@tangle-network/agent-bench 0.11.2 → 0.13.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (174) hide show
  1. package/CHANGELOG.md +28 -0
  2. package/HARNESS.md +6 -2
  3. package/README.md +1 -4
  4. package/dist/benchmarks/swe-bench.js +4 -9
  5. package/dist/benchmarks/swe-bench.js.map +1 -1
  6. package/package.json +5 -5
  7. package/scripts/run-package-tests.mjs +2 -2
  8. package/src/benchmarks/swe-bench.test.mts +49 -0
  9. package/src/benchmarks/swe-bench.ts +4 -9
  10. package/src/quant-arena/README.md +0 -144
  11. package/src/quant-arena/backtest.test.mts +0 -135
  12. package/src/quant-arena/backtest.ts +0 -218
  13. package/src/quant-arena/data.test.mts +0 -44
  14. package/src/quant-arena/data.ts +0 -141
  15. package/src/quant-arena/driver.test.mts +0 -253
  16. package/src/quant-arena/driver.ts +0 -219
  17. package/src/quant-arena/fixtures/data/PROVENANCE.md +0 -26
  18. package/src/quant-arena/fixtures/data/holdout/IDX.csv +0 -523
  19. package/src/quant-arena/fixtures/data/holdout/S01.csv +0 -523
  20. package/src/quant-arena/fixtures/data/holdout/S02.csv +0 -523
  21. package/src/quant-arena/fixtures/data/holdout/S03.csv +0 -523
  22. package/src/quant-arena/fixtures/data/holdout/S04.csv +0 -523
  23. package/src/quant-arena/fixtures/data/holdout/S05.csv +0 -523
  24. package/src/quant-arena/fixtures/data/holdout/S06.csv +0 -523
  25. package/src/quant-arena/fixtures/data/holdout/S07.csv +0 -523
  26. package/src/quant-arena/fixtures/data/holdout/S08.csv +0 -523
  27. package/src/quant-arena/fixtures/data/holdout/S09.csv +0 -523
  28. package/src/quant-arena/fixtures/data/holdout/S10.csv +0 -523
  29. package/src/quant-arena/fixtures/data/insample/IDX.csv +0 -2087
  30. package/src/quant-arena/fixtures/data/insample/S01.csv +0 -2087
  31. package/src/quant-arena/fixtures/data/insample/S02.csv +0 -2087
  32. package/src/quant-arena/fixtures/data/insample/S03.csv +0 -2087
  33. package/src/quant-arena/fixtures/data/insample/S04.csv +0 -2087
  34. package/src/quant-arena/fixtures/data/insample/S05.csv +0 -2087
  35. package/src/quant-arena/fixtures/data/insample/S06.csv +0 -2087
  36. package/src/quant-arena/fixtures/data/insample/S07.csv +0 -2087
  37. package/src/quant-arena/fixtures/data/insample/S08.csv +0 -2087
  38. package/src/quant-arena/fixtures/data/insample/S09.csv +0 -2087
  39. package/src/quant-arena/fixtures/data/insample/S10.csv +0 -2087
  40. package/src/quant-arena/fixtures/demo-campaign/cost-ledger.jsonl +0 -16
  41. package/src/quant-arena/fixtures/demo-campaign/notebook.jsonl +0 -5
  42. package/src/quant-arena/fixtures/demo-campaign/rollout-manifest.json +0 -171
  43. package/src/quant-arena/fixtures/demo-campaign/strategies/cand-001-default-author/strategy.ts +0 -119
  44. package/src/quant-arena/fixtures/demo-campaign/strategies/cand-002-default-author/strategy.ts +0 -119
  45. package/src/quant-arena/fixtures/demo-campaign/strategies/cand-003-quant-researcher/strategy.ts +0 -105
  46. package/src/quant-arena/fixtures/demo-campaign/strategies/cand-004-quant-researcher/strategy.ts +0 -102
  47. package/src/quant-arena/fixtures/demo-campaign-v2/cost-ledger.jsonl +0 -4
  48. package/src/quant-arena/fixtures/demo-campaign-v2/notebook.jsonl +0 -2
  49. package/src/quant-arena/fixtures/demo-campaign-v2/rollout-manifest.json +0 -84
  50. package/src/quant-arena/fixtures/demo-campaign-v2/strategies/cand-001-quant-researcher/strategy.ts +0 -117
  51. package/src/quant-arena/holdout-certify.mts +0 -206
  52. package/src/quant-arena/holdout-certify.test.mts +0 -82
  53. package/src/quant-arena/leak-audit.test.mts +0 -79
  54. package/src/quant-arena/leak-audit.ts +0 -95
  55. package/src/quant-arena/make-fixtures.mts +0 -161
  56. package/src/quant-arena/multiplicity.test.mts +0 -68
  57. package/src/quant-arena/multiplicity.ts +0 -87
  58. package/src/quant-arena/nautilus-certify.ts +0 -31
  59. package/src/quant-arena/oms.ts +0 -90
  60. package/src/quant-arena/profiles/quant-researcher.profile.json +0 -12
  61. package/src/quant-arena/python/pyproject.toml +0 -8
  62. package/src/quant-arena/python/uv.lock +0 -1297
  63. package/src/quant-arena/python/vbt-worker.py +0 -192
  64. package/src/quant-arena/quant-loop.mts +0 -840
  65. package/src/quant-arena/quant-loop.test.mts +0 -75
  66. package/src/quant-arena/strategies/buy-hold-index/strategy.ts +0 -11
  67. package/src/quant-arena/strategies/equal-weight/strategy.ts +0 -20
  68. package/src/quant-arena/strategies/sma-crossover/strategy.ts +0 -42
  69. package/src/quant-arena/types.ts +0 -133
  70. package/src/quant-arena/vbt-client.ts +0 -321
  71. package/src/quant-arena/vbt-parity.test.mts +0 -183
  72. package/src/quant-arena/windows.test.mts +0 -45
