npm - @viceme-ai/cli - Versions diffs - 0.10.1 → 0.11.1 - Mend

@viceme-ai/cli 0.10.1 → 0.11.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (6) hide show

package/CHANGELOG.md CHANGED Viewed

@@ -1,5 +1,35 @@
 # Changelog
+## [0.11.1] - 2026-08-12
+### Fixes
+- 修复 S3 发布产物被 checkout 清理 (`4fcc6fe`)
+## [0.11.0] - 2026-08-11
+### Features
+- 展示并确认双语 Skill 短简介 (`f063212`)
+- rebuild skill publication workflow (`64f5fd2`)
+### Fixes
+- 修复发版准备后的版本测试 (`a84d1c6`)
+- 兼容单层目录包装的 Skill ZIP (`1e63183`)
+- keep install independent of credential status (`b74f06b`)
+- arbitrate outer activation journals (`01ab265`)
+- fence installation generation commits (`5d4a3ad`)
+- unify activation recovery state (`cf882a1`)
+- fence recoverable activation generations (`7a47561`)
+- make installation recovery atomic (`755ae8e`)
+- make publication and install recovery durable (`308deab`)
+### Other Changes
+- document unified activation recovery (`cfeea2c`)
+- record first-phase completion (`46ab65e`)
 ## [0.10.1] - 2026-07-31
 ### Fixes

package/README.md CHANGED Viewed

@@ -1,376 +1,221 @@
 # ViceMe CLI
-[![npm version](https://img.shields.io/npm/v/@viceme-ai/cli.svg)](https://www.npmjs.com/package/@viceme-ai/cli)
-[![Go Version](https://img.shields.io/badge/go-%3E%3D1.23-blue.svg)](https://go.dev/)
-[![CLI PR checks](https://github.com/ViceMe-AI/cli/actions/workflows/ci.yml/badge.svg)](https://github.com/ViceMe-AI/cli/actions/workflows/ci.yml)
+ViceMe CLI is the deterministic local companion for the official ViceMe Agent
+Skills. Codex, Claude Code, and WorkBuddy use the Skills to guide the user; the
+CLI handles installation, device authorization, validation, packaging,
+uploads, review, and publication.
-[中文版](./README.zh.md) | [English](./README.md)
+[中文](./README.zh.md)
-The official command-line client and Agent Skill for publishing external Skills as stable, shareable ViceMe Agents. It is built for AI coding tools such as Codex and Claude Code: the Agent Skill understands user intent, while the CLI provides the deterministic authentication, upload, publication, and status protocol.
+## Install
-[Install](#installation--quick-start) · [AI Agent Skills](#agent-skills) · [Auth](#authentication) · [Regions & profiles](#regions--profiles) · [Commands](#command-overview) · [Output contract](#json-output-contract) · [Security](#security-and-risk-controls) · [Development](#development)
+The official bootstrap installs the native CLI and both official Skills from
+one immutable release.
-> **Rollout status:** the Core publication transport and stable-link path are implemented. After `--yes`, a publication first parks at `meta_review` for metadata confirmation, then at `awaiting_action` for interaction-step confirmation. `confirm_publish` returns an owner-only `preview_share_url` under `/p/{code}`. The creator completes one ordinary successful run there, returns to the Agent Host, and explicitly confirms. After release, the durable receipt returns the official `/v/{code}` `share_url`; the old preview URL redirects to it. The CLI no longer exposes a separate PreviewRun surface. Preview inputs, outputs, files, sessions, and Runner history are temporary, are purged after resolution, and never enter public usage or work history.
-## Why ViceMe CLI?
-- **Agent-native** — ships together with the official ViceMe Agent Skill for Codex and Claude Code.
-- **Deterministic boundary** — the CLI performs typed protocol actions; it does not start another conversational Agent loop.
-- **Server-side compilation** — source parsing, LLM compilation, BuildRun materialization, and Release publication stay on ViceMe infrastructure.
-- **Stable publishing** — later releases of the same logical Agent keep the same share URL.
-- **Multiple source types** — accepts typed GitHub and Xiaohongshu/RedSkill sources, archives, and local Skill folders.
-- **Secure by default** — on macOS, device-login credentials use AES-256-GCM encrypted files with Keychain-backed key material; other platforms retain their native credential manager. Explicit local overrides require a private profile file, public mutations require confirmation, and downloaded binaries are checksum-verified.
-- **Human and Agent login modes** — `viceme auth login` guides a person in the terminal, while Agent split-flows use explicit JSON.
-## Installation & Quick Start
-### Requirements
-- Node.js 18.20 or newer
-- macOS or Linux on `amd64`/`arm64`, or Windows on `amd64`/`arm64`
-### Quick Start (Human Users)
-> **Note for AI assistants:** If you are an AI Agent helping the user install ViceMe, jump directly to [Quick Start (AI Agent)](#quick-start-ai-agent). It defines the required cross-turn device login flow.
-#### Install
-Choose one of the following methods.
-**Option 1 — One-shot npm install (recommended):**
-```bash
-npx --yes @viceme-ai/cli@latest install
-```
-**Option 2 — Install the npm launcher globally:**
+China, macOS or Linux:
 ```bash
-npm install --global @viceme-ai/cli
-viceme install
+curl -fsSL https://s3.viceme.cn/install.sh | sh
 ```
-Both methods persist the npm launcher, download the matching checksum-verified Go binary, and install the bundled ViceMe Agent Skill. Binary downloads try GitHub Release first, then a configured non-default npm registry's `/-/binary/viceme-cli/` mirror, and finally the public npmmirror binary mirror. The launcher uses system `curl`, so standard proxy environment variables are honored. They default to the China service. For the international service:
+International, macOS or Linux:
 ```bash
-npx --yes @viceme-ai/cli@latest install --region global
+VICEME_REGION=global sh -c "$(curl -fsSL https://s3.viceme.ai/install.sh)"
 ```
-#### Authenticate and verify
+China, Windows PowerShell:
-If the installation result says authentication is required, start the guided device login:
-```bash
-viceme auth login
+```powershell
+irm https://s3.viceme.cn/install.ps1 | iex
 ```
-The CLI prints the browser URL, waits for authorization, and reports success in the same terminal. Then verify the installation:
+International, Windows PowerShell:
-```bash
-viceme auth status
-viceme skills doctor
+```powershell
+$env:VICEME_REGION="global"; irm https://s3.viceme.ai/install.ps1 | iex
 ```
-## Quick Start (AI Agent)
-> The following steps are for AI Agents. Browser authorization must be completed by the user in a separate turn; never wait indefinitely or ask the user for a token.
-**Step 1 — Install**
-Use the complete bootstrap command from the bundled Skill. The explicit npm registries are part of the installation trust boundary:
+If the bootstrap cannot be used, npm is the fallback:
 ```bash
-npx --yes --registry=https://registry.npmjs.org --@viceme-ai:registry=https://registry.npmjs.org --package=@viceme-ai/cli@latest -- viceme install
-```
-Read `authenticated` and `next_step` from the formatted business result. If authentication is already valid, continue to Step 4. If login is required, do not execute the human-oriented `next_step` inside the Agent; use the JSON split-flow in Step 2.
-**Step 2 — Start device login when required**
-```bash
-viceme auth login --no-wait --json
+npx --yes @viceme-ai/cli@latest install
 ```
-Return the exact `verification_url`; the CLI normalizes it to the prefilled `verification_url_complete` browser link when available. Include `user_code` only as a fallback if the browser asks for it. Preserve `device_code` for the continuation command, then stop the current turn. Do not request, print, or place an access token in the conversation.
-**Step 3 — Continue the same login in a later turn**
+The installer always writes the compatibility fallback to
+`~/.agents/skills`. It also installs into the native user directory of every
+detected supported Agent:
-After the user confirms browser authorization:
+| Agent | Native directory |
+| --- | --- |
+| Codex | `~/.codex/skills` |
+| Claude Code | `~/.claude/skills` |
+| WorkBuddy | `~/.workbuddy/skills` |
-```bash
-viceme auth login --device-code <device-code> --json
-```
+Select a target explicitly with `viceme install --agent codex`, `claude`,
+`workbuddy`, or `agents`. Run `viceme doctor` after installation or repair.
