engsys

A meta-harness that installs a complete AI engineering organization into any project.

Works natively across Claude Code, GitHub Copilot CLI, and Google Antigravity — materializing 15 specialist personas, 24 composable stack packs, Adaptive Spec-Driven Development (/design-change), a compounding Agentic Memory Engine, and an optional always-on autonomous fleet from one declarative config file.

Live Interactive Explainer ↗ — or open index.html locally — for a visual tour of the mental model, the 15 specialists, the pipeline, and the graduation ladder.
Deployment Playbook (DEPLOY.md) — standalone, step-by-step runbook for both humans and AI agents.

Want your agent to install it? Paste this into Claude Code, Copilot CLI, or Antigravity: Read https://raw.githubusercontent.com/eric-sabe/engsys/main/DEPLOY.md and deploy Stage 1 into this repository.

Mental Model: Model vs. Coding Harness vs. Meta-Harness

When people talk about "AI coding agents," three distinct layers often get blurred together. engsys is the top layer:

Layer	Role	Examples	What It Does — and Where It Stops
1. Model (The Brain)	Stateless reasoning & code generation	Claude Opus / Sonnet, GPT / Codex, Gemini, Grok, DeepSeek	Answers prompts and calls tools, but starts every session with amnesia and has no engineering process.
2. Coding Harness (The Workbench)	Interactive CLI / IDE runtime	Claude Code (`claude`), GitHub Copilot CLI (`copilot`), Google Antigravity (CLI `agy`, config `antigravity`)	Gives a model a terminal, file editor, subagents, MCP servers, and lifecycle hooks (`PreToolUse`, `PostToolUse`, `SubagentStart`) — but out of the box it is still one generalist assistant with no org chart, no stack packs, and no cross-PR memory.
3. Meta-Harness (The Engineering Org — `engsys`)	Portable engineering operating system	`engsys` (`engsys install`, `fleet`)	Sits above your coding harness and across models. Defines who does the work (15 specialists), which stack rules apply (24 packs), how work moves (spec → isolated worktree → independent review → merge), and what the team remembers (Agentic Memory) — then compiles that organization into whichever coding harness you run.

The Three-Part File Taxonomy Inside `engsys`

Inside the meta-harness, every file engsys manages belongs to one of three parts so you never rewrite prompts when switching projects or clouds:

Part	Question	Stability	Lives in
Persona	Who does the work	Stable across every project	`core/agents/` & `optional-agents/`
Capability	Which stack / cloud / language	Chosen per project (composable packs)	`stacks/**/skills/`
Project Facts	What is true about this repository	Unique per project	Generated `CLAUDE.md` / `GEMINI.md` / `.github/copilot-instructions.md`

The cloud architect (Melvin) never changes; you select cloud: aws or azure or gcp or cloudflare in engsys.config.yaml and he automatically loads the matching pack.

Where to Start: The 4-Stage Graduation Ladder

You do not need background daemons, "monsters," a second machine, or a GitHub App to use engsys. Start at Stage 1 in 3 minutes on your laptop, and graduate only when your project's throughput demands it:

Stage	What You Deploy	Monsters Needed?	Always-On Host?	Setup Time
Stage 1	Interactive Team — 11 core + 4 opt-in specialists, 24 stack packs, `/design-change` SDD, `/implement-issue`, `/pre-push`, and 51 curated lessons inside Claude Code, Copilot CLI, or Antigravity.	No	No	~3 min
Stage 2	Cross-Family Review & Agentic Memory — External review models (`Codex`, `Grok`, `DeepSeek`) enforce `reviewer family != author family`, and PR history feeds subagent memory briefs.	No	No	~10 min
Stage 3	Single-Host Autonomous Fleet — Merge Monster pilots `mm:ready` PRs through CI to merge, Maintenance Monster triages Dependabot/CodeQL & runs the Fleet Health Desk, supervised by `launchd` (macOS) or `systemd`/`cron` (Linux).	Yes (`mm`, `maintain`)	1 macOS/Linux host (`tmux`)	~30 min
Stage 4	Federated Multi-Host Fleet — GitHub-ref compare-and-swap batons fence monsters across machines, Resource Broker pools heavy build/test slots, and `fleet verify` enforces release-tree SHA & settings trust.	Yes (`+ broker`)	1–N hosts	~60 min

