mixdog 0.9.3 → 0.9.5

package/src/runtime/memory/lib/memory-cycle2-gate.mjs

@@ -0,0 +1,515 @@
+// Unified-gate cluster extracted from memory-cycle2.mjs: prompt formatting,
+// pipe-verdict parsing/validation, the rules-digest cache, the single LLM
+// gate pass (runUnifiedGate) and the Sonnet re-judge cascade. Facade
+// (memory-cycle2.mjs) re-exports the public members unchanged.
+import { existsSync, readFileSync } from 'fs'
+import { join } from 'path'
+import { resolveMaintenancePreset } from '../../shared/llm/index.mjs'
+import { callAgentDispatch } from './agent-ipc.mjs'
+import { listCore } from './core-memory-store.mjs'
+import { __mixdogMemoryLog, throwIfAborted, resourceDir } from './memory-cycle2-shared.mjs'
+
+export const CYCLE2_ACTIVE_TARGET_CAP = 100
+
+// Status-based verb whitelist. 3-tier policy: pending → active/archived,
+// active → active/archived/update/merge.
+const STATUS_ALLOWED_VERBS = {
+  pending: new Set(['active', 'archived']),
+  active:  new Set(['active', 'archived', 'update', 'merge']),
+}
+const NON_ARCHIVE_VERBS = new Set(['active', 'update', 'merge'])
+// Union of every primary (status) verb across all statuses, plus the two
+// non-verb line kinds. Used by the stray-index shift guard to decide whether
+// a `idx|id|verb` line had a leading row index prepended by the LLM.
+const ALL_PRIMARY_VERBS = new Set(['active', 'archived', 'update', 'merge'])
+const isShiftFollowToken = (tok) => {
+  const v = String(tok ?? '').trim().toLowerCase()
+  return ALL_PRIMARY_VERBS.has(v) || v === 'why' || v === 'core'
+}
+
+async function invokeLlm(prompt, mode, preset, timeout, llmCall = callAgentDispatch) {
+  return await llmCall({
+    agent: 'cycle2-agent',
+    taskType: 'maintenance',
+    mode,
+    preset,
+    timeout,
+    cwd: null,
+  }, prompt)
+}
+
+function buildPidMap(rowSets) {
+  const pids = [...new Set(rowSets.flat().map(r => r.project_id).filter(Boolean))].sort()
+  return new Map(pids.map((p, i) => [p, `P${i + 1}`]))
+}
+
+function formatEntriesForPromotePrompt(rows, pidMap, opts = {}) {
+  if (!rows || rows.length === 0) return '(none)'
+  const map = pidMap ?? buildPidMap([rows])
+  // When numbered, prefix each row with its 1-based prompt-order ordinal so the
+  // gate LLM can echo a row number it can see, instead of inventing one. The
+  // ordinal domain (1..N) and the 5-digit batch-id domain must stay disjoint —
+  // see the ordinalToId invariant in runUnifiedGate.
+  const numbered = opts.numbered === true
+  const lines = rows.map((r, i) => {
+    const tag = r.project_id ? (map.get(r.project_id) ?? 'C') : 'C'
+    const stat = r.status ? `[${r.status}]` : '[?]'
+    const prefix = numbered ? `${i + 1}. ` : '- '
+    return `${prefix}id:${r.id} ${stat} ${tag} ${r.category} s:${r.score ?? 'n'} el:${r.element} sm:${String(r.summary || '').slice(0, 100)}`
+  })
+  if (map.size === 0) return lines.join('\n')
+  const legend = [...map.entries()].map(([p, t]) => `${t}=${p}`).concat('C=COMMON').join(', ')
+  return `# pid: ${legend}\n` + lines.join('\n')
+}
+
+// User-curated rows from core_entries — id-less, no status, no score; the
+// LLM only needs element + summary + project tag to detect overlap with
+// candidate entries below. Format kept terse so the prompt budget stays small.
+function formatUserCoreForPrompt(rows, pidMap) {
+  if (!rows || rows.length === 0) return '(none)'
+  const map = pidMap ?? new Map()
+  return rows.map(r => {
+    const tag = r.project_id ? (map.get(r.project_id) ?? 'C') : 'C'
+    const sm = String(r.summary || '').slice(0, 200)
+    return `- ${tag} ${r.category}: ${r.element}${sm && sm !== r.element ? ` — ${sm}` : ''}`
+  }).join('\n')
+}
+
+// Parse pipe-format unified verdicts. Each line: <id>|<verb> [|...].
