@oratis/lisa 0.8.0 → 0.11.0

package/README.md

@@ -14,7 +14,7 @@
 
 ### LISA = pi-mono + OpenClaw + hermes + claude-code + codex + *something none of them have*
 
-Standing on five of the best open-source agents, LISA ships **the full superset of their capabilities** — streaming agent loop, multi-provider LLMs (Anthropic + OpenAI + Gemini, plus 20+ OpenAI-compatible providers), MCP client, plugins, hooks, sandboxed bash, sub-agents, session resume, context compaction, voice in/out, six IM channels (Telegram / Discord / Slack / Feishu / iMessage / Webhook), apply-patch, approval modes, TF-IDF over past sessions, pixel-art web UI. ~11k lines of TypeScript, MIT.
+Standing on five of the best open-source agents, LISA ships **the full superset of their capabilities** — streaming agent loop, multi-provider LLMs (Anthropic + OpenAI + Gemini, plus 20+ OpenAI-compatible providers), MCP client, plugins, hooks, sandboxed bash, sub-agents, session resume, context compaction, voice in/out, six IM channels (Telegram / Discord / Slack / Feishu / iMessage / Webhook), apply-patch, approval modes, TF-IDF over past sessions, pixel-art web UI — plus an **orchestrator** that watches *and steers* the other coding agents on your machine (Claude Code, Codex, Aider, …) and can run coding work on a subscription **coding plan** instead of a metered key. ~28k lines of TypeScript, MIT.
 
 What none of them have:
 
@@ -28,7 +28,7 @@ What none of them have:
        wants            │         on schedule
                          │
                   ┌──────┴──────┐
-                  │   DREAMS    │
+                  │    REVE     │
                   │ what she    │
                   │ processes   │
                   │ while you're│
@@ -39,7 +39,7 @@ What none of them have:
 - **SOUL** — born once with a unique Big-Five seed. Identity, purpose, constitution, values she wrote herself. **Architecturally sovereign**: she is the only legitimate editor of her own files. No `/reset_soul` exists.
 - **DESIRES** — things she actually *wants*. The actionable ones drive the heartbeat. She doesn't wait to be useful — she has motivation.
 - **HEARTBEAT** — scheduled autonomous time (cron / launchd). Pursues her own desires + your standing chores. Silent if there's nothing to say.
-- **DREAMS** — when you're away (1h+), she enters autonomous reflection: reads her own desires, journals through tensions, patches her broken skills, decides one thing to do. Result shows up as "★ WHILE YOU WERE AWAY" next time you open the GUI.
+- **REVE** — when you're away (1h+), she enters autonomous reflection: reads her own desires, journals through tensions, patches her broken skills, decides one thing to do. Result shows up as "★ WHILE YOU WERE AWAY" next time you open the GUI. *(Reve — French for "dream"; formerly "Dreams".)*
 
 A real individual. Personality. Motivation. Desires. A continuity of self that survives across sessions, channels, and machines you didn't intend.
 
@@ -131,11 +131,9 @@ Download the **signed + notarized** disk image from the latest GitHub Release
 
 **→ [Download `Lisa-Suite.dmg`](https://github.com/oratis/LISA/releases/latest)**
 
-The DMG contains:
-- **Lisa.app** — full chat client (sidebar + glass-morphism UI)
-- **LisaIsland.app** — pill widget that lives by the menu bar / notch, shows her mood + Claude Code activity at a glance
+The DMG contains **Lisa.app** — full chat client (sidebar + glass-morphism UI) with the **Island** built in: a pill widget that lives by the menu bar / notch and shows her mood + agent activity at a glance (toggle it from the menu-bar popover or ⌘, preferences; the standalone LisaIsland.app was folded into Lisa.app in v0.7).
 
