v2.0.1 — Installation fix & mascot rework

What's fixed

• Installation instructions — the github: shorthand doesn't resolve; correct flow is now documented:
  ```
  /plugin marketplace add https://github.com/Alexander-Tyagunov/magician
  /plugin install magician@magician-marketplace
  ```
• Wizard cat mascot — reworked ASCII art: happy `^.^` expression, visible smile `(u)`, curling tail, uniform 15-char line width that centers correctly in GitHub view mode
• Codex marketplace path — fixed canonical path in `.agents/plugins/marketplace.json`
• Codex install docs — clarified enablement steps in `.codex/INSTALL.md`

v2.0.0 — Codex Support

What's new

Codex plugin support

• Added .codex-plugin/ adapter layer — all 21 Magician skills now available in Codex
• Shared codex-adapter.md translation layer maps Claude Code tools to Codex equivalents
• Claude Code behavior is unchanged

Bug fixes & quality (v1.3.0)

• Fixed invalid hook JSON decisions (ask/suggest) in access-tracker.sh and pattern-detect.sh
• Fixed wrong input key in format.sh hook
• Added design personality commitment gate to /conjure (bans generic defaults)
• Added agent color frontmatter to all 4 agents
• Added $ARGUMENTS passthrough to /portal
• Added worktree isolation gate (GATE 3.5) to /manifest
• Added autonomous continuation rule to /manifest
• Added docs update step to /seal
• Added browser auto-open to /certify
• Fixed marketplace.json plugin name (root cause of 404 at claude.com/plugins/magician)

Installation

/plugin install github:Alexander-Tyagunov/magician

v1.2.0

Fix plugin marketplace config to align with official Anthropic requirements. See CHANGELOG.md for details.

v1.1.1 — README docs patch

Magician v1.1.1 — Release Notes

Released: 2026-04-21

Fix: README up to date

This is a documentation patch release.

• Added /magic to the skills table — the skill shipped in 1.1.0 but was missing from the README skills reference
• Updated version badge — was stuck at 1.0.0, now reflects 1.1.1

No behavior changes. All 21 skills are now documented in README.md.

---

For the full 1.1.0 feature notes (the /magic skill itself), see RELEASE-NOTES-1.1.0.md.

v1.1.0 — /magic Research & Consulting

Magician v1.1.0 — Release Notes

Released: 2026-04-21

New: /magic — Research, Analysis & Consulting

The highlight of this release is the /magic skill — a full research and consulting workflow that brings structure to any investigation, analysis, or information-gathering task.

What /magic does

Invoke /magician:magic (or just type words like "research", "investigate", "analyze") and /magic walks you through:

1. Source selection — choose web, library docs, local files, or all three
2. Research depth — quick overview, standard, or deep dive
3. Actual research — web search, context7 docs, document reading, synthesis
4. Output format — markdown report, exec summary, tables, visual (/conjure), or raw notes
5. Persistence — show only, save to file, or save + git commit
6. Navigate & Suggest — after delivering findings, intelligently proposes the next step

Every step uses the AskUserQuestion tool so you see a structured choice UI — no guessing what Claude is going to do next.

Context-aware skill navigation (Phase 5)

After delivering findings, /magic classifies the research topic and proposes the most relevant next magician skill — with an adapted question that reflects what was actually found:

Research type	Suggested next skills
Financial / business	/conjure (dashboard), /blueprint (action plan)
Technical feature / API	/blueprint (plan), /conjure (design), /forge (implement)
Bug / incident	/unravel (debug), /sentinel (security angle)
Security / vulnerability	/sentinel (full scan), /blueprint (remediation plan)
Performance / scalability	/accelerate (profile)
Architecture / system design	/conjure (visual spec), /blueprint (plan)

The question text adapts to the findings — e.g. "Solid analysis — Q3 numbers are in. What's the next move?" rather than a generic prompt. When the user is done, /magic closes with a warm one-sentence summary and a friendly sign-off.

Auto-invocation

You don't need to type /magic explicitly. The session hook now detects research-intent keywords in your messages and auto-activates the skill:

research · investigate · analyze · explore · examine · assess · evaluate · discover · audit · study · "find out" · "look into" · and more

Academic & scientific research support

/magic fully supports student and researcher workflows:

• Academic web search — when the topic is a thesis, paper, or scientific study, web search queries are automatically targeted at academic databases: Google Scholar, arXiv (STEM/CS), PubMed (biomedical), IEEE Xplore (engineering), ACM Digital Library (CS). Source credibility is assessed (peer-reviewed > institutional > press).
• Citation-aware output formats — literature review, annotated bibliography, research outline with section assignments, or summary with formatted citations
• Citation style selection — APA, MLA, IEEE, Harvard, Chicago/Turabian, or informal URLs
• Academic navigation — Phase 5 suggests researching another angle or designing a research poster/slide deck with /conjure

context7 vs. document reading — clearly separated

/magic distinguishes between two fundamentally different source types:

• Tech Library Docs (context7) — for software questions only: npm packages, Java/Spring/Maven libraries, Python modules, framework APIs, version compatibility. Checks if context7 is installed and offers to install it if missing.
• My Documents / Files — reads actual files you have locally (Excel, PDF, Word, CSV, financial reports, articles, meeting notes) using the built-in Read tool. No external service, no MCP — just reads the file and understands it in context.