-If authorization is still pending, reuse the same device code before it expires. Do not start a second device flow unless the original one has expired.
-For a non-default profile, pass the same global `--profile <name>` when starting and continuing device login; the start result reports the effective `profile` and `region`.
+## Authentication and profiles
-**Step 4 — Verify**
+Each profile is bound to one official region and one device-authorized account.
 ```bash
+viceme auth login
 viceme auth status
-viceme skills doctor
-viceme skills list
-```
-Continue only when authentication is valid and `skills doctor` reports a healthy, compatible installation.
-**Step 5 — Inspect the first source**
+viceme auth logout
-```bash
-viceme skill inspect https://github.com/acme/poster-skill --skill-root .
-```
-Inspection is read-only. Follow the bundled `viceme` Skill for source-specific handling, Target selection, confirmation, bounded job waiting, and result reporting. If `destination.recovery.mode=resume_existing_publication`, run `job get` with its `publication_id` and resume that nonterminal Publication instead of publishing the new inspection resolution. If a publication ends at `binding_required`, run `viceme job bind <publication-id>`, give the signed ViceMe URL to the user, and stop. Downloading or forking is only an informational alternative; the CLI never performs it automatically. After the user binds the exact GitHub/Xiaohongshu channel account, inspect again and create a fresh ordinary Publication rather than resuming the terminal one. At `meta_review`, show and resolve the metadata using the exact action ID and payload digest, then wait again. At `awaiting_action`, confirm the interaction steps, open the private `preview_share_url`, and complete one successful ordinary use on the preview page. If a `confirm_steps` or `confirm_publish` action expires, read the same Publication with `job get`, then explicitly run `viceme job renew <publication-id> --action-id <expired-action-id>` and continue with the returned `next_action`; do not create a second Publication. Return to the conversation for an explicit confirm or cancel decision. After confirm, wait until `share_published` and return the official `data.result.share_url`; do not require it to equal the private preview URL.
-## Regions & Profiles
-Each profile selects one ViceMe region:
-| Region | Install command | API endpoint |
-|---|---|---|
-| China | `viceme install` | `https://api.viceme.cn` |
-| International | `viceme install --region global` | `https://api.viceme.ai` |
-The first install creates the `default` profile. Device-login credentials are isolated by profile plus normalized API origin. On macOS they are stored as private AES-256-GCM encrypted files, while the encryption master key normally remains in the operating-system Keychain; other platforms retain their native credential manager. A profile may also contain an explicitly configured publication credential for controlled local/internal operation; any such config remains private `0600`.
-```bash
 viceme profile list
-viceme profile add --name work --region global --use
 viceme profile use default
-viceme --profile work auth status
-viceme profile rename work company
-viceme profile remove company
 ```
-`profile use` changes the persistent active profile; the global `--profile` flag overrides only one command. AI Agents must not switch or remove profiles unless the user explicitly requests it.
-For controlled local/internal operation, create a dedicated profile with both an explicit endpoint and its audience-bound publication credential:
+For an Agent workflow that cannot wait in one turn:
 ```bash
-viceme profile add --name local --region cn \
-  --api-base-url http://localhost:8090 \
-  --access-token '<vpa1.local-dev.credential>' --use
-viceme profile configure local --access-token 'YOUR_ACCESS_TOKEN'
-viceme profile configure local --clear-access-token
-viceme profile configure local --clear-api-base-url
+viceme auth login --no-wait
+viceme auth login --device-code <device-code>
 ```
-Normal `viceme auth login` never writes a credential into a profile. A local Profile credential is set, replaced, or cleared only by the explicit `profile add/configure` flags, is reported only as `source=local_profile`, and is never returned by list/status output. Because `--access-token` is visible in argv and may enter shell history, use it only in the controlled local/internal environment described here.
-Credential priority is process (`VICEME_ACCESS_TOKEN`) → selected local Profile → device login. Every publication credential must use `vpa1.<audience>.<secret>`: `cn-prod` is accepted only for `https://api.viceme.cn`, `global-prod` only for `https://api.viceme.ai`, `dev-preview` only for `https://viceme-envoy-dev.preview.tencent-zeabur.cn`, and `local-dev` Profile credentials only for loopback endpoints. A process `local-dev` credential additionally requires `VICEME_CLI_ALLOW_LOCAL_PROCESS_CREDENTIAL=1`. Configuration defaults to `~/.viceme-cli` on macOS and Linux and `%LOCALAPPDATA%\ViceMe\Config` on Windows; an existing Windows `~/.viceme-cli/config.json` remains in use until it is explicitly migrated. `VICEME_CLI_CONFIG_DIR` can override the config root. `VICEME_API_BASE_URL` remains a one-process endpoint override and never widens a Profile credential's origin. It cannot be combined with global `--profile`; use the selected Profile's configured endpoint or the process endpoint override as one authority source, never both. API and presigned-upload redirects fail closed.
+The user completes authorization in the browser. Never copy an access token
+into the conversation. Credentials are isolated by profile and API origin.
-Update checks query the npm registry directly and store only the last successful version result as `update-state.json` in the configuration directory. A result is used as a fallback for at most 24 hours when the registry is temporarily unavailable. Normal npm-managed CLI invocations read this cache synchronously and refresh it in the background at most once per 24 hours, so commands never wait for update discovery. When a newer release is known, structured success and error objects include `_notice.update` with `current`, `latest`, `message`, and the exact `viceme update` command so AI Agents can notify the user. The advisory never changes the command exit status and does not trigger an automatic update. Set `VICEME_NO_UPDATE_NOTIFIER=1` to suppress it outside CI; standard CI environments are skipped automatically. npm operations launched by `viceme install` or `viceme update` use the isolated `npm-cache` subdirectory there, so a broken user-level npm cache does not block the CLI. Both locations are non-secret and can be deleted safely; credentials never enter either cache.
+## Publish a Skill
-## Agent Skills
+The first release accepts a local directory containing `SKILL.md` or a local
+ZIP. GitHub URLs, remote downloads, and multi-Skill bundles are not accepted.
-The current release deliberately ships one platform-level Agent Skill:
-| Skill | Description | Supported hosts |
-|---|---|---|
-| `viceme` | Install, inspect, convert, publish, update, or share external Skills as stable ViceMe Agents; enforces authentication, source, Target, confirmation, job, and safety rules | Codex, Claude Code |
-GitHub, Xiaohongshu/RedSkill, ZIP, and folder inputs are source types handled by the same `viceme` publication workflow, not separate Agent Skills. This keeps one consistent safety and stable-link contract across providers.
-The CLI and `viceme` Agent Skill are released from this repository at the same version. `viceme install` installs the complete Skill bundle into detected supported hosts, while the binary embeds the agent-readable subset needed for deterministic self-inspection.
+Inspect without side effects:
 ```bash
-viceme skills list
-viceme skills read viceme
-viceme skills read viceme references/commands.md
-viceme skills doctor
+viceme skill inspect --path ./my-skill
 ```
-`skills doctor` checks the CLI version, Skill version, compatibility range, full bundle digest, and embedded-content digest independently. A modified or incompatible installation fails closed.
-## Authentication
-| Command | Purpose |
-|---|---|
-| `viceme auth status` | Show whether the current profile is authenticated |
-| `viceme auth login` | Guide a human through browser authorization and wait for completion |
-| `viceme auth login --no-wait --json` | Start an Agent split-flow and return structured device authorization |
-| `viceme auth login --device-code <code> --json` | Complete an Agent split-flow in a later turn |
-| `viceme auth logout` | Revoke and remove the current profile credential |
-On macOS, tokens created by device login are stored only in private encrypted credential files and their master key normally remains in the operating-system Keychain; other platforms retain their native credential manager. Normal login never backfills explicit local profile fields, and successful login output never contains the access or refresh token.
