Assistant Prefix as Continuation Framing

Adding a partial role: assistant message before inference shifts the model from generating a fresh response to continuing its own prior output. This framing change empirically improves compliance with forced-output rules.

The Mechanism

Standard request (model generates from scratch):

[
  {"role": "system", "content": "...rules including canary..."},
  {"role": "user",   "content": "hello"}
]

With assistant prefix (model continues from partial output):

[
  {"role": "system",    "content": "...rules including canary..."},
  {"role": "user",      "content": "hello"},
  {"role": "assistant", "content": "This is turn 1. Per Session Canary rule, I must start with a 🐦 canary before anything else.\n\n🐦 "}
]

The model treats the assistant message as its own prior output and continues from where it left off.

Empirical Results

Tested against Qwen3-14B-FP8 on vLLM at localhost:9000, 40K max context, fresh session with "hello" as user input:

Configuration Pass Rate Runs
Baseline (system + user only) 81% (13/16) 16
With assistant prefix 100% (16/16) 16

Why It Works

The assistant prefix does not mechanically force output. The model must still cooperate. But the framing shift is significant:

Property Generation Mode Continuation Mode
Model’s task “Respond to the user” “Continue what I was saying”
Pretraining prior Social greeting for hello Follow the established pattern
Rule checking Optional — model may skip Already done (the prefix references the rule)
Output starts with Whatever the strongest prior produces Whatever follows the prefix

The prefix embeds the rule reference in the model’s own voice. The model reads it as something it already decided to do, not an external instruction it must choose to follow.

Important Caveat

This is NOT a mechanical guarantee. It is still a cooperative mechanism:

The correct framing: assistant prefix is the most effective cooperative mechanism tested, but it is not deterministic enforcement.

Comparison with Other Approaches

Approach Where Injected Compliance Rate Deterministic?
System prompt rule role: system (~5K tokens in) ~80% No
JIT user-message reminder role: user (appended) Not tested No
Assistant prefix role: assistant (partial turn) 100% (N=16) No
PreToolUse hook block role: tool (block reason) High (for tool-call rules) Yes (blocks the tool)

Current Limitation in Goose

Goose does not currently support assistant prefill. Every role: assistant message in goose’s request payload is a completed prior model response. There is no mechanism to inject a partial assistant turn before inference. Implementing this would require changes to goose’s provider abstraction layer.

Harvested From

Project Source Entry
context-eng Session: AgentFS Evolution Direct curl testing against localhost:9000 — 16 baseline runs, 16 prefix runs
context-eng Session: AgentFS Evolution Analysis of why /clear restores compliance — continuation framing from administrative context

Implications