM,WAA6B,CAAC;AAGpC,aAAS;AAAA,MACP,YAAY;AAAA,QACV,YAAY;AAAA,QACZ,UAAU;AAAA,QACV,MAAM;AAAA,QACN,YAAY;AAAA,QACZ,OAAO,cAAc,MAAM,UAAU,aAAa,iBAAiB,uCAAuC,OAAO,MAAM,8DAA8D,OAAO,0GAC1L,UACI,KACA,mJACN;AAAA,QACA,oBAAoB,oWAAoW,OAAO;AAAA,QAC/X,eAAe,CAAC,EAAE,MAAM,YAAY,KAAK,QAAQ,CAAC;AAAA,QAClD,UAAU;AAAA,UACR,YAAY,MAAM;AAAA,UAClB,QAAQ;AAAA,UACR,cAAc;AAAA,UACd,YAAY,OAAO;AAAA,QACrB;AAAA,MACF,CAAC;AAAA,IACH;AAIA,eAAW,QAAQ,QAAQ;AACzB,UAAI,KAAK,MAAM,WAAW,EAAG;AAC7B,YAAM,eAAe,KAAK,MAAM,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,KAAK,IAAI;AAClE,YAAM,YAAY,KAAK,MACpB,IAAI,CAAC,MAAM;AACV,cAAM,SAAS,UAAU,EAAE,UAAU,eAAe,EAAE,UAAU,QAAQ,CAAC,CAAC,GACxE,EAAE,QAAQ,YAAY,SAAS,EAAE,OAAO,GAAG,CAAC,KAAK,EACnD,gBAAW,EAAE,OAAO;AACpB,cAAM,QAAQ,EAAE,MAAM,IAAI,CAAC,MAAM,SAAS,CAAC,EAAE,EAAE,KAAK,IAAI;AACxD,cAAM,OAAO,EAAE,iBAAiB;AAAA,kBAAgB,EAAE,OAAO,mBAAmB;AAC5E,eAAO,EAAE,MAAM,SAAS,IAAI,GAAG,MAAM;AAAA,EAAK,KAAK,GAAG,IAAI,KAAK;AAAA,MAC7D,CAAC,EACA,KAAK,IAAI;AAEZ,eAAS;AAAA,QACP,YAAY;AAAA,UACV,YAAY;AAAA,UACZ,UAAU,KAAK,YAAY,MAAM,aAAa;AAAA,UAC9C,MAAM;AAAA,UACN,YAAY;AAAA,UACZ,SAAS,KAAK;AAAA,UACd,OAAO,aAAa,KAAK,WAAW,qBAAqB,KAAK,UAAU,QAAQ,CAAC,CAAC,SAAS,KAAK,MAAM,MAAM,qBAAqB,YAAY,+BAA+B,KAAK,WAAW;AAAA,UAC5L,oBAAoB;AAAA,EAAyF,SAAS;AAAA,qDAAwD,KAAK,WAAW,kBAAkB,KAAK,WAAW;AAAA,UAChO,eAAe;AAAA,YACb,EAAE,MAAM,YAAY,KAAK,KAAK,YAAY;AAAA,YAC1C,GAAG,KAAK,MAAM;AAAA,cAAQ,CAAC,MACrB,EAAE,MAAM,IAAI,CAAC,OAAO,EAAE,MAAM,YAAqB,KAAK,EAAE,EAAE;AAAA,YAC5D;AAAA,UACF;AAAA,UACA,UAAU;AAAA,YACR,aAAa,KAAK;AAAA,YAClB,WAAW,KAAK;AAAA,YAChB,aAAa,KAAK;AAAA,YAClB,OAAO,KAAK,MAAM,IAAI,CAAC,OAAO;AAAA,cAC5B,YAAY,EAAE;AAAA,cACd,WAAW,EAAE;AAAA,cACb,SAAS,EAAE;AAAA,cACX,OAAO,EAAE;AAAA,cACT,GAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,IAAI,CAAC;AAAA,YACtC,EAAE;AAAA,UACJ;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AACF;AAKA,SAAS,aACP,UAUA,UACA,UACa;AACb,QAAM,cAAc,eAAe,SAAS,MAAM;