aaif-goose/goose · error

Failed to create token counter: {error}

Error message

Failed to create token counter: {error}

What it means

count_context_tokens estimates the token count of agent-visible messages (used to report retained context around compaction) by first building a token counter with create_token_counter. Any failure in counter construction is re-wrapped with this message, so the {error} suffix carries the real cause (e.g. tokenizer/model initialization failure).

Source

Thrown at crates/goose/src/context_mgmt/mod.rs:214

        Err(error) => {
            warn!("Failed to count retained context tokens, using billable output tokens: {error}");
            summarization_usage.usage.output_tokens.unwrap_or(0)
        }
    };

    Ok(CompactionResult {
        conversation,
        usage: summarization_usage,
        retained_context_tokens,
    })
}

/// Estimate the tokens of the agent-visible conversation, counted the same way
/// as the fallback estimation in `check_if_compaction_needed`.
pub(crate) async fn count_context_tokens(conversation: &Conversation) -> Result<i32> {
    let counter = create_token_counter()
        .await
        .map_err(|error| anyhow::anyhow!("Failed to create token counter: {error}"))?;
    let total: usize = conversation
        .messages()
        .iter()
        .filter(|message| message.is_agent_visible())
        .map(|message| counter.count_chat_tokens("", std::slice::from_ref(message), &[]))
        .sum();
    Ok(total.try_into()?)
}

/// Check if messages exceed the auto-compaction threshold
pub async fn check_if_compaction_needed(
    provider: &dyn Provider,
    conversation: &Conversation,
    threshold_override: Option<f64>,
    session: &crate::session::Session,
) -> Result<bool> {
    if provider.manages_own_context() {
        return Ok(false);

View on GitHub (pinned to 3810898a74)

Solutions

1. Read the wrapped {error} detail — it identifies the exact initialization failure
2. Fix the underlying cause: ensure network access to tokenizer assets on first use, or clear/repair the token-counter cache
3. Where possible rely on session.usage.total_tokens (session metadata) instead of estimating, as check_if_compaction_needed already does when usage is present

// Warm the token counter once at startup and surface its error early:
create_token_counter().await?;
// after this, count_context_tokens cannot fail on counter creation

match count_context_tokens(&conversation).await {
    Err(e) if e.to_string().contains("Failed to create token counter") => {
        // fall back to a rough estimate (chars/4) or skip reporting; do not abort the session
    }
    other => other?,
}

When it happens

Trigger: Calling count_context_tokens when create_token_counter() fails — typically tokenizer assets cannot be loaded (offline first run, missing cache) or the underlying model/tokenizer init errors out.

Common situations: Air-gapped or proxied environments where tokenizer files cannot be fetched; corrupted token-counter cache; calling the context-token helper before any provider has warmed up the tokenizer.

Related errors

• Failed to create token counter: {}
• Failed to compact: context limit exceeded even after removin
• Cannot resume with provider or model changes because provide
• /compact is not available for provider '{provider}' because
• No agent-visible tool pair found for tool id: {}

AI-assisted analysis of aaif-goose/goose@3810898a74 (2026-08-16). Data as JSON: /api/errors/73ac4bc89f90e7d3. Report an issue: GitHub.