Summary

The Apps extension's inner LLM call to generate HTML app content does not set max_tokens, so the API defaults to 8,192 output tokens. This is insufficient for generating any non-trivial HTML app. The response gets truncated (finish_reason: "length"), the html field is never written, and the user sees a cryptic error:

Error: Failed to parse tool response: missing field `html`

Reproduction

1. Use the Databricks provider with databricks-claude-opus-4-6 (or any provider where the default max output tokens is ~8K)
2. Ask goose to create a non-trivial app (e.g., "Create a figure skating game app")
3. The apps__create_app tool returns Error: Failed to parse tool response: missing field html every time

Root Cause

In crates/goose/src/agents/apps_extension.rs, generate_new_app_content() and generate_updated_app_content() call:

let (response, _usage) = provider
    .complete(session_id, &system_prompt, &messages, &tools)
    .await

This uses provider.complete() which inherits the provider's ModelConfig. When max_tokens is None in the config, no max_tokens is sent to the API, and the Databricks/Claude API defaults to 8,192 output tokens.

The inner LLM is asked to call a create_app_content tool where the html field (a complete HTML/CSS/JS app) is serialized as part of the JSON arguments. The LLM generates fields in order — name, description, width, height, resizable — and hits the 8,192 token ceiling before ever reaching html.

Evidence from diagnostic logs — all 3 inner LLM calls show the same pattern:

The returned arguments each time:

{"name":"figure-skating-championship","description":"Interactive figure skating game...","height":700,"resizable":false,"width":900}

No html field — the LLM ran out of tokens before starting it.

Suggested Fix

1. Set a higher max_tokens for the inner LLM call — use complete_with_model() with a ModelConfig that overrides max_tokens to something like 16,384+ (or even higher for complex apps)
2. Detect truncation — check finish_reason / usage in the response and return a meaningful error instead of the cryptic serde error
3. Both ideally

Environment

• Provider: Databricks
• Model: databricks-claude-opus-4-6
• max_tokens in config: null (not set)
• API default: 8,192 output tokens