AAClD,QAAM,UAAU,UAAU,WAAW;AACrC,QAAM,MAAmB,CAAC;AAC1B,aAAW,QAAQ,SAAS,OAAO;AACjC,UAAM,SAAS,OAAO,OAAO,KAAK,eAAe,CAAC,CAAC;AACnD,UAAM,YACJ,OAAO,WAAW,IACd,IACA,OAAO,OAAO,CAAC,KAAK,MAAM,OAAO,EAAE,aAAa,IAAI,CAAC,IAAI,OAAO;AACtE,QAAI,CAAC,KAAK,SAAS,aAAa,eAAgB;AAEhD,UAAM,UAAU,KAAK,aAAa,eAAe,KAAK,MAAM,CAAC;AAC7D,UAAM,gBAAgB,qBAAqB,SAAS,iBAAiB,KAAK,MAAM;AAChF,UAAM,aAAa,UAAU,eAAe,OAAO,IAAI,CAAC;AAGxD,UAAM,YAAY,CAAC,KAAK,SAAS,aAAa,GAAG,KAAK,SAAS,oBAAoB,CAAC;AACpF,UAAM,QAAQ,aAAa,CAAC,GAAG,YAAY,GAAG,eAAe,GAAG,SAAS,CAAC;AAE1E,QAAI,KAAK;AAAA,MACP,YAAY,KAAK;AAAA,MACjB,WAAW,OAAO,UAAU,QAAQ,CAAC,CAAC;AAAA,MACtC,GAAI,KAAK,QAAQ,EAAE,OAAO,KAAK,MAAM,IAAI,CAAC;AAAA,MAC1C;AAAA,MACA,OAAO,MAAM,MAAM,GAAG,QAAQ;AAAA,MAC9B,gBAAgB,MAAM,SAAS;AAAA,IACjC,CAAC;AACD,QAAI,IAAI,UAAU,SAAU;AAAA,EAC9B;AACA,SAAO;AACT;AAIA,SAAS,qBACP,iBACA,QACU;AACV,MAAI,CAAC,gBAAiB,QAAO,CAAC;AAC9B,QAAM,SAAS,GAAG,MAAM;AACxB,SAAO,OAAO,QAAQ,eAAe,EAClC,OAAO,CAAC,CAAC,GAAG,MAAM,IAAI,WAAW,MAAM,CAAC,EACxC,IAAI,CAAC,CAAC,EAAE,OAAO,MAAM,QAAQ,OAAO,CAAC;AAC1C;AAMA,SAAS,eAAe,KAAuB;AAC7C,QAAM,MAAgB,CAAC;AACvB,aAAW,SAAS,YAAY,GAAG,GAAG;AACpC,UAAM,OAAO,KAAK,KAAK,MAAM,IAAI;AACjC,QAAI,MAAM,OAAO,GAAG;AAClB,UAAI,KAAK,IAAI;AAAA,IACf,WAAW,CAAC,MAAM,eAAe,KAAK,MAAM,YAAY,GAAG;AACzD,iBAAW,OAAO,YAAY,IAAI,GAAG;AACnC,YAAI,IAAI,OAAO,EAAG,KAAI,KAAK,KAAK,MAAM,IAAI,IAAI,CAAC;AAAA,MACjD;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,YAAY,KAAuB;AAC1C,MAAI;AACF,WAAO,YAAY,KAAK,EAAE,eAAe,KAAK,CAAC;AAAA,EACjD,QAAQ;AACN,WAAO,CAAC;AAAA,EACV;AACF;AAIA,SAAS,eAAe,QAAwB;AAC9C,SAAO,OAAO,QAAQ,mBAAmB,GAAG;AAC9C;AAIA,SAAS,UAAU,QAAyB;AAC1C,SAAO,CAAC,OAAO,WAAW,QAAQ,KAAK,WAAW,MAAM;AAC1D;AAIA,SAAS,eAAe,QAAwB;AAC9C,SAAO,OAAO,WAAW,QAAQ,IAAI,SAAS,QAAQ,MAAM;AAC9D;AAEA,SAAS,aAAa,OAA2B;AAC/C,SAAO,CAAC,GAAG,IAAI,IAAI,KAAK,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM;AAExC,UAAM,KAAK,EAAE,MAAM,GAAG,EAAE,SAAS,SAAS,CAAC,EAAE