PatchXNote Agent

English | 简体中文

Official documentation: Public Chinese user guide on Feishu

Repository: https://github.com/ZsTs119/patchxnote-agent

PatchXNote Agent is the local AI assistant connector for PatchXNote. After installing it, you can let an AI assistant find PatchXNote records, inspect AI-generated results, create Markdown, and send user-approved messages to Feishu, DingTalk, or another webhook.

It helps your AI assistant search records synced from mobile or desktop, inspect the source text and AI output behind a processing run, save editable local drafts, and manually send the final text to a named webhook target.

PatchXNote server data access remains read-only. Agent does not operate hardware binding, read raw audio, handle payments, or expose Admin APIs. Webhook configuration and sending happen locally, and messages are sent only when you or your AI assistant explicitly call a send command.

Give this one-line prompt to a local-command-capable AI assistant:

Help me connect PatchXNote MCP and install the PatchXNote MCP Skill when the client supports Agent Skills. First, if this is a local command-capable environment, execute npx -y patchxnote-agent@latest skill install on this machine so future or long sessions remember the PatchXNote SOP; if skill installation is not supported, continue with MCP setup directly. Identify the MCP client you are running in; if this is a local MCP client such as Codex, Cursor, VS Code, Claude Desktop, Claude Code, or Windsurf, run npx -y patchxnote-agent@latest setup --client <client-id> on this machine in the same OS/runtime that will later launch the MCP server. If you are not sure which client ID to use, first run npx -y patchxnote-agent@latest mcp config and use the printed generic stdio config. For login, open the browser and let me complete PatchXNote phone-code authorization there; do not ask me to paste OTP codes, OAuth codes, authorization codes, access tokens, refresh tokens, or webhook secrets into chat. After setup, call patchxnote_get_current_user and patchxnote_list_memories with {"platform":"mobile","limit":5} to verify it works. References: https://github.com/ZsTs119/patchxnote-agent and https://patchx2025.feishu.cn/wiki/PnVRwYT7IirFPckairGcWPnHnCd

For a supported client, start with setup:

npx -y patchxnote-agent@latest setup --client cursor

Or print the generic MCP config manually:

npx -y patchxnote-agent@latest mcp config

PatchXNote MCP Skill

This repository also includes an Agent Skills package at skills/patchxnote-mcp/. It teaches compatible AI clients the PatchXNote setup, browser OAuth, runtime, verification, summary, memory, model-result, webhook, and evidence rules so the workflow survives fresh or long sessions.

patchxnote-agent on npm is the MCP installer/runtime package and also bundles the PatchXNote MCP Skill. Install or refresh the skill with:

npx -y patchxnote-agent@latest skill install

Use --agent <id> only after that client's local skills directory is verified; the default writes under the user's home directory at .agents/skills/patchxnote-mcp.

Useful skill installer options:

Option	Use
--dry-run --json	Preview target paths and conflict status without writing files.
--home <path> or PATCHXNOTE_AGENT_SKILL_HOME=<path>	Install into a test or alternate home directory.
`--agent universal	codex
--force	Replace an existing unmanaged or manually edited patchxnote-mcp skill after explicit user intent.

The installer is idempotent. If an existing patchxnote-mcp directory differs and is not managed by this package, it refuses to overwrite it unless --force is supplied. Managed installs include .patchxnote-agent-skill.json with the package version and source hash.

