Loading the presentation. If it does not open, download the complete HTML file and open it in a browser with JavaScript enabled.
MICROSOFT / HVE CORE / SEPTEMBER 2026

HVE Core
updates

RPI skills, HVE Builder
and plugin installation.

APRIL 01 → SEPTEMBER 10, 2026
  1. 01
    ResearchFind evidence and compare approaches
  2. 02
    PlanAgree on tasks and resolve critique findings
  3. 03
    ImplementMake changes and record the results
  4. 04
    ReviewCompare the work with the plan
microsoft/hve-coreIncludes scripted examples
01 / EVOLUTION

Changes since April

The starting snapshot and four changes from repository history.

APR 01

Task agents

Task prompts connect agents through handoffs. Research already uses a subagent.

Repository snapshot
JUN 24

RPI skills

Phase workflows move into reusable skills.

PR #1953
JUL 11

HVE Builder

A skill for creating, reviewing and testing Copilot customizations.

Introduced in PR #2438
AUG 16

One plugin

Existing plugin packages are combined into a single HVE Core plugin.

PR #2712
SEP 08

Workflow controls

RPI gains automatic loops. HVE Builder can reassess a corrected version.

PRs #2859 / #2858

Pacific dates. HVE Builder's first merged skill appears on July 11; plugin support predates August.

APRIL BASELINE / CURRENT CONTRACT

Task-agent handoffs became reusable skills

APRIL / SEPARATE TASK AGENTS
  1. Task Researcher/task-research
    Hand off to planning
  2. Task Planner/task-plan
    Hand off to implementation
  3. Task Implementor/task-implement

Task Researcher used a named research subagent. Agent and prompt files carried the workflow.

CURRENT / STANDALONE PHASE SKILLS
/rpi-research/rpi-plan/rpi-implement/rpi-review

Each skill defines its phase's work and evidence. The RPI Agent can coordinate the phases, but is optional.

RPI / CHAT SETUP

Switch Chat to RPI Agent

Before starting the coordinated workflow

  1. Open the agent dropdown below the VS Code Copilot Chat input.
  2. Select RPI Agent.
  3. Describe your task or invoke a phase skill such as /rpi-research.

The agent coordinates phase handoffs and asks how you want to participate.

RPI Agent is optional. The phase skills can also run independently. This dropdown is a reconstruction.

02 / RPI

What each RPI skill produces

01

Research

Evidence, alternatives,
readiness

/rpi-research
02

Plan

Goals, requirements,
dependencies

/rpi-plan
03

Implement

Source changes,
validation evidence

/rpi-implement
04

Review

Findings, decisions,
routed follow-up

/rpi-review

Planning includes one critique. rpi-plan-critique checks the completed plan.
The planning assistant addresses its findings before implementation.

RESEARCH

Check the research recommendation

M snack-queue-research.md Illustrative excerpt
## Executive Summary
Use a local queue for the first demo.

## Findings
Shared-device votes should clear on reset.
Evidence: C1, W1
Confidence: high for local behavior

## Recommendation and Alternatives
Session-only state / persistent storage

## Planning Readiness and Next Step
Ready to plan a local-only demo
Markdown Example using the current template

Questions to ask

  1. What evidence would change this recommendation?
  2. Why was the alternative rejected?
  3. What do we still need to learn before planning?
Wider→Deeper→Contrarian

C# identifies codebase evidence.
W# identifies an external source.

PLAN

Agree on what the plan will deliver

M snack-queue-plan.md Illustrative excerpt
### [ ] P01: Make snack voting visible
Goals:
* Teammates can see the current queue.

#### [ ] P01-T01: Add local voting
Goals:
* One vote changes the displayed count.
Requirements:
* FR-001: voting stays local.
* FR-002: reset clears the session.
Details:
* No account, backend, or live service.
References:
* Research: local-scope finding
Dependencies:
* None

Review the task before coding

Goals describe the result. Requirements make it checkable.

Details and References give the implementer context.

Use the dependencies and before/after diagrams to spot missing work.

Refer to task IDs when requesting changes.
CRITIQUE / CHANGES

Resolve plan gaps before implementation

plan-critique.mdBefore coding

PC-001: missing reset check

The plan requires Reset, but does not say how to check the displayed count.

Verdict
Revise
Planning assistant's action
Add a check that Reset clears votes and displays zero. Record PC-001 as resolved.
changes.mdAfter coding

Record what was done

## Completed Work
P01-T02: reset and render added.

## Validation Record
Reset clears votes: Passed
Displayed count is zero: Passed

Include the check results, plan changes and any unfinished work.

Example records. Critique examines the plan; the changes record documents implementation and its checks.

REVIEW

Review the work against the requirements

PlanCritique dispositionsChanges and check resultsReview record
Review worker's findingRV-001

Example: the count stays at 3

Expected
Reset displays zero.
Observed
The check records a visible count of 3.
Resolution
Correct the reset display and record a passing check.
Primary assistant's decision

Parent Decision Record

Finding
Accepted
Next action
Return the defect to implementation.
Review execution / outcome
Complete / Defects found

Separate failure example: Review is Complete, but its outcome is Defects found. Retained decisions still require user input.

OPTIONAL RPI AGENT / LIFECYCLE COORDINATION

Choose when the agent should ask you

Example selection. Safety confirmations, workflow gates and human-review requirements apply in every mode.

AUTOMATIC RPI

When an automatic session continues or stops

Research→Plan + critique→Implement→Review
↳ Required work remains within scopeStart a new child task at Research

Continue

The current phase's gates pass and the next action is authorized.

Pause

The agent needs a retained user decision, missing evidence or a safety confirmation.