,MAAM;AACnD,UAAM,KAAK,EAAE,MAAM,GAAG,EAAE,SAAS,SAAS,CAAC,EAAE,MAAM;AACnD,WAAO,OAAO,KAAK,SAAS,CAAC,EAAE,cAAc,SAAS,CAAC,CAAC,IAAI,GAAG,cAAc,EAAE;AAAA,EACjF,CAAC;AACH;AAEA,SAAS,SAAS,GAAW,GAAmB;AAC9C,SAAO,EAAE,UAAU,IAAI,IAAI,GAAG,EAAE,MAAM,GAAG,IAAI,CAAC,CAAC;AACjD;;;AC3SA,SAAS,qBAAqB;AAC9B;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OAGK;AACP;AAAA,EASE;AAAA,OACK;AACP,SAA4B,0BAA0B;AAkJtD,IAAM,yBAAyB;AAI/B,SAAS,iBAAiB,KAAwE;AAChG,SAAO,EAAE,SAAS,KAAK,SAAS,QAAQ,KAAK,OAAO;AACtD;AAIA,SAAS,oBACP,SACA,KAC6B;AAC7B,QAAM,QAAQ,KAAK,SAAS;AAC5B,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,aAAO,aAAa,EAAE,KAAK,iBAAiB,GAAG,GAAG,OAAO,QAAQ,sBAAsB,CAAC;AAAA,IAC1F,KAAK;AACH,aAAO,iBAAiB,EAAE,KAAK,iBAAiB,GAAG,GAAG,OAAO,QAAQ,uBAAuB,CAAC;AAAA,IAC/F,KAAK;AACH,aAAO,uBAAuB;AAAA,IAChC;AACE,aAAO;AAAA,EACX;AACF;AAKA,SAAS,mBACP,SACA,SACA,QACA,QACgB;AAChB,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,aAAO,QAAQ,QAAQ,gBAAgB;AAAA,IACzC,KAAK;AAGH,aAAO,QAAQ,YAAY,KAAK,UAAU,QAAQ,WAAW,UAAU,CAAC,CAAC;AAAA,IAC3E,KAAK;AACH,aAAO,KAAK,UAAU,QAAQ,SAAS,CAAC,CAAC;AAAA,IAC3C,KAAK;AACH,aAAO,KAAK,UAAU,QAAQ,OAAO,CAAC,CAAC;AAAA,IACzC,KAAK;AACH,aAAO,KAAK,UAAU,QAAQ,SAAS,CAAC,CAAC;AAAA,IAC3C,KAAK;AACH,aAAO,KAAK,UAAU,QAAQ,aAAa,CAAC,CAAC;AAAA,IAC/C,KAAK;AACH,aAAO,cAAc,OAAO;AAAA,IAC9B,KAAK;AACH,UAAI,CAAC,QAAQ;AACX,cAAM,IAAI,YAAY,2DAA2D;AAAA,MACnF;AACA,aAAO,OAAO;AAAA,IAChB,KAAK;AACH,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,EACJ;AACF;AAQA,SAAS,2BACP,gBAC4E;AAC5E,QAAM,MAAM;AACZ,SAAO,OAAO,UAAU;AACtB,UAAM,WAMD,CAAC;AACN,eAAW,aAAa,MAAM,YAAY;AACxC,iBAAW,WAAW,UAAU,SAAS,OAAO;AAC9C,cAAM,OAAO;AACb,cAAM,WAAW,OAAO,KAAK,cAAc,SAAS;AACpD,cAAM,QAAQ,OAAO,KAAK,UAAU,WAAW,KAAK,QAAQ;AAC5D,cAAM,cACJ,KAAK,eAAe,OAAO,KAAK,gBAAgB,WAC5C,OAAO;AAAA,UACL,KAAK;AAAA,QACP,IACA,CAAC;AACP,cAAM,YACJ,YAAY,WAAW,IACnB,IACA,YAAY,OAAO,CAAC,KAAK,MAAM,OAAO,EAAE,aAAa,IAAI,CAAC,IAAI,YAAY;AAChF,YAAI,CAAC,SAAS,aAAa,MAAO;AAClC,cAAM,QAAQ,YACX,IAAI,CAAC,MAAM,EAAE,KAAK,EAClB,OAAO,CAAC,MAAmB,OAAO,MAAM,YAAY,EAAE,SAAS,CAAC,EAChE,KAAK,IAAI,EACT,MAAM,GAAG,GAAG;AACf,cAAM,QAAQ,UAAU,QAAQ,YAAY,QAAQ,YAAY,MAAM,MAAM,GAAG,GAAG,CAAC,KAAK;AACxF,iBAAS,KAAK;AAAA,UACZ;AAAA,UACA,WAAW,OAAO,UAAU,QAAQ,CAAC,CAAC;AAAA,UACtC;AAAA,UACA,GAAI,QAAQ,EAAE,MAAM,IAAI,CAAC;AAAA,UACzB,GAAI,QAAQ,EAAE,OAAO,MAAM,MAAM,GAAG,GAAG,EAAE,IAAI,CAAC;AAAA,QAChD,CAAC;AAAA,MACH;AAAA,IACF;AACA,QAAI,SAAS,WAAW,EAAG,QAAO;AAClC,aAAS,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AACjD,WAAO,SAAS,MAAM,GAAG,GAAG;AAAA,EAC9B;AACF;AAKA,SAAS,0BACP,gBAC4E;AAC5E,QAAM,kBAAkB,2BAAiD,cAAc;AACvF,SAAO,OAAO,UAAU;AACtB,UAAM,QAAQ,MAAM,gBAAgB,KAAK;AACzC,WAAO,UAAU,iBAAiB,iBAAiB,CAAC,GAAG,gBAAgB,GAAG,KAAK;AAAA,EACjF;AACF;AAYA,eAAe,oBACb,OACA,SACA,SACgB;AAChB,QAAM,gBAA2B,CAAC;AAClC,WAAS,UAAU,GAAG,UAAU,GAAG,WAAW,GAAG;AAC/C,QAAI;AACF,YAAM,QAAQ;AAAA,IAChB,SAAS,cAAc;AACrB,oBAAc,KAAK,YAAY;AAC/B;AAAA,IACF;AACA,QAAI,cAAc,WAAW,EAAG,OAAM;AACtC,UAAM,IAAI,eAAe,CAAC,OAAO,GAAG,aAAa,GAAG,GAAG,OAAO,+BAA+B;AAAA,EAC/F;AACA,QAAM,IAAI,eAAe,CAAC,OAAO,GAAG,aAAa,GAAG,OAAO;AAC7D;AAEA,eAAe,wBACb,UACA,kBACA,OACgB;AAChB,SAAO;AAAA,IACL;AAAA,IACA,MAAM,SAAS,QAAQ,gBAAgB;AAAA,IACvC;AAAA,EACF;AACF;AAEA,SAAS,cAAc,SAA6D;AAClF,SAAO,OAAO,YAAY,YAAY,YAAY,QAAQ,QAAQ,SAAS;AAC7E;AAGA,eAAe,eAAe,MAAoD;AAChF,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,WACJ,KAAK,YACL,mBAAmB;AAAA,IACjB,UAAU,KAAK;AAAA,IACf,GAAI,KAAK,cAAc,EAAE,aAAa,KAAK,YAAY,IAAI,CAAC;AAAA,EAC9D,CAAC;AACH,QAAM,mBAAmB,MAAM,SAAS,OAAO,EAAE,SAAS,OAAO,qBAAqB,CAAC;AACvF,MAAI;AACF,UAAM,WAAW,MAAM,SAAS,SAAS,kBAAkB,yBAAyB;AACpF,QAAI,oBAAoB;AACxB,UAAM,YACJ,KAAK,aACL,iBAAiB;AAAA,MACf,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;AAAA,MAChD,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;AAAA,MAC7C,GAAI,KAAK,YAAY,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