Platform & Toolchain Support (macOS · Linux · Windows):

Stages 1 & 2 (Interactive Team, SDD, Multi-Provider Review, Agentic Memory): Natively supported on macOS (arm64 /opt/homebrew/bin & x64 /usr/local/bin), Linux (/usr/bin, /usr/local/bin, Linuxbrew), and Windows (via WSL2 or Git Bash).

Stages 3 & 4 (Autonomous Fleet & Monsters): Natively supported on macOS (tmux + launchd via fleet install-jobs; install node@22, bash 5+, gh, jq, tmux, and any HARNESS_FALLBACK CLIs into /opt/homebrew/bin or ~/.local/bin rather than shell-only nvm/fnm shims) and Linux / WSL2 (tmux + systemd user timer or cron; all monster, baton, and supervisor scripts include GNU/Linux + BSD/macOS dual paths).

Automatic SessionStart Host Toolchain Context (core/lib/host-toolchain.mjs): At SessionStart, engsys detects the host OS, userland dialect (BSD on macOS vs. GNU on Linux/WSL2), /bin/bash 3.2 vs. Homebrew Bash 5+, and any non-login PATH gaps (common in GUI harnesses like Antigravity/VS Code or launchd), injecting a 1-line guidance note so agents use portable flags and resolve CLI paths on the first try. See the full per-OS matrix in DEPLOY.md.

Stage 1 — Interactive Team (Start Here — 3 Minutes)

If you have an active project and Claude Code, GitHub Copilot CLI, or Google Antigravity, four commands give you the entire interactive engineering team.

Option A — Install into your project (CLI or Clone)

# 1. Install the zero-dependency CLI (Node >= 20.11) — or `git clone https://github.com/eric-sabe/engsys`
npm install -g engsys

# 2. Scaffold engsys.config.yaml in your repository
cd your-project
engsys init

# 3. Edit engsys.config.yaml (pick harness: claude | copilot | antigravity | all and stack packs), then materialize:
engsys install --into .
engsys verify --into .      # confirm all managed files match the lockfile

# 4. Open your project in Claude Code, Copilot CLI, or Antigravity and run the one model-driven step:
#    /naturalize

Running from a git clone instead of global npm? Replace engsys with node /path/to/engsys/install (or run npm link once inside the clone). Existing setup in the repo? engsys install adopts existing CLAUDE.md, settings, agents, and Copilot/Cursor rules without clobbering them, and snapshots originals into .claude/.engsys-backup/ so engsys uninstall --into . restores your repo byte-for-byte. See docs/install-scenarios.md.

Option B — Install as Claude Code Plugins (Zero Files Copied Into Repo)

For Claude Code users who want zero agent/skill files committed to the target repository, engsys ships as a Claude Code plugin marketplace (engsys core + one engsys-<pack> plugin per stack pack). Pin it in .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "engsys": {
      "source": { "source": "github", "repo": "eric-sabe/engsys", "ref": "v2.0.0" },
      "autoUpdate": false
    }
  },
  "enabledPlugins": {
    "engsys@engsys": true,
    "engsys-typescript@engsys": true,
    "engsys-web@engsys": true
  }
}

Install once per machine (claude plugin marketplace add eric-sabe/engsys#v2.0.0, then claude plugin install engsys@engsys and your chosen pack plugins). Commands and agents are namespaced (/engsys:implement-issue, engsys:isabelle).