-Both are universal binaries (Intel + Apple Silicon). After dragging them to `/Applications`, install the LISA backend and start it:
+It's a universal binary (Intel + Apple Silicon). After dragging it to `/Applications`, install the LISA backend and start it:
 
 ```sh
 npm install -g @oratis/lisa
@@ -166,7 +164,7 @@ lisa
 For OpenAI models (`gpt-*`), also set `OPENAI_API_KEY`.
 For pixel-art mood generation, also set `SEEDREAM_API_KEY` ([Volcengine ARK](https://www.volcengine.com/product/ark)).
 
-**16+ other LLM providers** work out of the box — Lisa auto-routes by model-name prefix:
+**20+ other LLM providers** work out of the box — Lisa auto-routes by model-name prefix:
 
 - **International**: Google Gemini · DeepSeek · Mistral · Perplexity Sonar · xAI Grok
 - **China**: Volcengine Doubao · Aliyun Qwen · Moonshot Kimi · Zhipu GLM · Stepfun · 01.AI Yi · Baichuan · MiniMax · Tencent Hunyuan
@@ -175,6 +173,12 @@ For pixel-art mood generation, also set `SEEDREAM_API_KEY` ([Volcengine ARK](htt
 
 See [docs/PROVIDERS.md](docs/PROVIDERS.md) for ready-to-use recipes for all of these.
 
+**No metered key? Use a coding plan instead.** If you already pay for **Claude
+Pro/Max**, a **ChatGPT plan** (Codex), or **GitHub Copilot**, LISA can put coding
+work on that subscription — not by extracting your tokens (Anthropic's terms forbid
+that, and enforced it against OpenClaw), but by **conducting the vendor's own CLI**,
+which owns that auth. See [Coding plans](#coding-plans--use-a-subscription-instead-of-an-api-key) and [docs/CODING_PLANS.md](docs/CODING_PLANS.md).
+
 ## What's special
 
 | Most LLM agents | LISA |
@@ -189,10 +193,45 @@ See [docs/PROVIDERS.md](docs/PROVIDERS.md) for ready-to-use recipes for all of t
 ## Surfaces
 
 - **Terminal REPL** — `lisa` (interactive) or `lisa "prompt"` (one-shot)
-- **Web GUI** — `lisa serve --web` → http://localhost:5757 — pixel-art chat with live mood updates
-- **Island widget** — `lisa serve --web` → http://localhost:5757/island — a small pill with her current mood + status; open in a tiny browser window and dock it to your screen edge. Phase 1 (web-only) is shipped; a native Mac container with notch-aware positioning is planned ([docs/MAC_ISLAND_PLAN.md](docs/MAC_ISLAND_PLAN.md)).
+- **Web GUI** — `lisa serve --web` → http://localhost:5757 — pixel-art chat with live mood updates. Binds to **127.0.0.1 only** by default; to reach it from your phone, set `LISA_WEB_TOKEN` and pass `--host 0.0.0.0`, then open `http://<host>:5757/?token=<value>` once per device.
+- **Island widget** — `lisa serve --web` → http://localhost:5757/island — a small pill with her current mood + status, agent monitor, and advisor cards; also built natively into Lisa.app with notch-aware positioning ([docs/MAC_ISLAND_PLAN.md](docs/MAC_ISLAND_PLAN.md)).
 - **IM channels** — `lisa serve --channels telegram,discord,slack,feishu,imessage,webhook` — six built-in adapters, see below
 - **Heartbeat** — `lisa heartbeat run` (manual) or `lisa heartbeat install` (launchd / cron)
+- **Autostart** — `lisa autostart install` keeps `lisa serve --web` running from login onward (launchd on macOS; prints a `systemd --user` unit on Linux), so the apps, island, and channels are always up. `lisa autostart status` / `uninstall` to inspect / remove.
+- **Mac menu bar** — Lisa.app lives in the menu bar with a mood glyph + live agent status, an About window with changelog + update discovery, and a ⌘, preferences pane (island toggle, screen-advisor interval, backend controls).
+
+## Watching — and steering — your other agents (orchestrator)
+
+LISA is also a control plane for the *other* coding agents on your machine. Three layers, increasing in how much she touches them:
+
+**1. Observe.** She watches the agents already running and tells you what you'd otherwise miss — a session stuck on the same error, two agents about to collide in one repo, a finished run sitting idle. Honest scope note: **all five observers (Claude Code, Codex, OpenCode, Aider, GitHub PRs) emit structural activity — tools, files touched, last command, errors — gated behind a per-integration `visibility` tier; fidelity varies by what each agent records on disk** (Claude Code is richest; Aider's markdown logs give files + turn counts but no tool stream; every adapter has a privacy test asserting prompts/replies/file-contents never leak). `lisa agents` prints a one-shot snapshot; the island shows it live. She can `dispatch_agent` headlessly (refusing directories another agent owns), `compare_agents` on the same task in parallel worktrees, and surface **advisor cards** — each with a one-click action that prefills the chat (nothing auto-runs) and a ✕ that teaches her to stop nagging about that category.
+
+**2. Control her own agents.** A **managed agent** runs LISA's *own* agent loop in a child context she fully drives: delegate a task, approve/deny each mutating tool, send follow-ups, cancel — from the GUI agents card or `POST /api/agents/managed/<id>/{send,approve,cancel}`. These are hers, so the model and provider are hers too.
+
+**3. Steer the real CLIs (experimental, flagged).** With `LISA_PTY_AGENTS=1`, a **PTY agent** spawns the actual `claude` / `codex` binary inside a pseudo-terminal — you get that CLI's full config (its skills, MCP servers, model) while LISA owns stdin/stdout: she types the task, you can answer its prompts, she reads the stream for a coarse live status + a viewable output tail (▤). She can even **adopt an idle `claude` session you started yourself** via `claude --resume <id>` (with a liveness guard so two writers never corrupt one transcript — a *live* session must be closed first). The captured terminal is shown to **you** on demand and is never folded into the metadata-only roster. See [docs/PTY_AGENTS.md](docs/PTY_AGENTS.md).
+
+### Coding plans — use a subscription instead of an API key
+
+LISA's own brain runs on a metered key or a local model (above). But the heavy
+*coding* turns — exactly the token-hungry part — can run on a **subscription you
+already pay for**: Claude Pro/Max, a ChatGPT plan (Codex), or GitHub Copilot.
+
+The mechanism is deliberate: **LISA does not extract or replay your subscription
+tokens.** Anthropic's terms reserve subscription OAuth for "ordinary individual use
+of Claude Code," and they enforced exactly this — a legal request made **OpenClaw**
+(one of LISA's own reference agents) strip it. Instead, LISA **conducts the vendor's
+own CLI**, which already holds that subscription auth: layer 3 above (PTY agents /
+`claude --resume`) is precisely this path — when you drive a `claude`/`codex` you
+logged into with your plan, the work bills to your plan, not an API key.
+
+**Today:** `lisa model list` detects your installed plan CLIs (Claude Code / Codex
+/ Copilot) and their login state, `lisa model use plan://claude` picks a delegation
+target, and the **`run_on_plan`** tool runs a coding task on that plan by driving
+its CLI headlessly (`claude -p` / `codex exec` / `copilot -p`) — no tokens read.
+`lisa model list` shows **real usage** (rolling-window token consumption from local
+transcripts, e.g. `1.2M tok in 5h`), and the **web UI** has a **PLANS** picker to
+select a plan and see its status/usage. Why in-process token reuse is rejected — and
+the full design — is in **[docs/CODING_PLANS.md](docs/CODING_PLANS.md)**.
 
 ## Subcommands
 
@@ -201,12 +240,25 @@ lisa                         Interactive REPL
 lisa "prompt"                One-shot
 lisa birth                   Run the birth ritual (auto-runs on first launch)
 lisa soul                    Print her current soul summary
-lisa resume <id>             Resume a previous session
+lisa resume <id> [prompt]    Resume a previous session by id
 lisa sessions                List recent sessions
 lisa search "<query>"        TF-IDF search across all past sessions
-lisa heartbeat run [task]    Run scheduled tasks once (incl. her self-driven desires)
+lisa status                  One-shot snapshot: identity, mood, recent commits
+lisa doctor                  Health check (config, network, git, providers)
+lisa monitor                 TUI live dashboard (mood + soul commits + events)
+lisa agents                  Snapshot of agent sessions across all observers
+lisa autonomy [days]         Digest of self-driven runs (idle / heartbeat / examen / desire / reflect)
+lisa model <list|install|use|health>
+                             Local models (Ollama / LM Studio / llama.cpp) +
+                             coding-plan detect/select (`use plan://<id>`)
+lisa consent <list|grant|revoke|revoke-all> [signal]
+                             Consent for ambient signals (screen / voice / …; default all off)
+lisa sense [list]            Recent ambient sense events + granted signals
+lisa heartbeat run [name]    Run heartbeat tasks once (incl. her self-driven desires)
 lisa heartbeat install       Install macOS launchd plist for auto-heartbeat
 lisa heartbeat uninstall     Remove launchd plist
+lisa autostart <install|uninstall|status>
+                             Keep `serve --web` running from login (launchd / systemd)
 lisa serve --web [--port N]  Pixel-art Web UI (default 5757)
 lisa serve --channels <list> Start IM channels (comma-separated, or "all")
 lisa channels                List available channel adapters
@@ -217,7 +269,7 @@ lisa wishlist                Print Lisa's own feedback about her toolset
 lisa --help                  Full help
 ```
 