,IACxD,CAAC;AACH,UAAM,UAAoC,kBAAkB,EAAE,UAAU,WAAW,QAAQ,CAAC;AAE5F,WAAO;AAAA,MACL;AAAA,MACA,UAAU;AAAA,MACV,MAAM,QAAQ,gBAAgB;AAC5B,cAAM,SAAoB,CAAC;AAC3B,cAAM,sBAAsB,cAAc,cAAc,IACpD,eAAe,cACf;AACJ,YAAI;AACF,gBAAM,SAAS,QAAQ,sBAAsB,CAAC,mBAAmB,IAAI,CAAC,CAAC;AAAA,QACzE,SAAS,OAAO;AACd,iBAAO,KAAK,KAAK;AAAA,QACnB;AACA,YAAI,CAAC,qBAAqB,wBAAwB,SAAS,aAAa;AACtE,cAAI;AACF,kBAAM,SAAS,QAAQ,gBAAgB;AACvC,gCAAoB;AAAA,UACtB,SAAS,OAAO;AACd,mBAAO,KAAK,KAAK;AAAA,UACnB;AAAA,QACF;AACA,YAAI,OAAO,SAAS,GAAG;AACrB,gBAAM,IAAI,eAAe,QAAQ,uDAAuD;AAAA,QAC1F;AAAA,MACF;AAAA,IACF;AAAA,EACF,SAAS,OAAO;AACd,WAAO,wBAAwB,UAAU,kBAAkB,KAAK;AAAA,EAClE;AACF;AAEA,SAAS,kBAAkB,SAAmD;AAC5E,MAAI,WAAW;AACf,MAAI;AACJ,SAAO,YAAY;AACjB,QAAI,SAAU;AACd,QAAI,SAAU,QAAO;AACrB,gBAAY,YAAY;AACtB,YAAM,QAAQ;AACd,iBAAW;AAAA,IACb,GAAG;AACH,QAAI;AACF,YAAM;AAAA,IACR,UAAE;AACA,iBAAW;AAAA,IACb;AAAA,EACF;AACF;AAKA,SAAS,gBAAmB,QAAgB,SAA4B;AACtE,MAAI;AACF,WAAO,KAAK,MAAM,MAAM;AAAA,EAC1B,SAAS,OAAO;AACd,UAAM,IAAI;AAAA,MACR,2BAA2B,OAAO,4EAC/B,MAAgB,OACnB;AAAA,IACF;AAAA,EACF;AACF;AAIO,SAAS,gCACd,SACA,SACA,QACc;AAKd,MAAI,OAAO,WAAW,SAAU,QAAO;AACvC,MAAI;AACJ,UAAQ,SAAS;AAAA,IACf,KAAK;AACH,kBAAY,EAAE,GAAG,SAAS,QAAQ,EAAE,GAAG,QAAQ,QAAQ,cAAc,OAAO,EAAE;AAC9E;AAAA,IACF,KAAK;AACH,kBAAY;AAAA,QACV,GAAG;AAAA,QACH,WAAW,EAAE,GAAG,QAAQ,WAAW,QAAQ,gBAAgB,QAAQ,OAAO,EAAE;AAAA,MAC9E;AACA;AAAA,IACF,KAAK;AACH,kBAAY,EAAE,GAAG,SAAS,OAAO,gBAAgB,QAAQ,OAAO,EAAE;AAClE;AAAA,IACF,KAAK;AACH,kBAAY,EAAE,GAAG,SAAS,KAAK,gBAAgB,QAAQ,OAAO,EAAE;AAChE;AAAA,IACF,KAAK;AACH,kBAAY,EAAE,GAAG,SAAS,OAAO,gBAAgB,QAAQ,OAAO,EAAE;AAClE;AAAA,IACF,KAAK;AACH,kBAAY,EAAE,GAAG,SAAS,WAAW,gBAAgB,QAAQ,OAAO,EAAE;AACtE;AAAA,IACF,KAAK;AACH,kBAAY,gBAAgB,QAAQ,OAAO;AAC3C;AAAA,IACF,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,EACX;AACA,QAAM,SAAS,mBAAmB,UAAU,SAAS;AACrD,MAAI,CAAC,OAAO,SAAS;AACnB,UAAM,IAAI;AAAA,MACR,2BAA2B,OAAO,mDAAmD,OAAO,