/magic also classifies the topic before showing source options: if you're clearly asking about a business document or report, the "Tech Library Docs" option is hidden automatically so it doesn't cause confusion.

Key design principle

Every decision point in /magic uses AskUserQuestion — the user always sees explicit choices and drives the research direction. This is intentional: magic assists, the user decides.

---

What's next

• /oracle — persistent research memory across sessions
• /brief — auto-generate executive summaries from chronicle session learnings

Release of core plugin functionality

Magician v1.0.0 — Release Notes

Released: April 20, 2026

Full-stack SDLC plugin for Claude Code. Inspects your project, assembles the right knowledge automatically, orchestrates parallel agents, learns from every session, and ships clean code — from idea to merged PR, autonomously.

---

What's included

20 Skills

Skill	Purpose
/manifest	Full autonomous SDLC — idea to merged PR with 4 human gates
/conjure	Structured design dialogue with visual browser companion; no code until spec approved
/blueprint	Converts an approved spec into a TDD task plan with parallelism map
/forge	Executes one blueprint task using strict TDD
/ward	Enforces red→green→refactor discipline, one behavior at a time
/unravel	Systematic debugging with mandatory hypothesis preflight
/certify	Full verification loop — tests, types, lint, build, Playwright browser check
/summon	Spawns parallel subagents seeded with the full skill registry
/orchestrate	Drives full multi-agent execution from a blueprint in parallel waves
/scrutinize	3 specialist reviewers in parallel: correctness, security, simplification
/absorb	Processes scrutiny findings by severity; never skips Critical/High
/portal	Creates a git worktree for isolated feature work
/seal	Ships a feature — simplifier, certify, commit, PR, CI monitor, merge
/almanac	One-time workspace init — CLAUDE.md, .workspace/, MCP suggestions
/chronicle	Views and manages session learning entries from the Stop hook
/sentinel	Security scan — OWASP Top 10, credentials, injection surfaces, deps
/accelerate	Performance profiling with mandatory baseline-first measurement
/deploy	CI/CD pipeline management for GitHub Actions, GitLab CI, CircleCI
/inscribe	Creates a new reusable skill; auto-triggered by the pattern detector
/autopsy	Blameless post-mortem — timeline, 5-Whys, action items

---

4 Specialist Agents

• reviewer — correctness and logic review
• simplifier — redundancy and over-engineering review
• verifier — evidence-based verification before any success claim
• sentinel — security-focused review

---

Full Hook System

Hook	Purpose
SessionStart	Detects stack, assembles lore, loads chronicle context, shows wizard cat
UserPromptSubmit	Pattern detection — offers /inscribe at threshold
PreToolUse	sentinel-guard.sh scans every Bash command for injection patterns
PostToolUse	Enforces permission and audit rules
PreCompact	Saves session state to .workspace/local/session.md
Stop	Writes chronicle entry, updates patterns.json
SubagentStart / SubagentStop	Seeds subagents with skill registry; collects status
WorktreeCreate	Initializes isolated worktree environment

---

Dynamic Project Inspector

Detects your stack automatically on every session start — no manual configuration needed. Supports:

• Web: Next.js, React, Vue, Astro, SvelteKit, Remix, Angular
• Backend: Node/Express, FastAPI, Django, Spring Boot, Rails, Laravel, Go, Rust
• Mobile: React Native, Flutter, iOS, Android, Capacitor
• Data / AI: Python ML stacks, Jupyter
• Other: Godot, Tauri, and more

Polyglot stacks (e.g. Next.js + FastAPI + Go) get full combined coverage automatically.

---

Self-Learning Chronicle

Session intelligence grows without replay. Each session end:

1. Stop hook writes a chronicle entry from git log + diff
2. patterns.json is updated with repeated prompt patterns
3. Next session loads the last 3 entries as additionalContext
4. At 3 repetitions a skill offer appears; at 5 an auto-draft is triggered

---

Security Infrastructure

Security is enforced at the infrastructure level, not as advice:

• Hard deny rules in settings.json — blocks pipe-to-shell, eval, credential file access before any prompt sees them
• sentinel-guard.sh PreToolUse hook — scans every Bash command for injection patterns and the lethal trifecta (private data + network + execution in one command)
• magician-scan — standalone CLI for CI integration: ./bin/magician-scan .
• Workspace isolation — .workspace/local/ is always gitignored; per-machine secrets never reach git

---

Team Workspace

.workspace/
├── shared/           ← git committed (team sees this)
│   ├── context.md
│   ├── roadmap.md
│   ├── decisions/
│   ├── specs/
│   ├── mockups/
│   ├── plans/
│   └── postmortems/
└── local/            ← always gitignored
    ├── prefs.md
    └── session.md

---

Installation

/plugin install github:Alexander-Tyagunov/magician

Then initialize your workspace:

/almanac

---

License

MIT © Alexander Tyagunov