[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-2073":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":12,"openIssues":14,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":14,"stars7d":14,"stars30d":14,"stars90d":14,"forks30d":14,"starsTrendScore":14,"compositeScore":15,"rankGlobal":10,"rankLanguage":10,"license":16,"archived":17,"fork":17,"defaultBranch":18,"hasWiki":19,"hasPages":17,"topics":20,"createdAt":10,"pushedAt":10,"updatedAt":21,"readmeContent":22,"aiSummary":23,"trendingCount":14,"starSnapshotCount":14,"syncStatus":24,"lastSyncTime":25,"discoverSource":26},2073,"Metis-agent","l3gi0nXXXX\u002FMetis-agent","l3gi0nXXXX","Metis-agent is a high performance and personal AI assistant which is build with Cangjie program language. And this project is based on magic-cli and CangjeMagic.","",null,"Cangjie",130,13,0,43.44,"MIT License",false,"main",true,[],"2026-06-12 04:00:13","# Metis— High Performance & Personal  AI Assistant\n\n![log](.\u002Fdocs\u002Ffigures\u002Fmetis.svg)\n\n\n\nMetis is a high performance and personal AI assistant which is build with Cangjie program language. And this project is based on magic-cli and CangjeMagic.\n\nIt combines an interactive CLI, a unified Gateway-first runtime, channel adapters, skills, persistent sessions, and operational surfaces for running agents in real workflows instead of one-off chat demos. Supported channels include: Telegram, QQ, Feishu.\n\n## Why Metis\n\nMetis is designed around one practical idea: the same agent runtime should be able to serve local CLI conversations, managed sessions, gateway APIs, and IM channels without splitting product behavior across unrelated code paths.\n\nThat gives the project a few defining characteristics:\n\n- Gateway-first runtime for default chat, automation, and channel delivery paths\n- Cangjie-based agent implementation with Cangjie Magic integration\n- Interactive CLI for development, debugging, and daily use\n- Skills system built on `SKILL.md` packages with workspace and user-level discovery\n- MCP integration for external tools and service extension\n- Persistent conversations, transcript-backed sessions, and workspace memory\n- Cron scheduling for recurring agent tasks\n- Channel delivery for Feishu, QQ, Telegram, and plugin-driven IM integrations\n- Control and observability surfaces for dashboard, HTTP, and runtime inspection\n\n## High performance\n\nWithout undergoing any specialized performance optimization, Metis has achieved a more than 20-fold leap in startup and shutdown efficiency compared with vendor baseline in the measured baseline scenario.\n\nAbsent any specialized performance optimization, Metis has realized a qualitative leap in its startup and shutdown efficiency:\n\n| Comparison Item               | vendor baseline Average Duration | Metis Average Duration | Magnification Factor |\n| ----------------------------- | ------------------------- | ---------------------- | -------------------- |\n| onboard（Cold Start）         | 2\\.591s                   | 0\\.122s                | 21\\.2 times          |\n| gateway restart（Cold Start） | 2\\.366s                   | 0\\.073s                | 32\\.4 times          |\n| gateway start（Cold Start）   | 2\\.174s                   | 0\\.074s                | 29\\.4 times          |\n| gateway stop（Cold Start）    | 1\\.979s                   | 0\\.073s                | 27\\.1 times          |\n\nThe detailed data is presented below; notably, these figures were achieved without any performance optimizations whatsoever (perhaps because Metis's feature set is not yet fully comprehensive?).