+// Verbs validated against the row's current status via STATUS_ALLOWED_VERBS.
+// Returns { actions, rejected } or null when no parseable lines.
+function parseUnifiedFormat(raw, statusById, ordinalToId = null) {
+  if (raw == null) return null
+  const text = String(raw).trim()
+  if (!text) return { actions: [], rejected: new Set() }
+  const lines = text.split('\n')
+  const actions = []
+  const rejected = new Set()
+  const support = new Map()
+  let sawValid = false
+  // Resolve a first-field/merge token to a real batch id. The gate may echo
+  // either the exact 5-digit batch id OR the 1-based row ordinal shown in the
+  // numbered Entries block. The two domains are disjoint (asserted in
+  // runUnifiedGate), so an exact-id hit always wins and an unmatched value
+  // falls back to ordinal lookup; anything else is NaN (line treated invalid).
+  const resolveId = (tok) => {
+    const n = Number(String(tok ?? '').trim())
+    if (!Number.isFinite(n)) return NaN
+    if (statusById.has(n)) return n
+    if (ordinalToId && ordinalToId.has(n)) return ordinalToId.get(n)
+    return NaN
+  }
+  for (const rawLine of lines) {
+    const line = rawLine.trim()
+    if (!line) continue
+    if (line.startsWith('//') || line.startsWith('#')) continue
+    if (line.startsWith('```')) continue
+    const parts = line.split('|')
+    if (parts.length < 2) continue
+    // LLM sometimes prefixes a row index, emitting `idx|id|verdict` instead of
+    // `id|verdict`; parts[0] (the index) is a stray token and the line must be
+    // shifted before parsing. Strict invariant so a real 2-field `id|verdict`
+    // is never shifted into a 1-field line (which would throw on parts[1]):
+    //   parts.length >= 3 AND parts[1] is a known batch id AND parts[2] is a
+    //   valid primary verb / why / core (the shifted verdict slot).
+    // Trigger when EITHER parts[0] is not a known id (classic stray index) OR
+    // parts[0] IS known but parts[1] is not itself a valid verb — that covers
+    // `1|1|active`, where the stray index collides with a real batch id and the
+    // un-shifted reading would verb-reject the wrong row.
+    if (
+      parts.length >= 3 &&
+      statusById.has(Number(parts[1].trim())) &&
+      isShiftFollowToken(parts[2]) &&
+      (!statusById.has(Number(parts[0].trim())) || !isShiftFollowToken(parts[1]))
+    ) {
+      parts.shift()
+    }
+    const entryId = resolveId(parts[0])
+    const action = parts[1].trim().toLowerCase()
+    if (!Number.isFinite(entryId) || !action) continue
+    const status = statusById.get(entryId)
+    if (!status) continue
+    // Only mark as parse-ok when the id is known to the batch; a response
+    // composed entirely of unknown ids would otherwise return parse-ok with
+    // zero actions/rejections, leaving the rows un-reviewed and re-queued.
+    sawValid = true
+    if (action === 'core') {
+      actions.push({ entry_id: entryId, action: 'core', core_summary: parts.slice(2).join('|').trim().slice(0, 120) })
+      continue
+    }
+    if (action === 'why') {
+      const kind = (parts[2] ?? '').trim().toUpperCase()
+      const reason = parts.slice(3).join('|').replace(/\s+/g, ' ').trim().slice(0, 240)
+      if ((kind === 'A' || kind === 'B') && reason) {
+        support.set(entryId, { kind, reason })
+      }
+      continue
+    }
+    const allowed = STATUS_ALLOWED_VERBS[status]
+    if (!allowed || !allowed.has(action)) {
+      __mixdogMemoryLog(`[cycle2] verb rejected: id=${entryId} status=${status} verb=${action}\n`)
+      rejected.add(entryId)
+      continue
+    }
+    if (action === 'update') {
+      actions.push({
+        entry_id: entryId, action,
+        element: (parts[2] ?? '').trim(),
+        summary: parts.slice(3).join('|').trim(),
+      })
+    } else if (action === 'merge') {
+      const targetId = resolveId(parts[2])
+      const sourceIds = [...new Set((parts[3] ?? '').split(',').map(s => resolveId(s)).filter(Number.isFinite))]
+      if (!Number.isFinite(targetId) || sourceIds.length === 0) {
+        __mixdogMemoryLog(`[cycle2] merge rejected: id=${entryId} invalid target/sources\n`)
+        rejected.add(entryId)
+        continue
+      }
+      if (targetId !== entryId && !sourceIds.includes(entryId)) {
+        __mixdogMemoryLog(
+          `[cycle2] merge rejected: id=${entryId} must be target or listed source (target=${targetId} sources=${sourceIds.join(',')})\n`,
+        )
+        rejected.add(entryId)
+        continue
+      }
+      actions.push({
+        entry_id: entryId, action,
+        target_id: targetId,
