Skip to content

Latest commit

 

History

144 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

מִשְׁכָּן · MISHKAN

A personal SWE R&D harness that lives inside Claude Code.

45 specialist agents · six teams · one research pipeline · one growing knowledge graph


MISHKAN turns Claude Code into a standing engineering organisation. Quality and security aren't requested from the model — they're enforced by the environment: path-scoped rules, pre-write security hooks, structural separation of generation from review. The knowledge graph (Cognee) persists what you learn so sessions pick up where the last one stopped. A code-structure graph (Graphify) answers "who calls X, who depends on Y" at 88× less token cost than loading source files.

It's personal, opinionated infrastructure built around one engineer's standards. To make it yours, replace docs/engineer/profile.md and re-sync — nothing else hardcodes the author.

v0.2.7 — agent fleet, rules, hooks, installer stable. Unified semantic mishkan <object> <verb> CLI (D-015) with engineer-gated curated promotion (knowledge curate, D-016), user-editable model-tier routing (model, D-017), and a confirm-gated knowledge reset. Cognee knowledge stack (per-project work · memory :7777 · curated :7730) + Graphify code graph. Observability stack (watchd + TUI) as two uv tool-installable packages.


Install

Requires Claude Code + Node ≥ 18.

npx mishkan-harness install
mishkan status
mishkan observability install      # optional: daemon + TUI (needs uv)

Full guide: docs/usage/01-installation.md.

First session

cd <project>
claude                    # starts in exploration mode — Nehemiah + Bezalel
/mishkan-init             # scaffold: spec chain → docs/ → Cognee → Sprint S0

/sprint-close at a milestone. /mishkan-resume restores state next session. Details: docs/usage/02-project-init.md.


The teams

Nehemiah (PM — scope, delivery, sprint) and Bezalel (CTO — architecture, standards, quality bar) route everything. Six teams, each Lead → Specialists → QA → Reporter:

flowchart TD
    You(["You · the engineer"]) -->|talk| MS[["MAIN SESSION = leadership<br/>Nehemiah · PM + Bezalel · CTO<br/>the one orchestrator — delegates one level deep, then synthesises"]]

    MS -->|Task| CH["Chosheb<br/>Design / UX"]
    MS -->|Task| PA["Panim<br/>Frontend"]
    MS -->|Task| YA["Yasad<br/>Backend & data"]
    MS -->|Task| MI["Migdal<br/>Infrastructure"]
    MS -->|Task| SE["Sefer<br/>Documentation"]
    MS -->|Task| RP["Research pipeline<br/>Jakin → Ezra → Caleb → Shaphan → Shemaiah → Baruch"]

    MM["Mishmar · Security<br/>cross-cutting — gates every team"]
    MM -. gates .-> CH
    MM -. gates .-> PA
    MM -. gates .-> YA
    MM -. gates .-> MI
    MM -. gates .-> SE

    MS -. reads / writes .-> KN
    subgraph KN["Knowledge — 4 surfaces"]
      direction LR
      GFY[("Graphify<br/>code structure · per-project")]
      WORK[("cognee work<br/>per-project · isolated")]
      MEM[("cognee-memory :7777<br/>session memory · shared")]
      CUR[("cognee-curated :7730<br/>reference library · shared")]
    end

    MS -. stateful ops stay yours .-> HANDS["your hands<br/>git push · prod · schema migrations"]
Loading

One leadership session (Nehemiah + Bezalel) delegates one level deep to the six teams + the research pipeline, then synthesises. Within each team: Lead → Specialists → QA → Reporter (QA & Reporter structurally separate — no agent grades its own work). Mishmar's security gate crosses every team. It reads/writes four knowledge surfaces, and stateful operations — git push, production, schema migrations — stay in your hands (the asymmetric AI/human boundary). Diagrams render on GitHub.

Team Hebrew Domain
Chosheb cunning work Design & UX
Panim face Frontend
Yasad foundation Backend & data
Mishmar guard Security (cross-cutting)
Migdal tower Infrastructure & ops
Sefer scroll Documentation (pull-based)

A shared research pipeline (Jakin → Ezra → Caleb → Shaphan → Shemaiah → Baruch) is invokable by any agent that hits an unknown. All 45 names + biblical sources: docs/design/MISHKAN_agent_aliases.md.


Knowledge stack

Wired by /mishkan-init into each project's .mcp.json:

Cognee — semantic knowledge graph. Per-project isolated work store (own port, Ladybug) + shared session memory (cognee-memory, :7777) + cross-project curated reference library (cognee-curated, :7730). Docker-based, pinned, SOPS-managed secrets. Three pillars wired per project by /mishkan-init (D-007 + D-012).

mishkan knowledge configure        # wizard: LLM provider + credentials + .env
mishkan knowledge-stack up         # memory :7777 + curated :7730 (guided; preflights config, seeds curated)

Guide: payload/mishkan/cognee/README.md · docs/usage/04-memory-layer.md.