\n\n| Num     | vendor baseline onboard | vendor baseline gateway restart | vendor baseline gateway start | vendor baseline gateway stop |\n| ------- | ---------------: | -----------------------: | ---------------------: | --------------------: |\n| 1       |           3.563s |                   1.989s |                 2.462s |                2.136s |\n| 2       |           2.470s |                   2.044s |                 2.397s |                1.955s |\n| 3       |           2.476s |                   2.555s |                 2.108s |                1.950s |\n| 4       |           2.473s |                   2.381s |                 2.089s |                1.962s |\n| 5       |           2.474s |                   2.391s |                 2.103s |                1.970s |\n| 6       |           2.493s |                   2.491s |                 2.118s |                1.957s |\n| 7       |           2.468s |                   2.421s |                 2.135s |                1.972s |\n| 8       |           2.505s |                   2.545s |                 2.099s |                1.966s |\n| 9       |           2.491s |                   2.432s |                 2.114s |                1.959s |\n| 10      |           2.500s |                   2.413s |                 2.117s |                1.961s |\n| average |           2.591s |                   2.366s |                 2.174s |                1.979s |\n\n| Num     | Metis onboard | Metis gateway restart | Metis gateway run | Metis gateway stop |\n| :------ | ------------: | --------------------: | ----------------: | -----------------: |\n| 1       |        0.562s |                0.072s |            0.075s |             0.071s |\n| 2       |        0.068s |                0.071s |            0.073s |             0.073s |\n| 3       |        0.076s |                0.074s |            0.077s |             0.073s |\n| 4       |        0.074s |                0.076s |            0.073s |             0.074s |\n| 5       |        0.076s |                0.073s |            0.074s |             0.073s |\n| 6       |        0.071s |                0.075s |            0.074s |             0.075s |\n| 7       |        0.071s |                0.075s |            0.073s |             0.077s |\n| 8       |        0.076s |                0.072s |            0.075s |             0.071s |\n| 9       |        0.072s |                0.074s |            0.073s |             0.073s |\n| 10      |        0.071s |                0.073s |            0.073s |             0.072s |\n| average |        0.122s |                0.073s |            0.074s |             0.073s |\n\nTest Environment:\n\n* MacBook Air 2026 M5 (10-core CPU + 10-core GPU) | 16GB RAM + 256GB SSD\n* macOS 26.4 (25E246)\n\n## Quick Start\n\nMetis is a source-first Cangjie project. Start from the core dependencies, build once, then install optional feature dependencies only when you use those features.\n\n### 1. Install Core Dependencies\n\n| Dependency | Required for | Setup |\n|---|---|---|\n| Cangjie SDK `1.0.0` | `cjpm build`, `cjpm run`, `cjpm test` | Download Cangjie LTS General Edition 1.0.0 from https:\u002F\u002Fcangjie-lang.cn\u002Fdownload\u002F1.0.0, install it locally, and source its `envsetup.sh` before building. |\n| CangjieMagic source checkout | Metis `magic` dependency | Clone\u002Fbuild CangjieMagic locally, then export `MAGIC_PATH` to its repository root. |\n| Cangjie stdx libraries bundled with CangjieMagic | stdx JSON\u002FHTTP\u002FTLS\u002Fruntime modules | Keep the `libs\u002Fcangjie-stdx-*` directories inside `MAGIC_PATH`. |\n| OpenSSL 3 | HTTPS\u002FTLS through Cangjie stdx | On macOS install `openssl@3` and export `DYLD_LIBRARY_PATH`. |\n| macOS SDK path (`SDKROOT`) | macOS Cangjie build\u002Flink step | On macOS, install Xcode or the macOS SDK from Apple Developer Downloads: https:\u002F\u002Fdeveloper.apple.com\u002Fdownload\u002Fall, then export `SDKROOT` to the installed `MacOSX.sdk` path. |\n| curl \u002F libcurl | CangjieMagic HTTP backend | Ensure `curl` and its runtime library are available on the host. |\n| Git | Fetching source dependencies | Install system Git. |\n| C compiler toolchain and `make` | Only when rebuilding FFI libraries | Not needed for normal build if the checked-in FFI libraries are present. Install Xcode Command Line Tools on macOS if rebuilding. |\n\n### 2. Configure Build Environment\n\n```bash\n# Cangjie SDK\nexport CANGJIE_HOME=\"\u002Fpath\u002Fto\u002Fcangjie-sdk-1.0.0\"\nsource \"$CANGJIE_HOME\u002Fenvsetup.sh\"\n\n# CangjieMagic dependency used by cjpm.toml\nexport MAGIC_PATH=\"\u002Fpath\u002Fto\u002FCangjieMagic\"\n\n# stdx TLS\u002FOpenSSL runtime on macOS\nbrew install openssl@3\nexport DYLD_LIBRARY_PATH=\"\u002Fopt\u002Fhomebrew\u002Fopt\u002Fopenssl@3\u002Flib:$DYLD_LIBRARY_PATH\"\n\n# macOS SDK used by the Cangjie build\u002Flink step\nexport SDKROOT=\"\u002Fpath\u002Fto\u002FXcode.app\u002FContents\u002FDeveloper\u002FPlatforms\u002FMacOSX.platform\u002FDeveloper\u002FSDKs\u002FMacOSX.sdk\"\n```\n\n### 3. Build CangjieMagic\n\n```bash\ncd \"$MAGIC_PATH\"\ncjpm clean\ncjpm build -i\n```\n\n### 4. Build and Verify Metis\n\n```bash\ncd \u002Fpath\u002Fto\u002FMetis\ncjpm clean\ncjpm build -i\ncjpm test --parallel 1\n```\n\n### 5. Configure Runtime Credentials\n\nThe build can pass without model or IM credentials, but real conversations need provider configuration under `~\u002F.metis\u002Fmetis.json`.\n\nRun the onboarding flow for first-time setup:\n\n```bash\ncjpm run --skip-build --name metis --run-args \"onboard\"\n```\n\nAt minimum, configure one chat model provider, for example `qwen`, `deepseek`, `openai`, or another provider supported by the model catalog:\n\n```json\n{\n  \"models\": {\n    \"providers\": {\n      \"qwen\": {\n        \"apiKey\": \"\u003Cyour-api-key>\",\n        \"baseUrl\": \"https:\u002F\u002Fcoding.dashscope.aliyuncs.com\u002Fv1\"\n      }\n    }\n  },\n  \"agents\": {\n    \"defaults\": {\n      \"model\": {\n        \"primary\": \"qwen\u002Fqwen3.6-plus\"\n      }\n    }\n  }\n}\n```\n\nFor Telegram image understanding, configure a dedicated image model whose model catalog `input` includes `image`. Qwen `qwen\u002Fqwen3.6-plus` is built in as `text,image`; unknown custom models must declare `input` metadata under `models.providers.\u003Cprovider>.models[]` or `gateway.media.image.models[]`.\n\nIM channels need their own credentials only when that channel is enabled:\n\n- Telegram: bot token from BotFather, configured under `gateway.telegram` or a Telegram account entry.\n- Feishu: app id\u002Fapp secret. The default long-connect path uses the Metis-owned vendor SDK-backed Node sidecar, so install its local Node dependencies before enabling `gateway.feishu.receiveMode = \"long_connect\"`.\n- QQ: app id\u002Fapp secret or the official gateway credentials required by the selected QQ mode.\n- Plugin IM adapters: plugin runtime files plus the Python dependencies and credentials for that adapter.\n\n### 6. Start Metis\n\nInteractive CLI:\n\n```bash\ncjpm run --name metis\ncjpm run --skip-build --name metis --run-args \"chat\"\n```\n\nGateway runtime:\n\n```bash\ncjpm run --skip-build --name metis --run-args \"gateway run\"\ncjpm run --skip-build --name metis --run-args \"gateway serve\"\n```\n\nOther entry points:\n\n```bash\ncjpm run --skip-build --name metis --run-args \"dashboard\"\ncjpm run --skip-build --name metis --run-args \"gateway help\"\n```\n\nLogs and runtime diagnostics:\n\n```bash\ncjpm run --skip-build --name metis --run-args \"logs path\"\ncjpm run --skip-build --name metis --run-args \"logs recent\"\ncjpm run --skip-build --name metis --run-args \"logs current\"\ncjpm run --skip-build --name metis --run-args \"logs tail --limit 100\"\n```\n\nUse `gateway run --verbose` for temporary runtime detail while diagnosing channel startup, inbound, outbound, or model failures. See [docs\u002Fuser\u002Flogging.md](docs\u002Fuser\u002Flogging.md) for the logging workflow.