mio-agent-runtime 0.5.17 → 0.5.23

package/observe/observer.js

@@ -65,6 +65,30 @@ function stateFile(home) {
   return path.join(home, 'observe-state.json')
 }
 
+// Drop repeated lines from stored context lists, preserving the most recent
+// occurrence order. Older builds could write the same line into every slot of
+// a workspace block (a generated digest headline broadcast to all projects),
+// so existing observe-state files still carry duplicates until they are
+// rewritten. Normalising on load heals that data without a migration step.
+function dedupeContexts(contexts) {
+  const out = {}
+  for (const [cwd, list] of Object.entries(contexts)) {
+    if (!Array.isArray(list)) {
+      out[cwd] = []
+      continue
+    }
+    const seen = new Set()
+    const deduped = []
+    for (const line of list) {
+      if (seen.has(line)) continue
+      seen.add(line)
+      deduped.push(line)
+    }
+    out[cwd] = deduped
+  }
+  return out
+}
+
 function loadState(home) {
   try {
     const parsed = JSON.parse(fs.readFileSync(stateFile(home), 'utf8'))
@@ -74,7 +98,7 @@ function loadState(home) {
         files: parsed.files || {},
         buffers: parsed.buffers || {},
         openTurns: parsed.openTurns || {},
-        contexts: parsed.contexts || {},
+        contexts: dedupeContexts(parsed.contexts || {}),
         codexFiles: parsed.codexFiles || {},
         codexBuffers: parsed.codexBuffers || {},
         codexTurns: parsed.codexTurns || {},
@@ -86,6 +110,9 @@ function loadState(home) {
         hmSeeded: parsed.hmSeeded || false,
         hmLastMsgId: parsed.hmLastMsgId || 0,
         hmTurns: parsed.hmTurns || {},
+        clFiles: parsed.clFiles || {},
+        clBuffers: parsed.clBuffers || {},
+        clTurns: parsed.clTurns || {},
       }
     }
   } catch (_) {}
@@ -106,6 +133,9 @@ function loadState(home) {
     hmSeeded: false,
     hmLastMsgId: 0,
     hmTurns: {},
+    clFiles: {},
+    clBuffers: {},
+    clTurns: {},
   }
 }
 
@@ -282,10 +312,23 @@ const CONTEXT_MAX = 6
 // Maintain a compact 'recent Mio context' block inside <workspace>/AGENTS.md,
 // which WorkBuddy's MemoryLoader auto-injects into every session. This makes
 // retrieval guaranteed (not dependent on the model deciding to read memory).
-function updateProjectContext(state, cwd, line) {
+function updateProjectContext(state, cwd, line, targetFile) {
   if (!cwd || !line) return
-  const target = path.join(cwd, 'AGENTS.md')
-  const list = state.contexts[cwd] || []
+  const target = path.join(cwd, targetFile || 'AGENTS.md')
+  // Normalise the whole list before inserting, not just the incoming line.
+  // Callers may load state without going through loadState (the digest
+  // write-back reads the file directly), so stale duplicates have to be
+  // cleared here or they keep consuming slots and push real history out.
+  const list = []
+  const seen = new Set()
+  for (const item of state.contexts[cwd] || []) {
+    if (seen.has(item)) continue
+    seen.add(item)
+    list.push(item)
+  }
+  // Re-inserting an existing line moves it to the front instead of duplicating.
+  const existing = list.indexOf(line)
+  if (existing !== -1) list.splice(existing, 1)
   list.unshift(line)
   if (list.length > CONTEXT_MAX) list.pop()
   state.contexts[cwd] = list
@@ -1133,6 +1176,275 @@ function runHermesCycle(home, state, sinks, opts) {
 }
 
