Struct AutoCompactConfig 

pub struct AutoCompactConfig {
    pub keep_first: usize,
    pub token_threshold: usize,
    pub tool_output_max_chars: usize,
    pub keep_last: usize,
    pub compact_strategy: CompactStrategy,
    pub hard_limit_tokens: Option<usize>,
    pub hard_limit_strategy: CompactStrategy,
    pub custom_compactor: Option<VmValue>,
    pub mask_callback: Option<VmValue>,
    pub compress_callback: Option<VmValue>,
    pub summarize_prompt: Option<String>,
    pub policy_strategy: String,
}

Configuration for automatic transcript compaction in agent loops.

Two-tier compaction: Tier 1 (token_threshold / compact_strategy): lightweight, deterministic observation masking that fires early. Masks verbose tool results while preserving assistant prose and error output. Tier 2 (hard_limit_tokens / hard_limit_strategy): aggressive LLM-powered summarization that fires when tier-1 alone isn’t enough, typically as the transcript approaches the model’s actual context window.

Fields

keep_first: usize

Number of earliest messages to keep verbatim before the compacted summary. The system prompt is not part of this list and is always preserved separately by the caller.

token_threshold: usize

Tier-1 threshold: estimated tokens before lightweight compaction.

tool_output_max_chars: usize

Maximum character length for a single tool result before microcompaction.

keep_last: usize

Number of recent messages to keep during compaction.

compact_strategy: CompactStrategy

Tier-1 strategy (default: ObservationMask).

hard_limit_tokens: Option<usize>

Tier-2 threshold: fires when tier-1 result still exceeds this. Typically set to ~75% of the model’s actual context window. When None, tier-2 is disabled.

hard_limit_strategy: CompactStrategy

Tier-2 strategy (default: Llm).

custom_compactor: Option<VmValue>

Optional Harn callback used when a strategy is custom.

mask_callback: Option<VmValue>

Optional callback for domain-specific per-message masking during observation mask compaction. Called with a list of archived messages, returns a list of Option<String> — Some(masked) to override the default mask for that message, None to use the default. This lets the host (e.g. an IDE or cloud runner) inject AST outlines, file summaries, etc. without putting language-specific logic in Harn.

compress_callback: Option<VmValue>

Optional callback for per-tool-result compression. Called with {tool_name, output, max_chars} and returns compressed output string. When set, used INSTEAD of the built-in microcompact_tool_output. This allows the pipeline to use LLM-based compression rather than keyword heuristics.

summarize_prompt: Option<String>

Optional prompt-template asset path used when LLM compaction is selected. The rendered template becomes the user message sent to the summarizer.

policy_strategy: String

User-facing policy label for replay and observability. This can be broader than the engine strategy, e.g. hybrid lowers to LLM summarization plus truncate fallback.

Trait Implementations

impl Clone for AutoCompactConfig

fn clone(&self) -> AutoCompactConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AutoCompactConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for AutoCompactConfig

fn default() -> Self

Returns the “default value” for a type.