dsh-omni-router 2.2.1 → 2.4.0

package/README.md

@@ -3,6 +3,10 @@
 > GitHub: https://github.com/qwe225380/dsh-omni-router
 > [中文说明](./README.zh-CN.md)
 
+> **Fast by default. Smart when needed. Proven when done.**
+> **Decide · Compose · Verify**
+> Install it. Keep coding.
+
 **Omni** 是一个 DeepSeek Harness 的 **reliability and orchestration control plane**，定位为 Intelligence Runtime：把复杂度路由、任务编译、项目上下文、能力编排、Mission DAG、Runtime 预算、证据验证、Memory/Learning、Benchmark 组合成一个统一的、自动运行的工程闭环。
 
 ## Why Omni Router?
@@ -74,6 +78,8 @@
 - **P1/P3 能力**：Real capability sandbox（已接入 `omni_mission_run`，按 role + capability 生成 toolFilter）、Capability Resolver v2（risk/reliability 综合打分）、OmniBench v2 runner（manifest schema + prepare + prompt generation + `--exec` 本地执行/结果采集 + DAG-to-plan compile）。
 - **2.1 收敛增强**：Mission Resume（`omni_mission_resume` 跨 session 继续 DAG，`runDagLoop` 支持 `onProgress` 周期保存）、Harness evidence extraction（`extractHarnessEvidence` 从 commands/tests/toolCalls/EVIDENCE_JSON 提取机器可验证证据）、context token budget（`maxContextTokens` 真正限制检索输出）、semantic failure retrieval + evidence-backed skill distillation、OmniBench v2 real runner。
 - **2.2 Capability Auto-Provisioning（优化7）**：Capability Auditor（baseline/task-time audit）、Gap Detection、Plugin/Skill Quality Scoring、Minimal Capability Set Solver、Discovery Adapters（marketplace/hub/static registry）、Trusted Auto Provisioning（recommend/auto-trusted/manual）、Post-install probe、Install rollback transaction、Capability performance learning / redundancy detection。
+- **2.3 Progressive Intelligence（优化8）**：L0 Direct / L1 Assisted / L2 Orchestrated / L3 Guarded 四级渐进智能；Context Capsule（最小相关上下文 + 动态 needs）；Strategy Shift（相同失败策略自动换策略，不再盲目 retry）；`omni_explain` / `omni_doctor` 用户入口。
+- **2.4 Reliability Kernel（优化9）**：TaskContract 统一任务契约；Intervention Gate（默认 NOOP，按 Expected Utility 决定是否介入）；OmniEvent / Host Interface（宿主无关事件与能力协商）；Evidence Trust（T0-T4）与 Invalidation；Context Freshness（workspace fingerprint）；Recovery Policy（retry/repair/expand/replan/escalate/stop）。
 - **TaskDecision**：`createTaskDecision` 生成唯一决策对象，Policy/Runtime 统一消费，避免重复 classify 分叉。
 - **Evidence Protocol**：`src/evidence.mjs` 用结构化 command/file/test/finding 证据判定 PASS/FAIL；`isQaPass` 已支持从 QA 输出解析 JSON evidence，不再只信文本。
 - **Task Compiler**：`compileTask` 生成 objective / constraints / non-goals / acceptance / hidden assumptions / ambiguities / invariants / risk / artifacts；`compileTaskWithLLM` 可用 LLM 增强。
@@ -155,7 +161,7 @@ Then restart DSH and select **Omni Router** in a new session.
 - `/omni direct` — enter direct mode.
 - `/omni mode spec|react|balanced` — set thinking mode.
 - `/omni reroute plan|direct` — adaptively reroute current task.
-- Model tools: `omni_status`, `omni_plan`, `omni_direct`, `omni_mode`, `omni_reroute`, `omni_delegate`, `omni_memory`, `omni_benchmark`, `omni_mission_run`, `omni_mission_resume`, `omni_capability_audit`, `omni_capability_provision`, `omni_capability_probe`, `omni_capability_performance`, `omni_visual_check`.
+- Model tools: `omni_status`, `omni_explain`, `omni_doctor`, `omni_plan`, `omni_direct`, `omni_mode`, `omni_reroute`, `omni_delegate`, `omni_memory`, `omni_benchmark`, `omni_mission_run`, `omni_mission_resume`, `omni_capability_audit`, `omni_capability_provision`, `omni_capability_probe`, `omni_capability_performance`, `omni_visual_check`.
 
