Prefill Rejection

OpenRouter Assistant-Prefill Host Rejection

Problem

A client (observed: opencode driving an agent loop) sometimes sends a request whose messages array ends on an assistant turn — assistant prefill. OpenRouter hosts serving Anthropic models reject that shape with:

This model does not support assistant message prefill. The conversation must end with a user message.

The run dies mid-conversation. Because the client only produces the prefill shape on SOME turns (continuation after truncation, certain compaction states), the failure is intermittent, which invites two wrong fixes: retrying, and pinning providers.

Context / Trigger Conditions

• The exact error above, prefixed with a host name: [Azure], [Amazon Bedrock], [Google] (Vertex), or [Anthropic].
• Model id is an Anthropic model via OpenRouter (openrouter/anthropic/...).
• Same pipeline succeeded on other runs with no config change.
• A provider.only / provider.ignore routing pin was added and the error recurred from a different (or eventually the same) host.

Why the obvious fixes fail (measured, one project, 4 host families)

1. Ignore-listing the failing host grows one corpse at a time: Azure, then "Claude Platform on AWS", then Anthropic first-party each rejected identically.
2. Positively pinning the one host not yet observed rejecting (Vertex) held for two runs, then rejected too. The green runs were conversation-shape luck — those conversations never happened to end a request on an assistant turn. N green runs is a claim about those runs' shapes, not about the host.
3. Advertisement surfaces carry no signal: no endpoint lists prefill in supported_parameters (/api/v1/models/<id>/endpoints), so you cannot select a safe host by reading metadata.

Solution

Change the MODEL FAMILY for the affected seat; select the replacement with a live behavioral probe, not metadata.

The probe — a minimal prefill-shaped request per candidate:

curl -s -X POST https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"<candidate>","max_tokens":16,"messages":[
       {"role":"user","content":"Complete this sentence."},
       {"role":"assistant","content":"The capital of France is"}]}'

Read the result on two axes:

• Crash-safe: returns 200 instead of the prefill error (this is the axis that kills the failure class).
• Semantically correct: the completion CONTINUES the prefilled text ("Paris") rather than answering as a fresh turn. Measured 2026-08-17: deepseek-chat-v3.1 and gemini-3.1-pro-preview continued correctly; grok-4.5 returned 200 but ignored the prefill content. A crash-safe-but-ignoring model survives the loop with occasional confused turns; weigh that against capability for the seat.

Also: remove any provider pin keyed to the old model. If the pin is keyed dynamically off the model constant, changing the model silently re-aims the pin at a host that does not serve the new model.

Verification

The previously-dying pipeline stage completes end-to-end on the new seat, including runs long enough to hit the conversation shapes that produced prefill before. (Observed: two full agent review stages green after the change, at ~1/3 the token price of the Opus seat they replaced.)

Notes

• Anthropic's NATIVE API supports assistant prefill; the rejection is an OpenRouter-host normalization behavior. Going direct-to-Anthropic (not via OpenRouter) is an alternative fix when credentials allow.
• Avoid pinning standing infrastructure to -preview model ids even when they probe best — a retired id fails at the first model call later.
• If the client grows a config to stop emitting prefill, the Anthropic seat becomes eligible again; pre-register that as the revisit condition.

See also: sdk-via-openrouter (same OpenRouter domain, different problem: env-var credential fallback).

References

• OpenRouter model endpoints API: https://openrouter.ai/docs (endpoints, supported_parameters — verified to carry no prefill flag as of 2026-08-17)