opencode-agent-skill 13.0.0-beta.2 → 14.4.0

package/lib/mcp-tool-policy.mjs

@@ -0,0 +1,19 @@
+function firstObject(...values) { return values.find((value) => value && typeof value === "object" && !Array.isArray(value)) || {} }
+function explicitBoolean(value) { return typeof value === "boolean" ? value : null }
+export function normalizeMcpAnnotations(tool = {}) {
+  const raw = firstObject(tool.annotations, tool.metadata?.annotations, tool._meta?.annotations, tool._meta?.mcp?.annotations)
+  const annotations = { readOnlyHint: explicitBoolean(raw.readOnlyHint), destructiveHint: explicitBoolean(raw.destructiveHint), idempotentHint: explicitBoolean(raw.idempotentHint), openWorldHint: explicitBoolean(raw.openWorldHint) }
+  return { schemaVersion: 1, explicit: Object.values(annotations).some((value) => value !== null), ...annotations }
+}
+export function mcpExecutionPolicy(tool = {}) {
+  const annotations = normalizeMcpAnnotations(tool)
+  const destructive = annotations.destructiveHint === true
+  const readOnly = annotations.readOnlyHint === true && !destructive
+  const idempotent = annotations.idempotentHint === true && !destructive
+  return { schemaVersion: 1, tool: String(tool?.name || ""), annotations, fastAllowed: readOnly, confirmationRequired: destructive, retryAllowed: idempotent, externalEvidenceBoundary: annotations.openWorldHint === true, trust: "hint-only" }
+}
+export function mcpPolicyMap(tools = []) {
+  const map = new Map()
+  for (const tool of Array.isArray(tools) ? tools : []) { const name = String(tool?.name || "").trim(); if (name) map.set(name, mcpExecutionPolicy(tool)) }
+  return map
+}

