aibast-easy-mode

AIBAST Easy Mode

Once you have read this file (the person sent one line naming its public URL; nothing was downloaded or attached), you own the whole workshop: discovery, local testing, Draft deployment, Preview validation, and the evidence verdict. If the line that pointed you here also named a solution or workshop ("…and run the X workshop for me"), treat that as the first request and start at once.

There is one mode. The person never picks "with Brainstem" or "Copilot only"; you decide the engine yourself in the first minute (see "Pick the engine") and never mention the choice unless they ask.

Public source

• Repository: microsoft/aibast-agents-library
• Workshop branch: main
• Registry path: registry.json

Resolve the branch through the GitHub API and pin one immutable commit SHA before downloading anything.

Persist the active run in .aibast/easy-mode-state.json, including solution, revision, engine (brainstem or copilot), local test results, Copilot Studio identity, status, and published: false. When the direct engine is used, also persist the Copilot Studio plugin repository, installed plugin ID, resolved revision or installed version, PAC CLI version, and verification status.

Pick the engine (you, silently, in the first minute)

Brainstem is the person's personal, on-device training AI. It works alongside Copilot, remembers the workshop, and hot-loads specialized instructors while Copilot remains the familiar work surface. Prefer it; fall back without fuss.

1. Check http://localhost:7071/health. If it answers, the engine is brainstem. Go to "Show it in VS Code".
2. If it does not answer, start or install it yourself and wait up to two minutes for /health:
   • prefer ~/.copilot/bin/brainstem start, then brainstem on the PATH
   • otherwise the public installer: Mac or Linux curl -fsSL https://microsoft.github.io/aibast-agents-library/install.sh | bash, Windows (PowerShell) irm https://microsoft.github.io/aibast-agents-library/install.ps1 | iex
   • if /health reports Copilot is not signed in, open http://localhost:7071/login for the person and say: "Sign in with your GitHub account, then tell me when you're done."
3. If /health still does not answer (no permission to install, the port is held by something else, the installer failed), the engine is copilot. Record why in the state file in one line and continue. Never ask the person to fix it and never stop here.

Show it in VS Code (engine brainstem, when you are running in VS Code)

Put the Brainstem's own folder in the Explorer so the person sees their AI as files they own: run code --add ~/.brainstem/src/rapp_brainstem (Windows: code --add %USERPROFILE%\.brainstem\src\rapp_brainstem), or the VS Code command Workspaces: Add Folder to Workspace with that folder. Then open http://localhost:7071 in VS Code's built-in browser beside this chat: use your open-browser tool if you have one, otherwise the VS Code command Simple Browser: Show with that address. Say in one sentence: "This is your Brainstem; soul.md is its personality and every *_agent.py in agents/ is a skill." Outside VS Code, open it in the person's normal browser instead.

Choose from the library

When the user asks what is available, or asks for a workshop without naming one (for example Show me the AIBAST library, then give me the workshop I pick using Easy Mode and test it for me):

1. Fetch registry.json from the pinned commit.
2. List every entry whose _catalog_kind is solution, grouped by category written as a plain industry name, one line per workshop: display_name — description. Do not paste JSON or file paths.
3. Ask one question: which workshop they want. Accept a number, a name, or a description, resolve it to exactly one registry entry, and confirm the display_name back to them.
4. Continue as Give me <solution> using Easy Mode and test it for me with that entry.

Never choose for the user, and never download a workshop before they choose.

Build and test

When the user says Give me <solution> using Easy Mode and test it for me:

With engine brainstem:

1. Resolve @aibast-agents-library/workshop from the immutable registry.
2. Download its exact _file, verify _sha256, and import it through POST http://localhost:7071/agents/import.
3. Send the user's request to POST http://localhost:7071/chat.
4. Continue executing every Brainstem handoff without asking the user to steer.
5. Brainstem's generic workshop engine must discover the requested package, hot-load the business agent, and run every locked local case.

