aphrodite-auto-expand-testing

name	aphrodite-auto-expand-testing
description	Protocol for testing auto-expand behavior - controlled by AUTO_EXPAND_LIMIT from TOML auto_expand_limit (default 5 → effectively OFF). APHRODITE_AUTO_EXPAND=1 enables aggressive. Context engine is the real async engine.
version	2.0.0
platforms	["macos"]
related_skills	["aphrodite-boundary-behaviors","aphrodite-dev-workflow"]

Aphrodite Auto-Expand Testing

Protocol for testing auto-expand behavior: whether <<<CCR:...>>> markers appear raw or get auto-resolved.

Historical note: the file references below (plugins/aphrodite/_hooks/session.py, _core/config.py, _engine.py) describe the pre-Rust-port Python plugin layout. The current plugin (plugins/aphrodite/) is a thin ctypes loader - all compression/engine/auto-expand logic now lives in the Rust dylib (crates/aphrodite/src/, crates/aphrodite-hermes/src/). The mechanism described here (TOML auto_expand_limit, APHRODITE_AUTO_EXPAND env var, the pre_llm hook resolving markers below the limit) still applies conceptually, but the specific Python file/line citations no longer exist in this tree and Step 1's from _core import AUTO_EXPAND_LIMIT will fail.

Mechanism

Auto-expand is controlled by AUTO_EXPAND_LIMIT (int). When > 0, the pre_llm hook scans conversation for CCR markers and resolves those where marker_size < AUTO_EXPAND_LIMIT.

Defaults (from aphrodite.toml):

• compression.auto_expand_limit = 5 → AUTO_EXPAND_LIMIT = 5
• With limit=5, only markers < 5 bytes get resolved → effectively OFF
• Set APHRODITE_AUTO_EXPAND=1 to enable aggressive (limit becomes 51200 = 50KB)

APHRODITE_NO_AUTO_EXPAND does NOT exist in the source code. It was fictional in old skills. Default (limit=5) already produces raw markers.

The Three Layers

1. Proxy response compression (Rust, :9798) - compresses provider RESPONSE messages. Produces <<<CCR:...>>> in model responses. Always active.

2. Context engine (Python, _engine.py) - compresses MIDDLE messages in conversation_history into a CCR marker. Fires on turn 2+ when threshold reached. This is the "async engine" - tool call outputs get compressed into CCR memory containers, LLM sees markers and polls via aphrodite_retrieve().

3. Auto-expand (pre_llm hook) - resolves CCR markers to full content before LLM sees them. Controlled by AUTO_EXPAND_LIMIT.

Protocol

Step 1: Check current config

python3 -c "
import sys; sys.path.insert(0, 'plugins/aphrodite')
from _core import AUTO_EXPAND_LIMIT
print(f'AUTO_EXPAND_LIMIT={AUTO_EXPAND_LIMIT}')
print('Auto-expand is', 'ON' if AUTO_EXPAND_LIMIT > 50 else 'effectively OFF')
"

Step 2: Test context engine with raw markers (default)

Start a multi-turn session. Default config already has auto-expand effectively OFF:

APHRODITE_ENGINE_THRESHOLD_PCT=1 \
	APHRODITE_ENGINE_MIN_MSGS=4 \
	APHRODITE_ENGINE_PROTECT_FIRST=1 \
	APHRODITE_ENGINE_PROTECT_LAST=1 \
	hermes

Run several tool calls. After turn 2+, the context engine compresses tool results. LLM sees <<<CCR:hash|context|N>>> - must use aphrodite_retrieve(hash) to poll from the CCR memory container.

Step 3: Test with auto-expand ON

APHRODITE_AUTO_EXPAND=1 hermes

Markers under 50KB get auto-resolved - LLM sees full content instead of markers.

Step 4: Verify retrieval always works

aphrodite_retrieve(hash="<hash>")

The retrieve tool always returns full content regardless of auto-expand setting.

Expected Behavior Matrix

Config	What LLM sees	Retrieve returns
Default (limit=5)	<<<CCR:hash|context|N>>> (raw)	Full content
APHRODITE_AUTO_EXPAND=1	Expanded content (<50KB)	Full content
Engine threshold not hit	Full content	Full content

Pitfalls

• APHRODITE_NO_AUTO_EXPAND is fictional - no such env var in source. Default (limit=5) already means raw markers.
• Single-turn (hermes -z): context engine won't fire (needs turn 2+). Use multi-turn.
• Proxy compresses responses, engine compresses messages: proxy handles provider responses; engine handles conversation middle messages.
• protect_first_n / protect_last_n: protected messages stay raw.