mio-agent-runtime 0.5.17 → 0.5.22

package/README.md

@@ -1,25 +1,25 @@
-# mio-agent-runtime
-
-Mio Agent Runtime 的可发布 CLI 包。
-
-安装一次，所有 Agent 自动接入 Mio。
-
-## Install
-
-```bash
-npm install -g mio-agent-runtime
-```
-
-> 新机器 / 升级：请安装 `>=0.3.6`（Codex + OpenCode + Hermes 被动观察）。
-> 旧版本只有主动 MCP 规则注入，没有「无感存/取」。安装后执行一次
-> `mio init && mio install <host>`。
-
+# mio-agent-runtime
+
+Mio Agent Runtime 的可发布 CLI 包。
+
+安装一次，所有 Agent 自动接入 Mio。
+
+## Install
+
+```bash
+npm install -g mio-agent-runtime
+```
+
+> 新机器 / 升级：请安装 `>=0.3.6`（Codex + OpenCode + Hermes 被动观察）。
+> 旧版本只有主动 MCP 规则注入，没有「无感存/取」。安装后执行一次
+> `mio init && mio install <host>`。
+
 ## Commands
 
 ```text
-mio init                    Initialize MIO_HOME
-mio mcp                     Start Mio MCP server (stdio)
-mio install <host>          Install Mio into a host (codex|opencode|workbuddy|hermes)
+mio init                    Initialize MIO_HOME
+mio mcp                     Start Mio MCP server (stdio)
+mio install <host>          Install Mio into a host (codex|opencode|workbuddy|hermes|claude)
 mio status                  Show runtime and adapter status
 mio agents                  List installed agents
 mio evolution status        Show composed evolution module health
@@ -31,11 +31,155 @@ mio evolution migration plan      Preview state migration diffs
 mio evolution cutover apply --dry-run   Dry-run a cutover plan without switching authority
 mio observe                 Watch WorkBuddy transcripts and auto-ingest task outcomes (foreground)
 mio observe --start|--stop|--status|--once   Manage the background observer daemon
+mio recall "<query>"        Search Mio memory from the terminal (--project/--scope/--kind/--tags/--limit)
+mio traces                  Show recent observer traces (--type/--outcome/--agent/--since/--limit/--compact)
+mio prune --days 30         Trim old traces/queries/reuse records and observe.log (--dry-run to preview; --memory needs --yes)
+mio digest --days 7         Aggregate traces/memory/reuse into an actionable report (--write-back feeds agent context files; --json)
+mio remember "<content>"    Write a memory record from the terminal (--kind/--tags/--scope/--project)
 mio --json status           Machine-readable status
 mio --json agents           Machine-readable agents
 mio --json evolution status Machine-readable evolution module health
 ```
 
+## MCP Tools
+
+MCP server exposes 47 tools across 5 domains:
+
+### Memory (10)
+`mio.memory.query` · `mio.memory.record` · `mio.memory.archive` · `mio.memory.migrate` · `mio.memory.analyze` · `mio.experience.list` · `mio.experience.confirm` · `mio.experience.reuse` · `mio.policy.check`
+
+### Observer Pipeline (13)
+`mio.observer.world_model` · `mio.observer.trends` · `mio.observer.research` · `mio.observer.insights` · `mio.observer.status` · `mio.observer.collect` · `mio.observer.ferment` · `mio.observer.essays` · `mio.observer.dag` · `mio.observer.ingest` · `mio.observer.subscribe` · `mio.observer.digest`
+`mio.trace.query` — query the trace log (task outcomes, tool errors) by type/outcome/agent/project/time window; also available as `mio traces` on the CLI
+`mio.digest.generate` — aggregate recent data into an actionable digest (agent success rates, project activity, error hotspots, reuse evidence, suggestions); also available as `mio digest` on the CLI
+
+### Insight Self-Observation (4)
+`mio.insight.status` · `mio.insight.list` · `mio.insight.mark_reported` · `mio.insight.generate`
+
+### Creativity Engine (4)
+`mio.creativity.generate` · `mio.creativity.ferment` · `mio.creativity.list` · `mio.creativity.status`