The skill does not authenticate PatchXNote or start MCP by itself. It complements the MCP setup command:

npx -y patchxnote-agent@latest setup --client <client-id>

For troubleshooting or rollback, pin the current published skill installer:

npx -y patchxnote-agent@0.2.11 skill install

At A Glance

Area	Agent V1 behavior
Runtime	Installs or verifies a versioned native patchxnote binary through an npm wrapper.
Browser MCP login	patchxnote mcp login opens the PatchXNote OAuth page, receives the loopback callback, and stores MCP connector credentials in the OS-native keychain.
Terminal CLI login	patchxnote login remains available for terminal-only use and the legacy local Agent path.
Local MCP service	patchxnote mcp serve runs the local stdio MCP server for editors and desktop agents.
Hosted MCP gateway	GoServer-hosted /mcp is the remote/platform path; platform-console acceptance is tracked separately.
Data access	Shows account, recorder cards, quota, records, and AI-generated results.
Webhook	Locally configures named targets and manually sends to Feishu, DingTalk, or another webhook.
Safety boundary	Server data is read-only; raw audio, hardware, payment, and Admin APIs are not exposed.
Package status	Current public beta release 0.2.11, defaulting to the PatchXNote public beta API.

Features

Capability	Available in 0.2.11	Notes
Browser OAuth MCP login	Yes	patchxnote mcp login opens the PatchXNote authorization page, completes phone OTP on the GoServer page, then stores MCP credentials locally.
Terminal phone OTP Agent login	Yes	patchxnote login keeps the terminal login path for CLI-first users and fallback environments.
Agent session refresh	Yes	Automatically rotates Agent access and refresh tokens from the local keychain.
Local MCP server	Yes	Lets MCP-capable AI assistants call PatchXNote Agent.
Hosted remote MCP gateway	Beta	Used for platform agents that cannot execute local npx; each platform still needs real console acceptance.
Local client setup wizard	Yes	patchxnote setup --client <id> plans, confirms, backs up, writes supported client config, and falls back to manual instructions where needed.
Account, recorder cards, quota	Yes	Shows account status, recorder-card list, quota, and current-month model usage.
Record list and search	Yes	Lists readable record entries by mobile or desktop, including saved results and model-generated outputs, then searches basics cached in the current MCP session.
Single record details	Yes	Shows safe basic information for one record.
AI processing lookup	Yes	Finds an AI processing run and the follow-up request_id.
Source text and AI result export	Yes	Explicitly inspect or export source text, AI response, parsed result, and final result.
Multiple webhook aliases	Yes	Configure multiple Feishu, DingTalk, or generic webhooks with custom names, including Chinese names and spaces.
Markdown drafts	Yes	Render a record into a local Markdown draft so the user can edit before sending.
Manual webhook send	Yes	Sends only when the user or AI explicitly invokes a send command; no background push.
Raw audio/audio download	No	Agent does not read raw audio and does not provide audio downloads.
Hardware bind/release/recovery	No	Owned by App/PC and MR20 flows.
Model execution/replay	No	Agent does not trigger new model runs or replay model calls.
Payment/purchase/Admin APIs	No	Quota purchase, payments, and Admin APIs are out of scope.

Requirements

• Node.js 18 or newer for the npm installer/launcher wrapper.
• Windows, macOS, or Linux on amd64 or arm64.
• A PatchXNote account that can receive the phone OTP login code.
• An MCP host that supports stdio MCP servers, such as Codex, Claude Desktop, Cursor, VS Code, or another compatible desktop agent.

0.2.11 is the current public beta release. The default server is the PatchXNote public beta API. Credentials are stored in the OS-native keychain by default.

Login And MCP Modes

PatchXNote Agent now has two productized login surfaces and two MCP service shapes:

Mode	Entry	Use when	Notes
Browser MCP login	npx -y patchxnote-agent@latest mcp login	The user is installing PatchXNote into a desktop editor or local MCP host.	Opens a browser to the PatchXNote authorization page, completes phone OTP there, receives the loopback callback, exchanges the code, and stores credentials in the OS keychain.
Terminal CLI login	npx -y patchxnote-agent@latest login	The user prefers a terminal-only flow, is on a headless runtime, or needs the legacy local Agent fallback.	Keeps the phone OTP conversation inside the terminal. Do not paste OTP codes or tokens into an AI chat.
Local MCP service	npx -y patchxnote-agent@latest mcp serve	VS Code, Cursor, Codex, Claude Desktop, Windsurf, Trae, Qoder, WorkBuddy, and other local stdio MCP clients.	MCP config stays secret-free. mcp serve does not open a browser during editor startup.
Hosted remote MCP gateway	https://ws-lab.patch-x.cn/patchnote-test-api/mcp	Feishu Aily, Doubao Work Partner, Tencent Agent Development Platform, enterprise WorkBuddy, and other platform clients that cannot run local commands.	Server route is the platform-facing MCP shape; real platform-console acceptance is still tracked separately.

