Context Envelope¶

The context envelope is CRP's core innovation — a dynamically packed payload that fills all available space in the LLM's context window with the most relevant facts.

Maximum Context Saturation¶

$$E_{\max} = C - S - T - G$$

Symbol	Meaning	Example (128K)
$C$	Context window size	131,072
$S$	System prompt tokens	~500
$T$	Task input tokens	~8,756
$G$	Generation reserve	16,384
$E_{\max}$	Envelope capacity	105,816

$G$ is automatically determined: user's max_output_tokens → provider's reported max → min(C // 4, 16384).

In practice

CRP achieves 0.939–1.021 saturation (mean 0.994) — virtually every available token is used for relevant context.

Envelope Sections¶

The envelope is structured with 11 priority-ordered sections. Higher-priority sections survive when space is limited:

Priority	Section	Tokens	Purpose
1	Critical State	100–500	GOAL, PHASE, BLOCKER, CONSTRAINT, WINDOW
2	LLM Synthesis	Adaptive	LLM's own curated understanding
3	Task Brief	Varies	What to do + output format
4	Discoveries	Bulk	Atomic facts with graph edges
5	Source Passages	Variable	Verbatim text for high-relevance facts
6	Decisions & Plan	Variable	Reasoning trail with justifications
7	Error Log	Small	What failed and why
8	Tool History	Small	Compact execution summaries
9	Expanded Context	Overflow	Full-fidelity data from warm state
10	CKF Retrievals	Variable	Cross-session knowledge
11	Reasoning Scaffold	Small	Step-by-step templates (weak models)

Fact Selection Algorithm¶

CRP uses a 3-phase pipeline to select which facts go into the envelope:

Phase 1: Multi-Aspect Task Decomposition¶

The task is broken into noun phrases / aspects. A fact matching any aspect scores high:

$$\text{score}(f) = \max_{a \in \text{aspects}} \cos\bigl(\text{embed}(f),\; \text{embed}(a)\bigr)$$

Phase 2: Bi-Encoder Fast Scoring¶

All facts scored using all-MiniLM-L6-v2 embeddings. For >1,000 facts, an HNSW ANN index provides $O(\log N)$ retrieval.

Composite score:

$$\text{final}(f) = \text{sim}(f) \times \text{recency}(f) \times \text{novelty}(f) + \text{dep_bonus}(f)$$

Factor	Formula	Range
Recency	$e^{-0.1 \times \text{age_in_windows}}$	0 → 1
Novelty	Unseen: 1.5×, <3 uses: 1.0×, 3+: 0.5×	0.5 → 1.5
Dependency	Graph-connected facts inherit relevance	0 → 0.5

Phase 3: Cross-Encoder Reranking¶

Top 200 candidates re-scored with ms-marco-MiniLM-L6-v2:

$$\text{blended} = 0.6 \times \text{CE_score} + 0.4 \times \text{BE_score}$$

Cache hit rate: 50–80% in continuation chains (saves 200–320 ms/window).

Packing Strategy¶

After scoring, facts are packed using greedy bin-packing with:

Dependency-aware graph pulling — up to 2 hops of connected facts
Bookend strategy — top 3 facts duplicated at envelope end (counters "lost in the middle" attention bias)
Progressive compression — truncation → summarization → tabular → reference replacement

Continuation Envelopes¶

When output is truncated, CRP builds a continuation envelope containing:

Component	Purpose
Extracted facts	From the truncated output
Structural state	Open blocks, list position, section headers
Task gap	Missing items from original task
Style anchor	Last natural paragraph for voice consistency
Voice profile	Sentence length, vocabulary, tone markers
Document map	Running TOC with section completion status

Note

Continuation envelopes use extraction results, not raw text overlap. This is key to CRP's quality preservation across windows.

Preview¶

Before dispatching, you can preview what the envelope will look like:

preview = client.preview_envelope(
    system_prompt="You are a technical writer.",
    task_input="Explain Kubernetes networking.",
)
print(f"Total tokens:      {preview.total_tokens}")
print(f"Envelope tokens:   {preview.envelope_tokens}")
print(f"Generation reserve: {preview.generation_reserve}")
print(f"Facts included:    {preview.facts_included}")
print(f"Facts available:   {preview.facts_available}")
print(f"Saturation:        {preview.saturation:.1%}")