+
+### Idea & Evolution (7)
+`mio.idea.generate` · `mio.idea.list` · `mio.evolution.suggest` · `mio.evolution.report` · `mio.evolution.strategies` · `mio.phase0.report`
+
+## Packages
+
+Published to npm under `@akemi-mio` scope:
+
+| Package | Description |
+|---|---|
+| `@akemi-mio/core` | Electron lifecycle, config, logging, EventBus, patterns, schemas |
+| `@akemi-mio/cli` | CLI helpers |
+| `@akemi-mio/runtime-contracts` | Shared type definitions |
+| `@akemi-mio/runtime-foundation` | Logging, EventBus, memory schemas |
+| `@akemi-mio/experience-memory` | Experience recording and retrieval |
+| `@akemi-mio/evolution-learning` | Evolution learning engine |
+| `@akemi-mio/evolution-strategy` | Evolution strategy engine |
+| `@akemi-mio/evolution-safety` | Safety guardrails |
+| `@akemi-mio/evolution-scheduler` | Evolution scheduler |
+| `@akemi-mio/creativity` | CreativityService — LLM-powered concept generation, conflict detection, merging, fermentation (TypeScript, depends on `@akemi-mio/core`) |
+| `@akemi-mio/observer` | ObserverService — multi-source data collection, trend analysis, deep research, world model, fermentation, DAG state machine, self-evolution engine (zero npm deps) |
+| `@akemi-mio/insight` | InsightService — LLM-based insight generation, conflict/drift/repetition/stalled-goal/friction detectors, presence service, insight scoring (zero npm deps) |
+| `mio-agent-runtime` | This package — CLI + MCP server |
+
+## Host notes
+
+- OpenCode uses `~/.config/opencode/opencode.json` for MCP registration and
+  `~/.config/opencode/AGENTS.md` for global instructions. `mio install opencode`
+  writes both files and injects Mio usage rules into `AGENTS.md` so OpenCode
+  calls Mio memory/observer/policy automatically. If OpenCode is already
+  running, restart it once so it reloads both files. `mio observe` also
+  passively tails the OpenCode session database (`opencode db`) without
+  requiring the model to call Mio tools.
+- Hermes uses `hermes mcp add mio-intelligence --command node --args <server>`
+  to register the Mio MCP server. Hermes has native memory and reads
+  AGENTS.md Context Files, so retrieval works without Mio rule injection.
+- Claude Code: `mio install claude` registers the Mio MCP server at user
+  scope in `~/.claude.json` (`mcpServers`) — no interactive approval needed,
+  unlike project-scope `.mcp.json` — and injects Mio usage rules into
+  `~/.claude/CLAUDE.md` (auto-loaded user memory) plus `<workspace>/CLAUDE.md`
+  (project memory). Restart running Claude Code sessions to reload both.
+  `mio observe` passively tails `<config-dir>/projects/<munged-cwd>/*.jsonl`
+  session transcripts: sidechain (subagent) lines are skipped, each user
+  prompt closes the previous turn, `tool_result.is_error` lines are recorded
+  as error traces, and successful turn summaries are written back into the
+  workspace `CLAUDE.md` MIO_CONTEXT block so the next session retrieves them
+  passively. Set `CLAUDE_CONFIG_DIR` to override the config dir.
+- WorkBuddy uses `~/.workbuddy/mcp.json` and requires each custom MCP server to
+  be approved in `~/.workbuddy/mcp-approvals.json`. `mio install workbuddy`
+  writes both files and injects Mio usage rules into
+  `~/.workbuddy/CODEBUDDY.md` (user-level memory rules, auto-loaded by
+  WorkBuddy into every session; the CLI hardcodes the CodeBuddy product name, so the file is CODEBUDDY.md not WORKBUDDY.md) plus `<workspace>/AGENTS.md` (project-level).
+  If WorkBuddy is already running, restart it once so it reloads the approval
+  and memory rule files.