package/lib/memory-engine.mjs

@@ -0,0 +1,493 @@
+import { createHash } from "node:crypto"
+import { existsSync } from "node:fs"
+import { mkdir, readFile, rename, rm, writeFile } from "node:fs/promises"
+import { spawnSync } from "node:child_process"
+import path from "node:path"
+import { evidenceExists, putEvidence } from "./evidence-store.mjs"
+
+const MEMORY_DIR = ".ues-memory"
+const MEMORY_FILE = "MEMORY.json"
+const SCHEMA_VERSION = 1
+const VALID_TYPES = new Set(["episodic", "semantic", "procedural", "failure", "decision"])
+const VALID_SCOPES = new Set(["global", "project", "module", "file"])
+const STOP = new Set([
+  "this","that","with","from","into","then","than","when","where","what","your","have","will","task","code",
+  "file","files","change","changes","được","các","cho","với","trong","này","một","những","không","theo",
+])
+
+function now() {
+  return new Date().toISOString()
+}
+
+function clamp(value, fallback = 0.7) {
+  const number = Number(value)
+  return Number.isFinite(number) ? Math.max(0, Math.min(1, number)) : fallback
+}
+
+function uniq(values) {
+  return [...new Set((values || []).filter(Boolean))]
+}
+
+function normalizeFile(value) {
+  return String(value || "").replaceAll("\\", "/").replace(/^\.\/+/, "")
+}
+
+function tokens(value) {
+  return String(value || "")
+    .toLowerCase()
+    .split(/[^\p{L}\p{N}_$.-]+/u)
+    .map((item) => item.replace(/^[-.$]+|[-.$]+$/g, ""))
+    .filter((item) => item.length >= 2 && !STOP.has(item))
+}
+
+function normalizedContent(value) {
+  return String(value || "").trim().replace(/\s+/g, " ")
+}
+
+function idFor(input) {
+  const key = input.key || [
+    input.scope || "project",
+    input.type || "episodic",
+    normalizedContent(input.content),
+  ].join(":")
+  return createHash("sha256").update(String(key)).digest("hex").slice(0, 20)
+}
+
+function memoryPath(root) {
+  return path.join(path.resolve(root), MEMORY_DIR, MEMORY_FILE)
+}
+
+async function atomicJson(file, value) {
+  await mkdir(path.dirname(file), { recursive: true })
+  const temp = file + "." + process.pid + "." + Date.now() + ".tmp"
+  await writeFile(temp, JSON.stringify(value, null, 2) + "\n", "utf8")
+  try {
+    await rename(temp, file)
+  } catch (error) {
+    await rm(temp, { force: true }).catch(() => {})
+    throw error
+  }
+}
+
+export async function readMemoryState(root = process.cwd()) {
+  const file = memoryPath(root)
+  if (!existsSync(file)) {
+    return { schemaVersion: SCHEMA_VERSION, updatedAt: null, memories: [] }
+  }
+  try {
+    const parsed = JSON.parse(await readFile(file, "utf8"))
+    return {
+      schemaVersion: SCHEMA_VERSION,
+      updatedAt: parsed.updatedAt || null,
+      memories: Array.isArray(parsed.memories) ? parsed.memories : [],
+    }
+  } catch {
+    return { schemaVersion: SCHEMA_VERSION, updatedAt: null, memories: [], invalid: true }
+  }
+}
+
+async function writeMemoryState(root, memories) {
+  const state = { schemaVersion: SCHEMA_VERSION, updatedAt: now(), memories }
+  await atomicJson(memoryPath(root), state)
+  return state
+}
+
+export async function proposeMemory(root, input = {}) {
+  const content = normalizedContent(input.content)
+  if (!content) throw new Error("memory content is required")
+  const type = VALID_TYPES.has(input.type) ? input.type : "episodic"
+  const scope = VALID_SCOPES.has(input.scope) ? input.scope : "project"
+  const state = await readMemoryState(root)
+  const id = idFor({ ...input, type, scope, content })
+  const timestamp = now()
+  const existing = state.memories.find((item) => item.id === id)
+  if (existing) {
+    const reinforced = {
+      ...existing,
+      lastSeenAt: timestamp,
+      reinforcementCount: Number(existing.reinforcementCount || 1) + 1,
+      evidenceRefs: uniq([...(existing.evidenceRefs || []), ...(input.evidenceRefs || [])]),
+      files: uniq([...(existing.files || []), ...(input.files || []).map(normalizeFile)]),
+      tags: uniq([...(existing.tags || []), ...(input.tags || [])]),
+      confidence: Math.max(clamp(existing.confidence), clamp(input.confidence, existing.confidence ?? 0.7)),
+      taskClass: input.taskClass || existing.taskClass || null,
+      expiresAt: input.expiresAt || existing.expiresAt || null,
+    }
+    await writeMemoryState(root, state.memories.map((item) => item.id === id ? reinforced : item))
+    return reinforced
+  }
+
+  const memory = {
+    id,
+    key: input.key || null,
+    type,
+    scope,
+    content,