 ## Configuration
 

package/README.zh-CN.md

@@ -3,6 +3,10 @@
 > GitHub：https://github.com/qwe225380/dsh-omni-router
 > [English](./README.md)
 
+> **默认快，必要时聪明，完成必须有证据。**
+> **Decide · Compose · Verify**
+> 安装即可使用，继续写代码。
+
 **Omni** 是一个 DeepSeek Harness 的 **reliability and orchestration control plane**，定位为 Intelligence Runtime：把复杂度路由、任务编译、项目上下文、能力编排、Mission DAG、Runtime 预算、证据验证、Memory/Learning、Benchmark 组合成一个统一的、自动运行的工程闭环。
 
 ## 为什么需要 Omni Router？
@@ -74,6 +78,8 @@
 - **P1/P3 能力**：Real capability sandbox（已接入 `omni_mission_run`，按 role + capability 生成 toolFilter）、Capability Resolver v2（risk/reliability 综合打分）、OmniBench v2 runner（manifest schema + prepare + prompt generation + `--exec` 本地执行/结果采集 + DAG-to-plan compile）。
 - **2.1 收敛增强**：Mission Resume（`omni_mission_resume` 跨 session 继续 DAG，`runDagLoop` 支持 `onProgress` 周期保存）、Harness evidence extraction（`extractHarnessEvidence` 从 commands/tests/toolCalls/EVIDENCE_JSON 提取机器可验证证据）、context token budget（`maxContextTokens` 真正限制检索输出）、semantic failure retrieval + evidence-backed skill distillation、OmniBench v2 real runner。
 - **2.2 Capability Auto-Provisioning（优化7）**：Capability Auditor（baseline/task-time audit）、Gap Detection、Plugin/Skill Quality Scoring、Minimal Capability Set Solver、Discovery Adapters（marketplace/hub/static registry）、Trusted Auto Provisioning（recommend/auto-trusted/manual）、Post-install probe、Install rollback transaction、Capability performance learning / redundancy detection。
+- **2.3 Progressive Intelligence（优化8）**：L0 Direct / L1 Assisted / L2 Orchestrated / L3 Guarded 四级渐进智能；Context Capsule（最小相关上下文 + 动态 needs）；Strategy Shift（相同失败策略自动换策略，不再盲目 retry）；`omni_explain` / `omni_doctor` 用户入口。
+- **2.4 Reliability Kernel（优化9）**：TaskContract 统一任务契约；Intervention Gate（默认 NOOP，按 Expected Utility 决定是否介入）；OmniEvent / Host Interface（宿主无关事件与能力协商）；Evidence Trust（T0-T4）与 Invalidation；Context Freshness（workspace fingerprint）；Recovery Policy（retry/repair/expand/replan/escalate/stop）。
 - **TaskDecision**：`createTaskDecision` 生成唯一决策对象，Policy/Runtime 统一消费，避免重复 classify 分叉。
 - **Evidence Protocol**：`src/evidence.mjs` 用结构化 command/file/test/finding 证据判定 PASS/FAIL；`isQaPass` 已支持从 QA 输出解析 JSON evidence，不再只信文本。
 - **Task Compiler**：`compileTask` 生成 objective / constraints / non-goals / acceptance / hidden assumptions / ambiguities / invariants / risk / artifacts；`compileTaskWithLLM` 可用 LLM 增强。
@@ -155,7 +161,7 @@ node scripts/install-preset.mjs
 - `/omni direct` — 进入直接执行模式。
 - `/omni mode spec|react|balanced` — 设置思维模式。
 - `/omni reroute plan|direct` — 动态切换当前任务路由。
-- 模型工具：`omni_status` / `omni_plan` / `omni_direct` / `omni_mode` / `omni_reroute` / `omni_delegate` / `omni_memory` / `omni_benchmark` / `omni_mission_run` / `omni_mission_resume` / `omni_capability_audit` / `omni_capability_provision` / `omni_capability_probe` / `omni_capability_performance` / `omni_visual_check`。
+- 模型工具：`omni_status` / `omni_explain` / `omni_doctor` / `omni_plan` / `omni_direct` / `omni_mode` / `omni_reroute` / `omni_delegate` / `omni_memory` / `omni_benchmark` / `omni_mission_run` / `omni_mission_resume` / `omni_capability_audit` / `omni_capability_provision` / `omni_capability_probe` / `omni_capability_performance` / `omni_visual_check`。
 