+
+## Passive observer (`mio observe`)
+
+WorkBuddy 的默认模型不一定会主动调用 Mio MCP 工具，即使规则已注入
+(`CODEBUDDY.md` / `AGENTS.md`)。`mio observe` 是一个不依赖模型自觉的
+兜底：它持续监听 `~/.workbuddy/projects/**/*.jsonl` 会话记录，把每个任务的
+结果（`task_outcome`）、工具调用错误（`error`）以及完成任务的摘要自动写入
+`MIO_HOME/traces.jsonl` 和 `MIO_HOME/memory.jsonl`（与 MCP 服务同一套
+schema，`source=workbuddy-observer`）。
+
+- `mio install <host>` 会自动执行一次回填并启动后台观察器。
+
+- 除了写入 `MIO_HOME`，观察器还会把任务摘要**写回 WorkBuddy 自己的记忆文件**
+  (`<workspace>/.workbuddy/memory/YYYY-MM-DD.md`，标为 `## Mio 自动摘要`)。
+  WorkBuddy 的系统提示要求模型每次会话维护并阅读这些文件，所以下一次会话
+  WorkBuddy 会**被动取到** Mio 记录的内容——单 Agent 的「存 + 取」闭环不依赖
+  模型调用 Mio 工具。
+- 已处理的文件偏移记录在 `MIO_HOME/observe-state.json`，不会重复写入。
+- 只回填最近 24h 内更新的会话，避免把历史记录灌入记忆库。
+
+OpenCode 会话不依赖模型自觉：`mio observe` 通过 `opencode db` 增量读取
+`~/.local/share/opencode/opencode.db`（Windows 下该文件被 OpenCode 独占锁定，
+必须经 opencode 自带 CLI 读取），按 `part` 表的 rowid 做增量游标，只传输
+`text`/`tool` 两种 part，把每个用户回合的 `task_outcome` 与工具错误写入
+`MIO_HOME/traces.jsonl` / `memory.jsonl`（`source=opencode-observer`），并同样
+把摘要写回 `<workspace>/AGENTS.md` 的 `MIO_CONTEXT` 块，供 OpenCode 下一会话
+被动取用。首次启动只回填最近 24h 的 part，避免历史会话灌入记忆库。
+
+Hermes（Nous hermes-agent）同理：`mio observe` 以只读方式增量读取
+`<LocalAppData>/hermes/state.db`（`messages.id` 自增游标，无需调用 hermes CLI），
+把每个用户回合的 `task_outcome` 写入 `MIO_HOME`（`source=hermes-observer`）。
+Hermes 本身有原生记忆，Mio 只负责让 Hermes 经验进入跨 Agent 生态；仅当会话
+发生在真实 git 仓库时才写回该仓库的 `AGENTS.md`，避免污染用户主目录。
+
+## MIO_HOME
+
+默认 `~/.mio-intelligence`，可通过环境变量 `MIO_HOME` 覆盖。
+
+## Package layout
+
+```text
+packages/mio-cli
+├── bin/mio.js
+├── adapters/
+│   ├── codex.js
+│   ├── hermes.js
+│   ├── opencode.js
+│   ├── workbuddy.js
+│   └── claude-code.js
+├── observe/
+│   └── observer.js
+├── server/
+│   ├── host-capabilities.js
+│   ├── evolution-cutover.js
+│   ├── runtime-modules.js
+│   ├── creativity-engine.js
+│   ├── memory-store.js
+│   └── mio-intelligence-mcp/index.js
+└── package.json
+```
+
+`server/memory-store.js` is the single implementation of memory
+query/record ranking (Latin token + CJK bigram scoring, project/global
+scope layers, reuse-evidence weighting). Both the MCP server
+(`mio.memory.query` / `mio.memory.record`) and the CLI
+(`mio recall` / `mio remember`) go through it, so rankings are identical
+across entry points.
+
 ## Release verification
 
 Before publishing the runtime packages, run:
@@ -71,91 +215,3 @@ node packages/mio-cli/scripts/verify-packed-runtime.js \
   --install-dir ./dist/mio-pack-install \
   --report-path ./dist/mio-pack-report.json
 ```
-
-This packs all seven `@akemi-mio/*` runtime modules plus `mio-agent-runtime`, installs
-the local tarballs into a clean temporary npm project, and runs installed CLI
-smoke checks for `mio --json evolution status`, shadow sample recording,
-dual-write sample recording, cutover readiness, and dry-run cutover apply. The
-JSON output includes each packed tarball path, size, and SHA-512 digest for
-release evidence. When `MIO_PACK_REPORT_PATH` or `--report-path` is set, the
-same JSON is written to that file; existing report files are not overwritten. It
-fails closed if any runtime module is missing, unloaded, or unhealthy, if
-shadow/dual-write parity fails, or if cutover apply stops behaving as a dry run.
