Release Notes - v0.5.4.dev2

Release Date: 2026-03-12
Changes: v0.5.5.dev0 → ci/dev-canary-release

Summary

This release introduces a full skills system — ingest, route, execute, observe and self-improve skills — plus a skills-aware MCP interface, CLI commands, LLM-based scoring for skill runs, and many parser, docs and testing improvements. It adds developer-friendly tools to build and iterate on reusable 'Skill' workflows and improves reliability with multiple bug fixes and expanded tests.

Highlights

• New Skills system: SKILL.md ingestion, skill routing, execution, and learnable preferences.
• MCP tools for skills: get_skill_context, load_skill, execute_skill, observe_skill_run and more (skills are exposed as tools/resources).
• Self-improving skills: automated inspect & amendify flows and LLM-based scoring (continuous learning with scores instead of binary pass/fail).
• New CLI commands for managing skills and richer docs + examples (skills README, demos, and e2e tests).
• Numerous reliability fixes and a large test suite expansion covering skills and e2e flows.

Breaking Changes

• Skills functionality is opt-in and requires the new cognee skills package to be installed (cognee.cognee_skills). MCP exposes skill tools only when that package is available — if you rely on MCP, install/upgrade the skills package to enable tools.
• SKILL.md format and routing behavior has evolved — if you have custom SKILL.md files, validate them against the new examples (see cognee/cognee_skills/README.md and agent_instructions.md). Some older SKILL.md layouts may require small adjustments due to parser improvements and flat-file support.
• CLI: new skills-related commands were added; if you have scripts that assume the previous CLI layout, check for command names and flags changes.

New Features

• Complete skills framework: add, parse, store, load, execute and observe skills defined as SKILL.md files (cognee.cognee_skills package).
• MCP skill tools: get_skill_context (semantic ranking), load_skill (full skill payload), execute_skill (run skill through an LLM and auto-record runs), observe_skill_run, promote_skill_runs, ingest_skills, upsert_skills, remove_skill and list_skills.
• Skill resources via MCP (skill:// agent-guide/index/{skill_id}) to allow progressive discovery from agents.
• CLI support for skills: new commands to ingest, list, load and manage skills from the CLI.
• Callable API: execute_skill function available from the skills client so you can programmatically run skills from Python code.
• Self-improvement features: inspect, preview_amendify and amendify flows that let the system propose and apply improvements to skills based on observed runs.
• LLM-based scoring for skill runs: skill evaluations now return a score (not just binary), enabling nuanced learning and routing.
• Flat-file SKILL.md parsing and improved parser flexibility to support different folder layouts and easier SKILL authoring.
• Example-rich demo and e2e tests: demo skills, example scenarios (team-comms, data-extraction), and multiple e2e tests added.

Improvements

• Documentation overhaul: simplified and clearer README, better instructions for skills, added agent_instructions and skill prompt templates, and examples for timecomponent and weight mechanisms.
• Parser improvements and naming refactors to make skill parsing more robust and predictable.
• More flexible testing paths and updated .gitignore entries (including .env.example) to reduce friction for local development.
• Enhanced task patterns, models and internal skill representations to enable richer skill metadata and tracking.
• Expanded automated tests covering client load/list, e2e amendify flows, execution, inspection and parser behavior to increase confidence in the new features.

Bug Fixes

• Fixed nodeset handling and preservation of original instructions when saving (prevents losing original content during skill processing).
• Fixed MCP availability issues so MCP skill tools are reliably discoverable when the skills package is installed.
• Improved parser robustness to handle more SKILL.md variants and avoid brittle parsing failures.
• Adjusted test paths and other minor fixes to stabilize the test suite.

Technical Changes

• Large new codebase added for skills: modules, models, tasks, client, execute/inspect/amendify logic and tests (see cognee/cognee_skills/ and cognee-mcp/src/server.py).
• Server-side MCP integration introduces multiple new tool endpoints and skill resource handlers in the MCP server.
• Refactors and naming changes across skill pipeline code to simplify steps and internal APIs.
• Significant test additions (many new unit and e2e tests) and example scripts under cognee/cognee_skills/example.

Generated by Cognee Release Notes Generator on 2026-03-12