Provider tool search

Roder normally sends explicit ToolSpec entries with a turn request. For providers that support native catalog search, Roder can instead defer loading large tool catalogs and let the model ask the provider for relevant tools during the turn.

Modes

• explicit keeps the default behavior and sends tool schemas directly.
• auto allows provider-native search when the active provider and model support it.
• provider_native requests provider-native search and falls back to explicit tools when configured.

[tool_search]
mode = "provider_native"
max_catalog_items = 200
include_mcp = true
include_skills = true
fallback_to_explicit_tools = true
provider_variant = "bm25"

The OpenAI Responses adapter maps supported GPT-5.4+ requests to the providertool_search tool and marks local function tools for deferred loading. Older or unsupported models continue to receive ordinary explicit tool schemas.

Why it matters

Provider-native search keeps model context focused when a distribution installs many built-in, MCP, skill, browser, workflow, or domain-specific tools. The runtime still owns execution, policy, approvals, and result routing after the model selects a tool.