+        source_ids: sourceIds,
+        element: (parts[4] ?? '').trim(),
+        summary: parts.slice(5).join('|').trim(),
+      })
+    } else {
+      actions.push({ entry_id: entryId, action })
+    }
+  }
+  if (!sawValid && rejected.size === 0) return null
+  return { actions, rejected, support }
+}
+
+// ─── Current rules digest cache ──────────────────────────────────────────────
+
+let _currentRulesDigest = null
+let _currentRulesDigestTs = 0
+export function loadCurrentRulesDigest() {
+  const now = Date.now()
+  if (_currentRulesDigest && now - _currentRulesDigestTs < 60_000) return _currentRulesDigest
+  const sources = [
+    join(resourceDir(), 'rules', 'shared', '01-general.md'),
+    join(resourceDir(), 'rules', 'shared', '01-tool.md'),
+    join(resourceDir(), 'rules', 'shared', '04-memory.md'),
+    join(resourceDir(), 'rules', 'shared', '06-team.md'),
+    join(resourceDir(), 'rules', 'shared', '07-workflow.md'),
+  ]
+  const parts = []
+  for (const p of sources) {
+    try {
+      if (!existsSync(p)) continue
+      const txt = readFileSync(p, 'utf8').trim()
+      if (txt) parts.push(`# Source: ${p}\n${txt}`)
+    } catch {}
+  }
+  const joined = parts.join('\n\n---\n\n')
+  const CAP = 40_000
+  _currentRulesDigest = joined.length > CAP ? joined.slice(0, CAP) + '\n…[truncated]' : joined
+  _currentRulesDigestTs = now
+  return _currentRulesDigest
+}
+
+function uniqueIds(values) {
+  return [...new Set(values
+    .map(id => Number(id))
+    .filter(id => Number.isFinite(id)))]
+}
+
+function validateUnifiedGate(parsed, statusById) {
+  const actions = Array.isArray(parsed?.actions) ? parsed.actions : []
+  const primary = actions.filter(a => a?.action !== 'core')
+  const verdictCounts = new Map()
+  for (const action of primary) {
+    const id = Number(action?.entry_id)
+    if (!Number.isFinite(id)) continue
+    verdictCounts.set(id, (verdictCounts.get(id) || 0) + 1)
+  }
+  const expectedIds = [...statusById.keys()]
+  const missingVerdictIds = expectedIds.filter(id => !verdictCounts.has(id))
+  const duplicateVerdictIds = [...verdictCounts.entries()]
+    .filter(([, count]) => count > 1)
+    .map(([id]) => id)
+  const support = parsed?.support instanceof Map ? parsed.support : new Map()
+  const coreIds = new Set(actions
+    .filter(a => a?.action === 'core')
+    .map(a => Number(a.entry_id))
+    .filter(id => Number.isFinite(id)))
+  const missingSupportIds = []
+  const missingCoreIds = []
+  for (const action of primary) {
+    if (!NON_ARCHIVE_VERBS.has(action?.action)) continue
+    const id = Number(action.entry_id)
+    if (!Number.isFinite(id)) continue
+    const coreId = action.action === 'merge' && Number.isFinite(Number(action.target_id))
+      ? Number(action.target_id)
+      : id
+    const hasSupport = support.has(id) || (action.action === 'merge' && support.has(coreId))
+    if (!hasSupport) missingSupportIds.push(id)
+    if (!coreIds.has(coreId)) missingCoreIds.push(id)
+  }
+  return {
+    missingVerdictIds: uniqueIds(missingVerdictIds),
+    duplicateVerdictIds: uniqueIds(duplicateVerdictIds),
+    missingSupportIds: uniqueIds(missingSupportIds),
+    missingCoreIds: uniqueIds(missingCoreIds),
+  }
+}
+
+function gateQualitySummary(quality) {
+  const parts = []
+  if (quality?.missingVerdictIds?.length) parts.push(`missing verdict ids=${quality.missingVerdictIds.join(',')}`)
+  if (quality?.duplicateVerdictIds?.length) parts.push(`duplicate verdict ids=${quality.duplicateVerdictIds.join(',')}`)
+  if (quality?.missingSupportIds?.length) parts.push(`missing why ids=${quality.missingSupportIds.join(',')}`)
+  if (quality?.missingCoreIds?.length) parts.push(`missing core ids=${quality.missingCoreIds.join(',')}`)
+  return parts.join('; ')
+}
+
+function stripUnsupportedPromotions(parsed, unsupportedIds) {
+  const ids = new Set(uniqueIds(unsupportedIds))
+  if (ids.size === 0) return parsed
+  const rejected = new Set(parsed?.rejected || [])
+  for (const id of ids) rejected.add(id)
+  const actions = (parsed?.actions || []).filter(a => {
+    if (a?.action === 'core') return true
+    return !ids.has(Number(a?.entry_id))
+  })
+  return { ...parsed, actions, rejected }
+}
+
+function requiredCoreIdForAction(action) {
+  if (action?.action === 'merge' && Number.isFinite(Number(action.target_id))) {
+    return Number(action.target_id)
+  }
+  return Number(action?.entry_id)
+}
+
+// ─── Unified gate ────────────────────────────────────────────────────────────
+
+// Single LLM pass over rows whose status is in {pending, active}.