 ## 配置
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-omni-router",
-  "version": "2.2.1",
+  "version": "2.4.0",
   "description": "Omni - a reliability and orchestration control plane for DeepSeek Harness, designed to turn fast models into dependable long-horizon software-engineering agents.",
   "type": "module",
   "main": "./lib/index.js",
@@ -41,7 +41,7 @@
     }
   },
   "scripts": {
-    "test": "node test/omni-router.test.mjs && node test/agent-chain.test.mjs && node test/skill-suggest.test.mjs && node test/methodology.test.mjs && node test/compat.test.mjs && node test/project-brain.test.mjs && node test/mission-planner.test.mjs && node test/agent-runtime.test.mjs && node test/judge.test.mjs && node test/memory.test.mjs && node test/memory-v3.test.mjs && node test/memory-engine.test.mjs && node test/mission-resume.test.mjs && node test/engineering-benchmark.test.mjs && node test/benchmark-runner.test.mjs && node test/benchmark-results.test.mjs && node test/continuous-eval.test.mjs && node --no-warnings test/project-brain-v2.test.mjs && node --no-warnings test/project-brain-v3.test.mjs && node test/visual-qa.test.mjs && node test/evidence.test.mjs && node test/evidence-store.test.mjs && node test/evidence-engine.test.mjs && node test/task-decision.test.mjs && node test/task-compiler.test.mjs && node test/capability-brain.test.mjs && node test/capability-manifest.test.mjs && node test/capability-sandbox.test.mjs && node test/capability-auditor.test.mjs && node test/capability-quality.test.mjs && node test/capability-solver.test.mjs && node test/capability-provisioner.test.mjs && node test/capability-performance.test.mjs && node test/planner-dag.test.mjs && node test/dynamic-context.test.mjs && node test/mission-dag.test.mjs && node test/failure-taxonomy.test.mjs && node test/context-expansion.test.mjs && node test/hybrid-retrieval.test.mjs && node test/dependency-graph.test.mjs && node test/ast-provider.test.mjs && node test/experience-router.test.mjs && node test/benchmark-tasks.test.mjs && node test/bundle-installer.test.mjs && node test/omnibench-runner.test.mjs",
+    "test": "node test/omni-router.test.mjs && node test/agent-chain.test.mjs && node test/skill-suggest.test.mjs && node test/methodology.test.mjs && node test/compat.test.mjs && node test/project-brain.test.mjs && node test/mission-planner.test.mjs && node test/agent-runtime.test.mjs && node test/judge.test.mjs && node test/memory.test.mjs && node test/memory-v3.test.mjs && node test/memory-engine.test.mjs && node test/mission-resume.test.mjs && node test/engineering-benchmark.test.mjs && node test/benchmark-runner.test.mjs && node test/benchmark-results.test.mjs && node test/continuous-eval.test.mjs && node --no-warnings test/project-brain-v2.test.mjs && node --no-warnings test/project-brain-v3.test.mjs && node test/visual-qa.test.mjs && node test/evidence.test.mjs && node test/evidence-store.test.mjs && node test/evidence-engine.test.mjs && node test/task-decision.test.mjs && node test/task-compiler.test.mjs && node test/capability-brain.test.mjs && node test/capability-manifest.test.mjs && node test/capability-sandbox.test.mjs && node test/capability-auditor.test.mjs && node test/capability-quality.test.mjs && node test/capability-solver.test.mjs && node test/capability-provisioner.test.mjs && node test/capability-performance.test.mjs && node test/progressive-intelligence.test.mjs && node test/context-capsule.test.mjs && node test/strategy-shift.test.mjs && node test/task-contract.test.mjs && node test/intervention-gate.test.mjs && node test/omni-event.test.mjs && node test/host-interface.test.mjs && node test/evidence-trust.test.mjs && node test/context-freshness.test.mjs && node test/recovery-policy.test.mjs && node test/planner-dag.test.mjs && node test/dynamic-context.test.mjs && node test/mission-dag.test.mjs && node test/failure-taxonomy.test.mjs && node test/context-expansion.test.mjs && node test/hybrid-retrieval.test.mjs && node test/dependency-graph.test.mjs && node test/ast-provider.test.mjs && node test/experience-router.test.mjs && node test/benchmark-tasks.test.mjs && node test/bundle-installer.test.mjs && node test/omnibench-runner.test.mjs",
     "benchmark": "node benchmark/run.mjs",
     "benchmark:engineering": "node benchmark/engineering-run.mjs",
     "benchmark:compare": "node benchmark/compare.mjs",