MAAM,OAAO;AAAA,IAC3G;AAAA,EACF;AACA,SAAO,OAAO;AAChB;AAeA,eAAsB,QACpB,SACA,UACA,MAC8C;AAC9C,QAAM;AAAA,IACJ,UAAU;AAAA,IACV,OAAO;AAAA,IACP;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG;AAAA,EACL,IAAI;AAEJ,QAAM,gBAAgB,mBAAmB,UAAU,OAAO;AAC1D,MAAI,CAAC,cAAc,SAAS;AAC1B,UAAM,IAAI;AAAA,MACR,iDAAiD,cAAc,MAAM,OAAO;AAAA,IAC9E;AAAA,EACF;AACA,MAAI,YAAY,YAAY,CAAC,aAAa,CAAC,QAAQ;AACjD,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,YAAY,YAAY,CAAC,QAAQ;AACnC,UAAM,IAAI,YAAY,2DAA2D;AAAA,EACnF;AACA,MAAI,YAAY,UAAU,WAAW;AACnC,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,QAAM,sBAAsB,CAAC,cAAc,YAAY,YAAY,YAAY;AAC/E,MAAI,qBAAqB;AACvB,uBAAmB,cAAc,KAAK,SAAS,wBAAwB,aAAa;AAAA,EACtF;AAEA,MAAI;AACJ,MAAI,YAAY,QAAQ;AACtB,QAAI,CAAC,MAAM;AACT,YAAM,IAAI;AAAA,QACR;AAAA,MACF;AAAA,IACF;AACA,mBAAe,MAAM,eAAe,IAAI;AAAA,EAC1C;AACA,QAAM,WACJ,cAAc,YAAY,aAAa,oBAAoB,SAAS,cAAc,GAAG;AACvF,MAAI,CAAC,UAAU;AACb,UAAM,IAAI;AAAA,MACR,uBAAuB,OAAO;AAAA,IAChC;AAAA,EACF;AAEA,QAAM,SACJ,SAAS,SAAS,EAAE,GAAG,cAAc,QAAQ,aAAa,EAAE,IAAI,EAAE,GAAG,cAAc,OAAO;AAE5F,MAAI;AACJ,MAAI;AACF,UAAM,MAAM,YAAkC;AAAA,MAC5C,GAAG;AAAA,MACH,iBACE,cAAc,YAAY,mBAAmB,SAAS,SAAS,QAAQ,MAAM;AAAA,MAC/E;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAI,kBAAkB,SAAY,EAAE,MAAM,cAAc,IAAI,CAAC;AAAA,MAC7D,GAAI,sBAAsB,OACtB,CAAC,IACD;AAAA,QACE,mBACE,sBACC,kBACG,kBAAwC,EAAE,kBAAkB,SAAS,CAAC,IACtE,YAAY,WACV,0BAAgD,QAAQ,IACxD,2BAAiD,QAAQ;AAAA,MACnE;AAAA,IACN,CAAC;AAAA,EACH,SAAS,OAAO;AACd,QAAI,CAAC,aAAc,OAAM;AACzB,WAAO;AAAA,MACL;AAAA,MACA,MAAM,aAAa,QAAQ;AAAA,MAC3B;AAAA,IACF;AAAA,EACF;AAEA,QAAM,UAAU,IAAI,iBAAiB;AACrC,QAAM,gBAAgB,IAAI,OAAO;AACjC,MAAI,cAAc;AAChB,QAAI;AACF,YAAM,aAAa,QAAQ,aAAa;AAAA,IAC1C,SAAS,cAAc;AACrB,UAAI;AACF,cAAM,aAAa,QAAQ;AAAA,MAC7B,SAAS,mBAAmB;AAC1B,cAAM,IAAI;AAAA,UACR,CAAC,cAAc,iBAAiB;AAAA,UAChC;AAAA,QACF;AAAA,MACF;AACA,YAAM,IAAI;AAAA,QACR,CAAC,YAAY;AAAA,QACb;