  73. package/src/quant-arena/windows.ts +0 -54
  74. package/src/rollout-ledger/backfill-swe-arena.mts +0 -610
  75. package/src/rollout-ledger/backfill-swe-arena.test.mts +0 -347
  76. package/src/rollout-ledger/settle-capture.mts +0 -448
  77. package/src/rollout-ledger/settle-capture.test.mts +0 -270
  78. package/src/swe-arena/activation.mts +0 -225
  79. package/src/swe-arena/activation.test.mts +0 -300
  80. package/src/swe-arena/analyze.ts +0 -211
  81. package/src/swe-arena/arms.ts +0 -862
  82. package/src/swe-arena/bootstrap-meta.mts +0 -188
  83. package/src/swe-arena/bootstrap-meta.test.mts +0 -51
  84. package/src/swe-arena/briefing.mts +0 -217
  85. package/src/swe-arena/briefing.test.mts +0 -179
  86. package/src/swe-arena/calibrate.ts +0 -217
  87. package/src/swe-arena/capabilities.mts +0 -76
  88. package/src/swe-arena/capabilities.test.mts +0 -57
  89. package/src/swe-arena/capacity.ts +0 -198
  90. package/src/swe-arena/cell-evidence.mts +0 -437
  91. package/src/swe-arena/cell-evidence.test.mts +0 -248
  92. package/src/swe-arena/diagnosis-ensemble.test.mts +0 -210
  93. package/src/swe-arena/diagnosis-ensemble.ts +0 -523
  94. package/src/swe-arena/execution.test.mts +0 -1171
  95. package/src/swe-arena/factory-command-container.ts +0 -284
  96. package/src/swe-arena/factory-judge-child.mts +0 -228
  97. package/src/swe-arena/factory.test.mts +0 -645
  98. package/src/swe-arena/fixtures/analyze.py +0 -80
  99. package/src/swe-arena/fixtures/excludes.txt +0 -8
  100. package/src/swe-arena/fixtures/factory/agent-eval-309/calibration.md +0 -51
  101. package/src/swe-arena/fixtures/factory/agent-eval-309/manifest.json +0 -29
  102. package/src/swe-arena/fixtures/factory/agent-eval-309/spec.md +0 -64
  103. package/src/swe-arena/fixtures/factory/agent-runtime-232/calibration.md +0 -48
  104. package/src/swe-arena/fixtures/factory/agent-runtime-232/manifest.json +0 -29
  105. package/src/swe-arena/fixtures/factory/agent-runtime-232/spec.md +0 -48
  106. package/src/swe-arena/fixtures/factory/loops-28/calibration.md +0 -47
  107. package/src/swe-arena/fixtures/factory/loops-28/manifest.json +0 -30
  108. package/src/swe-arena/fixtures/factory/loops-28/spec.md +0 -50
  109. package/src/swe-arena/fixtures/gen1-salvage/README.md +0 -45
  110. package/src/swe-arena/fixtures/gen1-salvage/cand0-e6d7361.diff +0 -116
  111. package/src/swe-arena/fixtures/gen1-salvage/cand1-76a8590.diff +0 -293
  112. package/src/swe-arena/fixtures/holdout-preregister.log +0 -12
  113. package/src/swe-arena/fixtures/holdout.json +0 -44
  114. package/src/swe-arena/fixtures/instances.json +0 -146
  115. package/src/swe-arena/fixtures/ledger.jsonl +0 -12
  116. package/src/swe-arena/fixtures/patches/pallets__flask-5014.solo.patch +0 -36
  117. package/src/swe-arena/fixtures/patches/pydata__xarray-4687.sup.patch +0 -33
  118. package/src/swe-arena/fixtures/rejudge.jsonl +0 -15
  119. package/src/swe-arena/fixtures/rematch.jsonl +0 -3
  120. package/src/swe-arena/fixtures/rematch2.jsonl +0 -3
  121. package/src/swe-arena/fixtures/rematch3.jsonl +0 -3
  122. package/src/swe-arena/fixtures/run-report/README.md +0 -43
  123. package/src/swe-arena/fixtures/run-report/factory-agent-eval-309-FSUP0.json +0 -173
  124. package/src/swe-arena/fixtures/run-report/factory-agent-eval-309-FSUP0.md +0 -100
  125. package/src/swe-arena/fixtures/run-report/gen3-rollup.json +0 -551
  126. package/src/swe-arena/fixtures/run-report/gen3-rollup.md +0 -64
  127. package/src/swe-arena/fixtures/sup-journal-true.json +0 -19
  128. package/src/swe-arena/fixtures/verify/astropy__astropy-13033.sh +0 -48
  129. package/src/swe-arena/fixtures/verify/django__django-11532.sh +0 -50
  130. package/src/swe-arena/fixtures/verify/matplotlib__matplotlib-20826.sh +0 -76
  131. package/src/swe-arena/fixtures/verify/pydata__xarray-4687.sh +0 -44
  132. package/src/swe-arena/fixtures/verify/pytest-dev__pytest-6197.sh +0 -32
  133. package/src/swe-arena/fixtures/verify/sphinx-doc__sphinx-9658.sh +0 -51
  134. package/src/swe-arena/fixtures/worker-tokens.json +0 -42
  135. package/src/swe-arena/fixtures.ts +0 -237
  136. package/src/swe-arena/gepa-seat.mts +0 -886
  137. package/src/swe-arena/gepa-seat.test.mts +0 -1136
  138. package/src/swe-arena/holdout-certify.mts +0 -408
  139. package/src/swe-arena/holdout-certify.test.mts +0 -160
  140. package/src/swe-arena/implementation-ref.test.mts +0 -64
  141. package/src/swe-arena/implementation-ref.ts +0 -62
  142. package/src/swe-arena/judge-child.mts +0 -37
  143. package/src/swe-arena/ledger-orphans.mts +0 -77
  144. package/src/swe-arena/ledger-orphans.test.mts +0 -149
  145. package/src/swe-arena/manifest.mts +0 -293
  146. package/src/swe-arena/manifest.test.mts +0 -169