Quickstart

For compatible Agent Skills clients, install or refresh the SOP skill first:

npx -y patchxnote-agent@latest skill install

Then run setup for the client you use:

npx -y patchxnote-agent@latest setup --client vscode
npx -y patchxnote-agent@latest setup --client cursor
npx -y patchxnote-agent@latest setup --client codex
npx -y patchxnote-agent@latest setup --client workbuddy

Setup checks the MCP browser OAuth login, keeps credentials in the OS keychain, writes or prints the client MCP configuration, and creates a backup before modifying supported config files. This is the recommended path for local editor integrations.

You can also log in explicitly before adding the client:

npx -y patchxnote-agent@latest mcp login
npx -y patchxnote-agent@latest mcp status

You can still print the generic MCP config and paste it into any local stdio MCP-capable assistant or editor:

npx -y patchxnote-agent@latest mcp config

The generated config starts PatchXNote Agent through:

npx -y patchxnote-agent@latest mcp serve

On first start, the npm wrapper downloads the matching patchxnote binary from GitHub Releases, verifies checksums.txt, installs it into a user-writable directory, and then delegates to patchxnote mcp serve. MCP stdout stays reserved for JSON-RPC.

The terminal Agent login remains available for terminal-only users and legacy local CLI/MCP fallback:

npx -y patchxnote-agent@latest login

If an MCP host times out during the first binary download, run the stable fallback once and paste the printed absolute-path config:

npx -y patchxnote-agent@latest install --print-config

To pin the current published public beta version for troubleshooting or rollback:

npx -y patchxnote-agent@0.2.11 install --print-config

The public beta build defaults to the PatchXNote public beta API:

https://ws-lab.patch-x.cn/patchnote-test-api

To target a different PatchXNote environment:

PATCHXNOTE_SERVER_BASE_URL=<PatchXNote API base URL> \
patchxnote mcp login

Common Workflows

Ask your AI assistant:

Find today's mobile records.
Show the AI result behind this record.
Send this Markdown to my Product Feishu webhook.
Export the AI response to a local JSON file.
Create a Markdown draft from this record so I can edit it before sending.

Client Setup

Local setup supports these P0 client IDs:

vscode, cursor, codex, claude-code, claude-desktop, windsurf, trae, qoder, workbuddy

vscode, cursor, codex, claude-desktop, and windsurf can write a local config file after confirmation. claude-code, trae, qoder, and workbuddy return manual commands or copyable config in V1. Platform clients such as Feishu Aily, Doubao Work Partner, Tencent Agent Development Platform, and enterprise WorkBuddy require the hosted remote MCP gateway and platform-console acceptance instead of local npx.

Useful setup flags:

patchxnote setup --client cursor --dry-run --print-config
patchxnote setup --client cursor --yes
patchxnote setup --client cursor --no-browser
patchxnote setup --all-local-supported --dry-run
patchxnote setup --client cursor --output json

Run setup in the same OS/runtime that will later launch MCP. For example, a Windows desktop editor needs Windows Credential Manager credentials, while a WSL or remote VS Code session needs credentials in that Linux runtime.

