Per-tool model routing: delegate lightweight tools to a cheaper/faster secondary model

Problem

When using a paid API plan (e.g. Z.AI Coding Plan, Claude Max, OpenAI), all tool calls consume the same model quota regardless of complexity. A web search, file grep, or browser action costs the same credits as complex reasoning or multi-file refactoring. There is no way to route lightweight tools to a cheaper/faster model while keeping the flagship model for hard tasks.

Concrete example with Z.AI Coding Plan credits:

Routing the first three rows to GLM-4.7 instead of GLM-5.2 saves ~15% credits per call on Z.AI. Over a week of agentic coding, this can meaningfully extend the weekly quota.

Proposed feature

Add a [[tools.model_override]] config table that routes specific tools to a different provider/model:

# Default model for all tools
# (set via provider/model selection as today)

# Override: route specific lightweight tools to a cheaper model
[[tools.model_override]]
tools = ["agentgrep", "websearch", "webfetch", "browser"]
provider = "zai-coding-plan"
model = "GLM-4.7"

# Keep heavy tools on the flagship
# (implicitly uses the default model)

Behavior

• When the agent decides to call a tool in the override list, that single tool call is sent to the overridden model/provider instead of the session model.
• Tool call inputs and outputs are still injected into the main conversation history for the primary model.
• The override applies to the tool execution itself (choosing parameters, interpreting results), not to the model deciding which tool to call.
• Falls back to the primary model if the override provider is unavailable.

Edge cases to consider

• Should MCP tool calls be overridable separately?
• Should there be a per-override concurrency limit to avoid hammering the cheaper provider?
• How does this interact with compaction (which model summarizes)?
• Swarm workers already support per-agent model selection. This feature is the single-agent analogue.

Alternatives considered

1. Manual model switching (Ctrl+Tab): Works but tedious and error-prone. Users forget to switch back.
2. Swarm delegation: Works for bulk search tasks but overkill for a single web search. Adds latency and complexity.
3. Prompt engineering: Asking the model to "use tools sparingly" reduces calls but does not change per-call cost.

Who benefits

Any user on a metered or flat-fee API plan who wants to maximize their quota by reserving expensive model capacity for tasks that actually need it. Especially relevant for Z.AI Coding Plan users (Lite/Pro/Max) and Claude Max users who hit weekly caps.

The issue names no files, tests, or entry points; start by locating the current provider/model selection and tool-execution paths. Define the configuration and behavior around per-tool overrides, fallback handling, conversation history, and the listed edge cases, then verify that lightweight tools use the override while the primary model handles the rest.