don't eval task output with llm unless the task lacks problem matchers

[meganrogge]

fix #266292

Skip the redundant ~10s LLM error assessment _assessOutputForErrors for tasks that have problem matchers.

Only pass taskProblemPollFn when the task has problem matchers; pass undefined otherwise. Skip _assessOutputForErrors when a _pollFn is provided, since it already supplies structured error info.

Tasks with problem matchers get fast structured errors. Tasks without problem matchers now correctly fall through to the LLM assessment instead of throwing.

[Copilot]

Pull request overview

This PR optimizes the performance of task output monitoring by eliminating redundant LLM calls for tasks that have problem matchers configured. The change addresses issue #266292 which identified that get_task_output was making two LLM requests that added 1.5-2+ seconds to tool execution time.

Changes:

• Skip LLM-based error assessment (_assessOutputForErrors) when a custom polling function is provided, as it already supplies structured error information
• Only pass taskProblemPollFn to OutputMonitor for tasks that have problem matchers configured, allowing tasks without matchers to use LLM assessment
• Prevent taskProblemPollFn from being called and throwing for tasks without problem matchers

Reviewed changes

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

File	Description
src/vs/workbench/contrib/terminalContrib/chatAgentTools/browser/tools/monitoring/outputMonitor.ts	Modified _handleIdleState to skip LLM error assessment when a custom poll function is provided
src/vs/workbench/contrib/terminalContrib/chatAgentTools/browser/taskHelpers.ts	Added problem matcher detection to conditionally pass taskProblemPollFn only for tasks with configured matchers