@tangle-network/agent-bench 0.3.7 → 0.4.0

package/src/rollout-ledger/backfill-swe-arena.mts

@@ -0,0 +1,606 @@
+/**
+ * swe-arena → rollout backfill — DOMAIN GLUE ONLY. Bench owns the JOIN
+ * (manifest + cached cells + judge.json + result.json workerCwds + proposer
+ * shot receipts → harness stores); the `tangle.rollout.v1` schema, line
+ * validation, serialization, readers, exporters, and release pipeline are
+ * owned by `@tangle-network/agent-eval/rollout` — bench never writes a
+ * rollout row through anything but that API.
+ *
+ * Pure READ over one round's outDir
+ * (the same artifact tree `swe-arena/manifest.mts` joins) plus the harness
+ * stores, emitting `tangle.rollout.v1` lines with capture:"backfill":
+ *
+ *   - one SUPERVISOR line per campaign cell (instance × rep × candidate):
+ *     reward = the official-judge verdict off cached-result.json, verdict =
+ *     raw judge.json, cost = the durable episode receipt. The loops brain
+ *     persists per-request STATS only (brain.jsonl has no message content),
+ *     so supervisor transcripts are honest gap lines until settle-time
+ *     capture lands.
+ *   - one WORKER line per opencode worker session, joined via result.json
+ *     recoveredSpend.workerCwds → opencode sqlite session.directory, with the
+ *     FULL transcript inlined and reward inherited from the episode.
+ *   - one PROPOSER line per proposer shot receipt, joined to the Claude
+ *     project transcript of the shot's worktree cwd by time window.
+ *
+ * A failed join is a labeled gap line (messages: [], provenance.gap) — never
+ * a silent drop: the ledger must account for every invocation it knows of.
+ *
+ *   tsx src/rollout-ledger/backfill-swe-arena.mts <outDir> <out.jsonl>
+ */
+
+import { randomUUID } from 'node:crypto'
+import { readdir, readFile } from 'node:fs/promises'
+import { join } from 'node:path'
+import { pathToFileURL } from 'node:url'
+import type { DatabaseSync } from 'node:sqlite'
+import { loadCampaignCellRecords } from '../swe-arena/cell-evidence.mts'
+import { buildRolloutManifest, type RolloutEntry } from '../swe-arena/manifest.mts'
+import {
+  DEFAULT_CLAUDE_PROJECTS_DIR,
+  DEFAULT_OPENCODE_DB,
+  findClaudeTranscripts,
+  findOpencodeSessionsByDirectory,
+  openOpencodeDb,
+  readClaudeTranscript,
+  readOpencodeSessionMessages,
+  ROLLOUT_SCHEMA,
+  type OpencodeSessionRow,
+  type RolloutLine,
+  writeRolloutLedger,
+} from '@tangle-network/agent-eval/rollout'
+import { candidateId } from './settle-capture.mts'
+
+const OFFICIAL_JUDGE = 'swe-arena-official-judge'
+
+const isRecord = (v: unknown): v is Record<string, unknown> =>
+  typeof v === 'object' && v !== null && !Array.isArray(v)
+
+/** Full cached-result.json record (superset of the EvidenceCell slice). */
+interface CellRecord {
+  scenarioId: string
+  rep: number
+  artifact: Record<string, unknown> | null
+  error?: string
+  judgeScores?: Record<string, { composite?: number }>
+  costUsd?: number
+  tokenUsage?: { input?: number; output?: number }
+  resolvedModel?: string
+  seed?: number
+  cached?: boolean
+}
+
+/** The same cell caches `loadCampaignCells` reads, kept verbatim: the backfill
+ *  needs fields (judgeScores, resolvedModel, seed) the EvidenceCell slice drops.
+ *  Traversal and the fail-loud identity check are owned by cell-evidence — one
+ *  rule for what counts as a campaign cell.