+// Returns { actions, rejected, parseOk } following parseUnifiedFormat shape.
+export async function runUnifiedGate(db, rows, activeContext, config = {}, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  if (!rows || rows.length === 0) return { actions: [], rejected: new Set(), parseOk: true }
+  const promptPath = join(resourceDir(), 'defaults', 'memory-promote-prompt.md')
+  if (!existsSync(promptPath)) {
+    throw new Error(`runCycle2: prompt file missing at ${promptPath}`)
+  }
+  const template = readFileSync(promptPath, 'utf8')
+  const userCoreRows = options.dataDir ? await listCore(options.dataDir, '*').catch(() => []) : []
+  throwIfAborted(signal)
+  const sharedPidMap = buildPidMap([activeContext ?? [], rows ?? [], userCoreRows ?? []])
+  const rulesDigest = loadCurrentRulesDigest() || '(no current rules digest available)'
+  const activeCount = activeContext?.length ?? 0
+  const activeCap = options.activeCap ?? CYCLE2_ACTIVE_TARGET_CAP
+
+  const prompt = template
+    .replace('{{CURRENT_RULES}}', rulesDigest)
+    .replace('{{USER_CORE}}', formatUserCoreForPrompt(userCoreRows, sharedPidMap))
+    .replace('{{CORE_MEMORY}}', formatEntriesForPromotePrompt(activeContext, sharedPidMap))
+    .replace('{{ITEMS}}', formatEntriesForPromotePrompt(rows, sharedPidMap, { numbered: true }))
+    .replace('{{ACTIVE_COUNT}}', String(activeCount))
+    .replace('{{ACTIVE_CAP}}', String(activeCap))
+
+  const preset = options.preset || resolveMaintenancePreset('memory')
+  const timeout = Number(config?.cycle2?.timeout ?? 600000)
+  const mode = 'cycle2-unified'
+
+  const previewRaw = (raw) => String(raw ?? '').replace(/\s+/g, ' ').slice(0, 200)
+  const callOnce = async (extraTag) => {
+    throwIfAborted(signal)
+    const p = extraTag ? `${prompt}\n\n[retry:${extraTag}]` : prompt
+    const raw = await invokeLlm(p, mode, preset, timeout, options.callLlm)
+    throwIfAborted(signal)
+    return raw
+  }
+
+  const statusById = new Map(rows.map(r => [Number(r.id), String(r.status)]))
+  // Ordinal → batch-id map, keyed by 1-based prompt order (the same order the
+  // numbered Entries block uses). The gate may echo either the real 5-digit
+  // batch id or the row ordinal 1..N; the parser resolves both. The two domains
+  // MUST be disjoint or an ordinal could shadow a real id (ids are 5-digit,
+  // ordinals are <= rows.length, so disjointness always holds in practice).
+  // On a violation a row-number line is indistinguishable from an exact-id
+  // line, so no safe interpretation exists — skip this batch (gate failure)
+  // rather than risk applying a verdict to the wrong entry. The cycle itself
+  // proceeds; the batch re-queues for a later run.