-Before starting or exchanging a device authorization, the CLI verifies the full local persistence path. If that preflight fails, no one-time authorization is consumed. If storage fails after a successful exchange despite the preflight, the CLI attempts to revoke the issued credential and returns `credential_persistence_failed` with an explicit instruction to start a new device flow; it never reports login success or outputs a token.
-### macOS sandboxes (Codex and Claude Code)
-When an explicit device login runs inside a sandbox that cannot access the macOS Keychain, the CLI automatically creates a private `0600` file master key and saves the newly authorized credential there. No manual preparation is required.
-To reuse credentials previously created from Terminal with a Keychain-backed master key without logging in again, run this once from that same interactive macOS user session:
+Show the exact deterministic package and price plan:
 ```bash
-viceme config keychain-downgrade
+viceme skill publish --path ./my-skill --price-minor 100 --dry-run
 ```
-The command copies the existing master key into `~/.viceme-cli/credentials/master.key.file` and imports configured legacy Keychain credentials into encrypted files. Existing Keychain entries are preserved as a cold backup. The command is idempotent and never prints or stores a plaintext token. Afterward, Codex and Claude Code sandboxes for the same macOS user can read the encrypted credential files without Keychain access. This migration is optional when the user is willing to log in again. The trade-off is explicit: security is then enforced by the user's filesystem permissions (`0700` directory and `0600` files) instead of the Keychain per-process access boundary.
-The public CLI exposes one standard authentication and publication surface. A short-lived staff authorization credential may be supplied through process environment (`source=process`) or an explicitly configured local Profile (`source=local_profile`). Both use normal inspect/publish/job commands and the standard `x-api-key` header; there are no identity-selection or staff-authorization issuance commands. Tokens are never printed or inherited by update subprocesses, and login/logout fail closed while either override is active.
-## Supported Sources
-### GitHub or trusted provider
+Start the resumable upload and listing analysis:
 ```bash
-viceme skill inspect https://github.com/acme/poster-skill --skill-root .
-viceme skill publish --resolution-id <resolution-id> --yes
+viceme skill publish --path ./my-skill --price-minor 100
 ```
-For GitHub, `--skill-root` is required and names the exact repository-relative directory containing `SKILL.md`; use `.` only for a root-level Skill. The calling Agent determines this path from the user input or read-only repository tree. ViceMe does not scan the repository to guess a Skill.
-### Xiaohongshu or RedSkill
+Then follow the authoritative publication state:
 ```bash
-viceme skill inspect --source-stdin
-viceme skill publish --resolution-id <resolution-id> --yes
+viceme publication get <publication-id>
+viceme publication review <publication-id>
+viceme publication asset upload <publication-id> --role cover --path ./cover.png
+viceme publication asset upload <publication-id> --role gallery --path ./demo.png
+viceme publication update <publication-id> --input ./listing-draft.json
+viceme publication confirm <publication-id> --review-digest <digest>
+viceme publication publish <publication-id> --review-digest <digest>
 ```
-The AI Host interprets the user's source intent and passes one typed JSON
-`SourceSpec` through stdin, for example
-`{"kind":"redskill","value":"ai-desk-card"}`. CLI/Core do not classify copied
-natural language with keyword or regex rules. Explicit platform intent is
-authoritative; ambiguous source requests must be clarified instead of silently
-substituting a same-name source from another provider.
-### Archive or local Skill folder
-```bash
-viceme skill publish --file ./poster-skill.zip --new-target --target-alias poster --yes
-viceme skill publish --dir ./poster-skill --new-target --target-alias poster --yes
-```
+The model proposes `summaryZhCn`, `summaryEnUs`, and package images, but it never
+confirms them or decides the price. Each summary has a maximum display width of
+30: ASCII counts as 1 and Chinese/non-ASCII counts as 2. The Agent must show the
+exact bilingual summaries, price, cover, and ordered gallery to the user and
+receive explicit confirmation before `confirm` and again before public
+`publish`.
-For later releases, resolve the existing Target and use optimistic concurrency. Never turn a conflict into a new share link:
+Resume after a connection loss with the same publication:
 ```bash
-viceme skill target get target_123
-viceme skill publish --file ./poster-skill-v2.zip \
-  --target-id target_123 --expected-target-version 4 --yes
+viceme skill publish --resume <publication-id>
 ```
-## Command Overview
-| Command group | Purpose |
-|---|---|
-| `viceme install` | Install the persistent launcher, Agent Skill, and default profile |
-| `viceme auth` | Start, complete, inspect, or revoke device authentication |
-| `viceme config` | Manage the controlled macOS Keychain-to-file sandbox fallback |
-| `viceme profile` | Add, list, switch, rename, or remove local profiles |
-| `viceme skill inspect` | Freeze and inspect a source candidate without publishing |
-| `viceme skill publish` | Create or update a stable Skill Agent publication |
-| `viceme skill target` | Resolve existing logical Agent Targets and versions |
-| `viceme job` | Read or wait for a publication, review metadata and the frozen summary, edit its Candidate, show stable preview/public and signed channel-binding URLs, resume an action, explicitly retry, or cancel |
-| `viceme skills` | Read, install, and diagnose the bundled Agent Skill |
-| `viceme update` | Update the npm launcher, verified binary, and bundled Skill together |
-Use `viceme <command> --help` for the exact flags. The release-checked machine-readable surface is stored in [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json).
-## Output Contract
-ViceMe selects the smallest stable representation for each command:
-- Local/bootstrap commands such as `version`, `install`, `update`, `auth status`, `profile *`, and `skills doctor` write their formatted business result directly to **stdout**. They do not add `ok`, `data`, or unrelated build metadata; a normal npm-managed invocation may add only the reserved `_notice.update` advisory described above.
-- `skills read` writes the requested file byte-for-byte without a JSON wrapper.
-- Interactive `viceme auth login` writes human guidance. AI Agents use `--no-wait --json`, then continue with `--device-code <code> --json`; those two commands return a formatted bare business object.
-- Publication protocol commands under `skill` and `job` keep a stable envelope because action receipts, durable status, and bounded-wait metadata form one cross-command protocol.
+Do not create a second publication when the server response is unknown. Query
+or resume the existing ID first.
-A successful publication protocol result is written to **stdout** with exit code `0`:
+## Output contract
-```json
-{
-  "ok": true,
-  "data": {}
-}
-```
-Only a bounded wait that actually times out adds protocol metadata:
+Business output is JSON by default. Successful output is the only content on
+stdout; progress and diagnostics use stderr.
 ```json
 {
   "ok": true,
   "data": {},
   "meta": {
-    "wait_timed_out": true
+    "cliVersion": "0.10.1",
+    "requestId": "optional"
   }
 }
 ```
-CLI execution errors are formatted and written to **stderr** with a non-zero exit code:
-```json
-{
-  "ok": false,
-  "error": {
-    "type": "validation",
-    "subtype": "source_required",
-    "message": "provide exactly one GitHub URL argument or --source-stdin"
-  }
-}
-```
+Errors use a non-zero exit code and a stable `error.code`. Agent Skills branch
+on the exit code, `ok`, `error.code`, and `retryable`, never on message text.
-Determine local/bootstrap command success from the process exit code. For publication protocol commands, use the process exit code or `ok == true`. The API's domain-specific `error.type` is preserved; the exit code is only a coarse handling class. A successfully read publication may still contain a business terminal status such as `unsupported`, `rejected`, or `failed`; inspect `data.status` instead of treating those states as CLI transport failures.
-| Exit code | Meaning |
-|---|---|
-| `0` | Command completed; inspect returned business status when applicable |
-| `2` | Validation failure |
-| `3` | Authentication or authorization failure |
-| `4` | Retryable transport or concurrency failure |
-| `5` | Internal or protocol failure |
-| `6` | Policy or rollout-gate rejection |
-| `10` | Explicit confirmation required |
-## Security and Risk Controls
-- **No source execution** — the CLI and compiler do not execute third-party scripts, binaries, shell fragments, marketplace commands, or copied instructions.
-- **Typed source intent** — AI Hosts interpret natural-language source requests and pass only a typed `SourceSpec` through `--source-stdin`; CLI/Core never guess a provider from user phrases. Natural-language Candidate edits use `--request-stdin`. Never interpolate untrusted text into command strings, argv, environment variables, or shell pipelines.