package/src/agent-runtime.mjs

@@ -11,6 +11,7 @@
 
 import { buildPhaseTasks, decideReplan } from './mission-planner.mjs'
 import { applyObservationToDag, getReadyTasks, isMissionDagComplete, markTaskDone, selectReadyBatch } from './mission-dag.mjs'
+import { nextStrategy, shouldShiftStrategy } from './strategy-shift.mjs'
 
 export function createRuntimeState(mission, options = {}) {
   const phases = mission?.phases || []
@@ -187,6 +188,7 @@ export async function runDagLoop(dag, {
   let cost = 0
   let toolCalls = 0
   const actions = []
+  const failureHistory = new Map()
   const startTime = Date.now()
 
   const statusFromBudget = () => {
@@ -219,12 +221,21 @@ export async function runDagLoop(dag, {
       tokenUsage += Number(result?.tokenUsage || 0)
       cost += Number(result?.cost || 0)
       toolCalls += Number(result?.toolCalls || 0)
-      const isFailure = observation?.type === 'test_failure' || observation?.type === 'build_failure'
+      const isFailure = observation?.type === 'test_failure' || observation?.type === 'build_failure' || observation?.type === 'strategy_shift'
       if (isFailure) {
         replanCount += 1
         repairCount += 1
         sameActionCount = sameActionCount + 1
-        current = applyObservationToDag(current, { ...observation, taskId: task.id }, task.id)
+        const logicalKey = task.logicalId || task.id
+        const history = failureHistory.get(logicalKey) || []
+        history.push({ category: observation.type, reason: observation.reason, hypothesis: observation.hypothesis })
+        failureHistory.set(logicalKey, history)
+        let appliedObservation = { ...observation, taskId: task.id }
+        if (shouldShiftStrategy(history)) {
+          appliedObservation = { type: 'strategy_shift', reason: nextStrategy(history), taskId: task.id }
+          failureHistory.delete(logicalKey)
+        }
+        current = applyObservationToDag(current, appliedObservation, task.id)
       } else {
         sameActionCount = 0
         current = markTaskDone(current, task.id, observation)

