AI configuration and usage

Configure a provider

Open Cointacted → AI → Settings. Enable AI, choose a preset, verify the endpoint and model, and configure access as public, signed-in, or administrator-only. Demo mode is local and deterministic; use it first to verify commands and limits without an external request.

Production credentials should be defined in wp-config.php with COINTACTED_AI_PROVIDER, COINTACTED_AI_ENDPOINT_URL, COINTACTED_AI_API_KEY, and COINTACTED_AI_MODEL. Constants take priority over saved fields and keep the API key out of the database. Supported adapters include OpenAI-compatible chat completions, Google Gemini, and Anthropic Messages. Presets are editable defaults, not availability guarantees from those providers.

Use Testing to run the primary or fallback connection check. ai-setup performs a dry run without calling the provider. Do not enable a fallback until its endpoint, key, and model pass independently.

Production activation checklist

1. Keep access on Administrator only and use Demo mode to confirm ai-help, ai-status, ai-setup, and ai-reset in the published Terminal.
2. Configure one real provider under Cointacted → AI → Settings. Save, open Testing, and run the primary connection test.
3. Set prompt length, output-token limit, cooldown, daily site limit, and per-user limit before changing access.
4. Enable command awareness only after the commands visible to a low-privilege test user have been reviewed. The model receives a sanitized command inventory, not permission to bypass command handlers.
5. Test one administrator, one ordinary signed-in user, and one signed-out visitor. Confirm both allowed and denied states match the selected access policy.
6. Open Diagnostics and confirm the successful request appears without prompt text, response text, or a provider credential.

If another official add-on registers AI tools, its controls and permissions remain in that add-on. Cointacted AI discovers the reviewed tool contract; it does not copy the add-on's business rules into the model prompt.

Commands

• ai QUESTION or ask QUESTION sends a question through the active provider.
• ai-help shows practical prompts and privacy guidance.
• ai-reset clears the current actor's short conversation context.
• ai-status shows administrator-safe readiness, limits, usage, and recent errors.
• ai-setup checks the saved provider shape without spending provider credit.
• ai-mode on|off|reset controls the browser AI mode where available.

Context, limits, and recommendations

Short context is optional and disabled by default. When enabled, a bounded number of recent turns is stored in expiring WordPress transients under a privacy-preserving actor hash. This is not durable chat history or long-term memory.

Set prompt length, output tokens, cooldown, daily site usage, and per-user usage before public access. Pro increases those ceilings and can retry a configured fallback. Command awareness supplies the model only with a sanitized list of commands the current user may access. Lite recommendations remain insert-and-review. Pro may run a reviewed read-only suggestion only after the server rechecks capability, access, allowlist, risk class, and a one-time browser approval where required.

Diagnostics and privacy

The Diagnostics tab shows aggregate requests/tokens/actors, sanitized recent failures, and safe activity records without prompt or response text. Downloaded support and migration bundles omit API keys. By default the plugin does not store prompts; the external provider still receives the prompt, system instruction, and bounded context required for a request.

Troubleshooting

• No provider: enable AI and complete endpoint, model, and credential configuration.
• 401/403: verify the API key, account access, endpoint protocol, and selected model.
• 429/quota: reduce limits or wait for the provider window; fallback is used only for retryable failures.
• Unsafe endpoint: use HTTPS and a publicly routable approved endpoint. Local/private targets are rejected by default.
• Command unavailable: ensure Social Terminal is active and the current terminal user matches the configured access mode.
• Recommendation will not run: confirm AI Pro, guarded runs, optional allowlist, read-only classification, and current-user command access.