+  const ordinalToId = new Map(rows.map((r, i) => [i + 1, Number(r.id)]))
+  const minBatchId = Math.min(...[...statusById.keys()])
+  if (Number.isFinite(minBatchId) && minBatchId <= rows.length) {
+    __mixdogMemoryLog(`[cycle2] batch id ${minBatchId} collides with ordinal range 1..${rows.length} — skipping batch (no safe id resolution)\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+
+  __mixdogMemoryLog(`[cycle2-diag] unified prompt=${prompt.length} bytes; rows=${rows.length}\n`)
+
+  let raw
+  try {
+    raw = await callOnce(null)
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    __mixdogMemoryLog(`[cycle2] unified LLM error: ${err.message}\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+  throwIfAborted(signal)
+  __mixdogMemoryLog(`[cycle2-diag] unified raw (first 1500): ${String(raw ?? '').replace(/\n/g, '⏎').slice(0, 1500)}\n`)
+
+  let parsed = parseUnifiedFormat(raw, statusById, ordinalToId)
+  let quality = parsed ? validateUnifiedGate(parsed, statusById) : null
+  const qualityIssue = () => gateQualitySummary(quality)
+  if (!parsed || qualityIssue()) {
+    throwIfAborted(signal)
+    const issue = parsed ? qualityIssue() : `unparseable (${previewRaw(raw)})`
+    __mixdogMemoryLog(`[cycle2] unified quality retry: ${issue}\n`)
+    // Preserve the first pass before retrying. A retry fired for a mere quality
+    // issue (e.g. a few missing verdicts) must not throw away an otherwise-valid
+    // first-pass parse if the retry comes back unparseable.
+    const firstParsed = parsed
+    const firstQuality = quality
+    try {
+      const retryTag = parsed
+        ? 'complete-verdicts-with-why-and-core-lines'
+        : 'first-field-must-be-the-listed-row-number'
+      const raw2 = await callOnce(retryTag)
+      const retryParsed = parseUnifiedFormat(raw2, statusById, ordinalToId)
+      if (retryParsed) {
+        parsed = retryParsed
+        quality = validateUnifiedGate(retryParsed, statusById)
+      } else if (firstParsed) {
+        __mixdogMemoryLog(`[cycle2] unparseable after retry — falling back to first-pass parse (${previewRaw(raw2)})\n`)
+        parsed = firstParsed
+        quality = firstQuality
+      } else {
+        __mixdogMemoryLog(`[cycle2] unparseable after retry — skipping batch (${previewRaw(raw2)})\n`)
+        return { actions: null, rejected: new Set(), parseOk: false }
+      }
+    } catch (err) {
+      if (signal?.aborted) throw signal.reason ?? err
+      if (firstParsed) {
+        __mixdogMemoryLog(`[cycle2] retry LLM error: ${err.message} — falling back to first-pass parse\n`)
+        parsed = firstParsed
+        quality = firstQuality
+      } else {
+        __mixdogMemoryLog(`[cycle2] retry LLM error: ${err.message}\n`)
+        return { actions: null, rejected: new Set(), parseOk: false }
+      }
+    }
+  }
+  const finalIssue = gateQualitySummary(quality)
+  // duplicateVerdictIds are genuinely ambiguous (the same row got two conflicting
+  // verbs) — keep the full-skip. missingVerdictIds, by contrast, used to skip the
+  // WHOLE batch, so a handful of persistently-missing poison rows could livelock
+  // the gate. Partial-apply instead: keep the valid verdicts we did receive, just
+  // log the missing ids and leave those rows for a later run.
+  if (quality?.duplicateVerdictIds?.length) {
+    __mixdogMemoryLog(`[cycle2] duplicate verdict coverage after retry — skipping batch (${finalIssue})\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+  if (quality?.missingVerdictIds?.length) {
+    __mixdogMemoryLog(`[cycle2] missing verdicts after retry — partial apply, leaving ids=${quality.missingVerdictIds.join(',')} for a later run (${finalIssue})\n`)
+  }
+  // A response made up solely of why/core lines parses "ok" yet carries zero
+  // primary (status-verb) verdicts. Without this guard parseOk stays true and
+  // the caller treats the batch as a clean no-op, masking the coverage failure
+  // and marking the rows reviewed. Fail the parse so the rows are re-queued.
