Constrain `run_factory` to registered factory names to prevent repeated guesses

[ulugbekna]

Summary

run_factory can be exposed with an unconstrained name argument even when no matching factory is registered. The model can interpret tool availability as factory availability, invent plausible names, and retry them serially.

Originally reported from the VS Code Agents Window in microsoft/vscode#329551.

Observed behavior

With Copilot CLI 1.0.78 and GPT-5.6 Sol, one session produced 13 distinct model-authored run_factory calls. Every call failed immediately with:

RpcResponseError: No factory registered with name "<name>"

The model first tried eight implementation-oriented names:

• apply_finding
• code_implementation_factory
• single_task
• implementation
• code-task
• fix-factory
• repo_task
• review_then_implement

It later tried five more names for video analysis:

• video_analysis
• analyze_video
• single_video_analysis
• video-analysis
• media_analysis

These were separate assistant-authored calls after each rejection, not host retries. No factory ran.

Steps to reproduce

1. Expose run_factory in a session where no matching factories are registered.
2. Ask the agent to perform a routine single-session task, such as a small implementation or media-analysis task.
3. Observe that the model may invent and repeatedly retry factory names.

Expected behavior

Invalid factory names should be impossible or strongly bounded. Ideally:

• Dynamically constrain name to registered factory names.
• Do not expose run_factory when there are no registered factories or resumable runs.
• If dynamic enum generation is not possible, require registry discovery before invocation and include available names in validation errors.
• Enforce exactly one of name and resumeFromRunId at the schema level.

Actual behavior

name is effectively a free-form string, so failures occur only after a costly tool round trip. Prompt guidance to use factories selectively did not prevent repeated guesses.