Conversation Compaction Configuration

When the conversation token count reaches 70% of the model's context window, Siada automatically compresses the message history by generating an LLM summary to replace older messages while preserving recent conversations.

Configuration

Edit ~/.siada-cli/conf.yaml:

# Compaction strategy (optional, auto-detected if not set)
compaction_strategy: null

Available Strategies

Strategy	Description	Default For
header_summary	Keeps first user-assistant pair + LLM summary + recent 30% messages	CLI / TUI mode
turn_prune_summary	Three-layer pipeline: turn pruning → tool result truncation → LLM summary, keeps last 3 turns	IM mode (Feishu)

When not configured, the system auto-selects based on session mode: CLI/TUI uses header_summary, IM uses turn_prune_summary.

Example

Force IM strategy in CLI mode:

compaction_strategy: "turn_prune_summary"