+ *
+ *  Only scenarioId and rep are proven by that check; `artifact` is normalized to
+ *  a record-or-null here so the `artifact !== null` guards downstream cannot be
+ *  handed an `undefined` that passes them. Every remaining field is optional and
+ *  read behind its own typeof guard. */
+async function readCellRecords(campaignDir: string): Promise<CellRecord[]> {
+  const records = (await loadCampaignCellRecords(campaignDir)).map((record) => ({
+    ...record,
+    artifact: isRecord(record.artifact) ? record.artifact : null,
+  })) as unknown as CellRecord[]
+  return records.sort((a, b) => a.scenarioId.localeCompare(b.scenarioId) || a.rep - b.rep)
+}
+
+async function readJson(path: string): Promise<unknown | null> {
+  const raw = await readFile(path, 'utf8').catch(() => null)
+  if (raw === null) return null
+  try {
+    return JSON.parse(raw) as unknown
+  } catch {
+    return null
+  }
+}
+
+function splitModelRef(ref: string | undefined): { provider: string | null; model: string | null } {
+  if (typeof ref !== 'string' || ref.length === 0) return { provider: null, model: null }
+  const slash = ref.indexOf('/')
+  if (slash === -1) return { provider: null, model: ref }
+  return { provider: ref.slice(0, slash), model: ref.slice(slash + 1) }
+}
+
+export interface BackfillStats {
+  lines: number
+  byRole: Record<string, number>
+  fullTranscripts: number
+  gapLines: number
+  rewardLabeled: number
+  workerSessionsJoined: number
+  /** Worker cwds with no opencode session row — a join failure. */
+  workerCwdsMissed: number
+  /** Sessions found but carrying no readable message parts — a store failure. */
+  workerSessionsEmpty: number
+  proposerShotsJoined: number
+  proposerShotsMissed: number
+  opencodeDbAvailable: boolean
+}
+
+export interface BackfillResult {
+  lines: RolloutLine[]
+  stats: BackfillStats
+}
+
+export interface BackfillOptions {
+  opencodeDb?: string
+  claudeProjectsDir?: string
+  /** Injected clock for deterministic tests. */
+  now?: () => Date
+}
+
+interface Ctx {
+  outDir: string
+  runId: string
+  instanceCount: number
+  db: DatabaseSync | null
+  claudeProjectsDir: string
+  capturedAt: string
+  stats: BackfillStats
+}
+
+function baseLine(ctx: Ctx): Pick<RolloutLine, 'schema' | 'run_id' | 'provenance'> {
+  return {
+    schema: ROLLOUT_SCHEMA,
+    run_id: ctx.runId,
+    provenance: { captured_at: ctx.capturedAt, capture: 'backfill' },
+  }
+}
+
+function push(ctx: Ctx, lines: RolloutLine[], line: RolloutLine): void {
+  lines.push(line)
+  ctx.stats.lines += 1
+  ctx.stats.byRole[line.role] = (ctx.stats.byRole[line.role] ?? 0) + 1
+  if (line.messages.length > 0) ctx.stats.fullTranscripts += 1
+  else ctx.stats.gapLines += 1
+  if (line.outcome.reward !== null) ctx.stats.rewardLabeled += 1
+}
+
+// ---------------------------------------------------------------------------
+// Supervisor episode + worker sessions for one campaign cell.
+// ---------------------------------------------------------------------------
+
+async function emitCellLines(
+  ctx: Ctx,
+  lines: RolloutLine[],
+  entry: RolloutEntry,
+  cell: CellRecord,
+): Promise<void> {
+  const artifact = cell.artifact
+  const runDir = artifact !== null && typeof artifact.runDir === 'string' ? artifact.runDir : null
+  const judgeVerdict = runDir !== null ? await readJson(join(runDir, 'judge.json')) : null
+  const resultJson = runDir !== null ? await readJson(join(runDir, 'result.json')) : null
+  const result = isRecord(resultJson) ? resultJson : null
+
+  const composite = cell.judgeScores?.[OFFICIAL_JUDGE]?.composite
+  const reward =
+    typeof composite === 'number'
+      ? composite
+      : artifact !== null && typeof artifact.resolved === 'boolean'
+        ? (artifact.resolved ? 1 : 0)
+        : null
+
+  const { provider, model } = splitModelRef(cell.resolvedModel)
+  const supStatus = result !== null && typeof result.sup_status === 'string' ? result.sup_status : null
+  const supervisorId = randomUUID()
+
+  push(ctx, lines, {
+    ...baseLine(ctx),
+    rollout_id: supervisorId,
+    parent_rollout_id: null,
+    candidate_id: candidateId(entry.state.generation, entry.state.candidateIndex),
+    generation: entry.state.generation,
+    candidate_index: entry.state.candidateIndex,
+    role: 'supervisor',
+    task: {
+      suite: 'swe-bench-verified',
+      instance_id: cell.scenarioId,
+      split: 'search',
+      seed: typeof cell.seed === 'number' ? cell.seed : null,
+      rep: cell.rep,
+    },
+    policy: {
+      harness: 'pi-loops',
+      harness_version: null,
+      model,
+      provider,
+      profile_commit:
+        artifact !== null && typeof artifact.commit === 'string' ? artifact.commit : entry.state.commit,
+      sampling: null,
+    },
+    messages: [],
+    tool_defs: [],
+    outcome: {
+      reward,
+      reward_source: reward === null ? null : OFFICIAL_JUDGE,
+      verdict: judgeVerdict,
+      metrics: {
+        resolved: artifact?.resolved ?? null,
+        verify_pass: artifact?.verifyPass ?? null,
+        patch_lines: artifact?.patchLines ?? null,
+        judge_attempts: artifact?.judgeAttempts ?? null,
+        judge_wall_s: artifact?.judgeWallS ?? null,
+        // Supervisor-tree spend (state.json spend tree); the durable episode
+        // receipt usd is on cost.usd. Worker-session tokens live on the
+        // worker lines — recovered_tokens here is the EPISODE total, so do
+        // not sum tokens across roles without dedup.
