AI assistant
Optional. Nothing on this screen is required for Walis to build anything.
Providers
| Provider | Transport | Default model |
|---|---|---|
| Claude | Official Anthropic SDK, adaptive thinking | claude-opus-5 |
| Codex | OpenAI REST endpoint | gpt-5-codex |
| Gemini | Google REST endpoint | gemini-2.5-pro |
Each has an API key, a model override, an optional custom endpoint, an enable switch and a Test button. One provider can be marked default; features that need a model use it unless told otherwise.
Keys are stored in the database and masked in API responses. Sending the masked value back means "unchanged", so you can edit a provider without re-entering its key.
What it is used for
- Check script — reviews a job's script and reports what looks wrong. It does not edit the job.
- Analyse build — summarises a failing build's output in the log window.
- Draft a plugin — generates a plugin from a description. Output has permissions cleared, type and version forced, and is never installed automatically. See Plugins.
The call log
Every call is recorded: time, provider, model, purpose, job, operator, duration, token usage and result — including the provider's error body when one fails, which is usually enough to tell a wrong key from a retired model from a rate limit.
Prompt content is never stored. The log holds statistics, not what was sent.
Cost and failure
Calls are made against your own keys, so usage lands on your bill. Every feature here degrades cleanly: if a provider is unreachable or the key is wrong, the build itself is unaffected — only the assistance is missing.