  147. package/src/swe-arena/materialize.ts +0 -142
  148. package/src/swe-arena/outer-loop.mts +0 -2854
  149. package/src/swe-arena/outer-loop.test.mts +0 -714
  150. package/src/swe-arena/parity.test.mts +0 -87
  151. package/src/swe-arena/premeasured-from-cells.mts +0 -296
  152. package/src/swe-arena/premeasured-from-cells.test.mts +0 -201
  153. package/src/swe-arena/proc.test.mts +0 -172
  154. package/src/swe-arena/proc.ts +0 -260
  155. package/src/swe-arena/profiles/deepseek-author.profile.json +0 -12
  156. package/src/swe-arena/profiles/default-author.profile.json +0 -12
  157. package/src/swe-arena/proposer-fanout.mts +0 -736
  158. package/src/swe-arena/proposer-fanout.test.mts +0 -660
  159. package/src/swe-arena/proposer-provenance.mts +0 -176
  160. package/src/swe-arena/proposer-provenance.test.mts +0 -106
  161. package/src/swe-arena/reconcile.ts +0 -0
  162. package/src/swe-arena/replay.mts +0 -183
  163. package/src/swe-arena/replay.test.mts +0 -300
  164. package/src/swe-arena/run-experiment.mts +0 -729
  165. package/src/swe-arena/run-report.mts +0 -75
  166. package/src/swe-arena/run-supervisor.mjs +0 -297
  167. package/src/swe-arena/run-supervisor.test.mts +0 -539
  168. package/src/swe-arena/score-split.mts +0 -140
  169. package/src/swe-arena/score-split.test.mts +0 -123
  170. package/src/swe-arena/scratch-worktree-serialization.test.mts +0 -72
  171. package/src/swe-arena/scratch-worktree.test.mts +0 -56
  172. package/src/swe-arena/scratch-worktree.ts +0 -64
  173. package/src/swe-arena/serialized-judge.ts +0 -414
  174. package/src/swe-arena/types.ts +0 -218
@@ -1,62 +0,0 @@
1
- import { createHash } from 'node:crypto'
2
- import { lstat, readFile, readdir, readlink, realpath } from 'node:fs/promises'
3
- import { join, relative } from 'node:path'
4
- import { canonicalCandidateDigest } from '@tangle-network/agent-interface'
5
-
6
- const sha256Content = (value: Uint8Array): string =>
7
- `sha256:${createHash('sha256').update(value).digest('hex')}`
8
-
9
- export async function fileTreeImplementationRef(root: string): Promise<string> {
10
- const canonicalRoot = await realpath(root)
11
- const files: Array<{ path: string; sha256?: string; symlink?: string }> = []
12
-
13
- const walk = async (directory: string): Promise<void> => {
14
- const entries = await readdir(directory, { withFileTypes: true })
15
- entries.sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0))
16
- for (const entry of entries) {
17
- const path = join(directory, entry.name)
18
- const relativePath = relative(canonicalRoot, path)
19
- if (entry.isDirectory()) {
20
- await walk(path)
21
- } else if (entry.isSymbolicLink()) {
22
- files.push({ path: relativePath, symlink: await readlink(path) })
23
- } else if (entry.isFile() || (await lstat(path)).isFile()) {
24
- files.push({ path: relativePath, sha256: sha256Content(await readFile(path)) })
25
- }
26
- }
27
- }
28
-
29
- await walk(canonicalRoot)
30
- return canonicalCandidateDigest(files)
31
- }
32
-
33
- export async function pythonDistributionImplementationRef(
34
- distribution: string,
35
- runPython: (script: string, args: string[]) => Promise<string>,
36
- ): Promise<string> {
37
- const output = await runPython(
38
- [
39
- 'import hashlib, importlib, importlib.metadata, json, pathlib, sys',
40
- 'name = sys.argv[1]',
41
- 'module = importlib.import_module(name)',
42
- 'root = pathlib.Path(module.__file__).resolve().parent',
43
- 'files = []',
44
- 'for path in sorted(p for p in root.rglob("*") if p.is_file()):',
45
- ' if "__pycache__" in path.parts or path.suffix == ".pyc":',
46
- ' continue',
47
- ' files.append({"path": str(path.relative_to(root)), "sha256": hashlib.sha256(path.read_bytes()).hexdigest()})',
48
- 'try:',
49
- ' version = importlib.metadata.version(name)',
50
- 'except importlib.metadata.PackageNotFoundError:',
51
- ' version = None',
52
- 'print(json.dumps({"distribution": name, "version": version, "python": sys.version, "files": files}, sort_keys=True))',
53
- ].join('\n'),
54
- [distribution],
55
- )
56
- const line = output.trim().split('\n').at(-1)
57
- if (line === undefined) {
58
- throw new Error(`python distribution ${distribution} returned no implementation manifest`)
59
- }
60
- const manifest = JSON.parse(line) as unknown
61
- return canonicalCandidateDigest(manifest)
62
- }
@@ -1,37 +0,0 @@
1
- /**
2
- * Judge child entrypoint — the experiment's `judge.mts`, tracked. One patch,
3
- * one official swebench verdict, printed as a single `JUDGE_RESULT {...}` line
4
- * that serialized-judge.ts parses. Runs as a CHILD process (not in-process) so
5
- * the judge ceiling can SIGKILL a hung swebench run without taking the
6
- * experiment down, exactly like the bash `timeout N node judge.mts`.