+  const primaryCount = (parsed.actions || []).filter(a => a?.action !== 'core').length
+  if (rows.length > 0 && primaryCount === 0) {
+    __mixdogMemoryLog(`[cycle2] gate produced zero primary verdicts for ${rows.length} rows — failing parse\n`)
+    return { actions: null, rejected: new Set(), parseOk: false, missingIds: [...statusById.keys()] }
+  }
+  const incompletePromotionIds = uniqueIds([
+    ...(quality?.missingSupportIds || []),
+    ...(quality?.missingCoreIds || []),
+  ])
+  if (incompletePromotionIds.length > 0) {
+    __mixdogMemoryLog(`[cycle2] incomplete non-archive verdicts rejected after retry ids=${incompletePromotionIds.join(',')} (${finalIssue})\n`)
+    parsed = stripUnsupportedPromotions(parsed, incompletePromotionIds)
+  }
+  return {
+    actions: parsed.actions,
+    rejected: parsed.rejected,
+    parseOk: true,
+    missingIds: quality?.missingVerdictIds || [],
+  }
+}
+
+// ─── Sonnet cascade ──────────────────────────────────────────────────────────
+
+// Sonnet re-judge over first-pass keep verdicts. Sonnet sees rules + summary
+// and returns binary keep/drop. Failures fail-open (preserve first-pass).
+export async function sonnetCascade(candidates, rulesDigest, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  if (!candidates || candidates.length === 0) return new Map()
+  const lines = candidates.map(c =>
+    `id:${c.id} status:${c.status} verb:${c.verb} cat:${c.category} el:${c.element} sm:${String(c.summary || '').slice(0, 200)}${c.core ? ` core:${String(c.core).slice(0, 200)}` : ''}`,
+  ).join('\n')
+  const prompt = [
+    `Final gate over first-pass keep verdicts.`,
+    `Keep a candidate ONLY if it lands in one of three layers: L1 relationship/communication`,
+    `(user identity, address form, reply-style preferences, disliked patterns); L2 behavior rules`,
+    `(principles the user corrected/insisted on, hard safety boundaries, quality bars); or L3 current`,
+    `map (one-line project-landscape summaries, live long-running goals, environment anchors documented`,
+    `nowhere else). For a past decision/failure, keep only the one-line lesson that still constrains`,
+    `behavior, else drop. DROP anything whose source of truth is code, rules files, or skill docs, plus`,
+    `implementation specs, code-internal constants, measurements, resolved-bug stories, status snapshots,`,
+    `and duplicates of source-of-truth rules.`,
+    `When a candidate has a core: field, judge THAT extracted one-line lesson (the entry will live as`,
+    `that line), not the raw narrative in el:/sm:.`,
+    ``,
+    `Source-of-truth rules (excerpt — DO NOT duplicate in memory):`,
+    String(rulesDigest || '').slice(0, 4000),
+    ``,
+    `Candidates:`,
+    lines,
+    ``,
+    `Reply one line per id: "<id>|keep" to retain, "<id>|drop" to reject.`,
+    `NO prose, NO preamble, NO meta-commentary. First character must be a digit.`,
+  ].join('\n')
+
+  // Hardcoded — resolveMaintenancePreset falls back to first preset (HAIKU)
+  // when no binding exists, which would defeat the cascade. SONNET HIGH
+  // matches the worker pool's default preset id from agent-config.
+  const preset = options.cascadePreset || 'SONNET HIGH'
+  const llmCall = typeof options?.callLlm === 'function' ? options.callLlm : callAgentDispatch
+  let raw
+  try {
+    raw = await llmCall({
+      agent: 'cycle2-agent',
+      taskType: 'maintenance',
+      mode: 'cycle2-cascade',
+      preset,
+      timeout: 600000,
+      cwd: null,
+    }, prompt)
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    __mixdogMemoryLog(`[cycle2] cascade error: ${err.message} — fail-open\n`)
+    return new Map()
+  }
+  throwIfAborted(signal)
+
+  const verdicts = new Map()
+  for (const line of String(raw ?? '').split('\n')) {
+    throwIfAborted(signal)
+    const trimmed = line.trim()
+    if (!trimmed) continue
+    if (trimmed.startsWith('//') || trimmed.startsWith('#') || trimmed.startsWith('```')) continue
+    const parts = trimmed.split('|')
+    if (parts.length < 2) continue
+    const id = Number(parts[0].trim())
+    const v = parts[1].trim().toLowerCase()
+    if (Number.isFinite(id) && (v === 'keep' || v === 'drop')) verdicts.set(id, v)
+  }
+  __mixdogMemoryLog(`[cycle2] cascade evaluated=${candidates.length} drops=${[...verdicts.values()].filter(v => v === 'drop').length}\n`)
+  return verdicts
+}
+
+export { NON_ARCHIVE_VERBS, requiredCoreIdForAction }