Releases: fairley46/juno
Release list
v1.2.0 — Technical Track
What's new
Technical users now get met where they are.
Juno v1.2.0 introduces a personalized learning path for technical learners — assessing AI familiarity upfront and adapting the path instead of running everyone through the same 15 modules regardless of background.
Changes
Personalized path assessment — new upfront question captures role implicitly, two follow-ups build a personalized module path, path shown to learner before any module starts
Skippable modules (02, 04, 06, 09, 13) — in-module check-ins replace full teaching when learner demonstrates prior knowledge, exercises still run
Module 01 fast-track for technical users — skips psychological safety framing, straight to how Juno works + 4-part prompt + accountability habit, target 5 minutes not 8
Learning style evaluation — non-technical: 4 questions unchanged / technical: 1 question
Technical track exercise variations (05, 07, 08, 12) — scenario-based agent setup, threat model exercise, least-privilege design from scratch, deeper code discussion
PROGRESS.md — new fields: AI Familiarity and Skip Modules
v1.1.0 — Golden Paths + Getting Started
What's new in v1.1.0
Golden Paths
Juno now ships with org-configurable prompt templates — pre-built patterns for common tasks that learners can reach for instead of prompting from scratch.
org/golden-paths.mdscaffolded automatically onnpm run setup- 5 pre-built templates included: plain-language summary, side-by-side comparison, first draft for review, policy/compliance check, and meeting prep
- Juno walks learners through their org's golden paths in Module 14
- Juno surfaces the right template opportunistically throughout sessions when a learner's task matches a pattern
- Admins can add, remove, and customize paths — see CUSTOMIZATION.md for instructions
Module 5 renamed
"OpenCode Modes" → "AI Tool Modes: Plan, Edit, and Agent" — the content was always tool-agnostic, the name now reflects that
Getting started improvements
- Prerequisites now include install links for Git, Node.js, and all 6 supported agents
- "Not sure which to pick?" callout — non-technical users directed to Claude.ai Projects (no install, runs in browser)
- setup.js console output updated throughout
Docs and polish
- README sample dialogue uses
Juno:speaker label - All remaining stale OpenCode references cleaned
- PROJECT_CONTEXT.md and CUSTOMIZATION_BACKLOG.md updated
Getting started
git clone https://github.com/fairley46/juno.git
cd juno
npm run setupOpen in your AI agent of choice. See README for platform instructions.
v1.0.0 — Juno AI Tutor
Juno — AI Tutor v1.0.0
First stable release. Juno now runs inside any AI agent, not just OpenCode.
What's new
Tool-agnostic
- Works with OpenCode, Claude Code CLI, Gemini CLI, OpenAI Codex CLI, Cursor, VS Code, and Claude.ai Projects
- README has a platform quick start table — pick your agent and go
- Model recommendation: use the most capable model your tool supports; older models may not follow Juno's behavioral protocol as reliably
Engineering role adaptation
- Juno asks one role question at session start: technical/engineering or business/operations
- Technical learners get precise language, deeper mechanics, no beginner-level analogies — across all 15 modules
- Non-technical learners: no change to the existing experience
Vocabulary and prompt foundations woven in
- Module 2: key terms defined (prompt, context window, model, tokens) + plain-English section on model differences
- Module 14: iterative/collaborative prompting — how to refine with the AI, not at it
- Module 1: four-part prompt framework (goal, source, format, review boundary) introduced at session start so learners have it from day one
What's unchanged
15 modules. 15 exercises. The same tutor protocol that has been working since v0.4.0. Nothing removed, nothing reordered.
Getting started
git clone https://github.com/fairley46/juno-ai-tutor.git
cd juno-ai-tutor
npm run setupThen open in your AI agent of choice. See README for platform-specific instructions.
v0.5.0 — Learning Styles, Full Exercise Coverage, CLI Removal
What's New
Learning Style Evaluation
Juno now assesses how each learner learns — before any module begins. Using a 4-question Honey & Mumford evaluation (Activist, Reflector, Theorist, Pragmatist), Juno determines a learner's dominant style, saves it to PROGRESS.md, and adapts her teaching approach for every module that follows. No blending — evaluate first, then teach to that style throughout.
