nextclaw-self-manage

NextClaw Self-Management

Use this skill whenever the user asks to manage NextClaw itself (version, service status, diagnostics, channels, config, agents, projects, sessions, apps and App data, cron, remote, update, installed skills, marketplace skills, local HTTP/API/webhook addresses), report a NextClaw problem, or follow up on a reported problem.

Source of Truth

Always use the built-in NextClaw self-management guide as the operation guide.

1. Read the built-in guide path provided by the system prompt first.
2. If the packaged guide is unavailable in a repo source checkout, use repo docs path docs/USAGE.md.
3. Never treat workspace USAGE.md snapshots or copied artifacts as the source of truth.
4. If both guide paths are unavailable, use nextclaw --help and nextclaw <subcommand> --help as fallback and tell the user that guide file is missing.

Routing Rules

• Treat NextClaw self-management as a product-management intent, not a generic "create/install/publish" intent.
• Read the built-in self-management guide before opening unrelated generic skills.
• Example: "create a new Agent" maps to NextClaw agent management, not skill-creator.

Stable Execution Rules

• Map version lookup directly to nextclaw --version; do not substitute status for version queries.
• Treat nextclaw update --channel beta as opting into both preview and production candidates; the updater offers whichever compatible version is newer. The stable channel remains production-only.
• Prefer machine-readable output: use --json when available.
• Before calling local HTTP APIs or /webhook, run nextclaw status --json and read endpoints.uiUrl / endpoints.apiUrl; do not guess the service port.
• For webhook payload details, read the focused guide linked from the self-management guide only when you need to implement or debug a webhook caller.
• Execute only commands documented in the self-management guide or CLI help; do not invent commands or config paths.
• When the user asks to report a NextClaw problem, use the guide's private feedback commands yourself, without sending the user to a form. Submit only relevant, redacted reproduction details; never require GitHub login. Keep receipt keys private, reuse saved request IDs after uncertain submission, and query the original report for replies. A released fix does not authorize upgrading the user's environment.
• When the user asks about their reports, use feedback list/get/reply yourself and explain actual maintainer replies. Do not create an AI polling job by default. Distinguish ready from published; report query failures honestly and never infer resolution from silence. Suggest submitting a report when a relevant bug is observed, but obtain the user's intent to submit before sending diagnostic details.
• Prefer object-level CLI commands for configuration management. For providers/models/search, use nextclaw providers ..., nextclaw models ..., and nextclaw search ...; do not read/edit config files or substitute generic config set/unset or gateway config actions for these covered tasks.
• Generic config commands and gateway config tools remain transitional options only for documented capabilities not covered by object-level CLI or explicit manual recovery. A missing running host is not a reason to silently edit files instead.
• Read API keys from a named environment variable with --api-key-env; never print the variable or copy credentials into a command argument. Query saved settings with the relevant show command, and verify provider connectivity with providers test. An API apply failure is not success even if values were saved.
• Provider model discovery does not save models. providers models set replaces the full list; providers models configure replaces all capability overrides. Query existing settings first and include entries the user intends to keep.
• For provider authorization, follow the URI/code from providers auth start, poll the returned session id at the advertised interval, and treat only authorized as completion. providers auth import is available for supported external CLI credentials.
• In desktop-installed runtimes, still use the same nextclaw ... commands. The desktop launcher provides a managed command surface to AI command tools; do not ask the user to install the NPM CLI just to run self-management commands.
• Keep installed skills and marketplace catalog as two different domains:
  • local installed: nextclaw skills installed|info
  • marketplace catalog and installed marketplace lifecycle: nextclaw marketplace skills search|info|recommend|install|update
• After mutating operations, close the loop with:
  • nextclaw status --json
  • and nextclaw doctor --json when needed
• Be explicit about restart semantics after changes. When a restart is required, ask the user to run nextclaw restart in an external terminal; do not invoke a restart from the active agent session.
• nextclaw gateway starts a foreground gateway. It has no start, status, restart, or stop subcommands.
• After modifying a running Service App, run nextclaw app restart <app-id> --json before validating through the live product UI or panel-to-service action calls.
• Treat App code and managed App data as separate lifecycles. Uninstall/removal keeps managed data by default; delete it only when the user explicitly chooses the destructive data-removal option.
• For App data discovery, run nextclaw app data list --json against the running host. Treat the returned id, appId, lifecycle, usage breakdown, and instance path as authoritative.
• Independently delete only an entry whose latest list result says lifecycle: retained and actions.deleteRetainedData: true. Use nextclaw app data delete <data-id> --confirm <app-id> --json, where the data id comes from that list result and the confirmation exactly matches its appId.
• Never synthesize App data ids, delete an active entry through the retained-data command, or substitute direct recursive filesystem deletion for the managed API. After deletion, run nextclaw app data list --json again and verify the exact entry is absent.
• Reset a Service App development instance only when explicitly requested, using nextclaw app dev <service-app-dir> --reset-data --confirm <app-id> --json; the confirmation must match the manifest id.
• App data removal never authorizes deletion of documents or directories separately granted to the App outside its managed instance.
• For channel discovery before messaging, use nextclaw channels list --json and treat returned channels[].id values as authoritative.
• For cron notifications, do not add delivery flags to the cron command. Put the notification intent in the scheduled message and let the scheduled agent call the message tool with an explicit channel and recipient.
• For Agent creation/update/removal, treat nextclaw agents list|new|update|remove --json as the default path and follow the Agent management section in the self-management guide.
• For runtime context injection, use nextclaw agents runtime config <runtime-id> --json to inspect and --inject-nextclaw-context <true|false> to update. Run nextclaw restart in an external terminal after changing it.
• For project creation, discovery, or removal, use nextclaw projects list|templates|create|remove --json; project removal also requires --confirm <project-id> and preserves the local folder, sessions, and Project Work. Do not synthesize placeholder sessions or edit the project registry file directly.
• For session naming, project binding, or deletion, use nextclaw sessions rename|set-project|clear-project|delete --json; deletion also requires --confirm <session-id>. Do not edit session journal metadata directly.
• Do not edit config.json or agents.list directly for normal Agent CRUD; only do that when the user explicitly wants a manual recovery path.
• When creating an Agent, prefer an explicit non-text avatar and avoid text-based styles such as DiceBear initials.