7
- *
8
- * usage: node --import tsx src/swe-arena/judge-child.mts <instance_id> <patchPath>
9
- */
10
-
11
- import { readFileSync } from 'node:fs'
12
- import { createSweBenchAdapter } from '../benchmarks/swe-bench.ts'
13
-
14
- const [, , iid, patchPath] = process.argv
15
- if (!iid || !patchPath) {
16
- console.error('usage: node --import tsx judge-child.mts <instance_id> <patchPath>')
17
- process.exit(2)
18
- }
19
- const patch = readFileSync(patchPath, 'utf8')
20
- const adapter = createSweBenchAdapter()
21
- const [task] = await adapter.loadTasks({ ids: [iid], split: 'test' })
22
- if (!task) {
23
- console.log(JSON.stringify({ iid, error: 'no-task' }))
24
- process.exit(2)
25
- }
26
- const t0 = Date.now()
27
- const score = await adapter.judge(task, patch)
28
- console.log(
29
- 'JUDGE_RESULT ' +
30
- JSON.stringify({
31
- iid,
32
- resolved: score.resolved,
33
- score: score.score,
34
- secs: Math.round((Date.now() - t0) / 1000),
35
- patch_bytes: patch.length,
36
- }),
37
- )
@@ -1,77 +0,0 @@
1
- /**
2
- * Crash-orphan reconciliation for the improve-run CostLedger.
3
- *
4
- * A killed outer-loop process leaves its in-flight paid call as a durable
5
- * 'pending' record with no receipt. On resume the ledger's fail-closed
6
- * admission guard ("N unresolved call(s) must be reconciled before new paid
7
- * work") then refuses EVERY new paid call, so a single crash permanently
8
- * blocks the run. This module settles those orphans at startup through the
9
- * ledger's dedicated crash-recovery verb, `CostLedger.reconcile()`.
10
- *
11
- * Safety contract — never settle a live run's call:
12
- * - The caller must hold the outDir single-instance pid lock
13
- * (`acquireInstanceLock` in outer-loop.mts) before invoking this, so no
14
- * concurrent outer-loop shares the ledger file. Every pending record
15
- * restored from disk is therefore from a dead process.
16
- * - Belt and braces, only `state: 'interrupted'` pendings are touched:
17
- * restored-from-persistence calls this process's ledger instance never
18
- * started. An 'active'/'late' pending belongs to an in-flight
19
- * `runPaidCall` and settles through its own path.
20
- * - The ledger's persistence append is compare-and-swap on the file
21
- * revision, so even an unexpected concurrent writer makes reconciliation
22
- * fail loud (`CostCallConflictError`) instead of corrupting the log.
23
- *
24
- * The orphan settles as a FAILED receipt with zero usage and a fully-known
25
- * $0 actual cost (`actualCostUsd: 0`), reason 'process-crash-orphan'. A
26
- * `costUnknown`/`usageUnknown` receipt would be marginally more honest about
27
- * the crashed call's true spend, but an unknown-cost receipt marks the
28
- * ledger's accounting incomplete forever, which blocks all paid work in
29
- * capped runs — the same permanent lockout this module exists to fix.
30
- */
31
-
32
- import { existsSync } from 'node:fs'
33
- import { join } from 'node:path'
34
- import type { CostReceipt, PendingCostCallView } from '@tangle-network/agent-eval'
35
- import { createRunCostLedger, fsCampaignStorage } from '@tangle-network/agent-eval/campaign'
36
-
37
- export const CRASH_ORPHAN_REASON = 'process-crash-orphan'
38
-
39
- /** The two ledger verbs reconciliation needs, structural for testability. */
40
- export interface OrphanReconcilableLedger {
41
- listPending(): PendingCostCallView[]
42
- reconcile(
43
- callId: string,
44
- observed: {
45
- model: string
46
- inputTokens: number
47
- outputTokens: number
48
- actualCostUsd?: number
49
- },
50
- options?: { error?: string },
51
- ): CostReceipt
52
- }
53
-
54
- /** Settle every crash-orphaned ('interrupted') pending call as a $0 failure. */
55
- export function reconcileCrashOrphanCalls(ledger: OrphanReconcilableLedger): CostReceipt[] {
56
- return ledger
57
- .listPending()
58
- .filter((pending) => pending.state === 'interrupted')
59
- .map((pending) =>
60
- ledger.reconcile(
61
- pending.callId,
62
- { model: pending.model, inputTokens: 0, outputTokens: 0, actualCostUsd: 0 },
63
- { error: CRASH_ORPHAN_REASON },
64
- ),
65
- )
66
- }
67
-
68
- /**
69
- * Open the durable ledger under `improveRunDir` and settle crash orphans.