\n\n### Optional Feature Dependencies\n\n| Feature | Dependency | Install command | When to install |\n|---|---|---|---|\n| Control UI source rebuild and UI tests | Node.js + npm from the official Node.js download page: https:\u002F\u002Fnodejs.org\u002Fen\u002Fdownload | `npm --prefix ui install` | Required when changing `ui\u002F` source or running `npm --prefix ui run build` \u002F UI tests. The committed `assets\u002Fcontrol-ui` bundle is enough for normal Gateway startup. |\n| PDF fallback extraction | Node.js + npm from the official Node.js download page: https:\u002F\u002Fnodejs.org\u002Fen\u002Fdownload; `tools\u002Fpdf_extract` packages `pdfjs-dist` and `@napi-rs\u002Fcanvas` | `npm --prefix tools\u002Fpdf_extract install` | Required when PDF analysis uses non-native PDF models and Metis must extract text or render page images locally. Check with `metis models pdf-status`. |\n| Feishu vendor SDK-backed long-connect sidecar | Node.js + npm; `tools\u002Ffeishu-sidecar` packages `@larksuiteoapi\u002Fnode-sdk` and `https-proxy-agent`. | `npm --prefix tools\u002Ffeishu-sidecar install` | Required for `gateway.feishu.receiveMode = \"long_connect\"` with `gateway.feishu.transportMode = \"sidecar\"`. `tools\u002Ffeishu-sidecar\u002Fnode_modules` is intentionally ignored; do not force-add it or bypass `.gitignore`. |\n| Gateway plugin IM adapters | Python 3 + pip; per-channel requirements under `tools\u002Fgateway_plugin_tool\u002Frequirements\u002F` | `python tools\u002Fgateway_plugin_tool\u002Finstall.py deps all` or `python tools\u002Fgateway_plugin_tool\u002Finstall.py deps dingtalk` | Required only for plugin-style adapters such as DingTalk, WeChat, WeCom, plugin Feishu, or plugin QQ. Built-in Telegram\u002FQQ\u002FFeishu Cangjie adapters do not need these Python packages. |\n| Docker image build | Docker + Docker Compose; `CANGJIE_HOME` pointing to SDK | `scripts\u002Fbuild-docker-image.sh` | Required only when building the container image. |\n| Faster source search | ripgrep | `brew install ripgrep` | Optional. Metis\u002FMagic workflows can fall back to slower search tools, but `rg` is recommended for development. |\n\nRun optional verification only for the features you installed:\n\n```bash\nnpm --prefix ui run build\nnpm --prefix tools\u002Fpdf_extract run check\nnpm --prefix tools\u002Ffeishu-sidecar install\ncjpm run --skip-build --name metis --run-args \"models pdf-status\"\ncjpm run --skip-build --name metis --run-args \"gateway channels get feishu\"\n```\n\n### Common Build Failures\n\n| Symptom | Likely cause | Fix |\n|---|---|---|\n| `MAGIC_PATH` dependency cannot be resolved | `MAGIC_PATH` is missing or points to the wrong directory. | `export MAGIC_PATH=\"\u002Fpath\u002Fto\u002FCangjieMagic\"` and verify `$MAGIC_PATH\u002Fcjpm.toml` exists. |\n| `cjc` \u002F `cjpm` not found | Cangjie SDK env was not sourced. | `source \"$CANGJIE_HOME\u002Fenvsetup.sh\"`. |\n| macOS build fails during compile\u002Flink with missing SDK or system headers\u002Flibraries | `SDKROOT` is missing or points to the wrong macOS SDK. | Install Xcode or the macOS SDK from https:\u002F\u002Fdeveloper.apple.com\u002Fdownload\u002Fall, then `export SDKROOT=\"\u002Fpath\u002Fto\u002FXcode.app\u002FContents\u002FDeveloper\u002FPlatforms\u002FMacOSX.platform\u002FDeveloper\u002FSDKs\u002FMacOSX.sdk\"`. |\n| TLS\u002FOpenSSL errors on macOS | OpenSSL 3 runtime library is not visible to dyld. | `brew install openssl@3` and export `DYLD_LIBRARY_PATH=\"\u002Fopt\u002Fhomebrew\u002Fopt\u002Fopenssl@3\u002Flib:$DYLD_LIBRARY_PATH\"`. |\n| `stdx` `.dylib` fails with `library load disallowed by system policy` on macOS | macOS quarantine attribute on downloaded libraries. | Run `xattr -rd com.apple.quarantine \"$MAGIC_PATH\u002Flibs\"`; if the directory is not writable by your user, retry with `sudo`. |\n| PDF upload says extractor failed or `pdfjs-dist` is `not_loadable` | PDF