Full Exercise Coverage (15/15)
All 15 modules now have exercises. Added 8 new exercises:
- Module 03 — Trust calibration: apply the two-question framework to a realistic AI output
- Module 04 — Agent or assistant: classify 4 work scenarios and name your safeguard
- Module 05 — Mode selection: choose plan/edit/agent for a real scenario and describe your review step
- Module 06 — Data check: classify 6 information types as safe or not safe for a prompt
- Module 08 — Permission audit: apply least privilege to an email-access scenario
- Module 13 — Token budget: trim an over-stuffed prompt to its essential elements
- Module 14 — First workflow: write a real four-part prompt for something on your plate this week
- Module 15 — Next 30 days: commit to 3 use cases with review steps and an escalation contact
All 15 modules in manifest.json now have an exercise field so Juno looks up the right file directly.
CLI Removed
The JavaScript CLI (archive/cli/, test/) is gone. This was always a workspace-context product — Juno runs the onboarding conversationally in OpenCode. The CLI was dead weight. Only setup.js (scaffold org config) and export.js (generate completion reports) remain.
Open Items
- Org config still needs to be filled in before rolling out to a real team (
org/org-context.md,org/escalation.md,org/approved-mcps.json) - A few OpenCode URLs in
further_readingneed verification - Tool-agnostic variant (Claude.ai Projects, Cursor) is on the roadmap
v0.4.0 — Juno, Tutor Model, Three-Layer Architecture
What's New in v0.4.0
This release is a major upgrade across every layer of the product — the tutor model, the content, the information architecture, the repo, and the experience.
Juno
The tutor now has a name. She introduces herself as Juno at the start of every session.
Tutor Model — From Facilitator to Tutor
The AI is no longer a facilitator that presents content and moves on. Juno actively diagnoses understanding and adapts.
- Elicits prior knowledge before each module — asks what the learner already thinks, teaches to close that gap
- Names the wrong mental model first — clears the old picture before presenting the right one
- Finds analogies in the learner's own work context — asks before assuming what will land
- Replaced "does that make sense?" with explain-it-back and diagnostic scenario checks
- Reads engagement — slows down and tries a genuinely different approach when a learner seems lost
- Connects backward — each module opens by linking to the previous one
- Uses visuals proactively — draws ASCII diagrams, Mermaid diagrams, and tables whenever a concept has a relationship, flow, or comparison
- Explains her own teaching approach at session start so learners know what to expect
- Reminds learners of their options at natural pause points — questions welcome, pace is theirs
Three-Layer Information Architecture
Training, further reading, and live research are now three distinct layers — separated intentionally so the golden path stays stable while learners can access current context when they want it.
Layer 1 — The golden path
15 modules, stable, org-consistent, every learner gets the same sequence.
Layer 2 — Further reading
Curated links per module in manifest.json. Admin-controlled. Offered after each module rating as an optional offer, clearly separate from training. All 15 modules have annotated further reading links covering Anthropic docs, MCP registry, OWASP, NIST, Anthropic Cookbook, and more.
Layer 3 — Live research
After each module, Juno searches for current industry developments on demand. Explicitly framed as exploration, not instruction. Modules 05, 06, and 09 have search_topics in manifest.json for focused, relevant search queries. The offer is unconditional — Juno always makes it and discloses honestly if search fails.