70
- * No-op (empty result) when no ledger file exists yet — a fresh run.
71
- * PRECONDITION: the caller holds the outDir instance lock (see module doc).
72
- */
73
- export function reconcileCrashOrphansOnDisk(improveRunDir: string): CostReceipt[] {
74
- if (!existsSync(join(improveRunDir, 'cost-ledger.jsonl'))) return []
75
- const ledger = createRunCostLedger({ storage: fsCampaignStorage(), runDir: improveRunDir })
76
- return reconcileCrashOrphanCalls(ledger)
77
- }
@@ -1,149 +0,0 @@
1
- /**
2
- * Crash-orphan ledger reconciliation: a synthetic durable ledger with one
3
- * orphaned pending call (crashed process — call opened, no receipt) and one
4
- * resolved call. Reconciliation must settle ONLY the orphan, as a $0
5
- * failure receipt tagged 'process-crash-orphan', after which the ledger's
6
- * fail-closed admission guard passes new paid work again — the exact resume
7
- * path improve() rides.
8
- */
9
-
10
- import { execFileSync } from 'node:child_process'
11
- import { mkdtempSync, readFileSync } from 'node:fs'
12
- import { tmpdir } from 'node:os'
13
- import { join } from 'node:path'
14
- import { describe, expect, it } from 'vitest'
15
- import type { CostReceipt, PendingCostCallView } from '@tangle-network/agent-eval'
16
- import { createRunCostLedger, fsCampaignStorage } from '@tangle-network/agent-eval/campaign'
17
- import {
18
- CRASH_ORPHAN_REASON,
19
- reconcileCrashOrphanCalls,
20
- reconcileCrashOrphansOnDisk,
21
- type OrphanReconcilableLedger,
22
- } from './ledger-orphans.mts'
23
-
24
- const ATTRIBUTION = { channel: 'agent', phase: 'search.baseline', model: 'm' }
25
-
26
- /**
27
- * Write one settled receipt and then terminate a process with a real pending
28
- * call. The child uses the public ledger API so this fixture cannot shadow its
29
- * private JSONL event format.
30
- */
31
- function writeCrashedLedger(dir: string): void {
32
- const child = `
33
- import { createRunCostLedger, fsCampaignStorage } from '@tangle-network/agent-eval/campaign'
34
- const runDir = process.env.COST_LEDGER_RUN_DIR
35
- if (typeof runDir !== 'string' || runDir.length === 0) throw new Error('missing COST_LEDGER_RUN_DIR')
36
- const attribution = { channel: 'agent', phase: 'search.baseline', model: 'm' }
37
- const ledger = createRunCostLedger({ storage: fsCampaignStorage(), runDir })
38
- await ledger.runPaidCall({
39
- ...attribution,
40
- callId: 'settled-1',
41
- actor: 'worker:astropy#r0',
42
- execute: async () => 'settled',
43
- receipt: () => ({ model: 'm', inputTokens: 10, outputTokens: 5, actualCostUsd: 0.01 }),
44
- })
45
- let started
46
- const pending = new Promise((resolve) => { started = resolve })
47
- void ledger.runPaidCall({
48
- ...attribution,
49
- callId: 'orphan-1',
50
- actor: 'worker:xarray#r0',
51
- tags: { cellId: 'pydata__xarray-4687:0' },
52
- execute: async () => { started(); return await new Promise(() => {}) },
53
- receipt: () => ({ model: 'm', inputTokens: 0, outputTokens: 0, actualCostUsd: 0 }),
54
- })
55
- await pending
56
- `
57
- execFileSync(process.execPath, ['--input-type=module', '--eval', child], {
58
- cwd: process.cwd(),
59
- env: { ...process.env, COST_LEDGER_RUN_DIR: dir },
60
- })
61
- }
62
-
63
- const openLedger = (dir: string) =>
64
- createRunCostLedger({ storage: fsCampaignStorage(), runDir: dir })
65
-
66
- describe('reconcileCrashOrphansOnDisk', () => {
67
- it('settles only the orphan, and the admission guard passes afterwards', async () => {
68
- const dir = mkdtempSync(join(tmpdir(), 'ledger-orphans-'))
69
- writeCrashedLedger(dir)
70
-
71
- // The bug being fixed: with the orphan unresolved, ALL new paid work is refused.