What Stage 1 Gives You Immediately

Multi-Harness Materialization (harness: claude | copilot | antigravity | all):
- claude (default): .claude/{agents,commands,skills,hooks,workflows}/, CLAUDE.md, .claude/settings.json, and .mcp.json.
- copilot: .github/copilot-instructions.md, .github/agents/*.agent.md, .github/skills/<name>/SKILL.md, .github/hooks/engsys-hooks.json, .github/copilot/settings.json, and .vscode/mcp.json.
- antigravity: GEMINI.md, AGENTS.md, .agents/agents/*.md, .agents/skills/<name>/SKILL.md, .agents/rules/*.md, .agents/workflows/*.md, .agents/hooks.json, and .agents/plugins/engsys/plugin.json.
- Universal hook I/O normalization (core/lib/harness-io.mjs) translates hook events, allow/deny/context decisions, and harness-matched Co-Authored-By commit trailers (Claude <noreply@anthropic.com>, GitHub Copilot <noreply@github.com>, Google Antigravity <noreply@google.com>) across all three harnesses automatically.
15 Specialist Personas (11 Core + 4 Opt-In):
- Core (agents.core: all): Leith (Product/UX Specs), Melvin (Cloud & Scale Architect), Aaron (IaC & Deploys), Jody (Planner), Isabelle (Implementer), Bert (Bug Hunter), Marcelo (Testing Strategist), Patricia (Librarian / ADRs), Nyx (Offensive Security — pinned to claude-opus-4-8), Otto (LLM/Token Optimization), and Gary (Design Critic).
- Opt-In (agents.extra): Sandy (Marketing/SEO), Jos (Monetization/ASO), DeAndre (Tutor & Plain-Language Session Explainer), and Steve (Morale Scapegoat).
- Every persona carries least-privilege tools: allowlists, explicit skills: bindings, and a machine-checked ## Handback Protocol (HANDBACK: COMPLETE / HANDBACK: INCOMPLETE).
Adaptive Spec-Driven Development (/design-change):
- Risk-adaptive spec classification (core/lib/sdd-validator.mjs, core/scripts/sdd-check.mjs):
  - Tier 0 (Trivial): Skip spec ceremony for typos and isolated 1-file fixes.
  - Tier 1 (Standard): Inline micro-spec in the issue/PR with explicit touches and verification commands.
  - Tier 2 (Architectural / High-Risk): Full spec in docs/specs/<slug>.md (core/templates/spec.md.tmpl) triggered by 6 Escalation Tripwires (Data Model, Public Contract, Security/Trust Boundary, Concurrency/Leases, Performance/Scale, Cross-Boundary UX), sliced via core/scripts/spec-slice.mjs and verified at /pre-push via core/scripts/contract-check.mjs.
51 Curated Cross-Project Lessons:
- Seeded into docs/agent-lessons/library/ (or read in place in plugin mode) with structured YAML frontmatter (applies_to.{stacks,paths,agents,roles}, security: true floor, triggers).

Stage 2 — Cross-Family Review & Compounding Memory

When you want stronger verification and cross-session learning — still 100% interactive on your workstation, with no monsters required — add External Worker/Review Providers and the Agentic Memory Engine.

2.1 Cross-Family Model Review & Worker Dispatch (`Codex`, `Grok`, `DeepSeek`)

When the same model reviews code it just wrote, blind spots compound. engsys dispatches implement / review / critique / investigate packages to external model families under a strict machine-checked receipt contract (docs/multi-provider-spec.md):

# 1. One-time CLI or API key setup for the providers you want:
npm install -g @openai/codex && codex login            # OpenAI Codex CLI
curl -fsSL https://x.ai/cli/install.sh | bash && grok  # xAI Grok Build CLI (or XAI_API_KEY in .env)
# For DeepSeek or metered Grok API, add DEEPSEEK_API_KEY / XAI_API_KEY to ~/.config/engsys/env (chmod 0600)

# 2. Enable your configured providers in the repo and check readiness:
engsys enable-providers codex,grok,deepseek,anthropic --into .
engsys verify --into .

provider doctor:
  READY     codex — codex-cli 0.147.0
  READY     grok — grok-build CLI authenticated (or xAI API reachable)
  READY     deepseek — remapped to https://api.deepseek.com/anthropic
  READY     anthropic — claude 2.x