package/src/context-capsule.mjs

@@ -0,0 +1,113 @@
+/**
+ * Context Capsule.
+ *
+ * Gives a weak model the smallest useful slice of context for the current
+ * task: a few relevant files, their callers/callees, related tests, invariants,
+ * and recent failures. This is the opposite of dumping the whole repository
+ * map into the prompt.
+ */
+
+const STOP_WORDS = new Set([
+  'the', 'and', 'for', 'with', 'that', 'this', 'from', 'are', 'was', 'were',
+  'not', 'but', 'you', 'your', 'fix', 'add', 'make', 'use', 'should', 'when',
+])
+
+export function buildContextCapsule(taskText, options = {}) {
+  const entries = options.entries || []
+  const files = options.files || {}
+  const graph = options.graph || {}
+  const needs = options.needs || []
+  const invariants = options.invariants || []
+  const failures = options.failures || []
+  const maxFiles = Number(options.maxFiles) || 6
+  const maxTotalChars = Number(options.maxTotalChars) || 8000
+  const maxFileChars = Number(options.maxFileChars) || 1200
+
+  const keywords = extractKeywords([taskText, ...needs])
+  const fileNames = entries.filter((e) => e.type === 'file').map((e) => e.name)
+  const scored = fileNames.map((name) => {
+    const content = files[name] || ''
+    const nameScore = scoreText(name, keywords)
+    const contentScore = content ? scoreText(content.slice(0, 2000), keywords) * 0.5 : 0
+    return { name, score: nameScore + contentScore }
+  }).sort((a, b) => b.score - a.score)
+  const selected = scored.slice(0, maxFiles).map((s) => s.name)
+
+  const lines = []
+  lines.push(`Task: ${taskText}`)
+  if (needs.length) lines.push(`Needs: ${needs.join('; ')}`)
+  if (selected.length) {
+    lines.push('', 'Relevant files:')
+    for (const name of selected) {
+      const content = files[name] || ''
+      lines.push(`--- ${name} ---`)
+      lines.push(content ? content.slice(0, maxFileChars) : '(no content indexed)')
+    }
+  }
+
+  const edgeLines = []
+  for (const name of selected) {
+    for (const edge of graph[name] || []) {
+      edgeLines.push(`${name} ${edge.kind || '->'} ${edge.to || edge}`)
+    }
+  }
+  if (edgeLines.length) {
+    lines.push('', 'Callers / callees:')
+    lines.push(...edgeLines.slice(0, 20))
+  }
+
+  const tests = fileNames.filter((n) =>
+    /\.(test|spec)\./i.test(n) &&
+    (selected.includes(n) || keywords.some((k) => n.toLowerCase().includes(k))),
+  )
+  if (tests.length) {
+    lines.push('', 'Tests:')
+    lines.push(...tests.map((t) => `- ${t}`))
+  }
+
+  if (invariants.length) {
+    lines.push('', 'Invariants:')
+    lines.push(...invariants.map((i) => `- ${i}`))
+  }
+
+  if (failures.length) {
+    lines.push('', 'Recent failures:')
+    lines.push(...failures.map((f) => `- ${f}`))
+  }
+
+  let text = lines.join('\n')
+  if (text.length > maxTotalChars) {
+    text = `${text.slice(0, maxTotalChars)}\n… (context capsule truncated)`
+  }
+  return text
+}
+
+export function expandContextCapsule(current, request, options = {}) {
+  const requestText = Array.isArray(request) ? request.join('; ') : String(request || '')
+  if (!requestText) return current
+  const extra = buildContextCapsule(requestText, {
+    ...options,
+    maxFiles: Math.max(Number(options.maxFiles) || 6, 3),
+    maxTotalChars: Math.max(Number(options.maxTotalChars) || 8000, 4000),
+  })
+  return `${current}\n\n--- additional context requested ---\n${extra}`
+}
+
+function extractKeywords(texts = []) {
+  const words = new Set()
+  for (const text of texts) {
+    for (const match of String(text || '').toLowerCase().match(/[a-z0-9_]+/g) || []) {
+      if (match.length > 2 && !STOP_WORDS.has(match)) words.add(match)
+    }
+  }
+  return [...words]
+}
+
+function scoreText(text, keywords) {
+  const lower = String(text || '').toLowerCase()
+  let score = 0
+  for (const keyword of keywords) {
+    if (lower.includes(keyword)) score += 1
+  }
+  return score
+}

