zeroclaw-labs/zeroclaw · error · anyhow::Error

temperature unsupported by Grok CLI model provider

Error message

temperature unsupported by Grok CLI model provider: {temperature}. Supported values: 0.7 or 1.0 (not forwarded; CLI has no sampling flag)

What it means

The Grok CLI exposes no sampling flag, so the provider accepts only 0.7 and 1.0 (compared with TEMP_EPSILON) as validated no-ops and rejects every other finite value so silent divergence cannot go unnoticed. The message records the offending value. Validation runs per request, before the CLI is spawned.

Solutions

Set temperature to 0.7 or 1.0, or omit it (None falls back to the provider default)
Remove per-agent temperature overrides when the agent moves to grok_cli
If you need real sampling control, use an API-based provider (e.g. xAI or an OpenAI-compatible endpoint) instead of the local CLI wrapper

Example fix

# before
[agents.coder]
temperature = 0.3

# after
[agents.coder]
temperature = 0.7

Defensive patterns

Strategy: validation

Validate before calling

fn grok_temperature_ok(t: Option<f64>) -> bool {
    t.map(|v| (v - 0.7).abs() < 1e-9 || (v - 1.0).abs() < 1e-9).unwrap_or(true)
}

Type guard

fn is_supported_grok_temperature(t: f64) -> bool {
    (t - 0.7).abs() < 1e-9 || (t - 1.0).abs() < 1e-9
}

Try / catch

match provider.chat(req, model, temp).await {
    Err(e) if e.to_string().contains("temperature unsupported by Grok CLI") =>
        provider.chat(req, model, Some(0.7)).await, // fall back to supported value
    other => other,
}

Prevention

Do not carry temperature overrides across provider switches
Keep grok_cli agents at the default or pin 0.7/1.0
Remember the value is never forwarded — tuning it against the CLI is pointless

When it happens

Trigger: chat/chat_with_system on a grok_cli provider with temperature = 0.0, 0.2, 1.5, or any value outside epsilon of {0.7, 1.0}; an agent profile tuned for an OpenAI-style provider (arbitrary floats accepted) reused with model_provider = grok_cli.

Common situations: Switching an existing agent to grok_cli without revisiting its temperature override; shared per-agent config applied across heterogeneous providers; temperature-sweep scripts run against all providers.

Related errors

AI-assisted analysis of zeroclaw-labs/zeroclaw@88bb9c8533 (2026-08-23). Data as JSON: /api/errors/df39cd6724748611. Report an issue: GitHub.

Appendix: source

Thrown at crates/zeroclaw-providers/src/grok_cli.rs:655

    }

    fn should_forward_model(model: &str) -> bool {
        let trimmed = model.trim();
        !trimmed.is_empty() && trimmed != DEFAULT_MODEL_MARKER
    }

    fn supports_temperature(temperature: f64) -> bool {
        GROK_CLI_SUPPORTED_TEMPERATURES
            .iter()
            .any(|value| (temperature - value).abs() < TEMP_EPSILON)
    }

    fn validate_temperature(temperature: f64) -> anyhow::Result<()> {
        if !temperature.is_finite() {
            anyhow::bail!("Grok CLI model provider received a non-finite temperature value");
        }
        if !Self::supports_temperature(temperature) {
            anyhow::bail!(
                "temperature unsupported by Grok CLI model provider: {temperature}. \
                 Supported values: 0.7 or 1.0 (not forwarded; CLI has no sampling flag)"
            );
        }
        Ok(())
    }

    /// Build the documented ACP invocation. Permission and tool overrides are
    /// accepted only through explicit per-alias `extra_args`.
    fn build_cli_args(model: &str, extra_args: &[String]) -> Vec<String> {
        let mut args = Vec::with_capacity(16 + extra_args.len());
        args.push("--no-auto-update".to_string());
        args.push("--no-plan".to_string());

        if !Self::extra_args_set_any(extra_args, &["--sandbox"]) {
            args.push("--sandbox".to_string());
            args.push(DEFAULT_SANDBOX_PROFILE.to_string());
        }

View on GitHub (pinned to 88bb9c8533)