Jean-Pierre

Appearance

View Markdown

Tool style does not inherit field-by-field from the '*' defaults ​

  • Status: Done
  • Kind: Bug
  • Authors: jp
  • Date: 2026-08-11
  • Label: package=jp_config
  • Label: type=bug

[conversation.tools.'*'.style] is silently ignored for any tool that sets a single style field of its own.

ToolConfigWithDefaults::style resolves the whole struct at once (crates/jp_config/src/conversation/tool.rs:1215):

rust
pub fn style(&self) -> &DisplayStyleConfig {
    self.tool.style.as_ref().unwrap_or(&self.defaults.style)
}

Every neighbouring accessor fills field-by-field from the defaults. enable goes through Enable::effective, and run, result, format, and cancellation_response each unwrap_or a scalar. Only style falls back as a unit.

Why it bites ​

Setting a single style default across every tool looks like this and does nothing:

toml
[conversation.tools.'*'.style]
hidden = true

Every tool the skills enable already sets style.inline_results, so every one of them takes its own style wholesale and drops the default. The failure is silent: no error, no warning, the setting is simply absent from the resolved config.

Found while hiding read-only tool calls for the RFD pipeline personas, which ended up enumerating twenty-two tools by name because the one-line version does not work. That list will drift as tools are added.

Fix ​

DisplayStyleConfig implements FillDefaults, and PartialToolsDefaultsConfig::fill_from already resolves the '*' block's own style against the schematic defaults. There is no PartialToolConfig::fill_from; nothing fills a tool's style from the '*' block.

Adding that cross-key step is the fix. It cannot go in ToolConfigWithDefaults::style, because ToolConfig::style is an Option<DisplayStyleConfig>: once a tool declares one key the option is Some and every other field reads as deliberately set. By the time the config is resolved, which keys the tool asked for is no longer recoverable.

So the fill happens at partial-merge time, in PartialToolsConfig::fill_from, where the partial still records the tool's own keys. The alternative is to follow enable and keep the per-field optionality in the resolved type, splitting DisplayStyleConfig (stored, optional fields) from a concrete style returned at read time, the way EnableConfig splits from Enable. That changes the type every caller of style() reads.

Filling at merge time needs a second half. A resolved config goes back to a partial in two places that matter: the conversation config layer (ConversationStream::config) and the diff a query records into the stream (get_config_delta_from_cli). Both run through ToolsConfig::to_partial, and a filled-in style field is indistinguishable there from one the tool wrote itself, so a plain round-trip republishes it as a key of the tool's own, pinning it against every later '*' change. So to_partial subtracts the resolved '*' style from each tool's style, keeping only the fields that differ, the same way AppConfig::to_partial subtracts the assistant from conversation.inquiry.assistant.

Watch for ​

Whether anything currently relies on the whole-struct behaviour, for instance a tool that sets one style field expecting the others to fall back to the hardcoded defaults rather than to a '*' block.

Equality is the only signal the subtraction has. A tool that deliberately sets a style field to the value the '*' block already holds is indistinguishable from one inheriting it, and will follow a later '*'-only change instead of holding its own value.