Skip to content

ContextOS 2.3.0 & contextos-mcp 0.4.0

Choose a tag to compare

View all tags
@kok-o kok-o released this 03 Oct 14:28
v2.3.0
97ea455

Overview

ContextOS 2.3.0 and companion package contextos-mcp 0.4.0 shift context management from monolithic instruction manuals to a lean bootstrap with task-proportional selection.

  • Core package: contextos-agents 2.2.0 -> 2.3.0
  • MCP companion: contextos-mcp 0.4.0 (renamed from @contextos/mcp)
  • Runtime execution: Remains opt-in experimental
  • Client workflows: External auto-routing in Cursor remains in pilot status

Key Changes

  1. Lean Bootstrap: Heavy engineering workflow, roles, and minimalism guidelines are relocated to reference documents. In Cursor, only the bootstrap rules remain permanently active; additional skills attach on demand via ctx.js resolve. Automatic background routing remains under client pilot.
  2. Proportional Skills & Security Priority: Simple tasks receive a concise workflow; deep implementation attaches extended guidance. Security instructions survive soft-budget overflow and profile exclusions.
  3. Workspace Discovery Fix: Foreign clones and runtime worktrees are excluded from package discovery. On the verification checkout, package count dropped from 502 to 3, eliminating reproduced ENOBUFS buffer errors.
  4. Adapter & MCP Alignment: All six adapters honor custom entrypoints and prefer skill.v2.yaml. Manifestless local skills hash and load correctly. MCP context assembly preserves complete instruction bodies and reports sources, hashes, omissions, and budget overflows.
  5. Safe Read-Only MCP Inspection: Default status and workspace inspection avoid creating sessions, agents, or worktrees. Execution and filesystem mutations remain opt-in experimental.
  6. Verified Upgrade & Rollback: Upgrade from core 2.2.0 / MCP 0.3.1 and checkpoint rollback verified across Windows, Linux, and macOS. On the tested migration fixture, all 325 configuration files restored cleanly while preserving user rules.

Calibration Benchmarks (GPT-6.1 Sol / high reasoning)

  • Compact selection was 32.2% cheaper than full instruction injection ($0.76 vs $1.12) with 71.8% fewer input tokens.
  • Compact selection was 35.0% more expensive than vanilla prompts without instructions ($0.76 vs $0.56).
  • Accuracy: All variants resolved 60/60 test cases with no measured quality advantage (ceiling effect on this regression corpus).

Release Governance & Boundaries

  • Already in 2.2.0: 36 catalog skills, stack presets (--preset), CLI and offline distribution.
  • Not in 2.3.0: AST graphs and multi-agent swarm orchestration.
  • Publication Process: Sequential publication of both packages with automated retry and safe recovery after partial releases. Post-release acceptance verifies clean installation, upgrade, and rollback from the public npm registry across Windows, Linux, and macOS.
Assets 2
Loading