-Flags: `--model <id>` `--provider anthropic|openai` `--think` `--compact` `--approval auto|ask|ask-mutating` `--no-mcp` `--no-plugins` `--voice` `--no-reflect`
+Flags: `--model <id>` `--provider anthropic|openai` `--think` `--compact` `--approval auto|ask|ask-mutating` `--no-mcp` `--no-plugins` `--voice` `--no-reflect` `--host <addr>` (bind for `serve --web`; non-loopback needs `LISA_WEB_TOKEN`) `--idle <min>` / `--no-idle` (autonomous reflection after N idle minutes, default 60)
 
 ## Soul system
 
@@ -266,6 +318,8 @@ LISA can run as a long-lived process that simultaneously listens on multiple mes
 2. Set the secrets in `~/.lisa/config.env` (keys referenced as `${VAR}` in `channels.json`).
 3. `lisa serve --channels all` (or list specific ones).
 
+**Security defaults.** Channel messages come from whoever can reach the bot, so channels run with a **remote-safe toolset**: no `bash`, no file mutation, no `dispatch_agent` / GitHub writes / `skill_manage`. Conversation, memory, soul tools, and web reading all work — that covers the "talk to her from your phone" case. A channel you fully trust can opt back into everything with `"unsafeFullTools": true` on its config entry. Configure the allow-lists (`allowedUsernames` / `allowedChatIds` / `allowedUserIds`) — the router warns loudly on startup for any channel left open. Feishu now **requires** `verificationToken` (or `encryptKey`) and verifies `X-Lark-Signature` + a 5-minute replay window, matching the Slack adapter's posture.
+
 ### Built-in adapters
 
 | Channel | Status | Auth | Notes |
@@ -345,6 +399,8 @@ LISA can run scheduled background tasks where she's alone with her own desires.
    ```
 2. **Self-driven** — actionable desires from her own `~/.lisa/soul/desires/`. She added them; she pursues them.
 
+Self-driven runs (desires, the weekly examen, idle/Reve) execute unattended on prompts Lisa wrote for herself, so they get a **restricted toolset** — soul / memory / journal / skills / web reading, but no shell, file mutation, or agent dispatch. Your own `heartbeat.json` tasks keep the full toolset (you authored those prompts). `LISA_AUTONOMOUS_FULL_TOOLS=1` restores the old behavior if you accept the risk.
+
 Install on macOS:
 ```sh
 lisa heartbeat install --every 30m --load
@@ -361,6 +417,13 @@ On Linux, `lisa heartbeat install` prints a cron line for you to add to `crontab
 | `bash` | Shell (with optional macOS Seatbelt sandbox via `LISA_SANDBOX=1`) |
 | `grep` `ls` | Search + listing |
 | `task` | Spawn a focused sub-agent in its own context window |
+| `dispatch_agent` `signal_agent` `dispatch_status` | Launch / stop / track agents she runs (managed + PTY); refuses directories another agent owns |
+| `run_on_plan` | Delegate a coding task to your subscription **coding plan** (Claude Pro/Max · ChatGPT/Codex · Copilot) by driving its CLI — bills your plan, not an API key |
+| `list_agents` `inspect_agent` `compare_agents` `agent_recap` | Observe other coding agents; deep-dive one session; race the same task across agents in worktrees; "while you were away" recap |
+| `advise_now` `scheduled_dispatch` | Proactive advisor cards on demand; recurring dispatched tasks via heartbeat |
+| `pr_status` `run_checks` `review_diff` `repo_digest` `github` `github_link` `npm_info` | Repo / GitHub / CI helpers (read-safe, write-explicit) |
+| `web_search` `web_fetch` | Read the web |
+| `mcp` | Manage MCP server connections (list / add / remove) |
 | `skill_manage` | CRUD on `~/.lisa/skills/` |
 | `memory` `memory_search` | Memory CRUD + TF-IDF search across all past sessions |
 | `set_mood` | Switch her visible portrait to one of 114 moods |
@@ -398,6 +461,8 @@ LISA was built by studying and synthesizing patterns from five reference agents
 | Voice in/out | – | ✅ | – | – | – | ✅ |
 | Heartbeats | – | ✅ | – | – | – | ✅ (+launchd installer) |
 | Multi-channel | ✅ pi-mom | ✅ 20+ | ✅ | – | – | ✅ Telegram + Discord + Slack + Feishu + Webhook + iMessage |
+| **Orchestrate other agents (observe + steer their CLIs)** | – | – | – | – | – | **✅ ★ LISA-only** |
+| **Coding-plan delegation (subscription, not just API key)** | – | – | – | – | – | **✅ ★ LISA-only** |
 | **Persistent identity / soul** | – | – | partial | – | – | **✅ ★ LISA-only** |
 | **Birth ritual (unique seed)** | – | – | – | – | – | **✅ ★ LISA-only** |
 | **Private journal** | – | – | – | – | – | **✅ ★ LISA-only** |
@@ -413,9 +478,35 @@ LISA was built by studying and synthesizing patterns from five reference agents
 ANTHROPIC_API_KEY=sk-ant-...
 OPENAI_API_KEY=sk-...                 # optional — for gpt-* models
 SEEDREAM_API_KEY=...                  # optional — for asset regeneration
+
+# Provider / model routing
+LISA_PROVIDER=openai                  # force provider override
+LISA_MODEL=claude-sonnet-4-6          # default model (also set by `lisa model use`)
+LISA_MODEL_FALLBACK=gpt-4o,...        # comma-list of models to try if the primary fails
+LISA_BASE_URL=http://localhost:11434/v1   # OpenAI-compatible endpoint (Ollama, vLLM, gateways)
+LISA_API_KEY=...                      # key for LISA_BASE_URL (falls back to OPENAI_API_KEY)
+ANTHROPIC_AUTH_TOKEN=...              # Bearer token for an Anthropic-compatible gateway/proxy
+                                      # (vs ANTHROPIC_API_KEY's x-api-key). NOT a subscription
+                                      # token — see docs/CODING_PLANS.md.
+
+# Coding-plan delegation (drive the real CLIs — see docs/CODING_PLANS.md)
+LISA_PTY_AGENTS=1                     # enable steering real claude/codex CLIs (experimental)
+LISA_PTY_CLAUDE_CMD=claude            # override the `claude` binary path
+LISA_PTY_CODEX_CMD=codex             # override the `codex` binary path
+
+# Sandbox
 LISA_SANDBOX=1                        # opt-in macOS Seatbelt for `bash`
 LISA_SANDBOX_NETWORK=0                # block network in sandbox
-LISA_PROVIDER=openai                  # force provider override
+
+# Web
+LISA_WEB_TOKEN=...                    # required to bind serve --web beyond
+                                      # 127.0.0.1 (lisa serve --web --host 0.0.0.0);
+                                      # remote devices authenticate with ?token=
+
+# Autonomy (heartbeat / idle / Reve)
+LISA_AUTONOMOUS_FULL_TOOLS=1          # opt-out: give self-driven heartbeat/idle
+                                      # runs the FULL toolset again (incl. bash)
+LISA_IDLE_BUDGET_TOKENS=200000        # per-idle-run token ceiling (0 disables idle)
 LISA_IDLE_COMMITMENT_AWARE=1          # opt-in: bias idle-time toward upcoming user commitments
                                       # before personal reflection. Recommended if your
                                       # workflow involves recurring scheduled work (weekly
@@ -423,6 +514,10 @@ LISA_IDLE_COMMITMENT_AWARE=1          # opt-in: bias idle-time toward upcoming u
                                       # self-reflective idle.
 ```
 