+        spent_tokens: artifact?.spentTokens ?? null,
+        spent_usd: artifact?.spentUsd ?? null,
+        recovered_tokens: artifact?.recoveredTokens ?? null,
+        sup_status: supStatus,
+        sup_verdict: result !== null && typeof result.sup_verdict === 'string' ? result.sup_verdict : null,
+        spawned: result !== null && typeof result.spawned === 'number' ? result.spawned : null,
+        workers: result !== null && typeof result.workers === 'number' ? result.workers : null,
+        settled: result !== null && typeof result.settled === 'number' ? result.settled : null,
+        subtasks: result !== null && Array.isArray(result.subtasks) ? result.subtasks : null,
+        cached: cell.cached ?? null,
+      },
+      is_completed: artifact !== null && cell.error === undefined,
+      is_truncated: supStatus === 'running',
+      error: cell.error ?? null,
+    },
+    cost: {
+      // The campaign CostLedger receipt for the whole episode.
+      usd: typeof cell.costUsd === 'number' ? cell.costUsd : null,
+      // Brain-side token split is not recorded (brain.jsonl is stats-only);
+      // totals live in metrics.spent_tokens / recovered_tokens.
+      tokens_in: null,
+      tokens_out: null,
+      tokens_reasoning: null,
+      cache_read: null,
+      cache_write: null,
+      wall_s: artifact !== null && typeof artifact.wallS === 'number' ? artifact.wallS : null,
+    },
+    artifacts: {
+      patch_path: artifact !== null && typeof artifact.patchPath === 'string' ? artifact.patchPath : null,
+      run_dir: runDir,
+      transcript_ref: runDir !== null ? join(runDir, 'brain.jsonl') : null,
+    },
+    provenance: {
+      captured_at: ctx.capturedAt,
+      capture: 'backfill',
+      gap: 'supervisor brain transcript not persisted (brain.jsonl carries per-request stats only)',
+    },
+  })
+
+  // Worker sessions: result.json recoveredSpend.workerCwds → opencode store.
+  const recovered = result !== null && isRecord(result.recoveredSpend) ? result.recoveredSpend : null
+  // A respawned worker reuses its clone cwd, so the recovered list can name the
+  // same directory twice. Joining it twice would mint two rollout ids over one
+  // opencode session — duplicate training signal, inflated worker counts.
+  const workerCwds =
+    recovered !== null && Array.isArray(recovered.workerCwds)
+      ? [...new Set(recovered.workerCwds.filter((c): c is string => typeof c === 'string'))]
+      : []
+  for (const cwd of workerCwds) {
+    const sessions = ctx.db === null ? [] : findOpencodeSessionsByDirectory(ctx.db, cwd)
+    if (sessions.length === 0) {
+      ctx.stats.workerCwdsMissed += 1
+      push(ctx, lines, workerLine(ctx, entry, cell, supervisorId, reward, cwd, null, []))
+      continue
+    }
+    for (const session of sessions) {
+      const messages = ctx.db === null ? [] : readOpencodeSessionMessages(ctx.db, session.id)
+      // Two distinct failures: no session row for the cwd (join failure) versus
+      // a session row whose parts are unreadable (store integrity). One counter
+      // for both would hide store corruption behind a cwd-join metric.