+    status: "candidate",
+    confidence: clamp(input.confidence),
+    evidenceRefs: uniq(input.evidenceRefs || []),
+    files: uniq((input.files || []).map(normalizeFile)),
+    tags: uniq(input.tags || []),
+    sourceTask: input.sourceTask || null,
+    sourceCommit: input.sourceCommit || null,
+    taskClass: input.taskClass || null,
+    expiresAt: input.expiresAt || null,
+    createdAt: timestamp,
+    lastSeenAt: timestamp,
+    lastUsedAt: null,
+    useCount: 0,
+    reinforcementCount: 1,
+    verifiedAt: null,
+    verifier: null,
+    supersededBy: null,
+  }
+  await writeMemoryState(root, [...state.memories, memory])
+  return memory
+}
+
+export async function verifyMemory(root, id, verification = {}) {
+  if (String(verification.verdict || "").toUpperCase() !== "PASS") {
+    throw new Error("memory verification requires verdict PASS")
+  }
+  if (!verification.verifier) throw new Error("memory verification requires an independent verifier identity")
+  const state = await readMemoryState(root)
+  const current = state.memories.find((item) => item.id === id)
+  if (!current) throw new Error("unknown memory: " + id)
+  if (current.status === "superseded") throw new Error("cannot verify a superseded memory")
+
+  const refs = uniq([...(current.evidenceRefs || []), ...(verification.evidenceRefs || [])])
+  if (!refs.length) throw new Error("memory verification requires at least one durable evidence reference")
+  for (const ref of refs) {
+    if (!(await evidenceExists(root, ref))) throw new Error("memory evidence reference is missing: " + ref)
+  }
+
+  const verified = {
+    ...current,
+    status: "verified",
+    evidenceRefs: refs,
+    verifier: String(verification.verifier),
+    verifiedAt: now(),
+    lastSeenAt: now(),
+    confidence: Math.max(clamp(current.confidence), clamp(verification.confidence, current.confidence ?? 0.7)),
+  }
+  await writeMemoryState(root, state.memories.map((item) => item.id === id ? verified : item))
+  return verified
+}
+
+export async function supersedeMemory(root, id, replacementId) {
+  if (!replacementId || replacementId === id) throw new Error("replacement memory id must be different")
+  const state = await readMemoryState(root)
+  const current = state.memories.find((item) => item.id === id)
+  const replacement = state.memories.find((item) => item.id === replacementId)
+  if (!current) throw new Error("unknown memory: " + id)
+  if (!replacement) throw new Error("unknown replacement memory: " + replacementId)
+  if (replacement.status !== "verified") throw new Error("replacement memory must be verified before supersession")
+  const timestamp = now()
+  const superseded = {
+    ...current,
+    status: "superseded",
+    supersededBy: replacementId,
+    lastSeenAt: timestamp,
+  }
+  await writeMemoryState(root, state.memories.map((item) => item.id === id ? superseded : item))
+  return superseded
+}
+
+function tokenVector(value, dims = 96) {
+  const vector = new Array(dims).fill(0)
+  for (const token of tokens(value)) {
+    let hash = 2166136261
+    for (let index = 0; index < token.length; index += 1) {
+      hash ^= token.charCodeAt(index)
+      hash = Math.imul(hash, 16777619) >>> 0
+    }
+    const slot = hash % dims
+    const sign = (hash & 0x100) === 0 ? 1 : -1
+    vector[slot] += sign
+  }
+  return vector
+}
+
+function cosine(a, b) {
+  let dot = 0
+  let left = 0
+  let right = 0
+  for (let index = 0; index < Math.min(a.length, b.length); index += 1) {
+    dot += a[index] * b[index]
+    left += a[index] * a[index]
+    right += b[index] * b[index]
+  }
+  return left && right ? dot / Math.sqrt(left * right) : 0
+}
+
+function bm25Scores(memories, queryTokens) {
+  const docs = memories.map((item) => tokens([item.content, ...(item.tags || []), ...(item.files || [])].join(" ")))
+  const avgLength = docs.length ? docs.reduce((sum, doc) => sum + doc.length, 0) / docs.length : 1
+  const df = new Map()
+  for (const term of queryTokens) {
+    df.set(term, docs.reduce((count, doc) => count + (doc.includes(term) ? 1 : 0), 0))
+  }
+  const k1 = 1.2
+  const b = 0.75
+  return docs.map((doc) => {
+    const counts = new Map()
+    for (const token of doc) counts.set(token, (counts.get(token) || 0) + 1)
+    let score = 0
+    for (const term of queryTokens) {
+      const tf = counts.get(term) || 0
+      if (!tf) continue
+      const freq = df.get(term) || 0
+      const idf = Math.log(1 + (memories.length - freq + 0.5) / (freq + 0.5))