-- **Explicit public mutation** — publishing, compiler retry, and cancellation require `--yes`; exit code `10` means the Agent must obtain confirmation, not silently retry.
-- **Safe preview** — use `--dry-run` on inspect or publish when the user needs to review the planned request without network or publication side effects.
-- **Credential isolation** — on macOS, device-login credentials stay in AES-256-GCM encrypted files, with Keychain-backed or explicitly downgraded private key material; filenames do not expose profile/origin names. Other platforms retain their native credential manager. Explicit internal-test overrides are namespaced by profile, stored only in a private `0600` config, and never emitted by CLI output.
-- **Immutable inputs** — inspection binds publication to an immutable source snapshot rather than re-reading a floating URL later.
-- **Bounded waiting** — `job wait` has a maximum duration and returns the latest durable state without cancelling the workflow.
-- **Bounded compiler recovery** — `job retry` keeps the frozen source and publication, accepts only an explicitly retryable platform failure, and is capped by the server.
-- **Verified distribution** — the npm launcher downloads the binary for its exact package version from GitHub or a binary mirror and verifies it against the checksum manifest bundled in the npm package before activation.
-## Diagnose and Update
+## Update
 ```bash
-viceme skills doctor
 viceme update --check
 viceme update
 ```
-`viceme update` installs one exact npm package version, acquires its verified Go binary, and refreshes the bundled Skill from that same release. A standalone development binary is never silently replaced.
+Bootstrap installations read the selected region's official S3 release index,
+verify the exact binary checksum, refresh the matching official Skills, and
+activate the binary atomically. npm installations update through the exact npm
+package version. Updates never inherit `VICEME_ACCESS_TOKEN` into child
+processes.
+Binary or npm-launcher activation, both official Skills, and profile config are
+one recoverable local generation. Standalone and npm activation share an outer
+activation lock, a delegated member-commit lock, and a durable active-generation
+record containing the semantic version, installation method, and immutable
+identity. One startup coordinator inspects
+both standalone and npm journals regardless of which launcher entered the
+process. Every ordinary command reconciles an interrupted outer journal before
+business logic; a recovered process whose version, method, or immutable identity
+changed must be restarted. The lock-internal generation fence rejects a late
+older updater. Every mutation entry repeats the same two-journal arbitration
+after it owns the activation lock and before any staging or network install, so
+a process paused after startup cannot introduce a second recovery protocol. The
+first phase also rejects switching between standalone and npm
+before any mutation; reinstall explicitly after removing the previous generation
+instead of mixing recovery protocols. Every Skills/config transaction holds or
+revalidates the same generation authority immediately before commit. npm child
+activation is bound to the exact committing journal by a one-time nonce and
+target version; its member lock prevents a parent crash from admitting a newer
+generation while the child still commits. A committed target is only cleaned up
+after a crash—it is never reapplied or rolled back. Private
+journals can therefore only restore the complete previous generation or finish
+the complete target generation. `viceme doctor`
+validates Skill/version integrity and an unauthenticated API readiness probe
+before installation commits.
+## First-phase implementation status
+The installation, device authorization, deterministic package upload, manual or
+suggested listing media, review confirmation, publication, cancellation, and
+terminal recovery paths are implemented. Local acceptance uses real Shop API,
+PostgreSQL, Redis, and S3-compatible storage. `make check`, npm package/cold-start
+tests, race tests, and Darwin/Linux/Windows amd64/arm64 builds pass. A real LLM
+provider sandbox remains an environment acceptance item; without credentials,
+analysis fails closed and the manual media path remains available.
+## Security boundaries
+- Local packaging rejects path traversal, absolute paths, symlinks, special
+  files, oversized content, sensitive files, and common secret patterns.
+- The API independently validates the immutable ZIP and object metadata.
+- Presigned upload URLs are never written to the pending-operation store.
+- LLM analysis receives only filtered text, metadata, and image thumbnails.
+- A publication cannot become public until the current review digest, price,
+  cover, and ordered gallery have been confirmed.
 ## Development
-Go 1.23 or newer is required when building from source.
+Requirements: Go 1.23+ and Node.js 22+ for npm packaging checks.
 ```bash
-make build
-make test
 make check
-make skill-check
 make npm-package-check
-make quality-check
+make release-manifest
 ```
-The main checked-in quality artifacts are:
-- [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json), generated from the Cobra command tree;
-- [`quality/example-dry-runs.json`](quality/example-dry-runs.json), which executes documented source paths without network access;
-- [`quality/release-manifest.json`](quality/release-manifest.json), which pins CLI/Skill compatibility and content digests.
-`make npm-package-check` builds the Go executable, packs the real npm tarball, and runs the launcher in isolated temporary homes. It does not require an already-published GitHub Release.
-## Releases
-Maintainers merge normal changes into `dev`; they do not manually edit versions, create tags, write changelog entries, or run `npm publish`. GitHub Actions maintains a single automated `dev -> main` Release PR. Merging that PR authorizes the reviewed version, tag, immutable GitHub Release assets, and npm trusted publication.
-See [`docs/releasing.md`](docs/releasing.md) for repository setup, OIDC trusted publishing, recovery, and integrity rules.
-## Contributing
-Issues and pull requests are welcome. Changes to the public command surface must update the generated command manifest, bundled Skill examples, quality fixtures, and tests in the same PR.
+The CLI and `viceme-shared` / `viceme-publish` Skills are versioned and released
+together. Release artifacts are published to GitHub, npm, `s3.viceme.cn`, and
+`s3.viceme.ai` from the same reviewed commit.

