helpme

CapSift /helpme

Render help from the static message catalogs in references/locales/. Do not ask a model to translate catalog messages at runtime.

Before claiming an action is available, resolve its exact command in the plugin-root menu-map.json. An absent command uses the explicit * unavailable mapping. MCP entries call the bundled server, CLI entries require an installed full CapSift CLI, menu and navigation entries are static, and unavailable entries must display their mapped reason.

Language selection

Resolve the locale once in this order:

1. Locale explicitly supplied by the current command.
2. Current-task preference.
3. Project preference in .capabilityhub/config.json.
4. Global CapSift preference for the current operating system.
5. Codex or operating-system locale when the preference is auto.
6. en as the final fallback.

Supported locales are zh-CN and en. Normalize zh, zh-Hans, and Chinese locale variants to zh-CN; normalize English variants to en. Never translate capability names, commands, paths, error codes, or credential identifiers.

For /helpme language set <locale>, use task scope unless another scope was explicitly selected. A project setting stores only the locale key in .capabilityhub/config.json. A global setting stores only the locale key in the platform's CapSift config directory. Never overwrite unrelated configuration keys. preview never persists. When local CLI execution is available, read persistent settings with capsift language show --pretty and save project/global choices with capsift language set <locale> --scope <project|global> --pretty. Keep task scope in the current interaction only. Report a failed CLI write instead of claiming success.