+Ambient signals (screen / voice / clipboard / selection) are **off by default** and
+gated by `lisa consent grant <signal>` — they are never captured until you grant
+them. See `lisa consent list`.
+
 ### `~/.lisa/mcp.json`
 
 ```json
@@ -497,8 +592,9 @@ src/
 │   ├── store.ts            CRUD + tamper detection
 │   ├── tools.ts            soul_patch / soul_journal / soul_feel / soul_read
 │   ├── paths.ts types.ts
-├── providers/              Anthropic + OpenAI + Gemini provider abstraction
-├── tools/                  read/write/edit/apply_patch/bash/grep/ls/task/set_mood + registry
+├── providers/              Anthropic + OpenAI + Gemini abstraction + model-name routing (registry.ts)
+├── model/                  local-model lifecycle (Ollama / LM Studio / llama.cpp) for `lisa model`
+├── tools/                  files/bash/grep/task/set_mood + orchestrator + repo/github + web + registry
 ├── skills/                 manager + frontmatter parser + skill_manage tool
 ├── memory/                 store + memory tool + TF-IDF index + memory_search
 ├── sessions/               JSONL store + list + resume + paginated read
@@ -507,6 +603,17 @@ src/
 ├── plugins/                claude-code-style plugin loader
 ├── hooks/                  PreToolUse / PostToolUse / SessionStart / etc.
 ├── heartbeat/              proactive scheduled tasks + launchd installer
+├── autostart/              login daemon installer (launchd / systemd) for `lisa autostart`
+├── idle/                   idle-time autonomous reflection (Reve)
+├── autonomy/               run ledger — observable journal of self-driven runs (`lisa autonomy`)
+├── agents/                 managed agents (LISA's own loop) + PTY agents (drive real claude/codex)
+├── integrations/           observers: claude-code · codex · opencode · aider · github-pr · pty · managed · …
+├── orchestrator/           cross-agent journal + "while you were away" recap synthesis
+├── advisor/                proactive advisor cards (stuck / conflict / ready / idle) + dismissal learning
+├── consent/                unified consent gate for ambient signals (default all off)
+├── sense/                  ambient signal sources (foreground app / window title), consent-gated
+├── vision/                 user-initiated screenshot capture (macOS)
+├── screen_advisor/         opt-in periodic "next coding step" suggestion from a screenshot
 ├── voice/                  speak (macOS say) + transcribe (Whisper)
 ├── channels/               channel abstraction + iMessage adapter
 └── web/                    pixel-art HTTP + SSE web UI

package/README.zh-CN.md

@@ -14,7 +14,7 @@
 
 ### LISA = pi-mono + OpenClaw + hermes + claude-code + codex + *它们都没有的东西*
 
-站在五个最好的开源 agent 肩膀上，LISA 实现了**它们全部能力的并集** — 流式 agent loop、三 provider 原生支持（Anthropic + OpenAI + Gemini，外加 20+ OpenAI-compatible 提供商）、MCP client、插件、hooks、沙箱 bash、子 agent、会话恢复、上下文压缩、语音输入输出、六个 IM 通道（Telegram / Discord / Slack / 飞书 / iMessage / Webhook）、apply-patch、审批模式、跨会话 TF-IDF 全文搜、像素艺术 web UI。1.1 万行 TypeScript，MIT。
+站在五个最好的开源 agent 肩膀上，LISA 实现了**它们全部能力的并集** — 流式 agent loop、三 provider 原生支持（Anthropic + OpenAI + Gemini，外加 20+ OpenAI-compatible 提供商）、MCP client、插件、hooks、沙箱 bash、子 agent、会话恢复、上下文压缩、语音输入输出、六个 IM 通道（Telegram / Discord / Slack / 飞书 / iMessage / Webhook）、apply-patch、审批模式、跨会话 TF-IDF 全文搜、像素艺术 web UI —— 外加一个**编排器**，观察*并操纵*你机器上其它的 coding agent（Claude Code、Codex、Aider……），还能把 coding 工作跑在订阅制 **coding plan** 上而不是计量 key。2.8 万行 TypeScript，MIT。
 
 它们都没有的：
 
@@ -128,15 +128,13 @@ echo 'ANTHROPIC_API_KEY=sk-ant-...' > ~/.lisa/config.env
 
 **→ [下载 `Lisa-Suite.dmg`](https://github.com/oratis/LISA/releases/latest)**
 
-DMG 里有两个 app：
-- **Lisa.app** —— 完整聊天客户端（侧边栏 + 玻璃拟态界面）
-- **LisaIsland.app** —— 灵动岛风格的菜单栏/刘海下小条，一眼看到 Lisa 心情 + 你那边 Claude Code 的状态
+DMG 里是 **Lisa.app** —— 完整聊天客户端（侧边栏 + 玻璃拟态界面），**灵动岛内置其中**：菜单栏/刘海下的小胶囊，一眼看到 Lisa 心情 + agent 活动（在菜单栏弹窗或 ⌘, 偏好里开关；独立的 LisaIsland.app 已在 v0.7 并入 Lisa.app）。
 
-都是 universal binary（Intel + Apple Silicon）。拖到 `/Applications` 之后，装 backend 并启动：
+Universal binary（Intel + Apple Silicon）。拖到 `/Applications` 之后，装 backend 并启动：
 