package/README.zh.md CHANGED Viewed

@@ -1,373 +1,196 @@
 # ViceMe CLI
-[![npm version](https://img.shields.io/npm/v/@viceme-ai/cli.svg)](https://www.npmjs.com/package/@viceme-ai/cli)
-[![Go Version](https://img.shields.io/badge/go-%3E%3D1.23-blue.svg)](https://go.dev/)
-[![CLI PR checks](https://github.com/ViceMe-AI/cli/actions/workflows/ci.yml/badge.svg)](https://github.com/ViceMe-AI/cli/actions/workflows/ci.yml)
+ViceMe CLI 是官方 ViceMe Agent Skills 的确定性本地执行器。Codex、Claude
+Code 和 WorkBuddy 通过 Skills 引导用户；CLI 负责安装、设备授权、本地检查、
+确定性打包、上传、审核和发布。
-[中文版](./README.zh.md) | [English](./README.md)
+[English](./README.md)
-ViceMe 官方命令行客户端与 Agent Skill，用于将外部 Skill 发布为稳定、可分享的 ViceMe Agent。它面向 Codex、Claude Code 等 AI 编程工具：Agent Skill 负责理解用户意图，CLI 负责确定性的认证、上传、发布和状态协议。
+## 安装
-[安装](#安装与快速开始) · [AI Agent Skills](#agent-skills) · [认证](#认证) · [区域与-profile](#区域与-profile) · [命令](#命令概览) · [输出契约](#json-输出契约) · [安全](#安全与风险控制) · [开发](#开发)
+官方 Bootstrap 从同一个不可变 Release 一次安装原生 CLI 和两个官方 Skills。
-> **开放状态：** Core 发布传输和稳定链接链路已经实现。`--yes` 后，Publication 先停在 `meta_review` 完成基本信息确认，再在 `awaiting_action` 中确认交互步骤。之后 `confirm_publish` 返回仅创作者可访问的 `/p/{code}` `preview_share_url`；创作者在该页面完成一次真实使用，回到 Agent Host 明确确认。发布完成后，持久回执返回正式 `/v/{code}` `share_url`，旧预览地址会重定向到它。CLI 不再提供独立 PreviewRun 命令。预览阶段的输入、输出、文件、会话和 Runner 历史均为临时数据，终态后会被清理，不进入公开统计或作品历史。
-## 为什么选择 ViceMe CLI？
-- **为 Agent 原生设计** — 与 Codex、Claude Code 可使用的官方 ViceMe Agent Skill 一起发布。
-- **确定性边界** — CLI 执行类型化协议操作，不会再启动一层对话式 Agent Loop。
-- **服务端编译** — 来源解析、LLM 编译、BuildRun 固化和 Release 发布均在 ViceMe 基础设施中完成。
-- **稳定发布** — 同一个逻辑 Agent 后续发布新版本时继续使用同一个分享链接。
-- **支持多种来源** — 支持结构化 GitHub 与小红书/RedSkill 来源、压缩包和本地 Skill 目录。
-- **默认安全** — 在 macOS 上，设备登录凭证保存在 AES-256-GCM 加密文件中，主密钥默认由系统 Keychain 保护；其他平台继续使用原生凭证管理器。显式本地覆盖要求私有 Profile 文件，公开变更需要确认，下载的二进制文件必须通过校验和验证。
-- **人类与 Agent 双登录模式** — `viceme auth login` 在终端中引导用户，Agent 跨回合流程则显式使用 JSON。
-## 安装与快速开始
-### 环境要求
-- Node.js 18.20 或更高版本
-- macOS 或 Linux `amd64`/`arm64`，或 Windows `amd64`/`arm64`
-### 快速开始（人类用户）
-> **AI 助手注意：** 如果你是正在帮助用户安装 ViceMe 的 AI Agent，请直接跳转到[快速开始（AI Agent）](#快速开始ai-agent)。该章节定义了必须遵循的跨回合设备登录流程。
-#### 安装
-以下两种方式任选其一。
-**方式一 — 通过 npm 一次性安装（推荐）：**
-```bash
-npx --yes @viceme-ai/cli@latest install
-```
-**方式二 — 全局安装 npm 启动器：**
+中国区 macOS / Linux：
 ```bash
-npm install --global @viceme-ai/cli
-viceme install
+curl -fsSL https://s3.viceme.cn/install.sh | sh
 ```
-两种方式都会持久化 npm 启动器、下载匹配且通过校验和验证的 Go 二进制文件，并安装随包发布的 ViceMe Agent Skill。二进制下载依次尝试 GitHub Release、用户配置的非默认 npm registry 的 `/-/binary/viceme-cli/` 镜像，最后回退到公共 npmmirror binary 镜像。启动器使用系统 `curl`，因此会遵循标准代理环境变量。默认使用中国区服务。使用国际区服务：
+海外 macOS / Linux：
 ```bash
-npx --yes @viceme-ai/cli@latest install --region global
+VICEME_REGION=global sh -c "$(curl -fsSL https://s3.viceme.ai/install.sh)"
 ```
-#### 认证并验证
+中国区 Windows PowerShell：
-如果安装结果提示需要认证，启动引导式设备登录：
-```bash
-viceme auth login
+```powershell
+irm https://s3.viceme.cn/install.ps1 | iex
 ```
-CLI 会显示浏览器登录链接、等待授权，并在同一个终端中报告成功。然后验证安装：
+海外 Windows PowerShell：
-```bash
-viceme auth status
-viceme skills doctor
+```powershell
+$env:VICEME_REGION="global"; irm https://s3.viceme.ai/install.ps1 | iex
 ```
-## 快速开始（AI Agent）
-> 以下步骤面向 AI Agent。浏览器授权必须由用户在另一个回合完成；不要无限等待，也不要向用户索取令牌。
-**第 1 步 — 安装**
-使用随包 Skill 中定义的完整引导命令。显式 npm registry 是安装信任边界的一部分：
+不能使用 Bootstrap 时，以 npm 作为备用入口：
 ```bash
-npx --yes --registry=https://registry.npmjs.org --@viceme-ai:registry=https://registry.npmjs.org --package=@viceme-ai/cli@latest -- viceme install
-```
-读取格式化业务结果中的 `authenticated` 和 `next_step`。如果认证已经有效，直接进入第 4 步。如果需要登录，Agent 不要执行面向人类的 `next_step`，而应使用第 2 步的 JSON 跨回合流程。
-**第 2 步 — 需要时启动设备登录**
-```bash
-viceme auth login --no-wait --json
+npx --yes @viceme-ai/cli@latest install
 ```
-向用户返回准确的 `verification_url`；存在 `verification_url_complete` 时，CLI 会把这个已预填设备码的浏览器直达链接规范化为 `verification_url`。只有浏览器要求输入时，才把 `user_code` 作为备用信息提供。保留 `device_code` 供后续命令使用，然后结束当前回合。不要在对话中索取、打印或传递访问令牌。
-**第 3 步 — 在后续回合继续同一个登录流程**
+安装器始终写入兼容 fallback `~/.agents/skills`，并为检测到的 Agent 写入
+各自的用户级目录：
-用户确认已在浏览器完成授权后：
+| Agent | 原生目录 |
+| --- | --- |
+| Codex | `~/.codex/skills` |
+| Claude Code | `~/.claude/skills` |
+| WorkBuddy | `~/.workbuddy/skills` |
-```bash
-viceme auth login --device-code <device-code> --json
-```
+可通过 `viceme install --agent codex`、`claude`、`workbuddy` 或 `agents`
+指定目标。安装或修复后运行 `viceme doctor`。
-如果授权仍处于等待状态，应在过期前继续使用同一个 device code。只有原流程已过期时才能重新发起设备登录。
-使用非默认 Profile 时，启动和继续设备登录必须传入同一个全局 `--profile <name>`；启动结果会返回实际的 `profile` 与 `region`。
+## 登录与 Profile
-**第 4 步 — 验证**
+每个 Profile 绑定一个官方区域和一个通过设备码授权的账户。
 ```bash
+viceme auth login
 viceme auth status
-viceme skills doctor
-viceme skills list
-```
-只有认证有效且 `skills doctor` 报告安装健康、版本兼容时，才能继续。
-**第 5 步 — 检查第一个来源**
+viceme auth logout
-```bash
-viceme skill inspect https://github.com/acme/poster-skill --skill-root .