 
+// ---------------------------------------------------------------------------
+// Claude Code session observation (passive, transcript JSONL tailing)
+// ---------------------------------------------------------------------------
+
+// Claude Code writes append-only JSONL transcripts under
+// <configDir>/projects/<munged-cwd>/<sessionId>.jsonl. Each line carries
+// {type, message:{role, content}, cwd, sessionId, timestamp, uuid, ...}.
+// Assistant content items are text/thinking/tool_use; tool results come back
+// as user lines whose content array holds tool_result items with is_error.
+// A user prompt line (string content or a text item) closes the previous
+// turn; sidechain lines (subagent transcripts) are skipped as noise.
+
+function claudeConfigDir() {
+  return process.env.CLAUDE_CONFIG_DIR || path.join(os.homedir(), '.claude')
+}
+
+function claudeProjectsDir() {
+  return path.join(claudeConfigDir(), 'projects')
+}
+
+function listClaudeTranscripts() {
+  const root = claudeProjectsDir()
+  const out = []
+  if (!fs.existsSync(root)) return out
+  const walk = (dir) => {
+    let entries
+    try {
+      entries = fs.readdirSync(dir, { withFileTypes: true })
+    } catch (_) {
+      return
+    }
+    for (const entry of entries) {
+      const full = path.join(dir, entry.name)
+      if (entry.isDirectory()) walk(full)
+      else if (entry.isFile() && entry.name.endsWith('.jsonl')) out.push(full)
+    }
+  }
+  walk(root)
+  return out
+}
+
+function createClaudeTurn(meta) {
+  meta = meta || {}
+  return {
+    id: null,
+    ts: null,
+    sessionId: meta.sessionId || null,
+    cwd: meta.cwd || null,
+    project: meta.cwd ? path.basename(meta.cwd) : 'unknown',
+    lastText: null,
+    toolCount: 0,
+    errors: [],
+    apiError: false,
+  }
+}
+
+function claudeContentItems(obj) {
+  const content = obj && obj.message && obj.message.content
+  if (typeof content === 'string') return [{ type: 'text', text: content }]
+  return Array.isArray(content) ? content.filter(Boolean) : []
+}
+
+function finishClaudeTurn(turn, outcome) {
+  return {
+    id: turn.id,
+    ts: turn.ts,
+    sessionId: turn.sessionId,
+    cwd: turn.cwd,
+    project: turn.project,
+    lastText: turn.lastText,
+    toolCount: turn.toolCount,
+    errors: turn.errors.slice(),
+    apiError: turn.apiError,
+    outcome: outcome,
+  }
+}
+
+function feedClaudeLine(turn, obj) {
+  if (!obj || typeof obj !== 'object') return { next: turn }
+  // Subagent (sidechain) transcripts duplicate main-thread work; skip them.
+  if (obj.isSidechain === true) return { next: turn }
+  if (obj.sessionId) turn.sessionId = obj.sessionId
+  if (obj.cwd) {
+    turn.cwd = obj.cwd
+    turn.project = path.basename(obj.cwd)
+  }
+
+  if (obj.type === 'user') {
+    const items = claudeContentItems(obj)
+    let hasToolError = false
+    for (const item of items) {
+      if (item.type === 'tool_result' && item.is_error) {
+        turn.errors.push({ toolUseId: item.tool_use_id || 'unknown' })
+        hasToolError = true
+      }
+    }
+    if (turn.errors.length > 20) turn.errors.splice(0, turn.errors.length - 20)
+
+    // tool_result-only user lines continue the current turn; a string content
+    // or text item means a fresh prompt, which closes the previous turn.
+    const isPrompt =
+      typeof (obj.message && obj.message.content) === 'string' ||
+      items.some((item) => item.type === 'text')
+    if (isPrompt) {
+      let finalized = null
+      if (turn.id) {
+        finalized = finishClaudeTurn(turn, turn.apiError ? 'failure' : turn.lastText ? 'success' : 'aborted')
+      }
+      const next = createClaudeTurn({ sessionId: turn.sessionId, cwd: turn.cwd })
+      next.id = obj.uuid || null
+      next.ts = obj.timestamp ? Date.parse(obj.timestamp) || Date.now() : Date.now()
+      return { finalized, next }
+    }
+    return { next: turn }
+  }
+
+  if (obj.type === 'assistant') {
+    for (const item of claudeContentItems(obj)) {
+      if (item.type === 'text' && typeof item.text === 'string' && item.text.trim()) {
+        turn.lastText = item.text
+      } else if (item.type === 'tool_use') {
+        turn.toolCount += 1
+      }
+    }
+    if (obj.isApiErrorMessage === true || obj.error) turn.apiError = true
+    const ts = obj.timestamp ? Date.parse(obj.timestamp) : NaN
+    if (Number.isFinite(ts)) turn.ts = ts
+    return { next: turn }
+  }
+
+  return { next: turn }
+}
+
+function emitClaudeOutcome(turn, sinks, state) {