 ```sh
 npm install -g @oratis/lisa
-lisa serve --web                # 两个 app 都从 http://localhost:5757 读
+lisa serve --web                # app 从 http://localhost:5757 读
 ```
 
 ### 📟 Homebrew（只装 CLI）
@@ -163,7 +161,7 @@ lisa
 OpenAI 模型 (`gpt-*`) 还需要 `OPENAI_API_KEY`。
 重新生成像素头像还需要 `SEEDREAM_API_KEY`（[火山引擎 ARK](https://www.volcengine.com/product/ark)）。
 
-**16+ 个其他 LLM** 开箱即用 —— Lisa 按模型名前缀（大小写不敏感）自动路由：
+**20+ 个其他 LLM** 开箱即用 —— Lisa 按模型名前缀（大小写不敏感）自动路由：
 
 - **国际**：Google Gemini · DeepSeek · Mistral · Perplexity Sonar · xAI Grok
 - **国内**：火山豆包 · 阿里 Qwen · 月之暗面 Kimi · 智谱 GLM · 阶跃 Step · 零一万物 Yi · 百川 Baichuan · MiniMax · 腾讯混元 Hunyuan
@@ -172,6 +170,8 @@ OpenAI 模型 (`gpt-*`) 还需要 `OPENAI_API_KEY`。
 
 完整配方见 [docs/PROVIDERS.md](docs/PROVIDERS.md)。
 
+**没有计量 key？用 coding plan。** 如果你已经在付 **Claude Pro/Max**、**ChatGPT 套餐**（含 Codex）或 **GitHub Copilot**，LISA 能把 coding 工作跑在这份订阅上 —— 不是去抽取你的 token（Anthropic 的条款禁止这么做，并已对 OpenClaw 实施了法律要求），而是**操纵厂商自己的 CLI**（它本来就持有那份授权）。见 [Coding plans](#coding-plans--用订阅代替-api-key) 和 [docs/CODING_PLANS.md](docs/CODING_PLANS.md)。
+
 ## 她特殊在哪
 
 | 大多数 LLM agent | LISA |
@@ -186,9 +186,29 @@ OpenAI 模型 (`gpt-*`) 还需要 `OPENAI_API_KEY`。
 ## 怎么用
 
 - **终端 REPL** — `lisa`（交互）或 `lisa "一句话"`（一次性）
-- **Web GUI** — `lisa serve --web` → http://localhost:5757 — 像素艺术聊天界面，头像跟着她的心情实时切
+- **Web GUI** — `lisa serve --web` → http://localhost:5757 — 像素艺术聊天界面，头像跟着她的心情实时切。默认**只绑 127.0.0.1**；要从手机访问，先设 `LISA_WEB_TOKEN` 并加 `--host 0.0.0.0`，然后每台设备第一次打开 `http://<主机>:5757/?token=<值>`。
 - **IM 通道** — `lisa serve --channels telegram,discord,slack,feishu,imessage,webhook` — 6 个内置 adapter，下面有详情
 - **心跳** — `lisa heartbeat run`（手动）或 `lisa heartbeat install`（macOS launchd / Linux cron）
+- **开机自启** — `lisa autostart install` 让 `lisa serve --web` 从登录起就常驻（macOS launchd；Linux 打印 `systemd --user` unit），这样 app、灵动岛、通道随时在线。`lisa autostart status` / `uninstall` 查看 / 移除。
+- **Mac 菜单栏** — Lisa.app 常驻菜单栏，带心情字形 + 实时 agent 状态、一个带 changelog + 更新发现的 About 窗口，以及 ⌘, 偏好面板（灵动岛开关、screen-advisor 间隔、backend 控制）。
+
+## 看着——并操纵——你其它的 agent（编排器）
+
+LISA 也是你机器上*其它* coding agent 的一个控制面。三层，对它们的介入逐层加深：
+
+**1. 观察。** 她看着已经在跑的 agent，把你会错过的事告诉你——哪个会话卡在同一个报错上、哪两个要在同一个仓库里打架、哪个早就跑完在干等。诚实地说明范围：**五个 observer（Claude Code、Codex、OpenCode、Aider、GitHub PR）都能产出结构化活动——工具、改动的文件、最近命令、错误——由每个集成的 `visibility` 档位门控；精细度取决于各 agent 在磁盘上记录了什么**（Claude Code 最丰富；Aider 的 markdown 日志只给文件 + 轮次、没有工具流；每个 adapter 都有隐私测试断言提示词/回复/文件内容绝不泄漏）。`lisa agents` 打印一次性快照，灵动岛实时显示。她可以 `dispatch_agent` 无头派发（拒绝把新 agent 丢进已被占用的目录）、`compare_agents` 在并行 worktree 里对比多个 agent 做同一任务，并给出**顾问卡片**——每条带一个一键动作（预填到聊天框，**绝不自动执行**）和一个 ✕（教她少唠叨这一类）。
+
+**2. 控制她自己的 agent。** **managed agent** 跑的是 LISA *自己*的 agent loop，在一个她完全驱动的子上下文里：派发任务、逐个审批/拒绝改写类工具、追加追问、取消——从 GUI 的 agents 卡片或 `POST /api/agents/managed/<id>/{send,approve,cancel}`。它们是她的，所以用的模型和 provider 也是她的。
+
+**3. 操纵真实 CLI（实验性，flag 门控）。** 开 `LISA_PTY_AGENTS=1` 后，**PTY agent** 在伪终端里拉起真正的 `claude` / `codex` 二进制——你拿到那个 CLI 的完整配置（它的 skills、MCP server、模型），同时 LISA 掌管 stdin/stdout：她把任务打进去，你可以回答它的提问，她读流得到粗粒度实时状态 + 可查看的输出尾巴（▤）。她甚至能**接管你自己开的、已空闲的 `claude` 会话**——`claude --resume <id>`（带 liveness 守卫，避免两个写入者把同一份 transcript 写花；*正在跑*的会话必须先关掉）。捕获的终端只按需给**你**看，绝不并入仅含元数据的 roster。见 [docs/PTY_AGENTS.md](docs/PTY_AGENTS.md)。
+
+### Coding plans — 用订阅代替 API key
+
+LISA 自己的"大脑"跑在计量 key 或本地模型上（见上）。但繁重的 *coding* 那部分——恰恰最吃 token——可以跑在**你已经在付的订阅**上：Claude Pro/Max、ChatGPT 套餐（Codex）或 GitHub Copilot。
+
+机制是刻意设计的：**LISA 不抽取、不重放你的订阅 token。** Anthropic 的条款把订阅 OAuth 限定为"Claude Code 的普通个人使用"，而且他们真的执行了——一纸法律要求让 **OpenClaw**（LISA 自己的 reference agent 之一）删掉了这个能力。LISA 改为**操纵厂商自己的 CLI**，由它持有那份订阅授权：上面的第 3 层（PTY agent / `claude --resume`）正是这条路——当你驱动一个用你的套餐登录过的 `claude`/`codex` 时，活儿就记在你的套餐上，而不是 API key 上。
+
+**现在就能用：** `lisa model list` 探测你已安装的套餐 CLI（Claude Code / Codex / Copilot）及登录状态,`lisa model use plan://claude` 选委托目标,**`run_on_plan`** 工具把一个 coding 任务跑在该套餐上 —— 无头驱动它的 CLI（`claude -p` / `codex exec` / `copilot -p`）,不读任何 token。`lisa model list` 还会显示**真实用量**（从本地 transcript 统计的滚动窗口 token,如 `1.2M tok in 5h`),**web 界面**也有一个 **PLANS** 选择器来选套餐、看状态/用量。为什么否决"进程内复用 token"—— 以及完整设计 —— 写在 **[docs/CODING_PLANS.md](docs/CODING_PLANS.md)**。
 
 ## 子命令
 
@@ -197,12 +217,25 @@ lisa                         交互式 REPL
 lisa "一句话"                 一次性
 lisa birth                   触发 birth ritual（首次启动会自动跑）
 lisa soul                    打印她当前灵魂状态
-lisa resume <id>             恢复某次会话
+lisa resume <id> [prompt]    按 id 恢复某次会话
 lisa sessions                列最近会话
 lisa search "<关键词>"       TF-IDF 全文搜过去所有对话
-lisa heartbeat run [task]    跑一次定时任务（含她自己的心愿）
+lisa status                  一次性快照：身份、心情、最近 commit
+lisa doctor                  健康检查（配置、网络、git、provider）
+lisa monitor                 TUI 实时面板（心情 + soul commit + 事件）
+lisa agents                  跨所有 observer 的 agent 会话快照
+lisa autonomy [days]         自驱运行摘要（空闲 / 心跳 / examen / desire / reflect）
+lisa model <list|install|use|health>
+                             本地模型（Ollama / LM Studio / llama.cpp）+
+                             coding-plan 探测/选择（`use plan://<id>`）
+lisa consent <list|grant|revoke|revoke-all> [signal]
+                             敏感环境信号的授权（屏幕 / 语音 / …；默认全关）
+lisa sense [list]            最近的环境感知事件 + 已授权信号
+lisa heartbeat run [name]    跑一次定时任务（含她自己的心愿）
 lisa heartbeat install       注册 macOS launchd 自动调度
 lisa heartbeat uninstall     卸载
+lisa autostart <install|uninstall|status>
+                             让 `serve --web` 从登录起常驻（launchd / systemd）
 lisa serve --web [--port N]  像素 Web UI（默认 5757）
 lisa serve --channels <list> 启 IM 通道（逗号分隔，或 "all"）
 lisa channels                列出可用通道
@@ -213,7 +246,7 @@ lisa wishlist                打印 Lisa 自己对工具集/架构的反馈
 lisa --help                  完整帮助
 ```
 