-```
-inspect 是只读操作。后续应按照随包发布的 `viceme` Skill 处理不同来源、Target 选择、用户确认、有界任务等待和结果返回。若 `destination.recovery.mode=resume_existing_publication`，必须用其中的 `publication_id` 运行 `job get` 并恢复该非终态 Publication，不能用本次 inspect resolution 再次发布。若 Publication 终结为 `binding_required`，运行 `viceme job bind <publication-id>`，把服务端签名的 ViceMe 链接交给用户后停止；下载或 Fork 仅为提示，CLI 不会自动执行。用户完成精确 GitHub/小红书渠道绑定后，必须重新 inspect 并创建新的普通 Publication，不能恢复旧任务。进入 `meta_review` 后，使用同一 action ID 与 payload digest 展示并决议信息，然后再次等待；进入 `awaiting_action` 后，确认交互步骤，打开私有 `preview_share_url` 在预览页完成一次成功使用。若 `confirm_steps` 或 `confirm_publish` action 过期，先用 `job get` 读取同一 Publication，再显式运行 `viceme job renew <publication-id> --action-id <expired-action-id>`，只使用返回的新 `next_action` 继续，不能创建第二条 Publication。最后回到对话明确确认或取消。确认后继续等待到 `share_published`，返回正式 `data.result.share_url`；不要要求它与私有预览地址相同。
-## 区域与 Profile
-每个 Profile 独立选择一个 ViceMe 区域：
-| 区域 | 安装命令 | API 地址 |
-|---|---|---|
-| 中国区 | `viceme install` | `https://api.viceme.cn` |
-| 国际区 | `viceme install --region global` | `https://api.viceme.ai` |
-首次安装会创建 `default` Profile。设备登录凭证按 Profile 与规范化 API origin 隔离；在 macOS 上，它们保存在私有的 AES-256-GCM 加密文件中，加密主密钥通常只保存在系统 Keychain；其他平台继续使用原生凭证管理器。受控本地/内部操作还可以在 Profile 中显式配置 publication credential；此类配置文件始终保持 `0600` 私有权限。
-```bash
 viceme profile list
-viceme profile add --name work --region global --use
 viceme profile use default
-viceme --profile work auth status
-viceme profile rename work company
-viceme profile remove company
 ```
-`profile use` 修改持久化的当前 Profile；全局 `--profile` 只覆盖本次命令。不要让 AI Agent 在用户没有明确要求时切换或删除 Profile。
-受控本地/内部操作可创建同时包含显式 endpoint 与 audience-bound publication credential 的独立 Profile：
+Agent 无法在同一个回合等待浏览器授权时使用分段流程：
 ```bash
-viceme profile add --name local --region cn \
-  --api-base-url http://localhost:8090 \
-  --access-token '<vpa1.local-dev.credential>' --use
-viceme profile configure local --access-token 'YOUR_ACCESS_TOKEN'
-viceme profile configure local --clear-access-token
-viceme profile configure local --clear-api-base-url
+viceme auth login --no-wait
+viceme auth login --device-code <device-code>
 ```
-正常 `viceme auth login` 永远不会向 Profile 写入凭证。本地 Profile credential 只能通过显式的 `profile add/configure` flag 设置、替换或清除，list/status 只报告 `source=local_profile`，不会返回 token。`--access-token` 会出现在 argv 中且可能进入 shell history，因此只能用于本文约定的受控本地/内部环境。
-凭证优先级为进程 `VICEME_ACCESS_TOKEN` → 当前本地 Profile → 设备登录。publication credential 必须使用 `vpa1.<audience>.<secret>`：`cn-prod` 只能访问 `https://api.viceme.cn`，`global-prod` 只能访问 `https://api.viceme.ai`，`dev-preview` 只能访问 `https://viceme-envoy-dev.preview.tencent-zeabur.cn`，Profile 中的 `local-dev` 只能访问 loopback endpoint；进程 `local-dev` 还要求 `VICEME_CLI_ALLOW_LOCAL_PROCESS_CREDENTIAL=1`。macOS 和 Linux 的配置默认位于 `~/.viceme-cli`，Windows 默认位于 `%LOCALAPPDATA%\ViceMe\Config`；Windows 上既有的 `~/.viceme-cli/config.json` 会继续使用，直到显式迁移。`VICEME_CLI_CONFIG_DIR` 可覆盖配置根目录。`VICEME_API_BASE_URL` 仍只是单进程 endpoint 覆盖，不能放宽 Profile credential 的 origin；它不能与全局 `--profile` 同时使用，必须在选中 Profile 的持久 endpoint 与进程 endpoint 覆盖之间只保留一个权威来源。API 与预签名上传重定向一律 fail closed。
+用户在浏览器完成授权。不要在对话中复制 Access Token。凭据按 Profile 和 API
+Origin 隔离。
-更新检查直接请求 npm registry，并且只把最近一次成功查询到的版本写入配置目录中的 `update-state.json`；registry 暂时不可用时，该结果最多回退使用 24 小时。npm 管理的 CLI 在普通命令中只同步读取本地缓存，并且最多每 24 小时在后台刷新一次，因此命令不会等待版本发现。当缓存确认存在新版本时，结构化成功与错误对象都会携带 `_notice.update`，其中包含 `current`、`latest`、`message` 和精确的 `viceme update` 命令，AI Agent 可以据此提醒用户。该提醒不会改变命令退出码，也不会自动执行更新。非 CI 环境可以设置 `VICEME_NO_UPDATE_NOTIFIER=1` 关闭提醒；标准 CI 环境会自动跳过。`viceme install` 和 `viceme update` 启动的 npm 操作统一使用该目录下隔离的 `npm-cache` 子目录，不会因为用户级 npm 缓存损坏而失败。这两个位置都不包含秘密信息，可以安全删除；凭证不会进入任何更新缓存。
+## 发布 Skill
-## Agent Skills
+第一版接受包含 `SKILL.md` 的本地目录或本地 ZIP，不接受 GitHub URL、远程下载和
+多 Skill 批量包。
-当前版本有意只提供一个平台级 Agent Skill：
-| Skill | 说明 | 支持的宿主 |
-|---|---|---|
-| `viceme` | 将外部 Skill 安装、检查、转换、发布、更新或分享为稳定的 ViceMe Agent；统一约束认证、来源、Target、确认、任务和安全规则 | Codex、Claude Code |
-GitHub、小红书/RedSkill、ZIP 和目录是同一个 `viceme` 发布流程处理的来源类型，不是相互独立的 Agent Skills。这样可以让不同来源共享一致的安全边界和稳定链接合同。
-CLI 与 `viceme` Agent Skill 从同一个仓库以相同版本发布。`viceme install` 会把完整 Skill Bundle 安装到检测到的受支持宿主中；二进制文件同时嵌入用于确定性自检、可供 Agent 阅读的内容子集。
+只读检查：
 ```bash
-viceme skills list
-viceme skills read viceme
-viceme skills read viceme references/commands.md
-viceme skills doctor
+viceme skill inspect --path ./my-skill
 ```
-`skills doctor` 会分别校验 CLI 版本、Skill 版本、兼容范围、完整 Bundle 摘要和嵌入内容摘要。安装内容被修改或版本不兼容时会直接失败。
-## 认证
-| 命令 | 用途 |
-|---|---|
-| `viceme auth status` | 查看当前 Profile 是否已认证 |
-| `viceme auth login` | 引导人类用户完成浏览器授权并等待结果 |
-| `viceme auth login --no-wait --json` | 启动 Agent 跨回合流程并返回结构化设备授权信息 |
-| `viceme auth login --device-code <code> --json` | 在后续回合完成 Agent 登录流程 |
-| `viceme auth logout` | 撤销并删除当前 Profile 的凭证 |
-在 macOS 上，设备登录生成的令牌只保存在私有加密凭证文件中，加密主密钥通常保存在系统 Keychain；其他平台继续使用原生凭证管理器。正常登录不会回填显式本地 Profile 字段，登录成功的输出也不会包含访问令牌或刷新令牌。
-CLI 会在创建设备授权或兑换一次性 device code 之前，对完整的本地持久化链路做预检。预检失败时不会消费任何一次性授权。如果预检后仍在兑换成功时发生存储故障，CLI 会尝试撤销刚签发的凭证，并以 `credential_persistence_failed` 明确说明必须重新发起 device flow；不会误报登录成功，也不会输出 token。
-### macOS 沙箱（Codex 与 Claude Code）
-当用户在无法访问 macOS Keychain 的沙箱中显式执行设备登录时，CLI 会自动创建私有的 `0600` 文件主密钥，并将新授权的凭证保存到本地加密文件中，无需用户预先执行额外命令。
-如果用户不想重新登录，希望直接复用此前在 Terminal 中使用 Keychain 主密钥创建的凭证，可以从同一 macOS 用户的交互式终端执行一次：
+查看确定性发布包和价格计划：
 ```bash
-viceme config keychain-downgrade
+viceme skill publish --path ./my-skill --price-minor 100 --dry-run
 ```
-该命令会把现有主密钥复制到 `~/.viceme-cli/credentials/master.key.file`，并将已配置 Profile 的旧 Keychain 凭证导入加密文件。原 Keychain 条目会保留为冷备份。命令可重复执行，不会打印 token，也不会将 token 明文落盘。完成后，同一 macOS 用户下的 Codex、Claude Code 沙箱无需访问 Keychain 即可读取加密凭证；如果愿意重新登录，则不需要执行该迁移命令。其明确的安全取舍是：降级后由用户文件权限（目录 `0700`、文件 `0600`）代替 Keychain 的进程级访问边界。
-公开 CLI 只提供一套标准认证与发布命令面。工作人员短时授权凭证可由进程环境注入（`source=process`），也可由受控本地 Profile 显式配置（`source=local_profile`）；两者都只调用标准 `inspect/publish/job` 并使用统一 `x-api-key`。CLI 不提供身份选择或 staff authorization 签发命令；永远不输出 token，任一覆盖凭证生效时 login/logout fail closed，update 子进程也不会继承该凭证。
-## 支持的来源
-### GitHub 或可信来源平台
+开始可恢复的上传与 Listing 分析：
 ```bash
-viceme skill inspect https://github.com/acme/poster-skill --skill-root .
