Three request kinds route to providers independently, configured in ~/.cues/OPENCUES.md. Failures are written into the line, with their cause; that is their only surface.
| Scalar | Covers |
|---|---|
| cues-llm-provider | word and sentence suggestions |
| auditors-llm-provider | background prose rewrites, including agent tasks |
| blanks-llm-provider | everything requested with _ |
Each has a matching -model scalar. Splitting them lets the constant background work run on a fast, cheap model while requests you make directly go to a stronger one. A single global llm-provider covers any bucket left unset.
One line per bucket: the provider and model that will actually serve it, and, in the marker, which setting decided that. Precedence is bucket scalar over global scalar over automatic selection from available keys.
The failure is written into the line.
Whether no key is set at all or only the selected provider's key is missing, no model source can serve the _, and the failure is substituted into the line, selected, so cycling back restores the bare blank. The failure is also logged with the provider's own message (provider error: Invalid API Key for a rejected key). A key that is set but rejected behaves like the classified failures below.
Classified from the response; each class writes its own message into the line.
A classified, fixable failure is substituted into the line at the _ as a bracketed message naming the fix. The classes: rejected key (401/403), endpoint 404, a model the provider does not serve, out of credits, rate limit (429), network failure, bad request (400). Two of the messages, as shipped:
The message stays until you edit or cycle it away. Unclassified failures (a provider's own 5xx, malformed output) stay silent and retry on the next change, so a fixable failure names itself and a transient one never litters your text. A local Ollama gets its own messages: a missing model names the ollama pull command, an unreachable server names ollama serve.
max-thinking: on | off, default on.
on lets reasoning models use their configured effort ceiling. off reduces effort: answers are faster and cheaper, some are worse. The setting only lowers effort below the ceiling, never raises it, and models without a reasoning-effort parameter ignore it.
Provider value claude-code-cli.
With the claude CLI installed and signed in, a bucket set to claude-code-cli routes through an existing Claude Pro or Max subscription instead of an ANTHROPIC_API_KEY. Because routing is per bucket, word suggestions can stay on a faster provider while rewrites use the subscription.