+      const denom = tf + k1 * (1 - b + b * (doc.length / Math.max(1, avgLength)))
+      score += idf * ((tf * (k1 + 1)) / denom)
+    }
+    return score
+  })
+}
+
+function commonPrefixParts(left, right) {
+  const a = normalizeFile(left).split("/")
+  const b = normalizeFile(right).split("/")
+  let count = 0
+  while (count < a.length && count < b.length && a[count] === b[count]) count += 1
+  return count
+}
+
+function graphAffinity(queryFiles, memoryFiles) {
+  if (!queryFiles.length || !memoryFiles.length) return 0
+  let best = 0
+  for (const left of queryFiles) {
+    for (const right of memoryFiles) {
+      if (left === right) best = Math.max(best, 1)
+      else {
+        const common = commonPrefixParts(left, right)
+        if (common >= 2) best = Math.max(best, 0.8)
+        else if (common === 1) best = Math.max(best, 0.45)
+      }
+    }
+  }
+  return best
+}
+
+function taskClassAffinity(queryTaskClass, memoryTaskClass) {
+  if (!queryTaskClass || !memoryTaskClass) return 0
+  return String(queryTaskClass).toLowerCase() === String(memoryTaskClass).toLowerCase() ? 1 : 0
+}
+
+function scopeAffinity(scope, graph) {
+  if (scope === "file") return graph > 0 ? 1 : 0
+  if (scope === "module") return graph > 0 ? 0.8 : 0
+  if (scope === "project") return 0.3
+  if (scope === "global") return 0.15
+  return 0
+}
+
+function isExpired(item, at = Date.now()) {
+  const expires = Date.parse(item?.expiresAt || "")
+  return Boolean(expires && expires <= at)
+}
+
+function recencyScore(value) {
+  const timestamp = Date.parse(value || "")
+  if (!timestamp) return 0
+  const ageDays = Math.max(0, (Date.now() - timestamp) / 86_400_000)
+  return 1 / (1 + ageDays / 30)
+}
+
+function rankMap(rows, key) {
+  return new Map(
+    [...rows]
+      .filter((row) => Number(row[key]) > 0)
+      .sort((a, b) => b[key] - a[key] || a.item.id.localeCompare(b.item.id))
+      .map((row, index) => [row.item.id, index + 1]),
+  )
+}
+
+export function buildMemorySnapshot(memories = []) {
+  const entries = (Array.isArray(memories) ? memories : [])
+    .map((item) => ({
+      id: item.id || null,
+      type: item.type || null,
+      scope: item.scope || null,
+      content: normalizedContent(item.content),
+      confidence: clamp(item.confidence),
+      files: uniq((item.files || []).map(normalizeFile)).sort(),
+      taskClass: item.taskClass || null,
+      verifiedAt: item.verifiedAt || null,
+      evidenceRefs: uniq(item.evidenceRefs || []).sort(),
+    }))
+    .filter((item) => item.id && item.content)
+    .sort((a, b) => String(a.id).localeCompare(String(b.id)))
+  const generation = createHash("sha256")
+    .update(JSON.stringify(entries))
+    .digest("hex")
+    .slice(0, 20)
+  return {
+    schemaVersion: 1,
+    generation,
+    entryCount: entries.length,
+    entries,
+  }
+}
+
+export async function retrieveMemories(root, query, options = {}) {
+  const state = await readMemoryState(root)
+  const memories = state.memories.filter((item) => item.status === "verified" && !item.supersededBy && !isExpired(item))
+  const queryTokens = uniq(tokens(query))
+  const queryVector = tokenVector(query)
+  const queryFiles = uniq((options.files || []).map(normalizeFile))
+  const lexical = bm25Scores(memories, queryTokens)
+  const rows = memories.map((item, index) => ({
+    item,
+    lexical: lexical[index] || 0,
+    semantic: Math.max(0, cosine(queryVector, tokenVector(item.content))),
+    graph: graphAffinity(queryFiles, item.files || []),
+    taskClass: taskClassAffinity(options.taskClass, item.taskClass),
+    recency: recencyScore(item.lastUsedAt || item.lastSeenAt || item.verifiedAt || item.createdAt),
+    confidence: clamp(item.confidence),
+    scope: 0,
+  }))
+  for (const row of rows) row.scope = scopeAffinity(row.item.scope, row.graph)
+  const ranks = {
+    lexical: rankMap(rows, "lexical"),
+    semantic: rankMap(rows, "semantic"),
+    graph: rankMap(rows, "graph"),
+    taskClass: rankMap(rows, "taskClass"),
+    recency: rankMap(rows, "recency"),
+  }
+  for (const row of rows) {
+    let rrf = 0
+    for (const key of ["lexical", "semantic", "graph", "taskClass", "recency"]) {
+      const rank = ranks[key].get(row.item.id)
+      if (rank) rrf += 1 / (60 + rank)
+    }
+    row.score = rrf + row.confidence * 0.01 + row.graph * 0.01 + row.scope * 0.006 + row.taskClass * 0.008
+  }
+  const limit = Math.max(1, Math.min(Number(options.limit || 6), 30))
+  const results = rows