72
- const blocked = await openLedger(dir).runPaidCall({
73
- ...ATTRIBUTION,
74
- actor: 'worker:next',
75
- execute: async () => 'ok',
76
- receipt: () => ({ model: 'm', inputTokens: 1, outputTokens: 1, actualCostUsd: 0 }),
77
- })
78
- expect(blocked.succeeded).toBe(false)
79
- if (!blocked.succeeded) {
80
- expect(blocked.error.message).toContain('unresolved call(s) must be reconciled')
81
- }
82
-
83
- const receipts = reconcileCrashOrphansOnDisk(dir)
84
- expect(receipts.map((r) => r.callId)).toEqual(['orphan-1'])
85
- expect(receipts[0]).toMatchObject({
86
- status: 'settled',
87
- error: CRASH_ORPHAN_REASON,
88
- inputTokens: 0,
89
- outputTokens: 0,
90
- costUsd: 0,
91
- costUnknown: false,
92
- })
93
-
94
- // Durable: a fresh ledger instance sees zero unresolved calls, the
95
- // settled call untouched, and admits new paid work again.
96
- const reopened = openLedger(dir)
97
- const summary = reopened.summary()
98
- expect(summary.unresolvedCalls).toBe(0)
99
- expect(summary.pendingCalls).toBe(0)
100
- expect(summary.accountingComplete).toBe(true)
101
- const settled = reopened.list().find((r) => r.callId === 'settled-1')
102
- expect(settled?.costUsd).toBe(0.01)
103
- expect(settled?.error).toBeUndefined()
104
- const admitted = await reopened.runPaidCall({
105
- ...ATTRIBUTION,
106
- actor: 'worker:next',
107
- execute: async () => 'ok',
108
- receipt: () => ({ model: 'm', inputTokens: 1, outputTokens: 1, actualCostUsd: 0 }),
109
- })
110
- expect(admitted.succeeded).toBe(true)
111
-
112
- // The durable log carries the orphan's failure receipt verbatim.
113
- const persisted = readFileSync(join(dir, 'cost-ledger.jsonl'), 'utf8')
114
- expect(persisted).toContain(CRASH_ORPHAN_REASON)
115
- })
116
-
117
- it('is idempotent — a second startup pass finds nothing to settle', () => {
118
- const dir = mkdtempSync(join(tmpdir(), 'ledger-orphans-'))
119
- writeCrashedLedger(dir)
120
- expect(reconcileCrashOrphansOnDisk(dir)).toHaveLength(1)
121
- expect(reconcileCrashOrphansOnDisk(dir)).toHaveLength(0)
122
- })
123
-
124
- it('is a no-op on a fresh run dir with no ledger file', () => {
125
- const dir = mkdtempSync(join(tmpdir(), 'ledger-orphans-'))
126
- expect(reconcileCrashOrphansOnDisk(dir)).toEqual([])
127
- })
128
- })
129
-
130
- describe('reconcileCrashOrphanCalls', () => {
131
- it("never settles a live in-process call — only 'interrupted' pendings", () => {
132
- const base = { ...ATTRIBUTION, status: 'pending' as const, timestamp: 1 }
133
- const pendings: PendingCostCallView[] = [
134
- { ...base, callId: 'live-1', actor: 'a', state: 'active' },
135
- { ...base, callId: 'late-1', actor: 'b', state: 'late' },
136
- { ...base, callId: 'orphan-1', actor: 'c', state: 'interrupted' },
137
- ]
138
- const settled: string[] = []
139
- const ledger: OrphanReconcilableLedger = {
140
- listPending: () => pendings,
141
- reconcile: (callId) => {
142
- settled.push(callId)
143
- return { callId } as CostReceipt
144
- },
145
- }
146
- reconcileCrashOrphanCalls(ledger)
147
- expect(settled).toEqual(['orphan-1'])
148
- })
149
- })
@@ -1,293 +0,0 @@
1
- /**
2
- * Rollout manifest — a pure READER/JOIN over what the round already writes.
3
- * No new capture pipeline: the lib's improve() run emits loop-provenance.json,
4
- * per-cell cached results, and the durable cost log; the swe-arena dispatch
5
- * emits arm result/judge files, candidate patches, and proposer-shot receipts.
6
- * This module joins them into one versioned per-generation manifest of
7
- * (state, action, outcome, cost, artifactPaths) triples.
