Skip to content

sheawinkler/ContextLattice

v3.11.2 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

agent-framework agent-orchestration ai-infra ai-interface context-engineering context-management
+4 more
go long-horizon-agents long-horizon-intelligence rust

Summary

AI summary

Forced‑minimal Context Packs now retain up to four ranked evidence cards instead of empty lists.

Changes in this release

Feature Low

Added regression coverage proving sparse pack stays under contract while producing non‑empty deterministic synthesis findings and evidence trails.

Added regression coverage proving sparse pack stays under contract while producing non‑empty deterministic synthesis findings and evidence trails.

Source: llm_adapter@2026-07-15

Confidence: high

Bugfix Medium

Forced-minimal Context Packs now retain up to four compact ranked evidence cards instead of empty list.

Forced-minimal Context Packs now retain up to four compact ranked evidence cards instead of empty list.

Source: llm_adapter@2026-07-15

Confidence: high

Bugfix Medium

Lower-value and duplicated prompt-section structure is compacted around the evidence kernel.

Lower-value and duplicated prompt-section structure is compacted around the evidence kernel.

Source: llm_adapter@2026-07-15

Confidence: high

Bugfix Medium

Reference prompt is rebuilt from the evidence‑preserving sparse pack.

Reference prompt is rebuilt from the evidence‑preserving sparse pack.

Source: llm_adapter@2026-07-15

Confidence: high

Bugfix Medium

Compiler and token‑budget metadata now distinguish pre‑boundary selection counts from actual returned evidence.

Compiler and token‑budget metadata now distinguish pre‑boundary selection counts from actual returned evidence.

Source: llm_adapter@2026-07-15

Confidence: high

Full changelog

ContextLattice v3.11.2 - Sparse Evidence Preservation

v3.11.2 fixes the last-resort Context Pack boundary so constrained output remains useful, not merely schema-valid. When a full pack exceeds its output contract, ContextLattice now preserves a compact kernel of the highest-ranked evidence and builds deterministic synthesis from it.

What changed

  • Forced-minimal Context Packs retain up to four compact ranked evidence cards instead of replacing all evidence with an empty list.
  • Lower-value and duplicated prompt-section structure is compacted around that evidence kernel.
  • The reference prompt is rebuilt from the evidence-preserving sparse pack.
  • Compiler and token-budget metadata now distinguish pre-boundary selection counts from the evidence actually returned.
  • Added regression coverage proving the sparse pack stays under contract while producing non-empty deterministic synthesis findings and evidence trails.

Boundaries preserved

  • Output size, string length, and list limits remain enforced.
  • Evidence order remains the impact-ranked order selected by the context allocator.
  • Synthesis remains deterministic with llm_used=false.
  • No raw logs, provider overflow payloads, secrets, or unbounded artifacts are introduced.

Verify

contextlattice_synthesis_pack "what matters before the next decision?" --project contextlattice --pretty
contextlattice_context_boundary --pretty

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track sheawinkler/ContextLattice

Get notified when new releases ship.

Sign up free

About sheawinkler/ContextLattice

Private-by-default memory and context layer for agents with Go/Rust runtime, staged retrieval across fused data backends, and long-horizon context continuity.

All releases →

Related context

Earlier breaking changes

  • v3.17.3 Agent guidance now mandates the `/agents/tasks` route family.
  • v3.17.3 Task worker now checks approval before any execution steps.

Beta — feedback welcome: [email protected]