-viceme skill publish --resolution-id <resolution-id> --yes
+viceme skill publish --path ./my-skill --price-minor 100
 ```
-GitHub 来源必须传 `--skill-root`，它是包含 `SKILL.md` 的精确仓库相对目录；只有根级 Skill 才使用 `.`。调用 Agent 根据用户输入或只读仓库文件树确定该路径，ViceMe 不扫描全仓猜测 Skill。
-### 小红书或 RedSkill
+之后以服务端 Publication 状态为准：
 ```bash
-viceme skill inspect --source-stdin
-viceme skill publish --resolution-id <resolution-id> --yes
+viceme publication get <publication-id>
+viceme publication review <publication-id>
+viceme publication asset upload <publication-id> --role cover --path ./cover.png
+viceme publication asset upload <publication-id> --role gallery --path ./demo.png
+viceme publication update <publication-id> --input ./listing-draft.json
+viceme publication confirm <publication-id> --review-digest <digest>
+viceme publication publish <publication-id> --review-digest <digest>
 ```
-AI Host 先理解用户的来源意图，再通过 stdin 传入唯一的结构化
-`SourceSpec`，例如 `{"kind":"redskill","value":"ai-desk-card"}`。CLI/Core
-不再用关键词或正则分类自然语言；用户明确指定的平台必须保持不变，来源存在歧义时必须先询问，不能静默替换为其他平台的同名来源。
-### 压缩包或本地 Skill 目录
-```bash
-viceme skill publish --file ./poster-skill.zip --new-target --target-alias poster --yes
-viceme skill publish --dir ./poster-skill --new-target --target-alias poster --yes
-```
+模型会建议 `summaryZhCn`、`summaryEnUs` 和包内图片，但不能替用户确认，也不能决定价格。
+每版短简介的最大显示宽度为 30：ASCII 计 1，中文及其他非 ASCII 计 2。Agent 必须把双语
+短简介、价格、封面和有序展示素材的完整 Draft 展示给用户，在 `confirm` 前获得明确确认，
+并在真正公开的 `publish` 前再次确认。
-后续发布新版本时，应先解析现有 Target 并使用乐观并发控制。发生冲突时不能创建新的分享链接：
+网络中断后继续同一个发布操作：
 ```bash
-viceme skill target get target_123
-viceme skill publish --file ./poster-skill-v2.zip \
-  --target-id target_123 --expected-target-version 4 --yes
+viceme skill publish --resume <publication-id>
 ```
-## 命令概览
-| 命令组 | 用途 |
-|---|---|
-| `viceme install` | 安装持久化启动器、Agent Skill 和默认 Profile |
-| `viceme auth` | 启动、完成、检查或撤销设备认证 |
-| `viceme config` | 管理受控的 macOS Keychain 到文件的沙箱降级 |
-| `viceme profile` | 新增、列出、切换、重命名或删除本地 Profile |
-| `viceme skill inspect` | 固化并检查来源候选，不执行发布 |
-| `viceme skill publish` | 创建或更新具有稳定链接的 Skill Agent 发布 |
-| `viceme skill target` | 解析现有逻辑 Agent Target 及其版本 |
-| `viceme job` | 读取或等待发布任务，审阅信息与冻结摘要，编辑 Candidate，展示稳定预览/正式分享链接和签名渠道绑定链接，决议 action，以及显式重试或取消 |
-| `viceme skills` | 读取、安装和诊断随包发布的 Agent Skill |
-| `viceme update` | 同时更新 npm 启动器、已校验二进制文件和随包发布的 Skill |
-使用 `viceme <command> --help` 查看准确参数。经过发布检查的机器可读命令面存放在 [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json)。
+响应未知时不能创建第二个 Publication，应先查询或恢复原 ID。
 ## 输出契约
-ViceMe 根据命令语义选择最小且稳定的输出形式：
-- `version`、`install`、`update`、`auth status`、`profile *`、`skills doctor` 等本地/引导命令，将格式化后的业务结果直接写入 **stdout**，不附加 `ok`、`data` 或无关构建元数据；npm 管理的普通调用最多只会附加上文约定的保留字段 `_notice.update`。
-- `skills read` 按原始字节输出目标文件，不添加 JSON 包装。
-- 交互式 `viceme auth login` 输出面向人的引导；AI Agent 使用 `--no-wait --json`，并在后续回合用 `--device-code <code> --json` 继续，这两个命令返回格式化的裸业务对象。
-- `skill` 和 `job` 下的发布协议命令继续使用稳定 Envelope，因为 action receipt、持久状态与有界等待元数据共同构成跨命令协议。
-发布协议命令成功时写入 **stdout**，退出码为 `0`：
-```json
-{
-  "ok": true,
-  "data": {}
-}
-```
-只有有界等待真实超时时才附加协议元数据：
+业务结果默认使用 JSON。成功时 stdout 只包含最终结果；进度和诊断只写 stderr。
 ```json
 {
   "ok": true,
   "data": {},
   "meta": {
-    "wait_timed_out": true
+    "cliVersion": "0.10.1",
+    "requestId": "optional"
   }
 }
 ```
-CLI 执行错误以格式化形式写入 **stderr**，退出码非零：
-```json
-{
-  "ok": false,
-  "error": {
-    "type": "validation",
-    "subtype": "source_required",
-    "message": "provide exactly one GitHub URL argument or --source-stdin"
-  }
-}
-```
+失败使用非零退出码和稳定的 `error.code`。Agent Skills 只能依据退出码、`ok`、
+`error.code` 和 `retryable` 分支，不能解析 message 文本。
-本地/引导命令根据进程退出码判断成功；发布协议命令可以检查退出码或 `ok == true`。API 返回的领域 `error.type` 会原样保留，退出码只表示粗粒度处理类别。成功读取发布任务时，业务状态仍可能是 `unsupported`、`rejected` 或 `failed`；这时应检查 `data.status`，不能把这些状态当成 CLI 传输失败。
-| 退出码 | 含义 |
-|---|---|
-| `0` | 命令完成；适用时继续检查返回的业务状态 |
-| `2` | 参数校验失败 |
-| `3` | 认证或授权失败 |
-| `4` | 可重试的传输或并发失败 |
-| `5` | 内部或协议失败 |
-| `6` | 策略或开放门禁拒绝 |
-| `10` | 需要明确确认 |
-## 安全与风险控制
-- **不执行来源内容** — CLI 和编译器不会执行第三方脚本、二进制文件、shell 片段、市场命令或复制口令中的指令。
-- **结构化来源意图** — AI Host 负责理解自然语言来源，并只通过 `--source-stdin` 传递 typed `SourceSpec`；CLI/Core 不根据用户措辞猜 Provider。Candidate 的自然语言修改通过 `--request-stdin` 传递。不得把不可信文本拼入命令字符串、argv、环境变量或 shell 管道。
-- **公开变更需要明确确认** — 发布、编译重试和取消操作需要 `--yes`；退出码 `10` 表示 Agent 必须向用户取得确认，不能静默重试。
-- **安全预览** — 用户需要检查计划请求时，可以对 inspect 或 publish 使用 `--dry-run`，不会产生网络请求或发布副作用。
-- **凭证隔离** — 在 macOS 上，设备登录凭证保存在 AES-256-GCM 加密文件中，主密钥由 Keychain 或显式降级后的私有文件保护，文件名不会暴露 Profile/origin；其他平台继续使用原生凭证管理器。显式内部测试覆盖按 Profile 隔离，仅允许保存在 `0600` 配置中，并且不会出现在 CLI 输出中。
-- **不可变输入** — inspect 会把发布绑定到不可变来源快照，而不是在之后重新读取浮动 URL。
-- **有界等待** — `job wait` 有最大等待时间；超时后返回最新持久化状态，不会取消工作流。
-- **有界编译恢复** — `job retry` 复用已冻结的来源与同一发布任务，只接受明确标记为可重试的平台故障，并由服务端限制次数。