+      if (messages.length > 0) ctx.stats.workerSessionsJoined += 1
+      else ctx.stats.workerSessionsEmpty += 1
+      push(ctx, lines, workerLine(ctx, entry, cell, supervisorId, reward, cwd, session, messages))
+    }
+  }
+}
+
+function workerLine(
+  ctx: Ctx,
+  entry: RolloutEntry,
+  cell: CellRecord,
+  supervisorId: string,
+  reward: number | null,
+  cwd: string,
+  session: OpencodeSessionRow | null,
+  messages: RolloutLine['messages'],
+): RolloutLine {
+  const artifact = cell.artifact
+  return {
+    ...baseLine(ctx),
+    rollout_id: randomUUID(),
+    parent_rollout_id: supervisorId,
+    candidate_id: candidateId(entry.state.generation, entry.state.candidateIndex),
+    generation: entry.state.generation,
+    candidate_index: entry.state.candidateIndex,
+    role: 'worker',
+    task: {
+      suite: 'swe-bench-verified',
+      instance_id: cell.scenarioId,
+      split: 'search',
+      seed: typeof cell.seed === 'number' ? cell.seed : null,
+      rep: cell.rep,
+    },
+    policy: {
+      harness: 'opencode',
+      harness_version: null,
+      model: session?.model?.id ?? null,
+      provider: session?.model?.providerID ?? null,
+      profile_commit:
+        artifact !== null && typeof artifact.commit === 'string' ? artifact.commit : entry.state.commit,
+      sampling: null,
+    },
+    messages,
+    tool_defs: [],
+    outcome: {
+      // Episode-level credit: the worker has no verdict of its own.
+      reward,
+      reward_source: reward === null ? null : `${OFFICIAL_JUDGE}/inherited`,
+      verdict: null,
+      metrics: {
+        worker_cwd: cwd,
+        session_agent: session?.agent ?? null,
+        session_parent_id: session?.parentId ?? null,
+        has_session: session !== null,
+      },
+      // A session row with no readable parts is a gap line, not a completed
+      // invocation — training filters that key on is_completed alone would
+      // otherwise pull empty transcripts.
+      is_completed: session !== null && messages.length > 0,
+      is_truncated: false,
+      error: null,
+    },
+    cost: {
+      usd: session?.costUsd ?? null,
+      tokens_in: session?.tokensInput ?? null,
+      tokens_out: session?.tokensOutput ?? null,
+      tokens_reasoning: session?.tokensReasoning ?? null,
+      cache_read: session?.tokensCacheRead ?? null,
+      cache_write: session?.tokensCacheWrite ?? null,
+      wall_s: session !== null ? Math.round((session.timeUpdated - session.timeCreated) / 1000) : null,
+    },
+    artifacts: {
+      patch_path: null,
+      run_dir: artifact !== null && typeof artifact.runDir === 'string' ? artifact.runDir : null,
+      transcript_ref: session !== null ? `opencode:${session.id}` : null,
+    },
+    provenance: {
+      captured_at: ctx.capturedAt,
+      capture: 'backfill',
+      ...(messages.length === 0
+        ? {
+            gap:
+              session === null
+                ? ctx.db === null
+                  ? `opencode store unavailable; worker cwd ${cwd} unrecoverable`
+                  : `no opencode session found for worker cwd ${cwd}`
+                : `opencode session ${session.id} has no readable message parts`,
+          }
+        : {}),
+    },
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Proposer shots — receipts under proposer-shots/ (flat or per-author dirs),
+// joined to Claude transcripts of the shot worktree by time window.
+// ---------------------------------------------------------------------------
+
+interface ShotReceiptFile {
+  path: string
+  author: string | null
+  generation: number
+  candidateIndex: number
+  shot: number
+}
+
+async function listShotReceipts(outDir: string): Promise<ShotReceiptFile[]> {
+  const shotDir = join(outDir, 'proposer-shots')
+  const found: ShotReceiptFile[] = []
+  const parse = (name: string, author: string | null, path: string): void => {
+    const match = /^gen(\d+)-cand(\d+)-shot(\d+)\.json$/.exec(name)
+    if (match === null) return
+    found.push({
+      path,
+      author,
+      generation: Number(match[1]),
+      candidateIndex: Number(match[2]),
+      shot: Number(match[3]),
+    })
+  }
+  for (const entry of await readdir(shotDir, { withFileTypes: true }).catch(() => [])) {
+    if (entry.isDirectory()) {
+      const authorDir = join(shotDir, entry.name)
+      for (const name of await readdir(authorDir).catch(() => [])) {
+        parse(name, entry.name, join(authorDir, name))
+      }
+    } else {
+      parse(entry.name, null, join(shotDir, entry.name))
+    }
+  }
+  return found.sort((a, b) => a.path.localeCompare(b.path))
+}
+
+async function emitProposerLines(ctx: Ctx, lines: RolloutLine[], entries: RolloutEntry[]): Promise<void> {
+  const receipts = await listShotReceipts(ctx.outDir)
+  const entryFor = new Map<string, RolloutEntry>()
+  for (const entry of entries) {
+    entryFor.set(`${entry.state.generation}:${entry.state.candidateIndex}`, entry)
+  }
+
+  for (const shot of receipts) {
+    const parsed = await readJson(shot.path)
+    const receipt = isRecord(parsed) && isRecord(parsed.receipt) ? parsed.receipt : null
+    const entry = entryFor.get(`${shot.generation}:${shot.candidateIndex}`) ?? null
+
+    // The proposer worktree the arena dispatched this shot into.