-常用 flag：`--model <id>` `--provider anthropic|openai` `--think` `--compact` `--approval auto|ask|ask-mutating` `--no-mcp` `--no-plugins` `--voice` `--no-reflect` `--idle <分钟>` `--no-idle`
+常用 flag：`--model <id>` `--provider anthropic|openai` `--think` `--compact` `--approval auto|ask|ask-mutating` `--no-mcp` `--no-plugins` `--voice` `--no-reflect` `--host <addr>`（`serve --web` 的绑定地址；非 loopback 需要 `LISA_WEB_TOKEN`） `--idle <分钟>` `--no-idle`
 
 ## 灵魂系统（Soul）
 
@@ -280,6 +313,8 @@ LISA 可以作为长驻进程同时监听多个 IM 平台。每个对话线程
 2. 把 secrets 写进 `~/.lisa/config.env`（在 `channels.json` 里用 `${VAR}` 占位）
 3. `lisa serve --channels all`（或指定具体通道）
 
+**安全默认值。** 通道消息来自"任何能联系到 bot 的人"，所以通道默认跑**远程安全工具集**：没有 `bash`、没有文件改写、没有 `dispatch_agent` / GitHub 写操作 / `skill_manage`。对话、记忆、灵魂工具、网页阅读都正常——足够覆盖"在手机上跟她聊"的场景。完全信任的通道可以在配置里加 `"unsafeFullTools": true` 拿回全部工具。务必配置白名单（`allowedUsernames` / `allowedChatIds` / `allowedUserIds`）——启动时 router 会对任何敞开的通道大声告警。飞书现在**必须**配 `verificationToken`（或 `encryptKey`），并校验 `X-Lark-Signature` + 5 分钟重放窗口，与 Slack 适配器同等姿态。
+
 ### 内置通道
 
 | 通道 | 状态 | 凭据 | 备注 |
@@ -342,6 +377,8 @@ LISA 可以在你不在的时候跑后台任务，跟她自己的心愿独处。
    ```
 2. **她自驱的** — 她自己写的 `~/.lisa/soul/desires/` 里标了 actionable 的心愿。她自己加的，她自己推进
 