Stop

The request is complete, the user stops, or repeated attempts cannot make progress.

Follow-ups cover work required by the request. Optional cleanup stays separate.

RPI HELPERS

Select helpers for the work at hand

Available skill or subagent↓Name or description matches research↓Expertise fits the question and scope↓Primary assistant combines the evidence

You can supply an alternative

A research skill can provide domain criteria. A registered research subagent can investigate an independent question.

  • RPI Researcher writes one evidence file.
  • RPI Planner revises one assigned phase.
  • RPI Review Builder writes one review record.

The assignment does not expand permissions. The primary assistant keeps the decisions.

RPI / SCRIPTED WALKTHROUGH

Example: build a snack-voting app

03 / HVE BUILDER

Create reusable
customizations

HVE Builder helps author skills, instructions,
agents and prompts, then checks their quality.

HVE BUILDER HISTORY

From Prompt Builder to HVE Builder

HISTORICAL ENTRY POINT

Prompt Builder

Prompt-engineering work used a dedicated agent, prompt and instruction file.

JUL 11 / #2438 → JUL 26 / #2474

HVE Builder was introduced first. The old entry-point files were retired on July 26.

CURRENT SKILL

HVE Builder

Create or improve a customization, refactor it, or request a read-only review.

SkillsAgentsSubagentsInstructionsPrompts

The prompt-builder, prompt-refactor and prompt-analyze skills now route to HVE Builder.

HVE BUILDER WORKFLOW

How HVE Builder checks a customization

01Discover→02Author / revise→03Validate +
static review→04Freeze→05Assess
behavior

Gather requirements

Read repository conventions and relevant extensions. Use rpi-research when a decision needs more evidence.

Check the right version

Major changes need behavior evidence for the version being delivered. Supported skips for smaller changes are recorded.

Separate editing and testing

The primary assistant changes the source. The tester assesses the frozen version without editing it.

SEPTEMBER REFINEMENT / #2858

Keep test evidence tied to its version

CANDIDATE AFreeze

Assess this version
and report findings

Assistant corrects the source→Using the reported findings
CANDIDATE BRefreeze

Repeat affected checks
and required assessment

Keep A's report with A. After a justified correction, assess B using evidence for that version.

Behavior testing defaults to simulation. Native execution needs an explicit request and the required safety controls.

EXTENDING RPI

Add domain criteria to RPI research

S snack-evidence / SKILL.md Illustrative excerpt
---
name: snack-evidence
description: Use during research on
  office snack queues to assess inventory
  evidence, freshness and uncertainty.
---
## Evidence criteria
Separate counted stock from estimates.
Record the source and observation time.
Report missing counts as unknown.

## Boundary
Read-only; do not order snacks.
The primary assistant recommends next steps.

Describe when to use the skill

The description should name the research task and the evidence it helps assess.

For a separate investigation, provide a research subagent with a stable name, host registration and a specific output file.

Instructions match paths.
Skills match tasks.
The assistant assigns work to subagents.
HVE BUILDER / SCRIPTED WALKTHROUGH

Example: create a research skill

USING BOTH WORKFLOWS

Use HVE Builder to extend your RPI work

RPI

Build the application

Investigate the task, agree on the plan, implement it and review the results.

Example: the snack-voting app.
HVE BUILDER

Create a reusable skill

Write and assess the instructions that guide future research.

Example: Snack Evidence.
HVE Builder creates the research skill→RPI can use it on a matching task
04 / PLUGINS

Install HVE Core
as a plugin

Use its customizations in VS Code
and Copilot CLI.

August's change combined the existing packages into one plugin.

PLUGIN PACKAGING

Add the marketplace, then install the plugin

REGISTERmicrosoft/hve-core

Marketplace source

→
INSTALLhve-core@hve-core

Plugin @ marketplace

→
INVOKE/rpi-research

A skill in the enabled plugin

plugin.json

Lists the HVE Core components included in the plugin.

marketplace.json

Points to the repository-root plugin. Registering the marketplace makes it discoverable; installation is a separate action.

INSTALLATION WALKTHROUGH / RECONSTRUCTED UI

VS Code: install from source

COPILOT CLI / EXAMPLE COMMANDS

Install through Copilot CLI

Terminal Walkthrough only
01 / REGISTER THE MARKETPLACEcopilot plugin marketplace add microsoft/hve-core
02 / INSTALL THE PLUGINcopilot plugin install hve-core@hve-core
Review source and trust prompts. Confirm installed/enabled status in your client.

Use it in both clients
VS Code discovers CLI-installed plugins when it can read the same user-home directory.

Choose a version when needed
These commands follow the current repository source. Use a supported ref or tag to pin a version.

September 10, 2026 reference: Copilot CLI v1.0.83 and VS Code 1.137.0. Commands are shown only.

PLUGIN DISCOVERY

Where the installed plugin appears

INSTALL WITH COPILOT CLI

CLI + VS Code

~/.copilot/installed-plugins/

VS Code reads the CLI's installed-plugin directory in the same user home.

Supported by VS Code 1.137 source
INSTALL WITH VS CODE

VS Code only*

VS Code manages this installation separately. Its default flow does not register a CLI plugin.

Default installation behavior

* Manual path sharing can differ. Discovery requires access to the same home directory, chat.plugins.enabled, an enabled plugin and policy permission. It does not sync installations across machines.

GETTING STARTED

Try a skill
on your next task

/rpi-research/hve-builder

Use RPI to investigate and implement a change,
or HVE Builder to create a reusable customization.

microsoft/hve-coreSource snapshot / 10 SEP 2026