Progressive routing

• Bare /helpme: read only the resolved catalog's root object and render every group and item in order. Professional terms remain visible; their parenthesized descriptions explain them in plain language.
• Keep /helpme language visible on the root menu. It opens language settings directly.
• /helpme skills and selection 1: hand off to the myskills skill without loading the Skill inventory.
• Root shortcuts inventory, search, loaded, providers, routing, lifecycle, mcp, status, dashboard, budget, security, and settings go directly to that function. Numeric selections map to the visible root order only while this menu is the active interaction.
• For /helpme inventory, call capability.search with an empty query, include_inventory: true, include_cards: false, limit: 1, and a bounded output budget. Render inventory.active_total, all five values in inventory.active_by_kind, inventory.generation, and inventory.status; do not infer totals from cards or query-level kind_counts. If status is partial, show only non-zero excluded_by_reason counts. If it is stale, show last_refresh_error_code and say the last complete snapshot is being used.
• For /helpme search <task>, call capability.search for compact cards only. Show at most five results unless the user requests more, and do not load any result body.
• For /helpme loaded, use capsift loaded --limit 20 --pretty. Explain that this is a bounded list of recent successful loads reconstructed from redacted project audit, not a promise that full bodies remain resident in the chat context.
• For /helpme providers, use capsift providers --pretty. Show the real Provider name, discovered count, active count, and kinds. Do not replace Provider names with capability kinds.
• For /helpme routing <task>, use capsift routing <task> --limit 5 --pretty. Show rank and match_reason; state that the current method is deterministic lexical routing with zero model calls and does not expose hidden model reasoning.
• For /helpme status, report Inventory freshness, generation, and safe counts. Say clearly that configured or discovered MCP/API/RAG entries are not proof of a live network connection.
• For /helpme dashboard, give the local capsift dashboard command and explain that it shows live Inventory, local wiring, no-model compact search, project language, Providers, recent Loaded entries, Routing reasons, and enable/disable/quarantine activation controls, exact-intent approval decisions, resident Context metadata, and advisory Reasoning state. Explain that these controls do not delete or execute files. Do not claim that it opens a browser automatically or displays alerts.
• For /helpme runtime health, use capsift health --pretty when local CLI execution is available. This check must not scan the capability catalog; it checks only the project path, bundled Dashboard files, MCP SDK presence, local Codex config parsing, and package version.
• For /helpme runtime connections, use capsift connections --pretty when available. Explain configured_not_probed as "configured, but no connection test was made" and never turn a configured count into a reachable or healthy claim. Only when the user explicitly asks to test connections, use capsift connections --probe --pretty. Explain that this performs bounded DNS/TCP/TLS setup only: reachable is transport reachability, tls_verified is transport security, and application authentication and health remain unknown. It never invokes an MCP tool or API operation.
• If the user explicitly asks for the local API, provide capsift http-serve --project-root <absolute path>. Explain that it is numeric-loopback only, prints a one-process bearer token to the launching terminal, uses an immutable startup snapshot, and is not a remote or multi-user service. Never repeat that token into chat.
• For /helpme lifecycle, use capsift lifecycle list --pretty. A user may set a discovered coordinate to enabled, disabled, or quarantined through capsift lifecycle set <coordinate> <state> --scope <project|global> --pretty. This changes catalog activation only; it does not uninstall, delete, update, or run the capability. Require an exact coordinate and confirmation for quarantine.
• For staged update details, use capsift updates list --pretty. Explain the explicit stage → health-pass/health-fail → activate flow and retained rollback pointer. Never report a candidate healthy unless an operator supplied that result. Every one of those three transitions requires explicit local artifact bytes, publisher, registry, and a configured verifier; the default fails closed. Development mode permits unsigned local bytes only when explicitly selected and is not public-key trust.
• For /helpme security audit, use capsift audit --limit 50 --pretty. Explain that task identifiers are hashed and arguments, credentials, and provider output are not included. Do not claim tamper evidence, cross-machine history, or complete search history when the local MCP runtime was not connected.
• If CAPABILITYHUB_AUDIT_KEY is configured, offer capsift secure-audit verify --pretty; rotation and export are explicit operator actions. Never display or request the signing key value in chat.
• For /helpme security approvals, use capsift approvals list --status pending --pretty. Create a request with capsift approvals request <revision> <operation> --arguments <json> --pretty, then approve or deny its returned ID explicitly. Execute with --approval-id <id>. Never display stored argument digests, and never use the fixture-only --approved shortcut for a configured Provider.
• For /helpme consumption context, use capsift context list --pretty. Explain that this is metadata for sections already disclosed by load, not a copy of their bodies. pin, unpin, and remove take an exact returned key; remove only forgets residency metadata and never deletes the source capability.
• For /helpme consumption reasoning, use capsift reasoning state <task-id> --pretty. A new recommendation uses reasoning recommend; low is preferred unless risk/policy requires more or the caller explicitly records a failed attempt with new evidence. Explain that advice does not itself call a model or spend tokens.
• For /helpme consumption budget, use capsift budget-report --pretty. The report is restart-safe local accounting for bytes, portable tokens, reasoning tokens, loads, and executions; it is not an API-provider billing statement.
• For manifest portability, validate accepts bounded JSON or safe YAML, export-manifest emits canonical JSON, migrate-manifest previews a legacy migration, and compatibility checks the v1alpha1 feature handshake. Use activation-lock export to capture exact active revisions and activation-lock verify <file> to reject missing, extra, or drifted state. Unknown required security semantics must fail closed.
• /helpme <topic>: read only that topic from the resolved catalog.
• /helpme language: render the catalog's language menu.
• /helpme back: return to the parent of the last menu rendered in this interaction; when no parent is known, render the root menu.
• /helpme home: always render the root menu. Also recognize explicit localized intent such as 返回上一级, 返回主菜单, back, and home while a CapSift menu is active.
• After every non-root menu or result, append the resolved catalog's navigation items exactly once. Do not append them to the root menu.
• Unknown topic: render the localized unknown-topic message and the root menu once.
• Keep every displayed item in the form command (what it does) or command （作用说明）.
• Never claim that a planned management action is live. When a selected action is not connected to the current runtime, label it as unavailable and explain the current read-only or pre-alpha boundary in one sentence.
• Do not discover or preload capability definitions, Skill bodies, provider schemas, credentials, or live runtime state merely to render a menu.
• Fetch live state only after the user selects an item that requires it.
• Keep each response under 220 Chinese characters or 120 English words before commands and live values.