Loading the presentation. If it does not open, download the complete HTML file and open it in a browser with JavaScript enabled.

HVE Core
updates

RPI skills, HVE Builder
and plugin installation.

April 1 to October 21, 2026

  1. 01
    ResearchFind evidence and compare approaches
  2. 02
    PlanAgree on tasks, then critique the plan
  3. 03
    ImplementMake changes and record the results
  4. 04
    ReviewCompare the work with the plan
microsoft/hve-coreIncludes scripted examples

Changes since April

  1. APR 01

    Task agents

    Handoffs between task agents. Research already uses a subagent.

    Repository snapshot
  2. JUN 24

    RPI skills

    Phase workflows move into reusable skills.

    PR #1953
  3. JUL 11

    HVE Builder

    A skill for creating, reviewing and testing customizations.

    Introduced in PR #2438
  4. AUG 16

    One plugin

    The packages combine into one HVE Core plugin.

    PR #2712
  5. SEP 08

    Automatic RPI

    RPI Agent adds participation choices and automatic loops.

    PR #2859
  6. Since September 10SEP 18

    Advisory helpers

    Helpers only suggest. HVE Builder reviews instead of testing.

    PR #2902
  7. SEP 28

    Optional steps

    Skip the plan critique or Review if you want. No RPI step runs a script.

    PR #3061
  8. OCT 07

    No prompt files

    48 prompts retired. Their rules now live in skills and agents.

    PR #3111
  9. OCT 19

    Parallel lanes

    Helper lanes run side by side. One critique and one Review per request.

    PR #3238

Pacific dates, spaced evenly rather than to scale. Earlier plugin support predates August.

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.

Now: 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.

Research, plan,
implement, review

Four phase skills. Each one leaves
a document the next phase reads.

  1. 01Evolution
  2. 02RPI
  3. 03HVE Builder
  4. 04Plugins

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.

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 runs one critique by default. rpi-plan-critique checks the finished plan from four angles.
Pass critique=skip to leave it out. Fixing its findings does not start another critique.

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?
  1. Wider
  2. Deeper
  3. Contrarian

Each wave can hand its open questions to helper lanes that run in parallel. A search lane runs at low effort; a lane that reads docs or code runs at max.

C# marks codebase evidence and W# an external source. Research starts balanced; set posture=focused or expansive to change it.

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.

Resolve plan gaps before implementation

plan-critique.mdBefore coding

PC-001: missing reset check

The Coverage lens found that the plan requires Reset but never says 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 the work against the requirements

PlanCritique dispositionsChanges and check resultsReview record
RPI Reviewer's candidateSuggestion

Example: the count stays at 3

Expected
Reset displays zero.
Observed
The check records a visible count of 3.
Suggested route
Fix the reset display and record a passing check.
Primary assistant's decisionRV-001

Parent Decision Record

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

In this failure example, the review ran to completion and found a defect. The reviewer is optional; the assistant can compare the evidence itself.

Choose when the agent should ask you

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

When an automatic session continues or stops

ResearchPlan + critiqueImplementReview
Required work remains within scopeStart a child task at Research, without critique or Review

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.

Child tasks skip the critique and Review unless you ask for them. The closeout says which changes were validated but not reviewed.

Helpers work in parallel lanes

Research

rpi-research

RPI Researcher, one lane per question

  • Search for reset codelow
  • Read storage docsmax

The parent reads the sources
and assigns C# and W#.

Plan critique

rpi-plan-critique

RPI Plan Critic, one lane per lens

  • Coveragelow
  • Evidencemedium
  • Executabilitymedium
  • Riskmedium

The parent verifies each candidate
and assigns PC-xxx.

Review

rpi-review

RPI Reviewer, one lane per comparison

  • Check task markerslow
  • Compare P01 requirementshigh

The parent checks the evidence
and assigns RV-xxx.

Helpers only suggest and never write files. Tags show each lane's effort. With subagents=none, the main assistant does all of it.