+  if (!turn.id) return false
+  const summary = cleanText(turn.lastText)
+  if (!summary && !turn.apiError) return false
+  const outcome = turn.apiError && turn.outcome !== 'failure' ? 'failure' : turn.outcome
+  const turnToken = String(turn.id).slice(0, 8)
+  const tsToken = turn.ts ? Math.floor(turn.ts / 1000) : Math.floor(Date.now() / 1000)
+
+  sinks.trace({
+    trace_id: 'claude:' + (turn.sessionId || 'unknown') + ':' + turnToken + ':' + tsToken,
+    event_type: 'task_outcome',
+    outcome: outcome,
+    payload: {
+      host: 'claude-code',
+      project: turn.project,
+      sessionId: turn.sessionId,
+      turnId: turn.id,
+      toolCalls: turn.toolCount,
+      errors: turn.errors.length,
+      summary: summary.slice(0, 300),
+      source: 'claude-observer',
+    },
+    agent: 'claude-code',
+    host: 'claude-code',
+    project: turn.project,
+  })
+
+  if (outcome === 'success' && summary.length >= 10) {
+    sinks.memory({
+      kind: 'note',
+      content: '[claude-code] 任务完成: ' + summary.slice(0, 160),
+      tags: ['task-outcome', 'claude-code', turn.project],
+      project: turn.project,
+      source: 'claude-observer',
+    })
+    // Claude Code auto-loads <workspace>/CLAUDE.md, so write the context
+    // block there instead of AGENTS.md.
+    if (turn.cwd) {
+      var ctxTime = new Date(Number(turn.ts) || Date.now()).toLocaleString('zh-CN', { hour12: false })
+      updateProjectContext(state, turn.cwd, ctxTime + ' | ' + outcome + ' | ' + summary.slice(0, 80), 'CLAUDE.md')
+    }
+  }
+
+  for (const err of turn.errors) {
+    sinks.trace({
+      trace_id: 'claude:' + (turn.sessionId || 'unknown') + ':' + turnToken + ':' + String(err.toolUseId || 'err').slice(0, 12),
+      event_type: 'error',
+      outcome: 'error',
+      payload: {
+        host: 'claude-code',
+        project: turn.project,
+        sessionId: turn.sessionId,
+        turnId: turn.id,
+        toolUseId: err.toolUseId,
+      },
+      agent: 'claude-code',
+      host: 'claude-code',
+      project: turn.project,
+    })
+  }
+  return true
+}
+
+function runClaudeCycle(home, state, sinks, opts) {
+  const out = { files: 0, lines: 0, outcomes: 0 }
+  const files = listClaudeTranscripts()
+  const now = Date.now()
+  const maxAgeMs = (opts && opts.maxAgeMs) || 24 * 60 * 60 * 1000
+
+  for (const file of files) {
+    out.files += 1
+    let st
+    try {
+      st = fs.statSync(file)
+    } catch (_) {
+      continue
+    }
+    const info = state.clFiles[file] || { size: 0 }
+    if (!state.clFiles[file] && now - st.mtimeMs > maxAgeMs) continue
+    const size = st.size
+    if (size < info.size) {
+      info.size = 0
+      state.clBuffers[file] = ''
+    }
+
+    let turn = state.clTurns[file] || createClaudeTurn()
+
+    if (size > info.size) {
+      let chunk = ''
+      try {
+        const fd = fs.openSync(file, 'r')
+        const buf = Buffer.alloc(size - info.size)
+        fs.readSync(fd, buf, 0, buf.length, info.size)
+        fs.closeSync(fd)
+        chunk = buf.toString('utf8')
+      } catch (_) {
+        continue
+      }
+      const pending = (state.clBuffers[file] || '') + chunk
+      const lines = pending.split(/\r?\n/)
+      state.clBuffers[file] = lines.pop()
+      for (const line of lines) {
+        if (!line.trim()) continue
+        out.lines += 1
+        let obj
+        try {
+          obj = JSON.parse(line)
+        } catch (_) {
+          continue
+        }
+        const result = feedClaudeLine(turn, obj)
+        turn = result.next
+        if (result.finalized) {
+          if (emitClaudeOutcome(result.finalized, sinks, state)) out.outcomes += 1
+        }
+      }
+      state.clFiles[file] = { size: size, mtimeMs: st.mtimeMs }
+    }
+
+    // Idle finalize: the session has answered but no new prompt arrived yet.
+    if (turn.id && turn.lastText && now - st.mtimeMs > IDLE_MS) {
+      if (emitClaudeOutcome(finishClaudeTurn(turn, turn.apiError ? 'failure' : 'success'), sinks, state)) {
+        out.outcomes += 1
+      }
+      state.clTurns[file] = null
+    } else if (turn.id) {
+      state.clTurns[file] = turn
+    } else {
+      state.clTurns[file] = null
+    }
+  }
+
+  return out
+}
+
+
 // ---------------------------------------------------------------------------
 // cycle
 // ---------------------------------------------------------------------------
@@ -1225,6 +1537,11 @@ function runCycle(home, state, sinks, opts) {
   out.lines += hmRes.lines
   out.outcomes += hmRes.outcomes
 