8
- *
9
- * tsx src/swe-arena/manifest.mts <outDir> # writes <outDir>/rollout-manifest.json
10
- *
11
- * Sources (all optional — a missing file is a labeled gap, never a crash):
12
- * - <outDir>/improve-run/loop-provenance.json (lib)
13
- * - <outDir>/improve-run/cost-ledger.jsonl (lib, durable receipts)
14
- * - <outDir>/improve-run/baseline and gen-N/candidate-K (lib, per-cell caches)
15
- * - <outDir>/candidates/<commit10>.patch (ours, diff writing)
16
- * - <outDir>/proposer-shots/genN-candK-shotS.json (ours, shot receipts)
17
- * - per-cell artifact.runDir result.json + judge.json (ours, arm + judge)
18
- */
19
-
20
- import { readdir, readFile, writeFile } from 'node:fs/promises'
21
- import { existsSync } from 'node:fs'
22
- import { join } from 'node:path'
23
- import { pathToFileURL } from 'node:url'
24
- import {
25
- loadCampaignCells,
26
- loadCandidateCellGroups,
27
- perInstanceFromCells,
28
- replicateRunsFromCells,
29
- resolvedInstanceCount,
30
- sumWallSFromCells,
31
- type EvidenceCell,
32
- type StaircasePerInstance,
33
- } from './cell-evidence.mts'
34
-
35
- export const ROLLOUT_SCHEMA = 'swe-arena.rollout.v1'
36
-
37
- export interface RolloutCost {
38
- /** Σ campaign-cell CostLedger spend (worker receipts) across the cells. */
39
- cellCostUsd: number
40
- cellTokensIn: number
41
- cellTokensOut: number
42
- /** Σ runtime spend-tree usd/tokens off the artifacts (telemetry-gap nulls skipped). */
43
- armSpentUsd: number
44
- armSpentTokens: number
45
- wallS: number
46
- judgeWallS: number
47
- /** Durable cost-ledger receipts whose tags name this campaign dir. */
48
- receiptCallIds: string[]
49
- }
50
-
51
- export interface RolloutEntry {
52
- state: {
53
- generation: number
54
- /** -1 candidateIndex = the baseline campaign. */
55
- candidateIndex: number
56
- campaignDir: string
57
- commit: string | null
58
- }
59
- action: {
60
- /** Candidate diff written at dispatch time (null for the baseline / a
61
- * resumed-only run whose diff was not rewritten). */
62
- diffPath: string | null
63
- /** Persisted proposer-shot receipts for this (generation, candidate). */
64
- shotReceiptPaths: string[]
65
- }
66
- outcome: {
67
- perInstance: StaircasePerInstance[]
68
- resolvedCount: number
69
- cellsCached: number
70
- }
71
- cost: RolloutCost
72
- artifactPaths: {
73
- cells: string[]
74
- armResults: string[]
75
- judgeVerdicts: string[]
76
- patches: string[]
77
- }
78
- }
79
-
80
- export interface RolloutGeneration {
81
- generation: number
82
- rollouts: RolloutEntry[]
83
- }
84
-
85
- export interface RolloutManifest {
86
- schema: typeof ROLLOUT_SCHEMA
87
- outDir: string
88
- at: string
89
- /** The lib's durable provenance record, verbatim (null = not emitted). */
90
- provenance: unknown | null
91
- /** Improvement-set ids inferred from the baseline cells. */
92
- instances: string[]
93
- generations: RolloutGeneration[]
94
- }
95
-
96
- interface LedgerReceipt {
97
- callId: string
98
- tags?: Record<string, string>
99
- [k: string]: unknown
100
- }
101
-
102
- /** Settled receipts out of the lib's append-only cost-ledger.jsonl. */
103
- export async function loadLedgerReceipts(improveRunDir: string): Promise<LedgerReceipt[]> {
104
- const raw = await readFile(join(improveRunDir, 'cost-ledger.jsonl'), 'utf8').catch(() => '')
105
- const receipts: LedgerReceipt[] = []
106
- for (const line of raw.split('\n')) {
107
- if (!line.trim()) continue
108
- let event: Record<string, unknown>
109
- try {
110
- event = JSON.parse(line) as Record<string, unknown>
111
- } catch {
112
- continue
113
- }
114
- const record = event.record as Record<string, unknown> | undefined
115
- if (record && record.status === 'settled' && typeof record.callId === 'string') {
116
- receipts.push(record as unknown as LedgerReceipt)
117
- }
118
- }
119
- return receipts
120
- }
121
-
122
- function costFromCells(
123
- cells: EvidenceCell[],
124
- campaignDir: string,
125
- receipts: LedgerReceipt[],
126
- ): RolloutCost {
127
- let cellCostUsd = 0
128
- let cellTokensIn = 0
129
- let cellTokensOut = 0
130
- let armSpentUsd = 0
131
- let armSpentTokens = 0
132
- let judgeWallS = 0
133
- for (const cell of cells) {
134
- cellCostUsd += cell.costUsd ?? 0
135
- cellTokensIn += cell.tokenUsage?.input ?? 0
136
- cellTokensOut += cell.tokenUsage?.output ?? 0
137
- const a = cell.artifact
138
- if (a !== null && a.kind === 'swe-arm') {
139
- armSpentUsd += a.spentUsd ?? 0
140
- armSpentTokens += a.spentTokens ?? 0
141
- judgeWallS += a.judgeWallS ?? 0
142
- }
143
- }
144
- return {
145
- cellCostUsd: Number(cellCostUsd.toFixed(6)),
146
- cellTokensIn,
147
- cellTokensOut,
148
- armSpentUsd: Number(armSpentUsd.toFixed(6)),
149
- armSpentTokens,
150
- wallS: sumWallSFromCells(cells),
151
- judgeWallS,
152
- receiptCallIds: receipts
153
- .filter((r) => r.tags?.runDir === campaignDir)
154
- .map((r) => r.callId)
155
- .sort(),
156
- }
157
- }
158
-
159
- function artifactPathsFromCells(cells: EvidenceCell[], campaignDir: string): RolloutEntry['artifactPaths'] {
160
- const armResults: string[] = []
161
- const judgeVerdicts: string[] = []
162
- const patches: string[] = []
163
- for (const cell of cells) {
164
- const a = cell.artifact
165
- if (a === null || a.kind !== 'swe-arm') continue
166
- if (a.runDir) {
167
- const resultPath = join(a.runDir, 'result.json')
168
- const judgePath = join(a.runDir, 'judge.json')
169
- if (existsSync(resultPath)) armResults.push(resultPath)
170
- if (existsSync(judgePath)) judgeVerdicts.push(judgePath)
171
- }
172
- if (a.patchPath && existsSync(a.patchPath)) patches.push(a.patchPath)
173
- }
174
- return {
175
- // The campaign dir holds the per-cell cached-result.json score records.