Module Content
Module 01 — Welcome and Safety
- New "How This Program Works" section at the top — explains Juno's teaching approach before any content starts
- "What This Doesn't Mean" (job replacement framing) replaced with "What Gets Better" — acceleration, capability, time freed for higher-value work
Module 05 — OpenCode Modes
- New "Industry Context" section: the plan/edit/agent control question is a live governance debate across the entire industry — not just an OpenCode feature
Module 06 — Local vs Web Execution
- New "Industry Context" section: regulation tightening (GDPR, EU AI Act), on-premise AI becoming real, vendors competing on data handling
Module 07 — Data Safety and Shadow AI
- New "Enterprise Tenancy" section: org tenant vs personal account, side-by-side comparison table, how to verify you're inside the right environment
- Same model through the wrong access path = shadow AI, even if the model name is identical
Module 09 — MCP
- New "Industry Context" section: MCP ecosystem post-2024, rapid adoption, connector fragmentation coming, governance habits at scale
Module 13 — Tokens
- New worked token budget example: vendor contract scenario showing 13,550 token unfocused approach vs 950 token scoped approach — same question, 14x less noise
Exercises
Exercise 07 — Rewritten from a 4-line stub to a full two-part exercise:
- Part 1: paste test with six items requiring reasoned explanation
- Part 2: tenancy verification — learner checks their own actual access path
AGENTS.md Protocol
- No bullet-point summaries — Juno teaches section by section, never compresses to key ideas
- No invented commands — no "type X to advance," everything is conversational
- Live research offer is unconditional — always made, handles failure inline
search_topicsfrom manifest.json used for focused search queries
New: Completion Export
npm run exportReads PROGRESS.md and generates a plain text completion report — learner name, start date, completed modules, remaining count. Saved as a local file, gitignored.
New: Example Org Config
org/examples/ now contains two fully filled-in fictional examples (Meridian Analytics):
org-context.example.md— approved use, disallowed use, data classification, client data rules, MCP approval processescalation.example.md— full escalation contacts across AI program, privacy, security, legal, governance with response times
Admins can see what "filled in" actually looks like before editing their own files.
Repo
- README rewritten to world-class standard: badges, quick start at top, "What Makes This Different" as a scannable table, single consolidated diagram
- GitHub repo About section: description, topics, homepage
- SECURITY.md updated — removed CLI-era references, reflects current architecture
- THREAT_MODEL.md updated — removed CLI-era references, added wrong-access-path risk and policy drift risk
- Backlog fully refreshed
Getting Started
git clone https://github.com/fairley46/opencode-onboarding.git
cd opencode-onboarding
npm run setupOpen the folder in OpenCode. Juno takes it from there.
Requirements: OpenCode (desktop or terminal) · Node.js 20+
For Admins
Fill in org/org-context.md and org/escalation.md before distributing to learners. See org/examples/ for fully worked examples of what these files should look like.
Known Gaps
- OpenCode-specific URLs in
manifest.jsonfurther_reading should be verified (docs, extensions, community pages) - Full 15-module end-to-end test with all v0.4.0 changes not yet complete — testing in progress
See docs/CUSTOMIZATION_BACKLOG.md for the full open backlog.
v0.3.0 — Tutor Model, Three-Layer Architecture
What This Is
OpenCode Onboarding is an AI-facilitated training program that runs entirely inside OpenCode. The AI acts as a tutor — not a chatbot reading slides — guiding learners through 15 modules at their own pace, inside the tool they are learning to use.
This release represents a significant upgrade across three areas: the teaching model, the content depth, and the information architecture.
What's New in v0.3.0
Tutor Model Upgrade
The AI is no longer a facilitator that presents content and checks in. It is a tutor that actively diagnoses understanding and adapts.
- Elicits prior knowledge before each module — asks what the learner already thinks, then teaches to close that gap
- Names the wrong mental model first — "Most people assume X; that's not quite right" — because clearing the old picture is half the work
- Finds analogies in the learner's own work context — asks before assuming what will land
- Replaced "does that make sense?" with explain-it-back and diagnostic scenario checks — learners demonstrate understanding, not just acknowledge it
- Reads engagement — if answers are short or uncertain, tries a different kind of explanation, not the same one again
- Connects backward — each module opens by linking to the one before it
Three-Layer Information Architecture
Training content, further reading, and live research are now three distinct layers — separated intentionally so the golden path stays stable while learners can access current context when they want it.