package/src/context-freshness.mjs

@@ -0,0 +1,37 @@
+/**
+ * Context Freshness.
+ *
+ * Context Capsules and evidence records must be bound to a workspace
+ * fingerprint. After files change, old context/evidence is stale and cannot be
+ * used as reliable input.
+ */
+
+import crypto from 'node:crypto'
+
+export function fingerprintText(text) {
+  return crypto.createHash('sha1').update(String(text || '')).digest('hex').slice(0, 16)
+}
+
+export function fingerprintFiles(files = {}) {
+  const hash = crypto.createHash('sha1')
+  for (const [name, content] of Object.entries(files)) {
+    hash.update(name)
+    hash.update('\0')
+    hash.update(String(content || ''))
+    hash.update('\0')
+  }
+  return hash.digest('hex').slice(0, 24)
+}
+
+export function fingerprintWorkspace({ commit = '', files = {} } = {}) {
+  return fingerprintText(`${commit || ''}|${fingerprintFiles(files)}`)
+}
+
+export function attachFingerprint(context = {}, fingerprint = '') {
+  return { ...context, workspaceFingerprint: fingerprint }
+}
+
+export function isContextStale(context = {}, currentFingerprint = '') {
+  if (!currentFingerprint) return false
+  return Boolean(context.workspaceFingerprint && context.workspaceFingerprint !== currentFingerprint)
+}

package/src/evidence-trust.mjs

@@ -0,0 +1,63 @@
+/**
+ * Evidence Trust & Invalidation.
+ *
+ * Assigns every evidence record a trust level and a workspace fingerprint.
+ * If the workspace changes after evidence was produced, that evidence becomes
+ * stale and must not count toward completion.
+ */
+
+export const TRUST_LEVELS = {
+  T0: 0, // model claim
+  T1: 1, // agent observation
+  T2: 2, // host/tool output
+  T3: 3, // deterministic execution
+  T4: 4, // independent/hidden verifier
+}
+
+export function createEvidenceRecord({
+  id = `E-${Date.now().toString(36)}`,
+  criterionId = '',
+  workspaceFingerprint = '',
+  producer = 'unknown',
+  trustLevel = 'T0',
+  createdAt = new Date().toISOString(),
+  payload = {},
+} = {}) {
+  return {
+    evidenceId: id,
+    criterionId,
+    workspaceFingerprint,
+    producer,
+    trustLevel,
+    trustValue: TRUST_LEVELS[trustLevel] ?? 0,
+    createdAt,
+    payload,
+  }
+}
+
+export function isEvidenceStale(record = {}, currentFingerprint = '') {
+  if (!currentFingerprint) return false
+  return Boolean(record.workspaceFingerprint && record.workspaceFingerprint !== currentFingerprint)
+}
+
+export function invalidateEvidence(records = [], currentFingerprint = '') {
+  return records.map((record) => ({
+    ...record,
+    stale: isEvidenceStale(record, currentFingerprint),
+  }))
+}
+
+export function requiredTrustForRisk(risk = 'low') {
+  if (risk === 'high' || risk === 'critical') return { minimum: 'T3', independent: true, label: 'T3 + independent T4' }
+  if (risk === 'medium') return { minimum: 'T3', independent: false, label: 'T3' }
+  return { minimum: 'T1', independent: false, label: 'T1/T2' }
+}
+
+export function evidenceMeetsTrust(record = {}, requiredLevel = 'T0') {
+  const required = TRUST_LEVELS[requiredLevel] ?? 0
+  return (record.trustValue ?? 0) >= required
+}
+
+export function formatEvidenceTrust(record = {}) {
+  return `E${record.evidenceId || '?'} T${record.trustLevel || '0'} ${record.stale ? 'STALE' : 'fresh'} criterion=${record.criterionId || '-'}`
+}