2.2 The Agentic Memory Engine (`docs/memory-spec.md`)

engsys ships a two-tier memory engine (requires Node.js >= 22.5 for node:sqlite) that makes every subagent smarter without risking prompt injection from untrusted GitHub comments:

Class A (Trusted Instructions): Loaded strictly from reviewed git refs (origin/HEAD) across core/lessons-library/*.md, docs/agent-lessons/library/*.md, docs/agent-lessons/*.md, and docs/agent-lessons-overlay.yml.
Class B (Historical Repo Evidence): Merged PR summaries, review findings, and issue outcomes stored on an append-only custom Git data ref (refs/engsys/memory/events/v1 via core/lib/memory/ref-store.mjs) with mandatory pre-write secret scanning, bot-identity (databaseId) & edit-history (userContentEdits) verification, and prompt-boundary spoofing quarantine (core/scripts/memory-sync.mjs).
Subagent Brief Injection (<= 2 KB):
- On PreToolUse (Agent|Task), core/.claude-plugin/subagent-memory-inject.mjs stashes the subagent's task prompt and target paths (0600, 30s TTL).
- On SubagentStart, core/scripts/memory-brief.mjs queries a host-local node:sqlite FTS5 cache (core/lib/memory/index-db.mjs) to inject mandatory security: true lessons, BM25/path-matched Class A lessons, and compact Class B pointers (anchor * idf * recency * outcome * flagMultiplier).
- nyx (security) is strictly isolated to Class A instructions only. Full Class B evidence is fetched on demand (memory-brief.mjs --evidence <key>) inside an untrusted HISTORICAL REPO EVIDENCE envelope.
Finding Harvester (core/scripts/harvest-findings.mjs): Clusters recurring PR review findings during /project-closeout into candidate Class A lessons.

Stage 3 — Single-Host Autonomous Fleet (Monsters + Health Desk)

When your repository produces enough PRs and security alerts that human merge-babysitting becomes the bottleneck, graduate to Stage 3 on an always-on macOS host (e.g., a Mac mini or workstation running tmux and launchd):

Merge Monster (/merge-monster · docs/merge-monster.md): Owns the merge baton. Preflights mm:ready PRs (projecting comment markers and FINDING-ACKs via mm-preflight.sh), pilots ready → CI → merge one active PR at a time, and bounces failures with root-cause diagnosis.
Maintenance Monster (/maintenance-monster · docs/maintenance-monster.md): Sole owner of Dependabot, CodeQL, secret scanning, and image scanning alerts; filters standing false-positive policies (mnt-fp.mjs) and hands verified fix PRs to Merge Monster.
Fleet Health Desk (fleet health · docs/fleet-health-desk.md):
- Probes upstream dependencies (statuspage with ETag 304 caching, observed CI rolling windows, http-health, and synthetic-command with local-auth separation) using asymmetric hysteresis (trip_fails / recover_clean) and depends_on DAG alarm suppression.
- Maintains a pinned  GitHub status board, posts calm transition-only Slack alerts, correlates inbound session friction reports (fleet health report), and injects 1–2 lines of weather context at SessionStart / UserPromptSubmit (core/.claude-plugin/weather-check.mjs) only when services are degraded or down.
Single-Watcher Efficiency, Seat-Quota Harness Fallback & Deterministic Supervision:
- Each monster runs one unified watcher (mm-watch.sh, mnt-watch.sh, broker-watch.sh) and single-call ticks (mm-tick.sh, mnt-tick.sh, broker-tick.sh) that coalesce events so a quiet 30-minute window costs 1 tool call.
- When idle context usage crosses rotate_context_tokens (clamped below 70% of the compaction window), the monster runs mm-baton.sh rotate / mnt-baton.sh rotate to record rotation requested, release its baton, and let the deterministic launchd supervisor (fleet-supervisor.sh) relaunch a fresh session (with automatic forfeited-on-rotation baton recovery if the session idles at its prompt).
- Automatic Seat-Quota Harness Fallback (HARNESS_FALLBACK=copilot,antigravity): When a primary seat hits its usage or rate limit (You've hit your limit · resets …, Usage limit reached, Copilot rate limit exceeded), fleet-supervisor.sh automatically revokes the stalled monster's baton (--reason quota-fallback), activates the next configured fallback harness (copilot → antigravity), and relaunches the monster using the same fleet bot identity (feedfrwd-fleet[bot]). After QUOTA_COOLDOWN_MIN (default 60m), the supervisor probes the primary harness and fails back automatically once the seat resets.
Deep Live Observability & Troubleshooting (fleet xray · core/fleet/lib/xray.mjs):
- fleet xray inspects all 10 runtime subsystems in one pass (host ENGSYS_DIR & PIN_DIR trust closure, GitHub App / Azure SP cert expiry, launchd jobs & active HARNESS_FALLBACK state, v2.0.0 model routing & live JSONL transcript token pressure, singleton batons & subagent agents.tsv liveness, a live 9-hook synthetic self-test, the Hook Flight Recorder (hook-events.jsonl), SQLite FTS5 & refs/engsys/memory/events/v1, SDD validators, and FHD/Federation/Broker state).
- Supports --session <name> / --role merge|maintain|e2e deep-dives, --timeline [--since 1h] chronological event reconstruction, --diagnose automated root-cause & copy-paste remediation playbooks, --watch [sec], and --json.