Layer 1 — The golden path (15 modules)
Stable, authoritative, org-consistent. Every learner at your org gets the same sequence. This is what makes shared vocabulary and shared habits possible.
Layer 2 — Further reading
Curated links per module, controlled by admins in manifest.json. Offered after each module rating as an optional offer — clearly separate from the training. Each link includes a one-line description of what it is and why it matters. Covers Anthropic docs, MCP registry, OWASP, NIST, Anthropic Cookbook, and more.
Layer 3 — Live research
After each module, the tutor can look up current industry developments, real examples, and how other orgs are approaching the topic. Explicitly framed as exploration, not instruction. Does not affect module progression. Conditional on web search being available in the session.
Curriculum Expanded: 12 → 15 Modules
Three new modules added to cover topics that were missing:
- Module 3 — When to Trust AI Output: Two-question trust calibration framework, hallucination patterns, verification matrix (Checkability × Stakes)
- Module 5 — OpenCode Modes: Plan, Edit, and Agent: Control ladder for plan/edit/agent modes, governance framing, how to choose
- Module 7 — Data Safety and Shadow AI: Never-paste table, shadow AI explainer, what leaves the machine and what doesn't
Module Content Depth
Several modules were expanded from thin overviews to substantive lessons:
- Module 4 (Agent vs Assistant): Added vendor review worked example, "Why Agent Can Feel Scary" section, day-to-day mental model
- Module 6 (Local vs Web Execution): Added data boundary diagram, four questions to ask before a workflow, enterprise governance context
- Module 8 (Guard Rails and Permissions): Added least privilege explanation, permissions table, what review actually looks like before/during/after
- Module 11 (Asking Questions Across Tools): Full rewrite with concrete worked example, four-element framework, conflicting sources section
- Module 14 (First Useful Workflows): Full rewrite with four-part prompt framework and three copy-and-adapt workflow templates
- Module 15 (Troubleshooting and Next Steps): Five-category diagnostic framework, reset prompt template, when to escalate
Exercise Rewrites
Exercises 01, 02, 03, and 04 were rewritten from stubs to structured tasks with specific completion criteria and facilitator notes. Terminal handoff protocol added for hands-on exercises — learner steps out to the terminal, comes back to the same session.
AGENTS.md Protocol Hardening
Critical rules added to prevent common AI failure modes observed in live testing:
- No CLI — the AI will not attempt to run terminal commands to advance the program
- One module at a time — never presents multiple modules in a single response
- Teach, do not summarize — works through content section by section
- Exercises are not optional — does not advance until exercise is complete
Repo Cleanup
- CONTRIBUTING.md and LICENSE (MIT) added
- Removed CLI-era documentation: DESKTOP_TEST_RUNBOOK.md, TEST_STRATEGY.md
- Removed orphaned directories: examples/, schemas/, facilitator/
- Archived old Node.js CLI to archive/cli/ — preserved for reference, no longer primary
- All 15 module H1 headers corrected to match file numbering after renaming
Getting Started
git clone https://github.com/fairley46/opencode-onboarding.git
cd opencode-onboarding
npm run setupThen open the folder in OpenCode. The tutor takes it from there.
Requirements: OpenCode (desktop or terminal) and Node.js 20+
For Admins
Before distributing to learners, fill in the three org config files in org/:
org/org-context.md— approved uses, disallowed uses, governance notesorg/escalation.md— who learners contact for policy and security questionsorg/approved-mcps.json— which MCP servers are approved
See the org/*.template.* files for the expected format.
Known Gaps
- Org config files still contain placeholder text — fill these in before real learner use
further_readingURLs in manifest.json should be verified before distributing (OpenCode docs URL in particular)- End-to-end test with new tutor behaviors not yet completed
See docs/CUSTOMIZATION_BACKLOG.md for the full open backlog.