package/src/host-interface.mjs

@@ -0,0 +1,62 @@
+/**
+ * OmniHost Interface.
+ *
+ * The small contract a host (DSH, Codex, Claude, ...) implements so Omni's
+ * kernel stays host-independent. Omni never executes; it decides, prepares,
+ * and verifies through this boundary.
+ */
+
+export const HOST_FEATURES = [
+  'workflow',
+  'approvals',
+  'skills',
+  'plugins',
+  'subagents',
+  'toolEvents',
+  'testEvents',
+  'fileEvents',
+]
+
+export function createHostAdapter(host = {}) {
+  const missing = []
+  for (const method of ['describeHost']) {
+    if (typeof host[method] !== 'function') missing.push(method)
+  }
+  return {
+    ...host,
+    missing,
+    valid: missing.length === 0,
+  }
+}
+
+export function describeHostCapabilities(capabilities = {}) {
+  const result = {}
+  for (const feature of HOST_FEATURES) {
+    result[feature] = capabilities[feature] === true
+  }
+  return result
+}
+
+export function negotiateHost(hostCapabilities = {}, required = HOST_FEATURES) {
+  const caps = describeHostCapabilities(hostCapabilities)
+  const supported = []
+  const degraded = []
+  for (const feature of required) {
+    if (caps[feature]) supported.push(feature)
+    else degraded.push(feature)
+  }
+  return {
+    supported,
+    degraded,
+    capabilities: caps,
+    mode: degraded.length ? 'degraded' : 'full',
+  }
+}
+
+export function formatHostNegotiation(result = {}) {
+  return [
+    `Host mode: ${result.mode}`,
+    `Supported: ${result.supported.join(', ') || '(none)'}`,
+    `Degraded: ${result.degraded.join(', ') || '(none)'}`,
+  ].join('\n')
+}

package/src/intervention-gate.mjs

@@ -0,0 +1,84 @@
+/**
+ * Intervention Gate.
+ *
+ * Decides whether Omni should intervene at all. The default answer is NOOP:
+ * if the host can already do the job without extra tokens/latency/context
+ * pollution, Omni stays out of the way.
+ */
+
+export function expectedInterventionUtility({
+  successGain = 0,
+  tokenOverhead = 0,
+  latency = 0,
+  contextPollution = 0,
+  overlap = 0,
+  failureSurface = 0,
+  taskFrequency = 1,
+} = {}) {
+  const utility = (
+    successGain * 1.0 -
+    tokenOverhead * 0.5 -
+    latency * 0.3 -
+    contextPollution * 0.5 -
+    overlap * 0.4 -
+    failureSurface * 0.4
+  ) * Math.max(0.1, taskFrequency)
+  return Math.round(utility * 1000) / 1000
+}
+
+export function shouldIntervene(utility = 0, threshold = 0.1) {
+  return utility > threshold
+}
+
+export function decideIntervention({
+  successGain,
+  tokenOverhead,
+  latency,
+  contextPollution,
+  overlap,
+  failureSurface,
+  taskFrequency,
+  threshold = 0.1,
+} = {}) {
+  const utility = expectedInterventionUtility({
+    successGain,
+    tokenOverhead,
+    latency,
+    contextPollution,
+    overlap,
+    failureSurface,
+    taskFrequency,
+  })
+  return {
+    utility,
+    threshold,
+    intervene: shouldIntervene(utility, threshold),
+    mode: shouldIntervene(utility, threshold) ? 'intervene' : 'noop',
+  }
+}
+
+export function noOpPrecision(decisions = []) {
+  // decision: { predictedNoop: boolean, actuallyNeeded: boolean }
+  let predictedNoop = 0
+  let correctNoop = 0
+  for (const d of decisions) {
+    if (!d.predictedNoop) continue
+    predictedNoop += 1
+    if (!d.actuallyNeeded) correctNoop += 1
+  }
+  return predictedNoop ? Math.round((correctNoop / predictedNoop) * 1000) / 1000 : 1
+}
+
+export function interventionEfficiency({ rawSuccess = 0, omniSuccess = 0, tokenOverhead = 0 }) {
+  const gain = omniSuccess - rawSuccess
+  const overhead = Math.max(0.01, tokenOverhead)
+  return {
+    gain: Math.round(gain * 1000) / 1000,
+    tokenOverhead,
+    efficiency: Math.round((gain / overhead) * 1000) / 1000,
+  }
+}
+
+export function formatInterventionGate(result = {}) {
+  return `${result.intervene ? 'INTERVENE' : 'NOOP'} (utility=${result.utility}, threshold=${result.threshold})`
+}