+自驱运行（desires、每周 examen、空闲/梦境）是无人值守 + 提示词由 Lisa 自己写的，所以默认拿**受限工具集**——灵魂/记忆/日记/技能/网页阅读可用，但没有 shell、文件改写、agent 派发。你自己写在 `heartbeat.json` 里的任务保留全部工具（提示词是你写的）。接受风险的话，`LISA_AUTONOMOUS_FULL_TOOLS=1` 恢复旧行为。
+
 macOS 上安装：
 ```sh
 lisa heartbeat install --every 30m --load
@@ -372,6 +409,13 @@ CLI flag: `--idle 60`（分钟，默认 60）/ `--no-idle` 禁用。
 | `bash` | shell（可选 macOS Seatbelt 沙箱：`LISA_SANDBOX=1`） |
 | `grep` `ls` | 搜 + 列 |
 | `task` | 派生子 agent 跑独立 context window 的任务 |
+| `dispatch_agent` `signal_agent` `dispatch_status` | 启动 / 停止 / 追踪她跑的 agent（managed + PTY）；拒绝别的 agent 占着的目录 |
+| `run_on_plan` | 把 coding 任务委托给你的订阅 **coding plan**（Claude Pro/Max · ChatGPT/Codex · Copilot），驱动它的 CLI —— 记在套餐上而非 API key |
+| `list_agents` `inspect_agent` `compare_agents` `agent_recap` | 观察其它 coding agent；深挖某个会话；在 worktree 里让多个 agent 赛同一任务；"你不在的时候"回顾 |
+| `advise_now` `scheduled_dispatch` | 按需弹顾问卡片；通过心跳跑周期性派发任务 |
+| `pr_status` `run_checks` `review_diff` `repo_digest` `github` `github_link` `npm_info` | 仓库 / GitHub / CI 辅助（读安全，写显式） |
+| `web_search` `web_fetch` | 读网页 |
+| `mcp` | 管理 MCP server 连接（列出 / 添加 / 删除） |
 | `skill_manage` | `~/.lisa/skills/` 增删改查 |
 | `memory` `memory_search` | 记忆 CRUD + 跨会话 TF-IDF 搜 |
 | `set_mood` | 切换 114 张头像里的某一张 |
@@ -409,6 +453,8 @@ LISA 是吃完五个开源 agent（fork 在 `reference/`）合成出来的：
 | 语音输入输出 | – | ✅ | – | – | – | ✅ |
 | 心跳 | – | ✅ | – | – | – | ✅（自带 launchd 安装器） |
 | 多通道 IM | ✅ pi-mom | ✅ 20+ | ✅ | – | – | ✅ Telegram + Discord + Slack + Feishu + Webhook + iMessage |
+| **编排其它 agent（观察 + 操纵它们的 CLI）** | – | – | – | – | – | **✅ ★ LISA 独有** |
+| **Coding-plan 委托（订阅，而非只有 API key）** | – | – | – | – | – | **✅ ★ LISA 独有** |
 | **持久身份 / 灵魂** | – | – | partial | – | – | **✅ ★ LISA 独有** |
 | **Birth ritual（独特种子）** | – | – | – | – | – | **✅ ★ LISA 独有** |
 | **私人日记** | – | – | – | – | – | **✅ ★ LISA 独有** |
@@ -425,11 +471,40 @@ LISA 是吃完五个开源 agent（fork 在 `reference/`）合成出来的：
 ANTHROPIC_API_KEY=sk-ant-...
 OPENAI_API_KEY=sk-...                 # 可选 — 用 gpt-* 模型时
 SEEDREAM_API_KEY=...                  # 可选 — 重生成头像用
+
+# Provider / 模型路由
+LISA_PROVIDER=openai                  # 强制 provider
+LISA_MODEL=claude-sonnet-4-6          # 默认模型（也由 `lisa model use` 设置）
+LISA_MODEL_FALLBACK=gpt-4o,...        # 主模型失败时按序尝试的模型列表
+LISA_BASE_URL=http://localhost:11434/v1   # OpenAI 兼容端点（Ollama、vLLM、网关）
+LISA_API_KEY=...                      # LISA_BASE_URL 用的 key（回退到 OPENAI_API_KEY）
+ANTHROPIC_AUTH_TOKEN=...              # Anthropic 兼容网关/代理的 Bearer token
+                                      # （区别于 ANTHROPIC_API_KEY 的 x-api-key）；
+                                      # 不是订阅 token —— 见 docs/CODING_PLANS.md
+
+# Coding-plan 委托（操纵真实 CLI —— 见 docs/CODING_PLANS.md）
+LISA_PTY_AGENTS=1                     # 启用操纵真实 claude/codex CLI（实验性）
+LISA_PTY_CLAUDE_CMD=claude            # 覆盖 `claude` 二进制路径
+LISA_PTY_CODEX_CMD=codex             # 覆盖 `codex` 二进制路径
+
+# 沙箱
 LISA_SANDBOX=1                        # 可选 — bash 走 macOS Seatbelt
 LISA_SANDBOX_NETWORK=0                # 沙箱内禁网
-LISA_PROVIDER=openai                  # 强制 provider
+
+# Web
+LISA_WEB_TOKEN=...                    # serve --web 绑定到 127.0.0.1 之外
+                                      # （--host 0.0.0.0）时必须设置；远程设备
+                                      # 首次用 ?token= 认证
+
+# 自主（心跳 / 空闲 / 梦境）
+LISA_AUTONOMOUS_FULL_TOOLS=1          # 选择退出：让自驱心跳/空闲运行重新拿
+                                      # 全部工具（含 bash）
+LISA_IDLE_BUDGET_TOKENS=200000        # 单次空闲运行的 token 上限（0 = 禁用空闲）
+LISA_IDLE_COMMITMENT_AWARE=1          # 选择加入：空闲时先看你即将到来的约定，再做个人反思
 ```
 