MCP Configuration

For generic local stdio MCP hosts, use the pure JSON printed by:

npx -y patchxnote-agent@latest mcp config

The default config looks like this:

{
  "mcpServers": {
    "patchxnote": {
      "command": "npx",
      "args": ["-y", "patchxnote-agent@latest", "mcp", "serve"]
    }
  }
}

Some clients may require a wrapper-specific field such as type: "stdio" or a different top-level key, but the command and args stay the same. If a client rejects npx, kills slow cold starts, or requires allowlisted absolute paths, use the fallback printed by:

npx -y patchxnote-agent@latest install --print-config

The fallback config uses the installed binary path:

{
  "mcpServers": {
    "patchxnote": {
      "command": "/absolute/path/to/patchxnote",
      "args": ["mcp", "serve"]
    }
  }
}

MCP config never contains access tokens, refresh tokens, OTP codes, phone numbers, webhook secrets, or a base URL by default. PatchXNote Agent stores credential material in macOS Keychain, Windows Credential Manager, or Linux Secret Service when available. The explicit PATCHXNOTE_AUTH_INSECURE_FILE_KEYCHAIN=true file store remains for local development and CI smoke only.

MCP Tools

PatchXNote Agent 0.2.11 exposes the same 19 local MCP tools as the current public local server. End users can think of them as three groups; exact tool names are for MCP hosts and AI assistants.

Account And Record Lookup

Tool	Purpose
patchxnote_get_current_user	Show the current PatchXNote account status.
patchxnote_list_recorder_cards	List bound recorder cards with masked identifiers only.
patchxnote_get_quota_summary	Show current quota.
patchxnote_get_model_usage_summary	Show current-month AI usage and charged quota.
patchxnote_list_memories	List records for mobile or desktop.
patchxnote_search_memories	Search record basics cached in the current MCP session.
patchxnote_get_memory	Show safe basic information for one record.

Webhook Configuration And Sending

Tool	Purpose
patchxnote_list_webhook_targets	List local webhook aliases and masked metadata.
patchxnote_configure_webhook_target	Create or update a webhook alias; URL and secret inputs are write-only.
patchxnote_remove_webhook_target	Remove a webhook alias and best-effort clean up stored secrets.
patchxnote_list_webhook_templates	List built-in Markdown templates.
patchxnote_render_webhook_message	Render a record into Markdown and optionally save a local draft.
patchxnote_export_model_io	Export a complete AI processing record to a user-chosen local file.
patchxnote_send_webhook	Manually send Markdown, a draft, a rendered record, or a test message to target aliases.

AI Result Inspection

Tool	Purpose
patchxnote_list_model_io_traces	Find AI processing runs and the follow-up request_id.
patchxnote_get_model_io_source_text	Inspect or export the source text used for that run.
patchxnote_get_model_io_provider_response	Inspect or export the AI response.
patchxnote_get_model_io_parsed_result	Inspect or export the parsed AI result.
patchxnote_get_model_io_packaged_result	Inspect or export the final result.

Record tools require an explicit platform argument: mobile or desktop. The record list now includes formal saved results plus readable model-generated outputs when the server has model IO data. patchxnote model-io list remains the lower-level AI processing list for finding request IDs and filtering by task or state.

Webhook MCP tools share the same local config, keychain, templates, and sender modules as the CLI. They do not return full webhook URLs or signing secrets, and send calls perform external network requests only when the MCP client explicitly invokes the send tool.

AI result tools are explicit inspection tools. They may expose source text or AI payloads for the logged-in user, so use them only from trusted local MCP hosts. Large fields should be written to an explicit local out file.

CLI Commands

Browser MCP login and local MCP service:

patchxnote version
patchxnote mcp login
patchxnote mcp status
patchxnote mcp config
patchxnote setup --client cursor
patchxnote mcp serve
patchxnote mcp logout

Terminal CLI login:

patchxnote login
patchxnote auth status
patchxnote logout

List AI processing runs and export results:

patchxnote model-io list --platform mobile
patchxnote model-io source-text --request-id <request_id> --platform mobile --out ./source.txt
patchxnote model-io provider-response --request-id <request_id> --platform mobile --out ./provider-response.json
patchxnote model-io parsed-result --request-id <request_id> --platform mobile --out ./parsed-result.json
patchxnote model-io packaged-result --request-id <request_id> --platform mobile --out ./packaged-result.json
patchxnote model-io export --request-id <request_id> --platform mobile --out ./model-io.json

Get request_id from patchxnote model-io list --platform mobile|desktop when you need a lower-level AI processing run. MCP patchxnote_list_memories returns id and platform for record rendering, drafts, webhook workflows, and model IO field tools; for model-generated entries, that id can be the same value as request_id.