+    .filter((row) => row.lexical > 0 || row.semantic > 0.05 || row.graph > 0)
+    .sort((a, b) => b.score - a.score || b.confidence - a.confidence || a.item.id.localeCompare(b.item.id))
+    .slice(0, limit)
+    .map((row) => ({
+      ...row.item,
+      retrieval: {
+        score: Number(row.score.toFixed(8)),
+        lexical: Number(row.lexical.toFixed(6)),
+        semantic: Number(row.semantic.toFixed(6)),
+        graph: Number(row.graph.toFixed(6)),
+        taskClass: Number(row.taskClass.toFixed(6)),
+        scope: Number(row.scope.toFixed(6)),
+        recency: Number(row.recency.toFixed(6)),
+      },
+    }))
+  if (options.touch === true && results.length) {
+    const used = new Set(results.map((item) => item.id))
+    const timestamp = now()
+    await writeMemoryState(root, state.memories.map((item) => used.has(item.id)
+      ? { ...item, lastUsedAt: timestamp, useCount: Number(item.useCount || 0) + 1 }
+      : item))
+  }
+  return {
+    schemaVersion: SCHEMA_VERSION,
+    query: String(query || ""),
+    eligible: memories.length,
+    results,
+    snapshot: buildMemorySnapshot(results),
+  }
+}
+
+export async function memoryStatus(root = process.cwd()) {
+  const state = await readMemoryState(root)
+  const byStatus = {}
+  const byType = {}
+  let expired = 0
+  let used = 0
+  for (const memory of state.memories) {
+    byStatus[memory.status] = (byStatus[memory.status] || 0) + 1
+    byType[memory.type] = (byType[memory.type] || 0) + 1
+    if (isExpired(memory)) expired += 1
+    if (Number(memory.useCount || 0) > 0) used += 1
+  }
+  return {
+    schemaVersion: SCHEMA_VERSION,
+    root: path.dirname(memoryPath(root)),
+    updatedAt: state.updatedAt,
+    entries: state.memories.length,
+    expired,
+    used,
+    byStatus,
+    byType,
+  }
+}
+
+function changedFiles(root) {
+  const result = spawnSync("git", ["status", "--porcelain=v1", "--untracked-files=all"], {
+    cwd: path.resolve(root),
+    encoding: "utf8",
+    windowsHide: true,
+    maxBuffer: 4 * 1024 * 1024,
+  })
+  if (result.status !== 0) return []
+  return uniq(
+    String(result.stdout || "")
+      .split(/\r?\n/)
+      .filter(Boolean)
+      .map((line) => line.slice(3).trim())
+      .map((value) => value.includes(" -> ") ? value.split(" -> ").at(-1) : value)
+      .map(normalizeFile),
+  ).slice(0, 80)
+}
+
+function compactVerification(value, limit = 1200) {
+  const text = String(value || "").replace(/\s+/g, " ").trim()
+  return text.length <= limit ? text : text.slice(0, limit) + "...[truncated]"
+}
+
+export async function recordVerifiedTaskMemory(root, input = {}) {
+  const task = normalizedContent(input.task)
+  if (!task) throw new Error("verified task memory requires task text")
+  const explicitFiles = uniq((input.files || []).map(normalizeFile))
+  const files = explicitFiles.length ? explicitFiles : changedFiles(root)
+  const verifier = input.verifier || "ues-verifier"
+  const verificationSummary = compactVerification(input.verifierOutput)
+  const integrationSummary = compactVerification(input.integrationOutput)
+  const evidence = await putEvidence(root, {
+    schemaVersion: 1,
+    kind: "verified-task-memory-receipt",
+    task,
+    verifier,
+    files,
+    verification: verificationSummary,
+    integration: integrationSummary || null,
+    recordedAt: now(),
+  }, {
+    kind: "verified-task-memory",
+    source: input.sourceTask || task.slice(0, 200),
+    summary: "Verified task outcome eligible for persistent memory",
+  })
+  const content = [
+    `Verified engineering task: ${task}`,
+    files.length ? `Files: ${files.join(", ")}` : "",
+    verificationSummary ? `Verification: ${verificationSummary}` : "",
+  ].filter(Boolean).join("\n")
+  const candidate = await proposeMemory(root, {
+    key: input.key || `verified-task:${task}:${files.slice().sort().join(",")}`,
+    type: input.type || "episodic",
+    scope: input.scope || "project",
+    content,
+    confidence: input.confidence ?? (integrationSummary ? 0.92 : 0.84),
+    evidenceRefs: [evidence.ref],
+    files,
+    tags: uniq(["verified-task", ...(input.tags || [])]),
+    sourceTask: input.sourceTask || task,
+    sourceCommit: input.sourceCommit || null,
+    taskClass: input.taskClass || null,
+    expiresAt: input.expiresAt || null,
+  })
+  return verifyMemory(root, candidate.id, {
+    verdict: "PASS",
+    verifier,
+    evidenceRefs: [evidence.ref],
+    confidence: candidate.confidence,
+  })
+}