176
- cells: cells.length > 0 ? [campaignDir] : [],
177
- armResults: [...new Set(armResults)].sort(),
178
- judgeVerdicts: [...new Set(judgeVerdicts)].sort(),
179
- patches: [...new Set(patches)].sort(),
180
- }
181
- }
182
-
183
- async function shotReceiptPathsFor(
184
- outDir: string,
185
- generation: number,
186
- candidateIndex: number,
187
- ): Promise<string[]> {
188
- const shotDir = join(outDir, 'proposer-shots')
189
- const names = await readdir(shotDir).catch(() => [])
190
- return names
191
- .filter((n) => n.startsWith(`gen${generation}-cand${candidateIndex}-`) && n.endsWith('.json'))
192
- .sort()
193
- .map((n) => join(shotDir, n))
194
- }
195
-
196
- /** Build the rollout manifest for one round outDir. Pure join — reads only. */
197
- export async function buildRolloutManifest(outDir: string): Promise<RolloutManifest> {
198
- const improveRunDir = join(outDir, 'improve-run')
199
- const provenanceRaw = await readFile(join(improveRunDir, 'loop-provenance.json'), 'utf8').catch(() => null)
200
- const provenance = provenanceRaw === null ? null : (JSON.parse(provenanceRaw) as unknown)
201
- const receipts = await loadLedgerReceipts(improveRunDir)
202
-
203
- const baselineDir = join(improveRunDir, 'baseline')
204
- const baselineCells = await loadCampaignCells(baselineDir)
205
- const instances = [...new Set(baselineCells.map((c) => c.scenarioId))].sort()
206
-
207
- const generations = new Map<number, RolloutGeneration>()
208
- const genOf = (g: number): RolloutGeneration => {
209
- const existing = generations.get(g)
210
- if (existing) return existing
211
- const fresh: RolloutGeneration = { generation: g, rollouts: [] }
212
- generations.set(g, fresh)
213
- return fresh
214
- }
215
-
216
- if (baselineCells.length > 0) {
217
- genOf(-1).rollouts.push({
218
- state: { generation: -1, candidateIndex: -1, campaignDir: baselineDir, commit: baselineCells[0]?.artifact?.commit ?? null },
219
- action: { diffPath: null, shotReceiptPaths: [] },
220
- outcome: {
221
- perInstance: perInstanceFromCells(baselineCells),
222
- resolvedCount: resolvedInstanceCount(
223
- replicateRunsFromCells(baselineCells),
224
- instances,
225
- maxRep(baselineCells) + 1,
226
- ),
227
- cellsCached: baselineCells.length,
228
- },
229
- cost: costFromCells(baselineCells, baselineDir, receipts),
230
- artifactPaths: artifactPathsFromCells(baselineCells, baselineDir),
231
- })
232
- }
233
-
234
- for (const group of await loadCandidateCellGroups(improveRunDir)) {
235
- const diffPath = group.commit ? join(outDir, 'candidates', `${group.commit.slice(0, 10)}.patch`) : null
236
- genOf(group.generation).rollouts.push({
237
- state: {
238
- generation: group.generation,
239
- candidateIndex: group.candidateIndex,
240
- campaignDir: group.dir,
241
- commit: group.commit,
242
- },
243
- action: {
244
- diffPath: diffPath !== null && existsSync(diffPath) ? diffPath : null,
245
- shotReceiptPaths: await shotReceiptPathsFor(outDir, group.generation, group.candidateIndex),
246
- },
247
- outcome: {
248
- perInstance: perInstanceFromCells(group.cells),
249
- resolvedCount: resolvedInstanceCount(
250
- replicateRunsFromCells(group.cells),
251
- instances,
252
- maxRep(group.cells) + 1,
253
- ),
254
- cellsCached: group.cells.length,
255
- },
256
- cost: costFromCells(group.cells, group.dir, receipts),
257
- artifactPaths: artifactPathsFromCells(group.cells, group.dir),
258
- })
259
- }
260
-
261
- return {
262
- schema: ROLLOUT_SCHEMA,
263
- outDir,
264
- at: new Date().toISOString(),
265
- provenance,
266
- instances,
267
- generations: [...generations.values()].sort((a, b) => a.generation - b.generation),
268
- }
269
- }
270
-
271
- /** Replicate count inferred from the cells themselves (reps = max rep index
272
- * + 1). The manifest is a reader with no config in scope; an incomplete
273
- * candidate under-infers reps and its resolvedCount stays fail-closed. */
274
- const maxRep = (cells: EvidenceCell[]): number => cells.reduce((m, c) => Math.max(m, c.rep), 0)
275
-
276
- export async function writeRolloutManifest(outDir: string): Promise<string> {
277
- const manifest = await buildRolloutManifest(outDir)
278
- const path = join(outDir, 'rollout-manifest.json')
279
- await writeFile(path, JSON.stringify(manifest, null, 2))
280
- return path
281
- }
282
-
283
- const isMain = process.argv[1] !== undefined && import.meta.url === pathToFileURL(process.argv[1]).href
284
-
285
- if (isMain) {
286
- const outDir = process.argv[2]
287
- if (!outDir) {
288
- console.error('usage: tsx src/swe-arena/manifest.mts <outDir>')
289
- process.exit(2)
290
- }
291
- const path = await writeRolloutManifest(outDir)
292
- console.log(`rollout manifest → ${path}`)
293
- }