+  const clRes = runClaudeCycle(home, state, sinks, opts)
+  out.files += clRes.files
+  out.lines += clRes.lines
+  out.outcomes += clRes.outcomes
+
   saveState(home, state)
   return out
 }
@@ -1311,12 +1628,17 @@ module.exports = {
   workbuddyConfigDir,
   loadState,
   saveState,
+  dedupeContexts,
   updateProjectContext,
   runCycle,
   opencodeDbPath,
   runOpencodeCycle,
   runHermesCycle,
   hermesDbPath,
+  claudeProjectsDir,
+  runClaudeCycle,
+  createClaudeTurn,
+  feedClaudeLine,
   createSinks,
   startForeground,
   startBackground,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mio-agent-runtime",
-  "version": "0.5.17",
+  "version": "0.5.23",
   "description": "Mio Agent Runtime installer and MCP intelligence layer",
   "license": "MIT",
   "type": "commonjs",
@@ -17,7 +17,8 @@
     "!server/mio-intelligence-mcp/__tests__"
   ],
   "scripts": {
-    "check": "node --check bin/mio.js && node --check adapters/host-utils.js && node --check adapters/codex.js && node --check adapters/opencode.js && node --check adapters/workbuddy.js && node --check adapters/hermes.js && node --check server/runtime-modules.js && node --check server/host-capabilities.js && node --check server/evolution-cutover.js && node --check server/mio-intelligence-mcp/index.js && node --check server/mio-intelligence-mcp/phase0.js && node --check observe/observer.js && node --check scripts/verify-packed-runtime.js",
+    "check": "node --check bin/mio.js && node --check adapters/host-utils.js && node --check adapters/codex.js && node --check adapters/opencode.js && node --check adapters/workbuddy.js && node --check adapters/hermes.js && node --check adapters/claude-code.js && node --check server/runtime-modules.js && node --check server/memory-store.js && node --check server/retention.js && node --check server/digest.js && node --check server/host-capabilities.js && node --check server/evolution-cutover.js && node --check server/mio-intelligence-mcp/index.js && node --check server/mio-intelligence-mcp/phase0.js && node --check observe/observer.js && node --check scripts/verify-packed-runtime.js",
+    "test": "node --test \"__tests__/*.test.js\" \"server/mio-intelligence-mcp/__tests__/*.test.js\"",
     "prepack": "npm run check",
     "verify:pack-install": "node scripts/verify-packed-runtime.js"
   },
@@ -34,5 +35,11 @@
     "@akemi-mio/evolution-scheduler": "^0.1.0",
     "@akemi-mio/observer": "^0.1.0",
     "@akemi-mio/insight": "^0.1.0"
-  }
+  },
+  "main": "index.js",
+  "directories": {
+    "doc": "docs"
+  },
+  "keywords": [],
+  "author": ""
 }