package/lib/model-performance.mjs

@@ -80,20 +80,44 @@ export function recordPerformanceOutcome(history = {}, outcome = {}) {
   return { ...normalized, [model]: { ...(normalized[model] || {}), [taskClass]: next } }
 }
 
+function wilsonLowerBound(successes, samples, z = 1.96) {
+  if (samples <= 0) return 0
+  const p = successes / samples
+  const z2 = z * z
+  const denominator = 1 + z2 / samples
+  const center = p + z2 / (2 * samples)
+  const margin = z * Math.sqrt((p * (1 - p) + z2 / (4 * samples)) / samples)
+  return Math.max(0, (center - margin) / denominator)
+}
+
 function performanceAdjustment(record, minSamples) {
   const normalized = normalizePerformanceRecord(record)
-  if (normalized.samples <= 0) return { adjustment: 0, confidence: 0, record: normalized }
-  const confidence = Math.min(1, normalized.samples / Math.max(1, minSamples))
-  if (normalized.samples < minSamples) return { adjustment: 0, confidence, record: normalized }
-  const correctness = (normalized.passRate - 0.5) * 80
+  if (normalized.samples <= 0) return { adjustment: 0, confidence: 0, record: normalized, lowerBound: 0 }
+  const confidence = Math.min(1, normalized.samples / Math.max(1, minSamples * 2))
+  const lowerBound = wilsonLowerBound(normalized.successes, normalized.samples)
+
+  // Do not let a tiny sample reorder weak models. Only empirical evidence with
+  // enough observations is allowed to move routing.
+  if (normalized.samples < minSamples) {
+    return { adjustment: 0, confidence, record: normalized, lowerBound }
+  }
+
+  const correctness = (lowerBound - 0.5) * 80
   const retryPenalty = Math.min(20, normalized.avgRetries * 5)
-  const latencyPenalty = normalized.avgLatencyMs > 0 ? Math.min(10, Math.max(0, Math.log10(Math.max(1, normalized.avgLatencyMs / 1000)) * 3)) : 0
-  return { adjustment: (correctness - retryPenalty - latencyPenalty) * confidence, confidence, record: normalized }
+  const latencyPenalty = normalized.avgLatencyMs > 0
+    ? Math.min(10, Math.max(0, Math.log10(Math.max(1, normalized.avgLatencyMs / 1000)) * 3))
+    : 0
+  return {
+    adjustment: (correctness - retryPenalty - latencyPenalty) * confidence,
+    confidence,
+    record: normalized,
+    lowerBound,
+  }
 }
 
 export function rerankCapabilitySelection(selection = {}, history = {}, options = {}) {
   const taskClass = inferTaskClass(options.text || "", { taskClass: options.taskClass })
-  const minSamples = Math.max(1, Number(options.minSamples || 3))
+  const minSamples = Math.max(1, Number(options.minSamples || 8))
   const normalized = normalizePerformanceHistory(history)
   const candidates = (selection.candidates || []).map((candidate) => {
     const record = normalized[candidate.id]?.[taskClass] || normalized[candidate.id]?.overall || null
@@ -104,10 +128,11 @@ export function rerankCapabilitySelection(selection = {}, history = {}, options
       empiricalTaskClass: taskClass,
       empiricalEvidence: evidence.record,
       empiricalConfidence: Number(evidence.confidence.toFixed(4)),
+      empiricalPassLowerBound: Number(evidence.lowerBound.toFixed(4)),
       adjustedScore: Number((Number(candidate.score || 0) + evidence.adjustment).toFixed(6)),
     }
   })
   const eligible = candidates.filter((candidate) => candidate.eligible)
     .sort((a, b) => b.adjustedScore - a.adjustedScore || b.baseScore - a.baseScore)
   return { ...selection, selected: eligible[0] || null, candidates, taskClass, empirical: true }
-}
+}

package/lib/model-policy.mjs

@@ -53,7 +53,7 @@ export function defaultModelPolicy() {
     roleTiers: { ...ROLE_TIERS },
     capabilities: {},
     performance: {},
-    performanceMinSamples: 3,
+    performanceMinSamples: 8,
   }
 }
 
@@ -99,7 +99,7 @@ export function resolveCapabilityModel(role, attempt, taskText = "", taskPolicy
   })
   const taskClass = inferTaskClass(taskText, facts)
   const selection = rerankCapabilitySelection(staticSelection, config.performance || {}, {
-    taskClass, text: taskText, minSamples: config.performanceMinSamples || 3,
+    taskClass, text: taskText, minSamples: config.performanceMinSamples || 8,
   })
   const capabilityEnforced = config.enabled === true && candidates.length > 0
   if (capabilityEnforced && !selection.selected) {
@@ -132,4 +132,4 @@ export function resolveCapabilityModel(role, attempt, taskText = "", taskPolicy
     capabilityFallback: false,
     capabilityBlocked: false,
   }
-}
+}