mirror of
https://github.com/coder/coder.git
synced 2026-09-22 05:05:20 +08:00
Ethan
a779320d87
`coder exp scaletest llm-mock` used to return only canned text, so Coder Agents scaletests pointed at it never exercised the path that matters most: the agentic tool-call loop where the model asks for a tool, the workspace runs it, the result is fed back, and the model is re-prompted before it finally answers. This teaches the mock to reproduce that loop deterministically so scaletests actually drive real tool execution and hold streams open the way a live model would. It adds `--tool-calls-per-turn` and `--tool-call-command` so the OpenAI Chat Completions endpoint emits a controllable number of `execute` tool calls per turn (only when the request advertises an `execute` tool, otherwise it falls back to text, so it stays a safe drop-in). It also adds paced streaming with `--min-stream-duration`/`--max-stream-duration` (randomized per response) and `--response-payload-size`, so runs can simulate slow or long-lived responses instead of flushing everything at once. Closes CODAGT-307 Closes GRU-48