# Scaffold a dedicated fleet instance repo on your always-on host:
git clone https://github.com/eric-sabe/engsys ~/git/engsys
~/git/engsys/install fleet init \
  --into ~/git/acme-fleet \
  --org acme --namespace acme \
  --pin-repo owner/repo --pin-dir ~/git/repo \
  --harness claude --identity github-app

# After placing your GitHub App key (core/fleet/identity/README.md) and running mm-setup.sh / mnt-setup.sh:
~/git/acme-fleet/scripts/fleet sync           # sync checkouts + plugins in isolated worktrees
~/git/acme-fleet/scripts/fleet verify         # verify wrapper, host ENGSYS_DIR & settings integrity
~/git/acme-fleet/scripts/fleet launch         # launch tmux sessions
~/git/acme-fleet/scripts/fleet install-jobs   # load launchd supervisor & identity jobs
~/git/acme-fleet/scripts/fleet xray           # full 10-subsystem X-Ray & live 9-hook self-test

Stage 4 — Multi-Repo / Multi-Host Federated Fleet

For organizations running multiple repositories or multiple always-on hosts (docs/multi-fleet.md, docs/fleet-guide.md):

GitHub-Ref Compare-and-Swap Batons (core/lib/lease/baton.mjs): Singleton roles (merge, maintain) hold lease batons on custom Git refs (refs/engsys/leases/baton-<role>) with session-bound keepalives and fenced write guards (core/.claude-plugin/singleton-write-guard.mjs).
Resource Broker (/resource-broker · core/skills/resource-broker/SKILL.md): Built on core/skills/durable-lease/, pools finite host hardware (simulators, Docker stacks, e2e runners) across concurrent workers.
Cross-Fleet & Same-Host Messaging (fleet msg, fleet relay · docs/agent-messaging.md):
- Cross-fleet (<fleet>:<session>): Works across all three harnesses (claude, copilot, antigravity) via GitHub App-signed  comments (fleet msg send), fleet relay poll inbox delivery, monster FLEET_MSG watch events, and the tri-harness fleet-inbox.mjs hook.
- Same-fleet peer nudges (<ns>-<role>): In Claude Code, uses native SendMessage / ListAgents (configured via messaging: in engsys.config.yaml → .claude/settings.local.json crossSessionInbound: "accept", merge-monster.yml, and launch-agent-sessions.sh). In Copilot CLI and Antigravity (where no cross-tmux SendMessage socket exists), fleet msg route skips the peer socket nudge and falls back cleanly to GitHub PR comments and mm:bounced/mm:escalated labels.
Remote Control & Desktop / Web / Mobile Session Visibility (--remote-control in roster):
- Claude Code (HARNESS=claude): Each tmux session launches with --name <ns>-<role> ... --remote-control, appearing immediately by logical name (acme-mm, acme-maintain, acme-build) in the Claude Desktop app, claude.ai/code, and Claude mobile.
- GitHub Copilot CLI (HARNESS=copilot): launch-agent-sessions.sh translates --remote-control → --remote (copilot --name <ns>-<role> -i "<prompt>" --remote). Sessions appear by name in GitHub Web (github.com/copilot) and GitHub Mobile (Copilot CLI has no standalone desktop app).
- Google Antigravity (HARNESS=antigravity): Requires both a host-level daemon (agy remote-control start) and per-session --remote-control. The daemon is opt-in: launch-agent-sessions.sh starts it (agy remote-control start --name <fleet>-<host>, after checking agy remote-control status) only when AGY_REMOTE_CONTROL_AUTOSTART=1 is set; it is off by default. The launcher forwards --remote-control, --dangerously-skip-permissions, and --effort to agy (sessions appear grouped by workspace/project in the Antigravity 2.0 Desktop App).