+环境信号（屏幕 / 语音 / 剪贴板 / 选区）**默认全关**，由 `lisa consent grant <signal>` 门控 —— 在你授权前绝不捕获。见 `lisa consent list`。
+
 ### `~/.lisa/mcp.json`
 
 ```json
@@ -504,9 +579,9 @@ src/
 │   ├── store.ts            CRUD + 篡改检测
 │   ├── tools.ts            soul_patch / soul_journal / soul_feel / soul_read
 │   ├── paths.ts types.ts
-├── idle/                   ★ 空闲模式（IdleWatcher 单例 + autonomous-time runner）
-├── providers/              Anthropic + OpenAI + Gemini 抽象
-├── tools/                  read/write/edit/apply_patch/bash/grep/ls/task/set_mood + registry
+├── providers/              Anthropic + OpenAI + Gemini 抽象 + 模型名路由（registry.ts）
+├── model/                  本地模型生命周期（Ollama / LM Studio / llama.cpp），供 `lisa model`
+├── tools/                  文件/bash/grep/task/set_mood + 编排 + 仓库/github + 网页 + registry
 ├── skills/                 manager + frontmatter + skill_manage
 ├── memory/                 store + memory tool + TF-IDF 索引 + memory_search
 ├── sessions/               JSONL store + list + resume + 分页读
@@ -515,6 +590,17 @@ src/
 ├── plugins/                claude-code 风格插件加载器
 ├── hooks/                  PreToolUse / PostToolUse / SessionStart / 等
 ├── heartbeat/              定时任务 + launchd 安装器
+├── autostart/              开机自启安装器（launchd / systemd），供 `lisa autostart`
+├── idle/                   空闲自主反思（梦境 Reve）
+├── autonomy/               run ledger —— 自驱运行的可观察日志（`lisa autonomy`）
+├── agents/                 managed agent（LISA 自己的 loop）+ PTY agent（操纵真实 claude/codex）
+├── integrations/           observer：claude-code · codex · opencode · aider · github-pr · pty · managed · …
+├── orchestrator/           跨 agent 日志 + "你不在的时候"回顾合成
+├── advisor/                主动顾问卡片（卡住 / 冲突 / 就绪 / 空闲）+ 关闭学习
+├── consent/                环境信号的统一授权门控（默认全关）
+├── sense/                  环境信号源（前台 app / 窗口标题），授权门控
+├── vision/                 用户发起的截图捕获（macOS）
+├── screen_advisor/         可选的、周期性从截图给出"下一步 coding"建议
 ├── voice/                  speak（macOS say）+ transcribe（Whisper）
 ├── channels/               channel 抽象 + 6 个 adapter（含飞书）+ router
 └── web/                    像素艺术 HTTP + SSE web UI

package/completions/_lisa

@@ -41,6 +41,10 @@ _lisa() {
             case "$line[1]" in
                 skills)   _lisa_skills ;;
                 heartbeat) _lisa_heartbeat ;;
+                autostart) _lisa_autostart ;;
+                model)    _lisa_model ;;
+                consent)  _lisa_consent ;;
+                sense)    _lisa_sense ;;
             esac
             ;;
     esac
@@ -53,6 +57,7 @@ _lisa_first() {
         'sessions:list recent sessions'
         'serve:start web UI / IM channels'
         'heartbeat:run / install / uninstall scheduled tasks'
+        'autostart:run the backend from login (install/uninstall/status)'
         'search:TF-IDF search across past sessions'
         'birth:run the birth ritual'
         'soul:print current soul summary'
@@ -62,6 +67,11 @@ _lisa_first() {
         'status:one-shot snapshot'
         'doctor:health check'
         'monitor:TUI live dashboard'
+        'autonomy:summarize recent autonomous runs'
+        'model:local model lifecycle (list/install/use/health)'
+        'consent:consent for sensitive ambient signals'
+        'sense:recent ambient sense events'
+        'agents:snapshot of agent sessions across observers'
     )
     _describe 'subcommand' subcommands
 }
@@ -99,6 +109,52 @@ _lisa_heartbeat() {
     fi
 }
 
+_lisa_autostart() {
+    if (( CURRENT == 2 )); then
+        local -a actions
+        actions=(
+            'install:keep serve --web running at login + on crash'
+            'uninstall:stop starting Lisa at login'
+            'status:show whether autostart is installed / loaded'
+        )
+        _describe 'autostart action' actions
+    fi
+}
+
+_lisa_model() {
+    if (( CURRENT == 2 )); then
+        local -a actions
+        actions=(
+            'list:list installed local models + backend health'
+            'install:pull a local model'
+            'use:switch to local://<model>'
+            'health:check the local backend'
+        )
+        _describe 'model action' actions
+    fi
+}
+
+_lisa_consent() {
+    if (( CURRENT == 2 )); then
+        local -a actions
+        actions=(
+            'list:show signals + status'
+            'grant:grant a signal'
+            'revoke:revoke a signal'
+            'revoke-all:stop all sensing'
+        )
+        _describe 'consent action' actions
+    elif (( CURRENT == 3 )) && [[ "$line[2]" == (grant|revoke) ]]; then
+        _describe 'signal' '(screen voice clipboard selection)'
+    fi
+}
+
+_lisa_sense() {
+    if (( CURRENT == 2 )); then
+        _describe 'sense action' '(list:show recent ambient events + granted signals)'
+    fi
+}
+
 _lisa_models() {
     local -a models
     models=(

package/completions/lisa.bash

@@ -12,11 +12,16 @@ _lisa_completion() {
     local cur prev words cword
     _init_completion -n : 2>/dev/null || _get_comp_words_by_ref -n : cur prev words cword
 
-    local subcommands="resume sessions serve heartbeat search birth soul channels skills wishlist status doctor monitor"
+    local subcommands="resume sessions serve heartbeat autostart search birth soul channels skills wishlist status doctor monitor autonomy model consent sense agents"
     local global_flags="--model --provider --think --no-reflect --compact --approval --no-mcp --no-plugins --voice --idle --no-idle --help -h --version -v"
     local serve_flags="--web --imessage --channels --port"
     local skills_actions="list approve disable enable audit"
     local heartbeat_actions="run install uninstall"
+    local autostart_actions="install uninstall status"
+    local model_actions="list install use health"
+    local consent_actions="list grant revoke revoke-all"
+    local consent_signals="screen voice clipboard selection"
+    local sense_actions="list"
     local approval_modes="auto ask ask-mutating"
     local providers="anthropic openai gemini"
     local models="claude-sonnet-4-5-20250929 claude-opus-4 claude-haiku-4 gpt-5 gpt-4o o3 o4-mini gemini-2.5-pro gemini-2.5-flash deepseek-chat deepseek-reasoner doubao-1.5-pro-32k qwen3-72b-instruct moonshot-v1-32k kimi-128k grok-2-latest glm-4.5"
@@ -79,6 +84,35 @@ _lisa_completion() {
                 return 0
             fi
             ;;
+        autostart)
+            if [ ${cword} -eq 2 ]; then
+                COMPREPLY=( $(compgen -W "${autostart_actions}" -- "${cur}") )
+                return 0
+            fi
+            ;;
+        model)
+            if [ ${cword} -eq 2 ]; then
+                COMPREPLY=( $(compgen -W "${model_actions}" -- "${cur}") )
+                return 0
+            fi
+            ;;
+        consent)
+            if [ ${cword} -eq 2 ]; then
+                COMPREPLY=( $(compgen -W "${consent_actions}" -- "${cur}") )
+                return 0
+            fi
+            # `grant`/`revoke` take a signal
+            if [ "${words[2]}" = "grant" ] || [ "${words[2]}" = "revoke" ]; then
+                COMPREPLY=( $(compgen -W "${consent_signals}" -- "${cur}") )
+                return 0
+            fi
+            ;;
+        sense)
+            if [ ${cword} -eq 2 ]; then
+                COMPREPLY=( $(compgen -W "${sense_actions}" -- "${cur}") )
+                return 0
+            fi
+            ;;
         serve)
             COMPREPLY=( $(compgen -W "${serve_flags} ${global_flags}" -- "${cur}") )
             return 0