+    const worktreeCwd =
+      shot.author !== null
+        ? join(ctx.outDir, 'proposer-wt', `gen${shot.generation}-cand${shot.candidateIndex}-${shot.author}`)
+        : join(ctx.outDir, 'proposer-wt', `gen${shot.generation}-cand${shot.candidateIndex}`)
+
+    const startedAt = receipt !== null && typeof receipt.startedAt === 'string' ? Date.parse(receipt.startedAt) : NaN
+    const completedAt =
+      receipt !== null && typeof receipt.completedAt === 'string' ? Date.parse(receipt.completedAt) : NaN
+
+    let transcript: Awaited<ReturnType<typeof readClaudeTranscript>> | null = null
+    let transcriptRef: string | null = null
+    const candidates = await findClaudeTranscripts(worktreeCwd, ctx.claudeProjectsDir)
+    const windowed: Array<{ path: string; startMs: number }> = []
+    for (const ref of candidates) {
+      const t = await readClaudeTranscript(ref.path)
+      if (t.startedAt === null) continue
+      const startMs = Date.parse(t.startedAt)
+      const inWindow =
+        Number.isNaN(startedAt) || Number.isNaN(completedAt)
+          ? candidates.length === 1
+          : startMs >= startedAt - 120_000 && startMs <= completedAt + 120_000
+      if (inWindow) windowed.push({ path: ref.path, startMs })
+    }
+    if (windowed.length > 0) {
+      windowed.sort(
+        (a, b) => Math.abs(a.startMs - (startedAt || a.startMs)) - Math.abs(b.startMs - (startedAt || b.startMs)),
+      )
+      transcriptRef = windowed[0].path
+      transcript = await readClaudeTranscript(transcriptRef)
+      ctx.stats.proposerShotsJoined += 1
+    } else {
+      ctx.stats.proposerShotsMissed += 1
+    }
+
+    // The proposer's reward is its candidate's official resolved fraction —
+    // fail-closed null when the candidate has no scored entry.
+    const reward =
+      entry !== null && ctx.instanceCount > 0 ? entry.outcome.resolvedCount / ctx.instanceCount : null
+
+    push(ctx, lines, {
+      ...baseLine(ctx),
+      rollout_id: randomUUID(),
+      parent_rollout_id: null,
+      candidate_id: candidateId(shot.generation, shot.candidateIndex),
+      generation: shot.generation,
+      candidate_index: shot.candidateIndex,
+      role: 'proposer',
+      task: {
+        suite: 'swe-arena-proposer',
+        instance_id:
+          shot.author !== null
+            ? `${candidateId(shot.generation, shot.candidateIndex)}-${shot.author}`
+            : candidateId(shot.generation, shot.candidateIndex),
+        split: 'search',
+        seed: null,
+        rep: shot.shot,
+      },
+      policy: {
+        harness: receipt !== null && typeof receipt.harness === 'string' ? receipt.harness : null,
+        harness_version: null,
+        model:
+          transcript?.model ?? (receipt !== null && typeof receipt.model === 'string' ? receipt.model : null),
+        provider: null,
+        profile_commit: entry?.state.commit ?? null,
+        sampling: null,
+      },
+      messages: transcript?.messages ?? [],
+      tool_defs: [],
+      outcome: {
+        reward,
+        reward_source: reward === null ? null : `${OFFICIAL_JUDGE}/candidate-resolved-fraction`,
+        verdict: null,
+        metrics: {
+          resolved_count: entry?.outcome.resolvedCount ?? null,
+          instance_count: ctx.instanceCount,
+          exit_code: receipt !== null && typeof receipt.exitCode === 'number' ? receipt.exitCode : null,
+          timed_out: receipt !== null && typeof receipt.timedOut === 'boolean' ? receipt.timedOut : null,
+          author: shot.author,
+          max_shots: receipt !== null && typeof receipt.maxShots === 'number' ? receipt.maxShots : null,