Graphify — deterministic code-structure graph (D-008 + D-009). Indexes a project's full AST into a queryable graph. For structural questions ("who calls X", "what depends on Y") it costs ~1.8k tokens per query — 88× cheaper than loading the source tree. Runs as a PreToolUse advisory: before every structural Read or Grep, agents see a palette of four surfaces (Graphify, Cognee work, Cognee curated, literal content) with token costs and staleness signals so they pick the cheap path first. Auto-detected and wired by /mishkan-init.

mishkan code-graph scan            # build/refresh for the current project
mishkan code-graph status          # node/edge count, last scan time

Observability

Two Python packages (uv tool-installable): a daemon (mishkan-watchd) that tails every session's event bus and a Textual TUI (mishkan-watch) with 8 tabs — Live, Agents, Workflows, Knowledge, Activity, Org-Ref, Usage, Skills. Cross-session, cross-project, near-zero overhead.

mishkan observability install      # install both packages
mishkan-watch                      # opens TUI, auto-starts daemon if absent
mishkan-watchd start|stop|status   # manual daemon control

Guide + event schema: docs/design/MISHKAN_observability.md.

Workflows

Beyond the agents, MISHKAN ships dynamic JavaScript workflows that orchestrate multiple subagents in parallel — fan-out/synthesize, pipeline, judge panel, adversarial verify, loop-until-X.

Org-level (10): mishkan-sprint-close, mishkan-deep-research, mishkan-codebase-audit, mishkan-migration-wave, mishkan-architecture-panel, mishkan-release-readiness, mishkan-init, mishkan-blast-radius, mishkan-knowledge-gap-discovery, mishkan-standards-rollout.

Team-level (8): chosheb-feature-ship, panim-ds-rollout, yasad-data-migration-wave, yasad-schema-evolution, mishmar-security-gate, migdal-infra-change, migdal-dr-drill, sefer-release-notes.

Governed by hard caps (10 org + 4 per team) and PM+CTO co-ownership per ADR D-010. Catalogue + cost expectations: payload/mishkan/workflows/README.md.


Slash commands (inside a Claude Code session)

Command Purpose
/mishkan-init Scaffold a project — spec chain, docs, Cognee, Sprint S0
/mishkan-resume Restore sprint state + open blockers
/sprint-close Team reporters → aggregate → docs pull → graph promote
/mishkan-org-reference Print the 45-agent org inline
/code-graph status|open|scan Inspect / open / refresh Graphify graph
/skills <task> Skill-discovery router (3-bucket result)
/mishkan-skills-reindex Rebuild skill index from disk
/mishkan-skills-misses Aggregate miss-log for threshold tuning
/eval-baruch Run Baruch contract eval
/dep-audit Cross-project dependency + supply-chain audit
/promote Promote a learning into Cognee by blast radius
/sefer-pull Trigger documentation pull

CLI commands (from any terminal)

mishkan help                                        # full reference
mishkan install                                     # install/refresh into ~/.claude
mishkan uninstall                                   # remove harness (keeps CLAUDE.md + rules)
mishkan uninstall --purge                           # also remove y4nn-standards.md
mishkan knowledge configure                         # wizard: LLM provider + Cognee .env
mishkan knowledge curate                            # approve research-found resources into curated (D-016)
mishkan knowledge reset                             # wipe stores → re-seed curated baseline (destructive)
mishkan model show|set|reset                        # re-tier agents per-agent/team/all — survives updates (D-017)
mishkan observability install                       # install daemon + TUI only (needs uv)
mishkan status                                      # install state, profile, version
mishkan org show [--json]                                # print the 45-agent org
mishkan code-graph [status|open|scan]               # inspect the project's Graphify graph
mishkan-watch                                       # open observability TUI (auto-starts daemon)
mishkan-watch --no-autostart                        # TUI only, no daemon fork
mishkan-watchd start|stop|status                    # manual daemon lifecycle

Customisation

The harness serves the engineer described in docs/engineer/profile.md. Swap in your own (keep the section structure), then:

~/.claude/mishkan/scripts/sync-profile.sh

Refreshes the runtime copy and audits references. Nothing else hardcodes the author. See docs/engineer/README.md.


Repository layout

bin/mishkan.js              installer (dependency-free)
payload/
  mishkan/                    agents, skills, rules, hooks, commands, templates, config, scripts, ontology
  user/                       user-level CLAUDE.md + standards rule (placed if absent)
  install/                    hook fragment merged into settings.json
docs/
  engineer/                   canonical engineer profile (replaceable)
  design/                     architecture, decisions, ontology, token model, observability
  usage/                      01-install … 12-skill-discovery

Key design docs

Doc Covers
Architecture 5 layers, 6 teams, knowledge model
Agent aliases 45 agents + biblical sources
Decisions Locked build decisions
Cognee ontology Knowledge graph schema
Token optimisation Context cost model
Observability Daemon + TUI event schema
Workflows Dynamic workflow catalogue

Usage guides: docs/usage/.


License

MIT — use it, fork it, make it serve your own engineering.

Built by >_theY4NN · github.com/Y4NN777

About

Run a complex SDLC with AI as a disciplined, multi-agent team — design, spec, architecture, security, build, review, and docs — instead of a single chat.

Topics

Resources

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages