feat(chat): add 'model' parameter to runSubagent tool for call-time model override by RockyWearsAHat · Pull Request #306841 · microsoft/vscode

RockyWearsAHat · 2026-03-31T16:27:56Z

Summary

Add an optional model parameter to the runSubagent built-in tool, allowing callers to specify which language model a subagent invocation should use at call time, independent of the agent's model: frontmatter configuration.

Motivation

Currently, subagent model assignment is static — determined by the agent definition's model: frontmatter and resolved at agent load time. This works for fixed agent pipelines, but leaves orchestrator agents unable to route individual tasks to the most appropriate model based on task complexity.

Use cases:

Cost optimization: An orchestrator routes simple lookups to a cheaper model (claude-haiku-4-5) and reserves the parent model for complex reasoning.
Capability matching: A task requiring strong code generation gets a code-specialized model, while a summarization task gets a general-purpose one.
Dynamic routing: MCP tools that surface available models (e.g., list_language_models) enable the LLM to make informed model selection decisions.

Design

Resolution Priority

The model resolution in resolveSubagentModel() now follows this priority order:

Call-time model parameter (highest precedence) — from the tool invocation
Agent's model: frontmatter — qualified names from .agent.md
Parent model (default) — inherited from the orchestrator

Model Resolution

The new resolveModelHint() helper resolves a model hint string by:

Trying qualified name lookup via lookupLanguageModelByQualifiedName() (e.g., "Claude Sonnet 4.6 (Copilot)")
Falling back to direct model ID lookup via lookupLanguageModel() (e.g., "claude-sonnet-4-6")
Returning undefined if neither resolves, with a warn-level log

Cost Safety

The existing multiplier-based cost cap continues to apply — the resolved model's multiplierNumeric must not exceed the parent agent's multiplier. This cap was moved outside the if (subagent) block to apply uniformly to all resolution paths.

Changes

`runSubagentTool.ts` (+65 lines)

Add model?: string to IRunSubagentToolInputParams
Add model to the tool's input schema in getToolData()
Extract resolveModelHint() helper for dual-lookup model resolution
Extend resolveSubagentModel() with optional callTimeModelHint parameter
Move multiplier cap to apply uniformly to all resolution paths
Update invoke() no-subagent branch to handle model override from cache
Update prepareToolInvocation() to pass args.model through

`runSubagentTool.test.ts` (+273 lines)

Schema test: model property present in tool data
6 new tests in call-time model parameter suite:
- Override via qualified name without a subagent
- Override via direct model ID without a subagent
- Call-time model takes precedence over agent model: frontmatter
- Call-time model respects multiplier cap (falls back to parent)
- Unknown model name falls back gracefully to parent model
- Override with cheaper model on a named agent

Testing

All 24 tests in the RunSubagentTool test suite pass, including the 7 new tests.

…odel override Add an optional `model` parameter to the `runSubagent` built-in tool that allows callers to specify which language model should be used for a subagent invocation, overriding the agent's default model from frontmatter. ## Changes ### runSubagentTool.ts - Add `model?: string` to `IRunSubagentToolInputParams` - Add `model` property to the tool's input schema in `getToolData()` - Extract `resolveModelHint()` helper that resolves a model hint string by trying qualified name lookup first, then direct model ID lookup - Extend `resolveSubagentModel()` with optional `callTimeModelHint` parameter that takes highest precedence over agent frontmatter - Move multiplier cap logic outside the `if (subagent)` block so it applies uniformly to both agent frontmatter and call-time overrides - Update `invoke()` no-subagent branch to read `modeModelId` from cache and handle call-time model override without a named agent - Update `prepareToolInvocation()` to pass `args.model` through ### runSubagentTool.test.ts - Add schema test: `model` property appears in tool data - Add 6 call-time model parameter tests: - Override via qualified name without a subagent - Override via direct model ID without a subagent - Call-time model takes precedence over agent frontmatter - Call-time model respects multiplier cap - Unknown model name falls back gracefully - Override with cheaper model on a named agent ## Design Resolution priority: 1. Call-time `model` parameter (highest precedence) 2. Agent's `model:` frontmatter qualified names 3. Parent model (default) After resolution, the existing multiplier-based cost cap still applies: a subagent cannot use a more expensive model than the parent agent regardless of how the model was specified. Fixes microsoft#306836

Copilot

Pull request overview

Adds an optional call-time model override to the built-in runSubagent tool so orchestrator agents can route individual subagent invocations to specific language models, while still enforcing the existing multiplier-based cost cap.

Changes:

Extend runSubagent tool input schema and params with optional model?: string.
Implement model-hint resolution (qualified name → model id) and update model resolution priority (call-time override → agent frontmatter → parent).
Add a dedicated test suite covering precedence, qualified-name/id resolution, cost-cap behavior, and fallback cases.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/vs/workbench/contrib/chat/common/tools/builtinTools/runSubagentTool.ts	Adds `model` param support, resolves model hints, and applies multiplier cap across all resolution paths.
src/vs/workbench/contrib/chat/test/common/tools/builtinTools/runSubagentTool.test.ts	Adds schema coverage and multiple call-time model override behavior tests.

Copilot · 2026-03-31T16:32:52Z

+		if (modeModelId && modeModelId !== mainModelId) {
+			const mainModelMetadata = mainModelId ? this.languageModelsService.lookupLanguageModel(mainModelId) : undefined;
+			const resolvedModelMetadata = this.languageModelsService.lookupLanguageModel(modeModelId);
+			const mainMultiplier = mainModelMetadata?.multiplierNumeric;
+			const resolvedMultiplier = resolvedModelMetadata?.multiplierNumeric;
+			if (mainMultiplier !== undefined && resolvedMultiplier !== undefined && resolvedMultiplier > mainMultiplier) {
+				const source = callTimeModelHint ? 'Call-time' : subagent ? `Subagent '${subagent.name}'` : 'Unknown';
+				this.logService.warn(`[RunSubagentTool] ${source} requested model '${resolvedModelMetadata?.name}' (multiplier: ${resolvedMultiplier}) which exceeds the main agent model '${mainModelMetadata?.name}' (multiplier: ${mainMultiplier}). Falling back to the main agent model.`);
+				modeModelId = mainModelId;


In resolveSubagentModel, the multiplier-cap warning uses source = callTimeModelHint ? 'Call-time' : .... If a call-time model hint is provided but doesn’t resolve (so the override is ignored), the subsequent multiplier-cap warning will still attribute the model request to “Call-time”, even though the selected model may have come from the subagent frontmatter. Consider tracking whether the call-time override actually applied (e.g., a boolean or enum resolutionSource) and use that for the warning message.

Copilot

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

Copilot · 2026-03-31T16:56:08Z

+	suite('call-time model parameter', () => {
+		function createMetadata(name: string, multiplierNumeric?: number): ILanguageModelChatMetadata {
+			return {
+				extension: new ExtensionIdentifier('test.extension'),
+				name,


This suite redefines helper functions (createMetadata/createTool/createAgent) that already exist in the earlier "model fallback behavior" suite in the same file. Consider extracting these helpers once (e.g., to file scope or a shared helper) to avoid duplication and keep future changes (like new required metadata fields) from needing updates in multiple places.

Previous commit exposed the runSubagent `model` schema unconditionally, which conflicted with the feature toggle that controls custom-agent usage. This change moves the property description under the `customAgentInSubagent` gate so the schema only offers `model` when agents are available, and updates the schema test to enable that flag to keep coverage aligned.

roblourens · 2026-03-31T21:47:34Z

Thanks for the PR, but this is already in progress in #298161, and we should drive that forward cc @digitarald

Copilot AI review requested due to automatic review settings March 31, 2026 16:27

RockyWearsAHat mentioned this pull request Mar 31, 2026

Chat: add 'model' parameter to runSubagent tool for call-time model override #306836

Closed

Copilot started reviewing on behalf of RockyWearsAHat March 31, 2026 16:29 View session

Copilot AI reviewed Mar 31, 2026

View reviewed changes

vs-code-engineering Bot added the triage-needed label Mar 31, 2026

vs-code-engineering Bot assigned connor4312 Mar 31, 2026

This was referenced Mar 31, 2026

feat(chat): add tiers frontmatter and tier parameter to runSubagent tool #306846

Open

Chat: add 'tiers' frontmatter and 'tier' parameter to runSubagent for semantic model tier routing #306717

Open

RockyWearsAHat requested review from Copilot March 31, 2026 16:50

Copilot started reviewing on behalf of RockyWearsAHat March 31, 2026 16:51 View session

Copilot AI reviewed Mar 31, 2026

View reviewed changes

connor4312 assigned roblourens and unassigned connor4312 Mar 31, 2026

roblourens closed this Mar 31, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(chat): add 'model' parameter to runSubagent tool for call-time model override#306841

feat(chat): add 'model' parameter to runSubagent tool for call-time model override#306841
RockyWearsAHat wants to merge 2 commits intomicrosoft:mainfrom
RockyWearsAHat:feature/runsubagent-model-param

RockyWearsAHat commented Mar 31, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 31, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 31, 2026

Uh oh!

roblourens commented Mar 31, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

RockyWearsAHat commented Mar 31, 2026

Summary

Motivation

Design

Resolution Priority

Model Resolution

Cost Safety

Changes

runSubagentTool.ts (+65 lines)

runSubagentTool.test.ts (+273 lines)

Testing

Related

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Copilot AI Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

roblourens commented Mar 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

`runSubagentTool.ts` (+65 lines)

`runSubagentTool.test.ts` (+273 lines)

roblourens commented Mar 31, 2026 •

edited

Loading