Configure and send webhooks:

patchxnote webhook set "Product Feishu" --type feishu --url-stdin
patchxnote webhook list
patchxnote webhook test "Product Feishu"
patchxnote webhook draft --memory-id <memory_id> --platform mobile --out ./patchxnote-drafts/example
patchxnote webhook send --target "Product Feishu" --file ./message.md
patchxnote webhook send --target "Product Feishu" --draft ./patchxnote-drafts/example
patchxnote webhook remove "Product Feishu"

Useful global flags:

--server-base-url <url>   PatchXNote API base URL
--profile <name>          local profile name
--output json             machine-readable output where supported
--config <path>           non-secret config file path

The npm package is a small installer/launcher wrapper:

npx -y patchxnote-agent@latest mcp login
npx -y patchxnote-agent@latest mcp status
npx -y patchxnote-agent@latest mcp config
npx -y patchxnote-agent@latest mcp serve
npx -y patchxnote-agent@latest mcp logout --local-only
npx -y patchxnote-agent@latest skill install
npx -y patchxnote-agent@latest login
npx -y patchxnote-agent@latest setup --client cursor
npx -y patchxnote-agent@latest install
npx -y patchxnote-agent@latest update
npx -y patchxnote-agent@latest uninstall

Webhook URLs and optional Feishu/DingTalk signing secrets are stored in the local secure credential store, not in the non-secret config file. --url-stdin and --secret-stdin avoid shell history. CLI and MCP webhook sending is manual only, does not follow redirects, and surfaces provider errors directly.

patchxnote model-io export is the preferred complete AI processing export command. patchxnote webhook export-model-io remains available for compatibility.

Security And Risk Notice

PatchXNote Agent gives an AI assistant access to account and record information for the logged-in PatchXNote user. Treat the MCP host as trusted software and review prompts, tool calls, local files, and logs that may contain account context, source text, or AI results.

Default safety boundaries:

• Agent auth is separate from App/PC mobile and desktop installations.
• Agent calls only dedicated read-only /v1/agent/** server routes for PatchXNote server data.
• MCP webhook tools can write local non-secret target metadata, write URL/secret material to local secure storage, and manually send external webhook HTTP requests.
• Webhook sending never happens in the background and is not scheduled automatically.
• MCP runs locally over stdio; stdout is reserved for JSON-RPC.
• MCP config does not store bearer tokens, refresh tokens, OTPs, SK, or full MAC values.
• Recorder-card identifiers are masked; live BLE state, battery, storage, and recording status are not exposed.
• Content is platform-scoped. The Agent does not merge mobile and desktop content.
• Tool outputs are bounded and validated before being returned to the MCP client.
• Webhook target URLs and signing secrets stay in local secure storage; normal webhook payloads never include access tokens, refresh tokens, or exported AI processing JSON.
• AI result tools return only the requested field. They do not replay model calls and do not include unrelated fields in single-field responses.
• Agent does not read raw audio and does not provide audio downloads. Source text and AI results are available only through explicit tool calls, and large/sensitive content should be exported to local files.

Do not paste access tokens, refresh tokens, OTP codes, raw phone numbers, full MAC values, SK values, raw audio, source text, prompts, or provider payloads into public issues. Use the private process in SECURITY.md for vulnerability reports.

Current Limitations

0.2.11 is the current public beta release.