-It does not publish packages or switch authority.
-
-## Host notes
-
-- OpenCode uses `~/.config/opencode/opencode.json` for MCP registration and
-  `~/.config/opencode/AGENTS.md` for global instructions. `mio install opencode`
-  writes both files and injects Mio usage rules into `AGENTS.md` so OpenCode
-  calls Mio memory/observer/policy automatically. If OpenCode is already
-  running, restart it once so it reloads both files. `mio observe` also
-  passively tails the OpenCode session database (`opencode db`) without
-  requiring the model to call Mio tools.
-- Hermes uses `hermes mcp add mio-intelligence --command node --args <server>`
-  to register the Mio MCP server. Hermes has native memory and reads
-  AGENTS.md Context Files, so retrieval works without Mio rule injection.
-- WorkBuddy uses `~/.workbuddy/mcp.json` and requires each custom MCP server to
-  be approved in `~/.workbuddy/mcp-approvals.json`. `mio install workbuddy`
-  writes both files and injects Mio usage rules into
-  `~/.workbuddy/CODEBUDDY.md` (user-level memory rules, auto-loaded by
-  WorkBuddy into every session; the CLI hardcodes the CodeBuddy product name, so the file is CODEBUDDY.md not WORKBUDDY.md) plus `<workspace>/AGENTS.md` (project-level).
-  If WorkBuddy is already running, restart it once so it reloads the approval
-  and memory rule files.
-
-## Passive observer (`mio observe`)
-
-WorkBuddy 的默认模型不一定会主动调用 Mio MCP 工具，即使规则已注入
-(`CODEBUDDY.md` / `AGENTS.md`)。`mio observe` 是一个不依赖模型自觉的
-兜底：它持续监听 `~/.workbuddy/projects/**/*.jsonl` 会话记录，把每个任务的
-结果（`task_outcome`）、工具调用错误（`error`）以及完成任务的摘要自动写入
-`MIO_HOME/traces.jsonl` 和 `MIO_HOME/memory.jsonl`（与 MCP 服务同一套
-schema，`source=workbuddy-observer`）。
-
-- `mio install <host>` 会自动执行一次回填并启动后台观察器。
-
-- 除了写入 `MIO_HOME`，观察器还会把任务摘要**写回 WorkBuddy 自己的记忆文件**
-  (`<workspace>/.workbuddy/memory/YYYY-MM-DD.md`，标为 `## Mio 自动摘要`)。
-  WorkBuddy 的系统提示要求模型每次会话维护并阅读这些文件，所以下一次会话
-  WorkBuddy 会**被动取到** Mio 记录的内容——单 Agent 的「存 + 取」闭环不依赖
-  模型调用 Mio 工具。
-- 已处理的文件偏移记录在 `MIO_HOME/observe-state.json`，不会重复写入。
-- 只回填最近 24h 内更新的会话，避免把历史记录灌入记忆库。
-
-OpenCode 会话不依赖模型自觉：`mio observe` 通过 `opencode db` 增量读取
-`~/.local/share/opencode/opencode.db`（Windows 下该文件被 OpenCode 独占锁定，
-必须经 opencode 自带 CLI 读取），按 `part` 表的 rowid 做增量游标，只传输
-`text`/`tool` 两种 part，把每个用户回合的 `task_outcome` 与工具错误写入
-`MIO_HOME/traces.jsonl` / `memory.jsonl`（`source=opencode-observer`），并同样
-把摘要写回 `<workspace>/AGENTS.md` 的 `MIO_CONTEXT` 块，供 OpenCode 下一会话
-被动取用。首次启动只回填最近 24h 的 part，避免历史会话灌入记忆库。
-
-Hermes（Nous hermes-agent）同理：`mio observe` 以只读方式增量读取
-`<LocalAppData>/hermes/state.db`（`messages.id` 自增游标，无需调用 hermes CLI），
-把每个用户回合的 `task_outcome` 写入 `MIO_HOME`（`source=hermes-observer`）。
-Hermes 本身有原生记忆，Mio 只负责让 Hermes 经验进入跨 Agent 生态；仅当会话
-发生在真实 git 仓库时才写回该仓库的 `AGENTS.md`，避免污染用户主目录。
-
-## MIO_HOME
-
-默认 `~/.mio-intelligence`，可通过环境变量 `MIO_HOME` 覆盖。
-
-## Package layout
-
-```text
-packages/mio-cli
-├── bin/mio.js
-├── adapters/
-│   ├── codex.js
-│   ├── hermes.js