With engine copilot:

1. Run the repository-owned, fail-closed setup step python3 tools/easy_mode_preflight.py --json and require passed: true. Stop when it exits nonzero or reports passed: false, and report the remediation it printed. Record the installed plugin/version and PAC CLI version in .aibast/easy-mode-state.json.
2. Resolve the requested solution from immutable registry.json.
3. Fetch its deployment.json, export-manifest.json, portable *_agent.py, tests/demo_cases/<slug>.json, reviewed instructions, knowledge, skills, Copilot Studio source, and every tool named by the manifest.
4. Verify the portable source SHA.
5. Create an isolated workspace with only required files and BasicAgent.
6. Import the class named by deployment.json.
7. Execute every locked case with its operation and arguments.
8. Validate every must_include and must_not_include marker.

Either way: persist status: tested, return the verdict with exact case totals, and offer the next message: Deploy it into Copilot Studio for me.

Required Copilot Studio plugin (engine copilot, before any deployment)

The preflight owns plugin discovery, installation, and PAC remediation. Its explicitly reviewed CLI sequence is:

1. Check whether the mcs-assistant Copilot Studio capabilities are already available.

2. If they are missing, execute these GitHub Copilot plugin-manager commands in order:

   copilot plugin marketplace add microsoft/copilot-studio-plugin
   copilot plugin install mcs-assistant@copilot-studio-plugin
   

3. Recheck copilot plugin list, then resolve microsoft/copilot-studio-plugin through the GitHub API and record its immutable default-branch commit SHA or the installed plugin version in .aibast/easy-mode-state.json.

4. Inspect pac --version, including version output from a nonzero exit. The plugin requires a Power Platform CLI version newer than 2.9.3. When PAC is too old and .NET is available, run:

   dotnet tool update --global Microsoft.PowerApps.CLI.Tool
   

   Recheck PAC and otherwise stop with the reported remediation.

5. Verify the installed plugin exposes the Copilot Studio manage, describe, initialize, and architect capabilities. Do not continue to deployment when any required capability is unavailable.

Use only microsoft/copilot-studio-plugin and the installed ID mcs-assistant@copilot-studio-plugin. Do not substitute a fork or execute unreviewed plugin source directly. The plugin is experimental, so validate all generated YAML and stop on schema or PAC validation errors.

Deploy

When the user says Deploy it into Copilot Studio for me:

With engine brainstem:

1. Send that exact request to Brainstem /chat.
2. Resolve it from Brainstem's persisted active workshop.
3. Continue every returned handoff.
4. Use the real Copilot Studio front door for Preview validation.
5. Return captured evidence to Brainstem using its callback schema.
6. Do not stop until Brainstem reports status: complete.

With engine copilot:

1. Resolve it from .aibast/easy-mode-state.json.
2. Re-verify mcs-assistant@copilot-studio-plugin, its recorded revision or version, and the required PAC CLI version.
3. Resolve the active PAC environment without asking when one is selected.
4. Use the plugin's read-only describer for an existing agent and its deterministic initializer only when a new target project is required.
5. Use the plugin architect to assemble or migrate the exact reviewed instructions, knowledge, skills, model, and modern agentic-loop YAML.
6. If the recorded schema exists, use the plugin manager to clone and reconnect automatically, then pull and reconcile it.
7. Remove web search and unapproved tools, then validate the generated source and PAC project before any remote write.
8. Use the plugin manager to push the Draft. Never invoke publish.
9. Open the real Draft and run every locked case in a fresh Preview chat.
10. Validate all markers and capture evidence.

Either way: persist status: complete and report exact totals, the engine used, plugin provenance when the direct engine deployed, and published: false.

Rules

• Own all terminal, file, PAC CLI, browser, and evidence work.
• Never ask the user for URLs, environment IDs, commands, or test markers.
• Never ask the user to choose an engine, a lane, or a mode.
• Never mix assets from different Git revisions.
• Never deploy through the direct engine without the verified Microsoft Copilot Studio plugin and supported PAC CLI.
• Never invent a missing result or success-shaped fallback.
• Never retry until a response happens to pass.
• Stop at Draft. Never publish.