Crate clark_agent_compaction 

Provider-agnostic context compaction primitives for agent transcripts.

The crate owns the deterministic parts of compaction: checkpoint threshold checks, transcript rendering budgets, recent user-message retention, summary handoff finalization, and policy for append-only episodic histories. Callers own provider I/O, persistence, cancellation, telemetry, and conversion back into their typed message model.

Structs

CharHeuristic

Fast four-characters-per-token heuristic.

CompactedTranscript

Final model-visible compaction output.

CompactionConfig

Configuration for compaction planning.

CompactionPlan

Deterministic data needed to install a model-produced summary.

CompactionRequest

Prompt to send to the caller’s summarization model.

EpisodicCompactionConfig

Default policy for long-lived, append-only conversation histories.

EpisodicCompactionPressure

Storage-neutral observations for one append-only conversation episode.

PlainToolCall

Tool-call rendering helper for PlainMessage.

PreparedCompaction

Caller-owned model request and deterministic replacement plan.

Enums

EpisodicCompactionTrigger

Durable reason an episodic boundary was written.

PlainMessage

Simple test/example transcript shape.

Constants

DEFAULT_AUTO_COMPACT_TOKEN_LIMIT

Default estimated-token threshold for automatic compaction.

DEFAULT_COMPACTION_PROMPT

Default prompt for the caller-owned summarization model request.

DEFAULT_COMPACT_REQUEST_TOKEN_LIMIT

Default source budget for the compaction request itself.

DEFAULT_PRIVATE_REASONING_EXCERPT_CHARS

Default character cap for one private-reasoning excerpt.

DEFAULT_PRIVATE_REASONING_SUMMARY_GUIDANCE

The behavior the default prompt requires for private reasoning excerpts. Exported so callers with a custom prompt can preserve the same contract.

DEFAULT_RECENT_USER_TOKEN_BUDGET

Default budget for recent real user messages retained after compaction.

DEFAULT_SUMMARY_PREFIX

Default runtime-context marker for model-visible compaction handoffs.

Traits

TokenEstimator

Approximate token estimator used by compaction budgets.

TranscriptMessage

Runtime-owned transcript adapter.

Functions

append_private_reasoning_for_compaction

Append one bounded private-reasoning excerpt to a compaction transcript.

episodic_compaction_trigger

Select the highest-priority automatic episodic-compaction trigger.

estimate_transcript_tokens

Estimate the rendered transcript token count.

finalize_compaction

Combine a model-produced summary with a previously prepared plan.

prepare_compaction

Build a compaction prompt plus deterministic finalization plan.

should_compact

Return true when compaction should run for this transcript.

truncate_to_token_budget

Truncate text by the default character heuristic while preserving the tail.

truncate_to_token_budget_with_estimator

Truncate text according to estimator.max_chars_for_tokens, preserving the tail.

with_private_reasoning_summary_guidance

Add the private-reasoning handoff contract to a caller-supplied prompt.