fallback dependencies are not installed. | Install Node.js from https:\u002F\u002Fnodejs.org\u002Fen\u002Fdownload, run `npm --prefix tools\u002Fpdf_extract install`, restart Gateway, then run `metis models pdf-status`. |\n| Feishu long-connect status reports `sidecarDependencyStatus=missing` or `invalid` | The sidecar script, runtime package files, Node\u002Fnpm, or `tools\u002Ffeishu-sidecar\u002Fnode_modules` are missing. | Install Node.js from https:\u002F\u002Fnodejs.org\u002Fen\u002Fdownload, run `npm --prefix tools\u002Ffeishu-sidecar install`, keep `node_modules` untracked, then restart Gateway and check `metis gateway channels get feishu` or `metis gateway status`. |\n| Feishu long-connect transport startup fails after dependencies are ready | Feishu app credentials are missing or the sidecar transport rejected startup. | Verify `gateway.feishu.appId`, `gateway.feishu.appSecret`, `gateway.feishu.receiveMode = \"long_connect\"`, and `gateway.feishu.transportMode = \"sidecar\"`, then restart Gateway. |\n\n## Common Workflows\n\n### Interactive Work\n\nUse the CLI when you want a local conversational interface with memory, commands, skills, and project instructions.\n\n```bash\ncjpm run --name metis\n```\n\nInside the shell, common commands include:\n\n- `\u002Fhelp`\n- `\u002Fconversation list`\n- `\u002Fconversation save \u003Cname>`\n- `\u002Fconversation resume \u003Cname>`\n- `\u002Fmemory`\n- `\u002Fskills list`\n- `\u002Fmcp`\n- `\u002Fcron`\n\n### Gateway-Backed Agent Calls\n\nUse the gateway runtime when you want a long-lived operational surface for CLI chat, APIs, managed sessions, channels, cron jobs, and delivery workflows.\n\n```bash\nmetis gateway status\nmetis gateway health\nmetis gateway discover\n```\n\n### Skills\n\nSkills are packaged as directories containing `SKILL.md` and can be discovered from bundled, user, and workspace locations.\n\nExamples:\n\n```text\n\u002Fskills list\n\u002Fskills info weather\n\u002Fskills search weather\n\u002Fweather Shanghai\n```\n\n### MCP\n\nMetis can manage MCP servers and expose its unified runtime through an MCP bridge.\n\n```bash\nmetis mcp list\nmetis mcp serve --url ws:\u002F\u002F127.0.0.1:18788\u002Fws\n```\n\n### Sessions and Agents\n\nMetis supports managed agent sessions, bindings, and operational inspection through the gateway path.\n\n```bash\nmetis agents list\nmetis sessions list\nmetis subagents list\n```\n\nAgentTeam can be managed today with agent CLI commands plus Gateway RPC team calls:\n\n```bash\nmetis agents add --agent content-writer --name \"Content Writer\" --model openai:gpt-4o-mini\nmetis gateway call agents.teams.create '{\"id\":\"content\",\"displayName\":\"Content Team\",\"template\":\"pm-writer-reviewer\"}'\nmetis agents bind --agent content-writer --bind telegram:bot-a\n```\n\nSee `docs\u002Fuser\u002Fagent-team.md` for Gateway startup, team RPC usage, per-agent files\u002Fmodels, IM account binding, and current limitations.\n\nBackground subagents can be started from CLI or IM sessions without blocking the current conversation:\n\n```text\n\u002Fsubagents spawn explorer \"analyze the gateway runtime and notify this session when done\"\n```\n\nSee `docs\u002Fuser\u002Fsubagents.md` for CLI, Telegram, natural-language, custom agent, policy, and control-ui usage.\n\n### Control UI Slash Commands\n\nThe dashboard chat input supports the same slash menu shown by `\u002Fhelp` and `\u002Fcommands`.\nLocal Control UI commands return readable chat summaries and do not print raw JSON by default.\n\nCommon Control UI commands:\n\n```text\n\u002Fhelp\n\u002Fcommands\n\u002Fstop\n\u002Fkill \u003Cid|all>\n\u002Fsteer [id] \u003Cmessage>\n\u002Fredirect [id] \u003Cmessage>\n```\n\n`\u002Fstop` aborts only the current Control UI chat turn. `\u002Fkill` targets matching sub-agent sessions in the current Control UI session subtree. `\u002Fsteer` soft-injects guidance into an active run without restarting it; use `\u002Fredirect` when the run should be aborted and restarted with a new message.