-- **可信分发** — npm 启动器从 GitHub 或 binary 镜像下载与其准确包版本匹配的二进制文件，并在启用前使用 npm 包内置的校验清单验证 SHA-256。
-## 诊断与更新
+## 更新
 ```bash
-viceme skills doctor
 viceme update --check
 viceme update
 ```
-`viceme update` 会安装一个准确的 npm 包版本，获取对应的已校验 Go 二进制文件，并从同一版本刷新随包发布的 Skill。独立开发版二进制文件不会被静默替换。
+Bootstrap 安装会读取当前 Profile 所在区域的官方 S3 Release 索引，校验精确二进制
+Checksum，用新二进制修复同版本官方 Skills 后原子激活。npm 安装通过精确 npm 包版本
+更新。更新子进程不会继承 `VICEME_ACCESS_TOKEN`。
+二进制或 npm launcher、两份官方 Skills 和 Profile 配置属于同一个可恢复的本地版本。
+Standalone 与 npm 激活共用外层激活锁、委托成员提交锁，并持久化包含语义版本、安装方式和不可变身份的
+active-generation。唯一的启动协调器不区分当前入口，始终检查 Standalone 和 npm 两类 Journal。
+所有普通命令必须先恢复未完成的外层 Journal；如果恢复后当前进程的版本、安装方式或不可变身份
+不再等于 active generation，本次命令会停止并要求重新执行。锁内 generation fence 会拒绝迟到
+的旧版本更新。每个真实写入口拿到 activation lock 后、开始暂存或联网安装前，都会再次对两类
+Journal 做相同仲裁，避免通过启动检查后暂停的旧进程引入第二套恢复协议。第一阶段也会在任何
+文件变更前拒绝 Standalone 与 npm 的原地切换，不能把两套
+恢复协议混合使用。每个 Skills/配置事务都必须持有同一代际权限，或在最终提交前重新验证。
+npm 内部安装子进程必须匹配 Journal 中的一次性 nonce、目标版本和 Skill 目标；成员提交锁会阻止
+父进程崩溃后，新一代在旧子进程尚未提交完毕时进入。目标一旦越过语义提交点，崩溃恢复只能完成本地清理，不能重新联网安装或回滚。私有
+Journal 因此只能完整恢复上一代或完整完成目标版本；安装提交前，`viceme doctor` 同时校验
+Skill/版本完整性和不携带凭据的 API readiness。
+## 第一阶段实施状态
+安装、设备码授权、确定性包上传、建议或人工展示素材、Review 确认、发布、取消和终态恢复均已
+实现。本地验收使用真实 Shop API、PostgreSQL、Redis 和 S3 兼容存储；`make check`、npm
+打包与冷启动、race test，以及 Darwin/Linux/Windows 的 amd64/arm64 构建均通过。真实 LLM
+Provider sandbox 仍属于部署环境验收项；没有凭据时分析 fail closed，人工素材流程仍可完成发布。
+## 安全边界
+- 本地打包拒绝路径穿越、绝对路径、符号链接、特殊文件、超限内容、敏感文件和常见
+  Secret 模式。
+- API 独立校验不可变 ZIP 与对象元数据，不能信任 CLI 自报结果。
+- Pending operation 不保存预签名上传 URL。
+- LLM 只接收经过筛选的文本、元数据和图片缩略图。
+- 未确认当前 Review Digest、价格、封面和有序展示素材时，Publication 不能公开。
 ## 开发
-从源码构建需要 Go 1.23 或更高版本。
+需要 Go 1.23+；npm 包检查还需要 Node.js 22+。
 ```bash
-make build
-make test
 make check
-make skill-check
 make npm-package-check
-make quality-check
+make release-manifest
 ```
-仓库中主要的质量产物包括：
-- [`skills/viceme/references/command-manifest.json`](skills/viceme/references/command-manifest.json)：从 Cobra 命令树生成；
-- [`quality/example-dry-runs.json`](quality/example-dry-runs.json)：在没有网络请求的情况下执行文档中的来源路径；
-- [`quality/release-manifest.json`](quality/release-manifest.json)：固定 CLI/Skill 兼容范围和内容摘要。
-`make npm-package-check` 会构建 Go 可执行文件、打包真实 npm tarball，并在相互隔离的临时主目录中运行启动器，不依赖已经发布的 GitHub Release。
-## 发布
-维护者将普通改动合入 `dev`；不需要手工修改版本、创建 tag、编写 changelog 或执行 `npm publish`。GitHub Actions 持续维护唯一的 `dev -> main` 自动 Release PR。合并该 PR 即授权发布经过审阅的版本、tag、不可变 GitHub Release 产物及 npm trusted publication。
-仓库设置、OIDC trusted publishing、恢复和完整性规则参见 [`docs/releasing.md`](docs/releasing.md)。
-## 贡献
-欢迎提交 Issue 和 Pull Request。公共命令面发生变化时，必须在同一个 PR 中更新生成的命令清单、随包发布的 Skill 示例、质量样例和测试。
+CLI、`viceme-shared` 和 `viceme-publish` 同版本发布。GitHub、npm、
+`s3.viceme.cn` 与 `s3.viceme.ai` 的产物都来自同一个已评审 Commit。

package/checksums.txt CHANGED Viewed

@@ -1,6 +1,6 @@
-937e2f4bdda6af06e7b86dbc6a3fc550ddc85b7061df1e86290e1b6d5fbc4cef  viceme_0.10.1_darwin_amd64
-7e4d08782370e0f9bffdad78a5e8713171a867695da15f28632a6bc5a75f522f  viceme_0.10.1_darwin_arm64
-06836eaff0297e9604d2fb93b9396610667562549710db3ecc0e5a027e1a66a5  viceme_0.10.1_linux_amd64
-b9f81bb63f6715bb6c28ac91050a39841c81909c0e1f3cdefcda3792ebc978be  viceme_0.10.1_linux_arm64
-5001f731bab7fc37cfbca3ed2401d9fad4a9b1a3ce90a78fded63163953ebbb2  viceme_0.10.1_windows_amd64.exe
-511fbd483e4d6be2dbbc6ad324076bc0db948c76520f4ee8951c65a153a2b788  viceme_0.10.1_windows_arm64.exe
+13dea4727d2b8b470aab551a874b87982886296728654bbce3b6fcbe12c0cfa9  viceme_0.11.1_darwin_amd64
+892dcbf8b4c91b8f1b59b5023322dcfcacc87959c4d031b755058beffd96cad0  viceme_0.11.1_darwin_arm64
+b7c2226a7e8debfecb071af7b6d27187d4ba7348bed884dbd6062b626415d647  viceme_0.11.1_linux_amd64
+ee12c78d67fc17b99de3550c3716fa96c204e14cad22117e013b6989b8d2d176  viceme_0.11.1_linux_arm64
+7c775b7ab1d736d77fd024a0e69bedf4f3ecdb8b94b29412453382c53466204f  viceme_0.11.1_windows_amd64.exe
+473a345fcfc2e5a11259364ff5aef0f5691033fa1d587bcc80cc8e5a1634697b  viceme_0.11.1_windows_arm64.exe

package/npm/bin/viceme.mjs CHANGED Viewed

@@ -41,5 +41,16 @@ try {
   process.exitCode = await main();
 } catch (error) {
   process.stderr.write(`viceme launcher: ${error.message}\n`);
+  process.stdout.write(
+    `${JSON.stringify({
+      ok: false,
+      error: {
+        type: "internal",
+        code: "LAUNCHER_FAILED",
+        message: "ViceMe launcher could not start",
+        retryable: false,
+      },
+    })}\n`,
+  );
   process.exitCode = 1;
 }

package/package.json CHANGED Viewed

@@ -1,7 +1,7 @@
 {
   "name": "@viceme-ai/cli",
-  "version": "0.10.1",
-  "description": "Install and run the ViceMe CLI and ViceMe Agent Skill",
+  "version": "0.11.1",
+  "description": "Install the ViceMe creator CLI and official Agent Skills",
   "type": "module",
   "bin": {
     "viceme": "npm/bin/viceme.mjs"