Repository navigation
Fresh worker fails before tool calls with assistant-prefill rejection #1895
Description
graff: 0.302.74.64 (latest: 0.302.74.64), managed engine
Harness: 0.2.189 (latest: 0.2.189)
Surface: Harness desktop
Platform: macOS arm64
Symptom
A freshly spawned worker fails before making any tool calls:
invalid_request_error: This model does not support assistant message prefill. The conversation must end with a user message.
The operation reports 2 rejected attempts and marks retrying as-is unsafe.
Repro
- Spawn a fresh worker on the affected backend.
- Observe the rejection above before any tool calls (0 tool calls; 2 rejected attempts).
Reproduced on the versions above. Root cause is not established.
Image
Synthetic illustration of the error and observed counts, not an application capture.
Fix
Construct a valid request history for backends that reject assistant prefill. If a request cannot be made valid, fail fast with an actionable error rather than making another rejected attempt.
Activity
justrach commentedon Nov 12, 2026
Second occurrence, same versions (graff 0.302.74.64, Harness 0.2.189, macOS arm64).
Two independent fresh-worker spawns in one session failed the same way: 0 tool calls, 2 rejected attempts,
invalid_request_error: This model does not support assistant message prefill. The conversation must end with a user message.New evidence on scope: re-running the identical task text on a different worker tier succeeded, so the rejection is specific to that backend route rather than to the task content. Both failures happened at worker spawn, before any tool call, which is consistent with the request shape rather than with anything the task did.
justrach commentedon Nov 14, 2026
Implemented in #1900 (not yet merged); current-head CI is green.