• The default server points to the PatchXNote public beta API and does not imply a production SLA.
• mcp serve never opens a browser during editor startup. Run mcp login first, or let setup --client <id> reuse that same OAuth flow.
• The GoServer website and authorization page own phone OTP input. The local Agent owns loopback callback, token exchange, secure storage, and stdio bridging.
• Remote platform MCP for Feishu Aily, Doubao Work Partner, Tencent Agent Development Platform, and enterprise WorkBuddy requires hosted platform-console acceptance; local npx setup only closes desktop/terminal clients.
• Linux headless environments may not have Secret Service available; use the explicit development file-store fallback only for local smoke.
• Public beta users should expect iterative improvements to setup guidance, MCP client examples, and webhook formatting.
• patchxnote_search_memories searches only record basics cached during the current MCP session.
• If the record list is empty, first check whether you selected mobile or desktop. Lower-level AI processing runs can also be listed separately with patchxnote model-io list.
• Raw audio, audio downloads, hardware write actions, model execution/replay, automatic webhook pushes, quota purchase/reward actions, payment, and Admin APIs are out of scope.

Troubleshooting

Problem	What to check
patchxnote is not found after install	Add the printed install directory to PATH, then open a new terminal.
Login says credential storage is unavailable	Check that macOS Keychain, Windows Credential Manager, or Linux Secret Service is available and unlocked. For local development only, set PATCHXNOTE_AUTH_INSECURE_FILE_KEYCHAIN=true.
MCP login expired or points at the wrong server	Run npx -y patchxnote-agent@latest mcp logout --local-only, then run npx -y patchxnote-agent@latest mcp login again in the same runtime.
MCP host cannot start the server	If first start is slow or the client rejects npx, run npx -y patchxnote-agent@latest install --print-config once and use the printed absolute command path.
Setup writes credentials in the wrong place	Run setup from the same OS/runtime that will launch MCP. Windows desktop apps, WSL terminals, and VS Code Remote do not automatically share keychain credentials.
Need to undo setup	Restore the timestamped .bak-YYYYMMDDTHHMMSSZ file printed by setup, or remove only the patchxnote MCP server entry from the client config.
Record list is empty	Check that you selected the correct platform: mobile or desktop; use model-io list for lower-level AI processing runs.
Webhook did not send	Confirm the alias exists, the target is enabled, and check the provider error returned by the command.
Checksum verification fails	Retry later or pin a known version; the installer refuses unchecked binaries.
skill install says the target already exists and differs	The target contains an unmanaged or manually edited patchxnote-mcp skill. Inspect or back it up first; rerun with --force only when you want PatchXNote Agent to replace that skill folder.
Wrong server environment	Set PATCHXNOTE_SERVER_BASE_URL=<PatchXNote API base URL>.

Verify The Install

npm view patchxnote-agent@latest version --registry https://registry.npmjs.org
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest skill install --dry-run --json
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest mcp config
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest mcp status --output json
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest mcp logout --local-only --output json
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest setup --client cursor --dry-run --print-config
npx -y --registry https://registry.npmjs.org patchxnote-agent@latest install --dry-run --print-config
patchxnote version

The release binary should report the npm package version and the commit attached to the matching GitHub Release tag.

0.2.11 Highlights

• Bundles the canonical PatchXNote MCP Skill inside the npm package.
• Adds npx -y patchxnote-agent@latest skill install for npm-based skill installation without relying on a separate skills CLI or GitHub clone.
• Makes skill installation idempotent, records a managed marker, and protects manually edited skill folders unless --force is explicitly used.
• Extends skill package sync and validation so OpenAI, Claude, and npm copies stay byte-identical to skills/patchxnote-mcp/.
• Updates the one-line setup prompt and discovery metadata to prefer the npm-bundled skill installer while preserving MCP setup, browser OAuth, and tool verification.