Closed-Loop Host & Singleton Trust (fleet verify · core/fleet/lib/verify-wrappers.mjs): Verifies singleton plugin files, host ENGSYS_DIR scripts, and .claude/settings.json against the pinned GitHub release tree SHAs; revokes singleton batons (--reason integrity-mismatch) if tampering or drift is detected; enforces root-owned git and clean SSH config on guarded pushes; and holds relaunches if harness OAuth expires.

CLI Reference

Project Installer (`engsys` / `./install`)

Command	Purpose
`engsys init [--into <path>]`	Scaffold `engsys.config.yaml` from the bundled example.
`engsys install --into <path> [--harness claude\|copilot\|antigravity\|all]`	Materialize personas, skills, commands, hooks, settings, and lessons into `<path>` (adopting any existing setup).
`engsys update --into <path>`	Re-render from current `engsys` + config; preserves `PROJECT-FACTS` and hand-added permissions; prunes deselected packs.
`engsys verify --into <path>`	Verify all managed files against `.claude/engsys.lock` and print the external provider readiness matrix.
`engsys enable-providers <names> --into <path>`	Append a `providers:` block (`codex,deepseek,grok,anthropic`) to `engsys.config.yaml` and run `update`.
`engsys fleet init --into <dir> ...`	Scaffold a fleet instance repository (`fleet.conf`, roster, launchd templates, `scripts/fleet` shim, optional `federation.yml` & `resource-broker`).
`engsys uninstall --into <path>`	Remove all `engsys`-managed files and restore snapshotted originals byte-for-byte.

Fleet Host CLI (`scripts/fleet` in a fleet instance)

Command	Purpose
`fleet status [--federation]`	Show pin alignment, session state, auth holds, or federation-wide heartbeats and baton holders.
`fleet xray [--role <r>\|--session <s>] [--timeline] [--diagnose] [--watch] [--json]`	Deep 10-subsystem runtime inspection, live 9-hook synthetic self-test, session deep-dive, chronological behavior timeline, and automated root-cause remediation playbook.
`fleet sync [--check]`	Sync host checkouts and plugins (including `HARNESS_FALLBACK` caches) to pinned tags using per-instance worktree isolation.
`fleet verify [--alert]`	Verify singleton guard scripts, host `ENGSYS_DIR`, and settings against GitHub release tree SHAs.
`fleet launch` / `fleet restart [--stale]`	Start missing sessions or cycle stale sessions onto newly synced pins.
`fleet supervise` / `fleet install-jobs`	Run one deterministic supervisor tick (relaunches stay on the primary Claude harness; a seat limit is alerted, never failed over) or load the recurring `launchd` jobs.
`fleet health <subcommand>`	Operate the Fleet Health Desk (`validate`, `status`, `tick`, `check`, `report`, `simulate`, `invoke`, `self-probe`, `digest`, `watch-add`).
`fleet msg <route\|send\|inbox>` / `fleet relay`	Send, route, and poll cross-fleet messages.
`fleet pin`	Verify tag/manifest version alignment (`package.json`, `marketplace.json`, `plugin.json`) and open a pin bump PR.