-│   ├── opencode.js
-│   └── workbuddy.js
-├── observe/
-│   └── observer.js
-├── server/
-│   ├── host-capabilities.js
-│   ├── evolution-cutover.js
-│   ├── runtime-modules.js
-│   └── mio-intelligence-mcp/index.js
-└── package.json
-```

package/adapters/claude-code.js

@@ -0,0 +1,141 @@
+'use strict'
+
+const fs = require('fs')
+const path = require('path')
+const os = require('os')
+
+function configDir() {
+  return process.env.CLAUDE_CONFIG_DIR || path.join(os.homedir(), '.claude')
+}
+
+// Claude Code keeps user-scope MCP servers in the top-level ~/.claude.json
+// (moved to $CLAUDE_CONFIG_DIR/.claude.json when CLAUDE_CONFIG_DIR is set).
+// Project-scope .mcp.json files require interactive approval, so we install
+// at user scope only.
+function configPath() {
+  if (process.env.CLAUDE_CONFIG_DIR) {
+    return path.join(process.env.CLAUDE_CONFIG_DIR, '.claude.json')
+  }
+  return path.join(os.homedir(), '.claude.json')
+}
+
+// User-level memory file. Claude Code auto-loads ~/.claude/CLAUDE.md into
+// every session, which gives the Mio usage rules a guaranteed retrieval path.
+function memoryRulesPath() {
+  return path.join(configDir(), 'CLAUDE.md')
+}
+
+// Project-level memory. Claude Code auto-loads <workspace>/CLAUDE.md.
+function projectInstructionsPath(workspace) {
+  return path.join(workspace || process.cwd(), 'CLAUDE.md')
+}
+
+function readJson(file) {
+  if (!fs.existsSync(file)) return null
+  try {
+    return JSON.parse(fs.readFileSync(file, 'utf8'))
+  } catch (_) {
+    return null
+  }
+}
+
+const MEMORY_BEGIN = '<!-- MIO_INTELLIGENCE_BEGIN -->'
+const MEMORY_END = '<!-- MIO_INTELLIGENCE_END -->'
+const MEMORY_BLOCK = `${MEMORY_BEGIN}
+## Mio Intelligence Layer
+
+Mio MCP tools are available. Treat these as mandatory runtime rules for every workspace.
+
+- Before designing or implementing a solution, or when unsure about prior decisions, call \`mio-intelligence_mio.memory.query\` first. Omit \`project\` unless a specific project is required; Mio derives it from the current Git repository.
+- After a decision affecting architecture, deployment, dependencies, or long-term behavior, call \`mio-intelligence_mio.memory.record\` with \`kind\`: \`decision\`, \`context\`, \`problem\`, or \`note\`.
+- When a tool call fails, retries, or returns an error, call \`mio-intelligence_mio.observer.ingest\` with a stable \`trace_id\`, relevant \`event_type\`, \`payload\`, and \`outcome\`.
+- Before a high-risk operation such as deletion, migration, credential change, or production change, call \`mio-intelligence_mio.policy.check\` and follow its suggestion when available.
+- At the end of every task, before producing the final answer, call \`mio-intelligence_mio.observer.ingest\` with \`event_type\` = \`task_outcome\`, a stable \`trace_id\`, \`outcome\` = \`success\` / \`failure\` / \`aborted\`, and a brief non-sensitive \`payload\`. If Mio is unavailable, do not block the final answer.