0.2.10 Highlights

• Adds the reusable PatchXNote MCP Skill at skills/patchxnote-mcp/ so compatible AI clients can keep the setup and usage SOP across fresh or long sessions.
• Adds OpenAI/Codex, Claude Code, Agent Skills, MCP Registry, Smithery, and third-party directory draft packaging and listing materials.
• Adds MCP Registry metadata through server.json and package.json#mcpName, plus local validation and stdio smoke scripts for release evidence.
• Updates the one-line setup prompt so agents install the skill when supported, then run MCP setup, browser OAuth, and tool verification without asking users to paste codes or tokens into chat.

0.2.9 Highlights

• Polishes the browser OAuth loopback success and failure pages shown after patchxnote mcp login.
• Keeps the post-login result page focused on plain user guidance, without exposing OAuth codes, state values, or token-shaped details.
• Adds regression coverage for the browser callback pages so future changes keep those sensitive details out of the UI.

0.2.8 Highlights

• Adds patchxnote setup --client <id> and npm wrapper delegation with dry-run, JSON output, confirmation, config printing, force repair, and local MCP smoke hooks.
• Adds a client registry for VS Code, Cursor, Codex, Claude Code, Claude Desktop, Windsurf, Trae, Qoder, WorkBuddy, Feishu/Doubao, Tencent platform, and P1 follow-up clients.
• Adds JSON and TOML config merge adapters with backup, conflict detection, rollback, and manual JSONC mode.
• Adds patchxnote mcp login/status/logout, browser OAuth with PKCE, MCP OAuth secure storage, and remote /mcp stdio proxy mode with local fallback.
• Adds website page specs, detail-page copy, and remote platform gateway design for product-style onboarding.

0.2.6 Highlights

• MCP expands to 19 tools across account/record lookup, webhook delivery, and AI result inspection.
• Webhook workflows are available to MCP: configure named aliases and manually send to Feishu, DingTalk, or generic webhooks.
• AI processing runs can be listed first, then inspected by request_id for source text, AI response, parsed result, and final result.
• Record lists can now include readable model-generated outputs returned by the server, so users can find a record first and then inspect its source text, AI response, parsed result, or final result.
• Webhook aliases containing dots, Chinese text, or spaces now persist and reload correctly from the local config file.
• README, npm README, and public visual assets have been refreshed for the new user-facing positioning.

Development

Local checks:

go test ./...
scripts/e2e/mvp-smoke.sh
node packages/npm/test/install.test.js
node docs/mcp-clients/validate-clients.mjs

The MVP smoke builds the CLI, runs installer dry-run, checks the npm universal MCP launcher plus mcp login/status/logout non-interactive paths, logs in against an in-process Agent V1 test server, checks auth status, starts patchxnote mcp serve, calls all 19 V1 MCP tools, exercises AI processing discovery, field export, and local webhook delivery, logs out, and scans evidence for secret-like values.

Before changing CLI behavior, installer logic, MCP tools, authentication, local cache, or release configuration, read:

• AGENTS.md
• docs/engineering-rules.md
• docs/release-and-maintenance-runbook.zh-CN.md
• docs/plans/2026-08-06-agent-v1-mvp.md

Release Notes For Operators

The detailed release and documentation maintenance checklist lives in docs/release-and-maintenance-runbook.zh-CN.md.

1. Confirm the target PatchXNote GoServer exposes the required /v1/agent/** routes.
2. Confirm packages/npm/package.json version matches the release tag without the leading v.
3. Push a clean tag, for example v0.2.11.
4. Wait for GitHub Release assets: checksums.txt plus Linux/macOS/Windows amd64 and arm64 binaries.
5. Configure npm Trusted Publishing for this GitHub Actions workflow before npm publish:
   • owner/user: ZsTs119
   • repository: patchxnote-agent
   • workflow filename: publish-npm.yml
   • allowed action: npm publish
6. Publish npm only after release assets exist and the trusted publisher is configured.
7. After a successful trusted publish, revoke the old npm automation token and disallow token-based publishing for this package.

License

This repository is currently published without an open-source license. Contact PatchXNote before redistributing or embedding it in another product.