package/scripts/verify-packed-runtime.js

@@ -76,12 +76,13 @@ function npmCommand(platform = process.platform) {
 
 function npmInvocation(args, env = process.env, platform = process.platform, node = process.execPath) {
   if (env.npm_execpath) {
-    return { command: node, args: [env.npm_execpath, ...args] }
+    return { command: node, args: [env.npm_execpath, ...args], shell: false }
   }
-  if (platform === 'win32') {
-    return { command: 'cmd.exe', args: ['/d', '/s', '/c', 'npm', ...args] }
-  }
-  return { command: npmCommand(platform), args }
+  // Fall back to the npm shim on PATH instead of routing through cmd.exe.
+  // cmd.exe is unavailable in some sandboxes, and a bare .cmd/.bat shim cannot
+  // be spawned directly (Node requires shell:true for those on Windows), so
+  // signal the shell requirement to the runner instead of guessing there.
+  return { command: npmCommand(platform), args: [...args], shell: platform === 'win32' }
 }
 
 function packageDirName(name) {
@@ -95,9 +96,12 @@ function tarballFileName(name, version) {
 
 function readPackageJson(workspaceRoot, name) {
   const packageDir = path.join(workspaceRoot, 'packages', packageDirName(name))
+  // Windows editors may write package.json with a UTF-8 BOM; JSON.parse
+  // rejects the leading U+FEFF, so strip it before parsing.
+  const raw = fs.readFileSync(path.join(packageDir, 'package.json'), 'utf8').replace(/^\uFEFF/, '')
   return {
     packageDir,
-    manifest: JSON.parse(fs.readFileSync(path.join(packageDir, 'package.json'), 'utf8')),
+    manifest: JSON.parse(raw),
   }
 }
 
@@ -211,6 +215,11 @@ function run(command, args, options = {}) {
     env: options.env || process.env,
     encoding: 'utf8',
     stdio: options.stdio || ['ignore', 'pipe', 'pipe'],
+    // Windows .cmd/.bat shims (npm.cmd) cannot be spawned directly by
+    // execFileSync since Node's CVE-2024-27980 hardening; they need a shell.
+    // This only kicks in for the npm shim path — node-based invocations pass
+    // an absolute node executable and still run shell-free.
+    shell: options.shell === true,
   })
   return output ? output.trim() : ''
 }
@@ -296,14 +305,19 @@ function verifyPackedRuntimeInstall(input = {}) {
     run(pack.command, pack.args, {
       cwd: pkg.packageDir,
       stdio: ['ignore', 'pipe', 'pipe'],
+      shell: pack.shell,
     })
     packageArtifactEvidence(pkg)
   }
 
   const init = npmInvocation(['init', '-y'])
-  run(init.command, init.args, { cwd: plan.installDir, stdio: ['ignore', 'pipe', 'pipe'] })
+  run(init.command, init.args, { cwd: plan.installDir, stdio: ['ignore', 'pipe', 'pipe'], shell: init.shell })
   const install = npmInvocation(plan.install.args)
-  run(install.command, install.args, { cwd: plan.install.cwd, stdio: ['ignore', 'pipe', 'pipe'] })
+  run(install.command, install.args, {
+    cwd: plan.install.cwd,
+    stdio: ['ignore', 'pipe', 'pipe'],
+    shell: install.shell,
+  })
   const smokeEnv = {
     ...process.env,
     MIO_HOME: path.join(plan.installDir, '.mio-home'),
@@ -360,6 +374,7 @@ module.exports = {
   createVerifyPackedRuntimeOptions,
   createPackedRuntimeInstallPlan,
   packageArtifactEvidence,
+  readPackageJson,
   npmInvocation,
   npmCommand,
   parseVerifyPackedRuntimeArgs,