package/src/mission-dag.mjs

@@ -103,7 +103,7 @@ export function insertAfter(dag, afterId, task) {
 
 export function applyObservationToDag(dag, observation = {}, failedTaskId = null) {
   const type = observation.type || ''
-  if (type === 'test_failure' || type === 'build_failure') {
+  if (type === 'test_failure' || type === 'build_failure' || type === 'strategy_shift') {
     const taskId = failedTaskId || observation.taskId || null
     const failedTask = taskId ? dag.tasks.find((t) => t.id === taskId) : null
     const baseId = taskId || 'T'
@@ -114,13 +114,16 @@ export function applyObservationToDag(dag, observation = {}, failedTaskId = null
     const tasks = (dag.tasks || []).map((t) => t.id === taskId ? { ...t, status: 'failed', attempt, failure: observation } : t)
     const doneIds = tasks.filter((t) => t.status === 'done').map((t) => t.id)
     const failure = classifyFailure(observation)
+    const isShift = type === 'strategy_shift'
     const repair = createTask({
       id: repairId,
-      goal: `Diagnose and repair: ${failure.category} — ${failure.recovery}`,
+      goal: isShift
+        ? `Strategy shift: ${observation.reason || 're-investigate with broader context and fresh capabilities'}`
+        : `Diagnose and repair: ${failure.category} — ${failure.recovery}`,
       dependencies: doneIds,
-      acceptance: ['verification passes after repair'],
+      acceptance: ['root cause is re-diagnosed and verification passes'],
       verification: ['run failing checks again'],
-      requiredCapabilities: ['debugging'],
+      requiredCapabilities: isShift ? ['debugging', 'repository.search'] : ['debugging'],
       forbiddenCapabilities: ['source.write'],
     })
     // Repair tasks should be allowed to write; fix forbidden.

package/src/omni-event.mjs

@@ -0,0 +1,70 @@
+/**
+ * OmniEvent.
+ *
+ * Host-independent event model. DSH, Codex, Claude, etc. normalize their own
+ * events into OmniEvent so the Evidence/Recovery/Memory kernels never depend on
+ * a specific host.
+ */
+
+export const OMNI_EVENT_TYPES = [
+  'model.started',
+  'model.completed',
+  'tool.started',
+  'tool.completed',
+  'command.completed',
+  'test.completed',
+  'file.changed',
+  'approval.requested',
+  'approval.completed',
+  'agent.started',
+  'agent.completed',
+]
+
+export function createOmniEvent({
+  type,
+  host = 'unknown',
+  sessionId = '',
+  taskId = '',
+  workspaceFingerprint = '',
+  payload = {},
+} = {}) {
+  if (!type) throw new Error('OmniEvent type is required')
+  return {
+    type,
+    host,
+    sessionId,
+    taskId,
+    workspaceFingerprint,
+    timestamp: new Date().toISOString(),
+    payload,
+  }
+}
+
+export function normalizeHostEvent(hostEvent = {}, host = 'unknown', context = {}) {
+  const rawType = String(hostEvent.type || hostEvent.event || hostEvent.name || 'unknown')
+  const type = mapHostEventType(rawType, host)
+  return createOmniEvent({
+    type,
+    host: hostEvent.host || host,
+    sessionId: hostEvent.sessionId || context.sessionId || '',
+    taskId: hostEvent.taskId || context.taskId || '',
+    workspaceFingerprint: hostEvent.workspaceFingerprint || context.workspaceFingerprint || '',
+    payload: hostEvent.payload || hostEvent.data || hostEvent,
+  })
+}
+
+function mapHostEventType(raw, host) {
+  const value = String(raw).toLowerCase()
+  if (host === 'dsh') {
+    if (/tool|command/.test(value) && /start/.test(value)) return 'tool.started'
+    if (/tool|command/.test(value) && /end|done|complete/.test(value)) return 'tool.completed'
+    if (/test/.test(value)) return 'test.completed'
+    if (/file|change/.test(value)) return 'file.changed'
+    if (/approval/.test(value)) return 'approval.requested'
+    if (/agent|subagent/.test(value) && /start/.test(value)) return 'agent.started'
+    if (/agent|subagent/.test(value) && /end|done|complete/.test(value)) return 'agent.completed'
+  }
+  if (/start|begin/.test(value)) return 'model.started'
+  if (/end|done|complete|finish/.test(value)) return 'model.completed'
+  return OMNI_EVENT_TYPES.includes(raw) ? raw : 'model.completed'
+}