Command → Agent → Skill → Hook → Crystal
- Agents hand off explicitly — none try to be everything.
- Hooks auto-load learnings and guard writes.
- Crystal survives compaction, session restart, and new repos.
A Jedi council for your codebase — 14 role-specialist agents, 24 domain commands, 11 skills, a SQLite-backed learning crystal that compounds your corrections across sessions, and a token economy that makes every $ of Claude Code usage visible.
14 Agents 24 Commands 11 Skills 13 Hook Scripts 6 Events SQLite + FTS5 Cost tracking MIT
NEW IN v0.3.0
Token economy. Every session's $, token counts, cache hit rate, and per-model split are logged and browsable via /holocron:cost. Set a /holocron:budget. Prune MCP overhead with /holocron:mcp-audit. Ranked savings proposals from @cost-analyst.
| @architect | Design, ADRs, tradeoffs |
|---|---|
| @frontend-engineer | React · Vue · Svelte · UX |
| @backend-engineer | APIs · data · services |
| @qa-engineer | Strategy · gaps · flakes |
| @a11y-auditor | WCAG 2.2 AA |
| @perf-engineer | Web Vitals · N+1 |
| @security-reviewer | OWASP · authz · crypto |
| @threat-modeler | STRIDE / LINDDUN |
| @sre-engineer | Docker · k8s · CI · SLOs |
| @data-engineer | SQL · dbt · pipelines |
| @cost-analyst | Ranked $ savings from usage data NEW |
| @debugger | Hypothesis-driven triage |
| @code-reviewer | Staff-level PR review |
| @docs-writer | READMEs · runbooks · ADRs |
Each has a rubric and a "not my jurisdiction" delegation table.
Every correction you capture becomes a durable rule — stored in SQLite with full-text search, indexed by bm25, loaded on every session start, and matched against every prompt you send.
/holocron:learn "<rule>"~/.holocron/learnings.db
Follows you across every project.
<repo>/.holocron/learnings.db
Lives with the repo. Commit or ignore per team.
"The holocron was the ultimate Jedi repository of knowledge. What the archives preserve, the archives protect."
/holocron:onboard | Draft a repo-appropriate CLAUDE.md |
|---|---|
/holocron:plan | Plan-mode scaffold |
/holocron:develop | Research → Plan → Implement → Review |
/holocron:commit | Quality-gated commit |
/holocron:review | Staff-level review of the diff |
/holocron:handoff | Session handoff doc |
/holocron:learn | Capture a correction |
/holocron:recall | FTS search the crystal |
/holocron:insights | Correction heatmap NEW |
/holocron:worktree | Parallel git worktree setup NEW |
/holocron:doctor | Self-check install |
/holocron:a11y | WCAG audit on UI diff |
|---|---|
/holocron:perf | Web Vitals · N+1 · indexes |
/holocron:test-gap | Prioritized missing tests |
/holocron:flaky | Flake-triage tree |
/holocron:sec-scan | SAST + secrets + CVE |
/holocron:threat-model | STRIDE on the diff NEW |
/holocron:adr | MADR-style decision record |
/holocron:permission-tune | Analyze denials · suggest allow/deny rules |
|---|---|
/holocron:llm-gate | AI verdict before destructive ops |
/holocron:cost | $ and tokens across sessions · cache hit rate NEW |
|---|---|
/holocron:budget | Set per-session / daily / monthly caps NEW |
/holocron:mcp-audit | MCP token footprint · prune unused servers NEW |
/holocron:context-size | Snapshot context · find trim candidates NEW |
The Stop hook parses every session's transcript and writes per-session token & cost records to ~/.holocron/sessions/. You can browse them, set budgets that warn you in-session, prune MCP overhead, and get ranked savings from @cost-analyst.
# $ + tokens + cache hit rate
/holocron:cost --breakdown
# Top sessions by cost
/holocron:cost --by-session
# Per-session + daily caps
/holocron:budget \
--session 2 --daily 10
# Stop hook warns at 80% / 100%
# MCP token footprint
/holocron:mcp-audit
# Biggest context offenders
/holocron:context-size
# Ranked savings
@cost-analyst
# Large Read, no offset/limit
# → nudge to use a range
# Tool output > ~8k tokens
# → nudge to narrow
# Cache hit rate low in /cost
# → tip on what's busting it
Pricing hardcoded for Claude 4.x (Opus / Sonnet / Haiku). Update scripts/holocron-usage.js when rates change.
/holocron:onboard
/holocron:plan
/holocron:develop <feature>
/holocron:commit
/holocron:learn "<rule>"
@frontend-engineer <task>
/holocron:a11y
/holocron:perf
/holocron:test-gap <file>
@backend-engineer <task>
/holocron:sec-scan
/holocron:perf
/holocron:adr "<title>"
@qa-engineer <task>
/holocron:test-gap <file>
/holocron:flaky <test>
@sre-engineer <task>
/holocron:llm-gate "push to main"
/holocron:permission-tune
@threat-modeler <feature>
/holocron:threat-model
/holocron:sec-scan
// 1 — add the marketplace
/plugin marketplace add atuljha23/holocron
// 2 — install the plugin
/plugin install holocron@holocron
// 3 — reload
/reload-plugins
// 4 — self-check
/holocron:doctor
First SessionStart runs npm install automatically if the SQLite native binding is missing. Node ≥ 18 required.
Structural inspiration: rohitg00/pro-workflow (MIT). Holocron reimplements the persistent-learning idea from scratch and builds role-specialist content on top.
Plugin authoring reference: Anthropic's Claude Code plugin docs.
Star Wars, Jedi, Sith, and Holocron are trademarks of Lucasfilm Ltd. This project is unaffiliated thematic homage.