+          prompt_sha256: receipt !== null && typeof receipt.promptSha256 === 'string' ? receipt.promptSha256 : null,
+        },
+        is_completed: receipt !== null && receipt.exitCode === 0,
+        is_truncated: receipt !== null && receipt.timedOut === true,
+        error: receipt !== null && typeof receipt.error === 'string' ? receipt.error : null,
+      },
+      cost: {
+        usd: receipt !== null && typeof receipt.costUsd === 'number' ? receipt.costUsd : null,
+        tokens_in: transcript !== null ? transcript.usage.tokensIn : null,
+        tokens_out: transcript !== null ? transcript.usage.tokensOut : null,
+        tokens_reasoning: null,
+        cache_read: transcript !== null ? transcript.usage.cacheRead : null,
+        cache_write: transcript !== null ? transcript.usage.cacheWrite : null,
+        wall_s:
+          receipt !== null && typeof receipt.durationMs === 'number' ? Math.round(receipt.durationMs / 1000) : null,
+      },
+      artifacts: {
+        patch_path: entry?.action.diffPath ?? null,
+        run_dir: worktreeCwd,
+        transcript_ref: transcriptRef ?? shot.path,
+      },
+      provenance: {
+        captured_at: ctx.capturedAt,
+        capture: 'backfill',
+        ...(transcript === null
+          ? { gap: `no claude transcript matched worktree ${worktreeCwd} in the shot's time window` }
+          : {}),
+      },
+    })
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Entry point.
+// ---------------------------------------------------------------------------
+
+export async function backfillSweArena(outDir: string, opts: BackfillOptions = {}): Promise<BackfillResult> {
+  const manifest = await buildRolloutManifest(outDir)
+  const provenance = isRecord(manifest.provenance) ? manifest.provenance : null
+  const runId = provenance !== null && typeof provenance.runId === 'string' ? provenance.runId : outDir
+
+  const db = await openOpencodeDb(opts.opencodeDb ?? DEFAULT_OPENCODE_DB)
+  const ctx: Ctx = {
+    outDir,
+    runId,
+    instanceCount: manifest.instances.length,
+    db,
+    claudeProjectsDir: opts.claudeProjectsDir ?? DEFAULT_CLAUDE_PROJECTS_DIR,
+    capturedAt: (opts.now?.() ?? new Date()).toISOString(),
+    stats: {
+      lines: 0,
+      byRole: {},
+      fullTranscripts: 0,
+      gapLines: 0,
+      rewardLabeled: 0,
+      workerSessionsJoined: 0,
+      workerCwdsMissed: 0,
+      workerSessionsEmpty: 0,
+      proposerShotsJoined: 0,
+      proposerShotsMissed: 0,
+      opencodeDbAvailable: db !== null,
+    },
+  }
+
+  const lines: RolloutLine[] = []
+  const entries: RolloutEntry[] = manifest.generations.flatMap((g) => g.rollouts)
+  try {
+    for (const entry of entries) {
+      for (const cell of await readCellRecords(entry.state.campaignDir)) {
+        await emitCellLines(ctx, lines, entry, cell)
+      }
+    }
+    await emitProposerLines(ctx, lines, entries)
+  } finally {
+    db?.close()
+  }
+
+  return { lines, stats: ctx.stats }
+}
+
+const isMain = process.argv[1] !== undefined && import.meta.url === pathToFileURL(process.argv[1]).href
+
+if (isMain) {
+  const [outDir, ledgerPath] = process.argv.slice(2)
+  if (!outDir || !ledgerPath) {
+    console.error('usage: tsx src/rollout-ledger/backfill-swe-arena.mts <outDir> <out.jsonl>')
+    process.exit(2)
+  }
+  const { lines, stats } = await backfillSweArena(outDir)
+  await writeRolloutLedger(ledgerPath, lines)
+  console.log(JSON.stringify(stats, null, 2))
+  console.log(`rollout ledger → ${ledgerPath} (${lines.length} lines)`)
+}