\n\n### Scheduled Jobs\n\nCron jobs let you run agent work on recurring schedules and optionally route successful results back into channels or session targets.\n\n```bash\nmetis gateway cron list\nmetis gateway cron run \u003Cjob-id>\n```\n\n## Configuration\n\nMetis stores user-facing runtime configuration under `~\u002F.metis\u002F`.\n\nImportant files and directories:\n\n- `~\u002F.metis\u002Fmetis.json` for model, provider, gateway, tool, and MCP configuration\n- `~\u002F.metis\u002Fskills\u002F` for user-managed skills\n- `~\u002F.metis\u002Fagents\u002F` for custom subagents\n- `~\u002F.metis\u002Fconversation-history\u002F` for saved conversations\n- project `AGENTS.md` for workspace-specific instructions\n\n## Architecture at a Glance\n\nThe repository is organized around a few major surfaces:\n\n- `src\u002Fapp\u002F` for CLI interaction and command handling\n- `src\u002Fgateway\u002F` for the unified runtime, channels, toolsets, transport, dashboard, and control UI\n- `src\u002Fcore\u002F` for conversations, memory, policies, and shared runtime behavior\n- `src\u002Fcron\u002F` for recurring job scheduling and delivery\n- `src\u002Fmcp\u002F` for MCP bridge and server support\n- `src\u002Flsp\u002F` for richer code-assistance infrastructure\n- `docs\u002F` for runtime, parity, control, skills, and integration documentation\n- `skills\u002F` for bundled skills\n- `tools\u002F` for channel plugin tooling and supporting integrations\n\n## Channel and Plugin Story\n\nMetis ships with built-in channel support and also supports plugin-style integrations for external IM systems. The gateway\u002Fplugin model is designed so channels can feed normalized inbound events into the runtime and reuse the same agent, session, and delivery infrastructure.\n\nThat makes Metis suitable for:\n\n- terminal-first agent workflows\n- workspace-aware coding assistants\n- private personal assistants with persistent context\n- channel-connected bots for teams or personal messaging\n- scheduled and semi-autonomous agent tasks\n\n## Documentation\n\nUseful project documents:\n\n- `README.md` for the current project intro and command overview\n- `docs\u002Fuser\u002Ftelegram.md` for Telegram channel configuration, media usage, and troubleshooting\n- `docs\u002Fuser\u002Fpdf-tool.md` for PDF reading, model configuration, fallback dependency installation, and Telegram PDF usage\n- `docs\u002Fuser\u002Fsubagents.md` for background subagent usage and control-ui operations\n- `docs\u002Fuser\u002Fskills-guide.md` for skills discovery, commands, and installation\n- `docs\u002Fuser\u002Fgateway-im-plugins.md` for IM plugin integration\n- `docs\u002Fmcp.md` for MCP-related configuration\n\n## Development Notes\n\nMetis is a source-first Cangjie project. In practice, the most important setup steps are making `MAGIC_PATH` valid, configuring provider credentials in `~\u002F.metis\u002Fmetis.json`, and deciding whether you are operating in interactive CLI mode, explicit local mode, or Gateway service mode.\n\nIf you are evaluating the codebase, start with:\n\n1. `cjpm.toml`\n2. `src\u002Fprogram_runner.cj`\n3. `src\u002Fapp\u002F`\n4. `src\u002Fgateway\u002F`\n\n## Thanks\n\nThis is a personal project based on magic-cli and CangjieMagic.\n\n* https:\u002F\u002Fgitcode.com\u002FCangjie-TPC\u002FCangjieMagic\n* https:\u002F\u002Fgitcode.com\u002Fcangjie-sig\u002Fmagic-cli\n","Metis-agent 是一个高性能的个人AI助手，使用Cangjie编程语言构建，并基于magic-cli和CangjeMagic。其核心功能包括一个交互式CLI、统一的网关优先运行时、多渠道适配器（如Telegram、QQ、飞书）、技能系统以及持久会话支持。该项目特别适合需要在本地命令行界面、管理会话、网关API及即时通讯渠道中无缝切换的应用场景。Metis的设计理念在于通过单一代理运行时服务多种交互方式，从而避免了不同代码路径之间的行为分裂。此外，尽管没有进行专门的性能优化，Metis在启动与关闭效率上相比基准实现了显著提升，部分测试条件下甚至达到了20倍以上的加速效果。",2,"2026-06-11 02:47:58","CREATED_QUERY"]