Settings you can pass

RPI settings, the skills that accept them and what they change
SettingUsed byWhat it changes
posture=focusedResearchStays inside a bounded task. expansive widens the search.
critique=skipPlanSkips the critique. deep asks for a deeper one.
effort=highResearch, critique, ReviewSets the reasoning effort for every helper lane.
critique-effort=lowPlanSets the effort the plan's critique lenses use.
subagents=noneEvery phase skillRuns the work in the main assistant, with no helpers.
Ask to skip ReviewRPI AgentMoves on to Follow-up using validation evidence.

A skipped critique or Review is recorded as skipped, never as a pass. Effort levels depend on what the host supports.

Scripted walkthrough

Example: build a snack-voting app

Create reusable
customizations

HVE Builder writes and reviews skills,
instructions, agents and prompts.

  1. 01Evolution
  2. 02RPI
  3. 03HVE Builder
  4. 04Plugins

From Prompt Builder to HVE Builder

Prompt Builder Retired

Prompt work had its own agent, prompt and instruction file. Three skills then kept the old names working.

Jul 26, #2474Oct 5, #3113

The agent files went first. The prompt-builder, prompt-analyze and prompt-refactor aliases were removed on October 5.

HVE Builder Current

Create, improve, refactor, replace, review or validate a customization. Name the mode directly.

SkillsAgentsSubagentsInstructionsPrompts

With no mode, it creates, improves and refactors together. Review or validate on its own never edits source.

Review replaced behavior testing

BeforeA separate tester checked a frozen version

  1. Discover
  2. Author
  3. Static review
  4. Freeze removed
  5. Assess behavior removed

NowThe assistant reviews its own candidate

  1. 01Scope
  2. 02Author
  3. 03Validate
  4. 04Review pass
  5. 05Resolve

Modes combine

Create, improve and refactor run together by default. Review or validate alone stays read-only.

Checks come first

Local checks pass before the review starts, so the review looks at a finished candidate.

Stop when stuck

Another correction cycle needs a clear purpose and evidence that it will make progress.

Removed on September 18 in #2902: the hve-builder-tester skill, HVE Artifact Tester, freeze, sandbox, fidelity levels and the behavior gate.

The assistant decides what passes

HVE Builder ReviewerOptional and read-only, in fresh contextSuggested verdict: ReviseAdvisory only
Reviewer findings and the assistant's decisions
FindingReviewer suggestionAssistant's decision
R-001High / requiredA stop rule in a reference file still contradicts the new skip-with-reason behavior.AcceptedFixed in the reference file. A targeted re-review closed it.
R-002Low / advisoryOne sentence reads as unclear.RejectedThe sentence fits the skill's purpose and repository conventions.

Recorded verdict: Pass for the revision after R-001 was closed.

Adapted from the example in PR #2902, where HVE Builder updates a code-review skill.

Add domain criteria to RPI research

S snack-evidence / SKILL.md Illustrative excerpt
---
name: snack-evidence
description: Use during research
  when a task involves office snack stock.
  Separates counted stock from estimates
  and reports missing counts as unknown.
---
## Evidence criteria
Record the source and observation time.
Report a missing count as unknown.

## Boundary
Read-only; do not order snacks.
The research phase decides what to accept.

The description is the contract

Research, planning and review look for skills and subagents whose description names their phase. They decide from the description alone.

A helper subagent returns sources and suggestions. It writes nothing; the phase assigns the IDs and writes the artifact.

Name the phase and the trigger.
Say what goes in and what comes back.

Scripted walkthrough

Example: create a research skill

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 review the instructions that guide future research.

Example: Snack Evidence

HVE Builder creates the research skillRPI can use it on a matching task

Install HVE Core
as a plugin

Use its customizations in VS Code
and Copilot CLI.

  1. 01Evolution
  2. 02RPI
  3. 03HVE Builder
  4. 04Plugins

August's change combined the existing packages into one plugin. Prompt files were retired in October.

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.

Reconstructed UI

VS Code: install from source

Install through Copilot CLI

Terminal Walkthrough only
# Register the marketplacecopilot plugin marketplace add microsoft/hve-core
# 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.

Updates follow main

Refresh the marketplace, then update the plugin. Re-add an older tag or release-branch registration.

Clients last checked September 10, 2026: Copilot CLI v1.0.83 and VS Code 1.137.0. The deck only shows these commands.

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.

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, October 21, 2026