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACA,QAAM,UAAU,kBAAkB,YAAY,cAAc,QAAQ,CAAC;AAIrE,QAAM,mBACJ,YAAY,YAAY,SAAS,SAAS,YAAY,YAAY,SAAS,SAAS;AACtF,MAAI,WAAW,kBAAkB;AAC/B,QAAI,OAAO,kBAAkB,UAAU;AACrC,YAAM,IAAI;AAAA,QACR,2BAA2B,OAAO;AAAA,MACpC;AAAA,IACF;AACA,UAAM,iBAAiB,YAAY,aAAa;AAAA,EAClD;AACA,QAAM,cACJ,WAAW,CAAC,mBACR,gCAAgC,SAAS,SAAS,aAAa,IAC/D;AAEN,SAAO;AAAA,IACL,SAAS;AAAA,IACT;AAAA,IACA,MAAM,IAAI;AAAA,IACV,cAAc,IAAI;AAAA,IAClB;AAAA,IACA;AAAA,EACF;AACF;","names":[]}
|
|
@@ -255,7 +255,7 @@ function harnessInvocation(harness, profile, taskPrompt, options = {}) {
|
|
|
255
255
|
if (reasoningEffort !== void 0 && invocation.reasoningArgs) {
|
|
256
256
|
args.push(...invocation.reasoningArgs(reasoningEffort));
|
|
257
257
|
}
|
|
258
|
-
return { command: invocation.command, args };
|
|
258
|
+
return { command: invocation.command, args, prompt: composedPrompt };
|
|
259
259
|
}
|
|
260
260
|
var DEFAULT_TIMEOUT_MS = 5 * 60 * 1e3;
|
|
261
261
|
var processKillGraceMs = 250;
|
|
@@ -280,7 +280,7 @@ async function runLocalHarness(options) {
|
|
|
280
280
|
const requestedCommand = options.invocation?.command ?? invocation.command;
|
|
281
281
|
const args = options.invocation ? [...options.invocation.args] : buildHarnessArgs(harness, taskPrompt, options);
|
|
282
282
|
if (options.codexReproducible) assertCodexReproducibleInvocation(requestedCommand, args);
|
|
283
|
-
const baseEnv = options.env ?? process.env;
|
|
283
|
+
const baseEnv = { ...options.env ?? process.env, PWD: cwd };
|
|
284
284
|
const isolated = options.codexReproducible ? await isolateCodexHome({
|
|
285
285
|
baseEnv,
|
|
286
286
|
cwd,
|
|
@@ -1097,4 +1097,4 @@ export {
|
|
|
1097
1097
|
runLocalHarness,
|
|
1098
1098
|
parseCodexTokenUsage
|
|
1099
1099
|
};
|
|
1100
|
-
//# sourceMappingURL=chunk-
|
|
1100
|
+
//# sourceMappingURL=chunk-EP6RVHMX.js.map
|