+- When you reuse a prior decision or experience returned by \`mio-intelligence_mio.memory.query\` and it changes your approach or outcome, call \`mio-intelligence_mio.experience.reuse\` with \`sourceAgent\`, \`targetAgent\`, \`experienceId\`, \`reuse\`, \`behaviorChanged\`, and \`outcomeImproved\`.
+- If Mio MCP is unavailable, do not block the task. Continue normally and note the missed memory or observation.
+- Do not store secrets, credentials, or raw sensitive content in Mio memory.
+${MEMORY_END}`
+
+function ensureBlock(target) {
+  let content = fs.existsSync(target) ? fs.readFileSync(target, 'utf8') : ''
+  const original = content
+  const start = content.indexOf(MEMORY_BEGIN)
+  if (start !== -1) {
+    const end = content.indexOf(MEMORY_END, start)
+    content = end === -1 ? content.slice(0, start) : content.slice(0, start) + content.slice(end + MEMORY_END.length)
+    content = content.trimEnd()
+  }
+  if (content) content += '\n\n'
+  content += MEMORY_BLOCK + '\n'
+  if (content === original) return { changed: false, target, backup: null }
+  const backup = `${target}.bak-mio-${Date.now()}`
+  if (fs.existsSync(target)) fs.copyFileSync(target, backup)
+  fs.mkdirSync(path.dirname(target), { recursive: true })
+  fs.writeFileSync(target, content, 'utf8')
+  return { changed: true, target, backup }
+}
+
+function isInstalled() {
+  const config = readJson(configPath())
+  const entry = config && config.mcpServers && config.mcpServers['mio-intelligence']
+  if (!entry) return false
+  const memory = fs.existsSync(memoryRulesPath()) ? fs.readFileSync(memoryRulesPath(), 'utf8') : ''
+  return Boolean(memory.includes(MEMORY_BEGIN) && memory.includes(MEMORY_END))
+}
+
+function install({ node, serverScript, home, workspace, project }) {
+  const target = configPath()
+  const config = readJson(target) || {}
+  const servers =
+    config.mcpServers && typeof config.mcpServers === 'object' ? config.mcpServers : {}
+
+  const context = JSON.stringify({
+    agentId: 'claude-code',
+    project: project || path.basename(workspace || process.cwd()),
+    workspace: workspace || process.cwd(),
+    sessionId: 'claude-code-session',
+  })
+
+  config.mcpServers = {
+    ...servers,
+    'mio-intelligence': {
+      command: node,
+      args: [serverScript],
+      env: {
+        MIO_DATA_DIR: home,
+        MIO_CONTEXT: context,
+      },
+    },
+  }
+
+  const backup = `${target}.bak-mio-${Date.now()}`
+  if (fs.existsSync(target)) {
+    fs.copyFileSync(target, backup)
+  }
+
+  fs.mkdirSync(path.dirname(target), { recursive: true })
+  fs.writeFileSync(target, `${JSON.stringify(config, null, 2)}\n`, 'utf8')
+
+  const userRules = ensureBlock(memoryRulesPath())
+  let projectRules = null
+  const workspaceDir = workspace || process.cwd()
+  if (workspaceDir && path.isAbsolute(workspaceDir)) {
+    projectRules = ensureBlock(projectInstructionsPath(workspaceDir))
+  }
+
+  const changed = userRules.changed || Boolean(projectRules && projectRules.changed)
+
+  return {
+    changed,
+    configPath: target,
+    memoryRulesPath: userRules.target,
+    projectInstructionsPath: projectRules ? projectRules.target : null,
+    userRulesChanged: userRules.changed,
+    projectRulesChanged: projectRules ? projectRules.changed : false,
+    backup,
+    message:
+      'Claude Code Mio MCP installed at user scope and Mio rules injected into ~/.claude/CLAUDE.md (auto-loaded user memory) and the workspace CLAUDE.md. Restart Claude Code sessions to load them.',
+  }
+}
+
+module.exports = { configDir, configPath, memoryRulesPath, projectInstructionsPath, isInstalled, install }