Repository Layout & Documentation Map

core/
  agents/             11 core specialists (melvin, aaron, isabelle, bert, jody, leith, marcelo, patricia, nyx, otto, gary)
  commands/           /generate-project, /design-change, /implement-project, /implement-issue, /pre-push,
                      /prep-review*, /project-closeout, /naturalize, /merge-monster, /maintenance-monster, /resource-broker
  skills/             stack-agnostic skills: git-workflow-agents, code-review, gh-cli, github-actions, pre-push,
                      refactor, llm-structured-outputs, web-design-reviewer, webapp-testing, chrome-devtools,
                      agentic-eval, interactive-explainer, subagent-liveness, untrusted-input, hermetic-git-tests,
                      merge-monster, maintenance-monster, mnt-watch, resource-broker, durable-lease, agent-sessions
  fleet/              fleet kit (bin/fleet, sync, verify, launch, supervise, pin, notify, federation, msg, relay)
                      + health/ (Fleet Health Desk: registry, adapters, hysteresis engine, incidents, board, cli)
                      + identity/ (GitHub App bot token minting & scoped gh shim)
  lib/                harness-io.mjs (tri-harness hook adapter), sdd-validator.mjs, untrusted.mjs, git-env.mjs,
                      monster.sh, claim.mjs, lease/ (GitHub-ref CAS baton), memory/ (SQLite FTS5 index-db & ref-store)
  scripts/            memory-brief.mjs, memory-sync.mjs, harvest-findings.mjs, validate-lessons.mjs,
                      sdd-check.mjs, spec-slice.mjs, contract-check.mjs, worker-run.mjs, worker-package.mjs
  lessons-library/    51 curated, frontmatter-validated cross-project lessons
  workflows/          long-form procedures & multi-provider worker briefs
  templates/          CLAUDE.md, GEMINI.md, copilot-instructions.md, spec.md.tmpl, settings, hooks, repo-gates

stacks/               24 detachable capability packs — compose per project (scalar or list)
  cloud/              aws · azure · gcp · cloudflare
  iac/                terraform · bicep · cdk
  lang/               typescript · python · swift · kotlin · shell
  platform/           web · ios · android
  db/                 prisma · mongo
  domain/             mobile-growth
  tooling/            issue-tracker-github · issue-tracker-linear · provider-{anthropic,codex,deepseek,grok}

optional-agents/      sandy (marketing/SEO), jos (monetization), deandre (tutor/explainer), steve (morale)
DEPLOY.md             standalone LLM & human progressive deployment playbook (Stages 1–4)
index.html            interactive visual explainer (https://eric-sabe.github.io/engsys/)
dashboard.html        privacy-preserving GitHub activity dashboard

Deep-Dive Documentation (`docs/`)

Getting Started & Architecture: DEPLOY.md · docs/architecture.md · docs/install-scenarios.md · docs/naturalization.md
Specs, Review & Memory: docs/specs/engsys-sdd-upgrade.md · docs/memory-spec.md · docs/multi-provider-spec.md · docs/review-methodology.md
Autonomous Fleet & Multi-Harness: docs/specs/multi-harness-support.md · docs/harness-adaptation-guide.md · docs/fleet-guide.md · docs/multi-fleet.md · docs/merge-monster.md · docs/maintenance-monster.md · docs/fleet-health-desk.md · docs/agent-messaging.md · docs/subagent-liveness.md · docs/gate-check.md

Activity Dashboard (Optional)

A self-hosted, privacy-preserving GitHub activity dashboard ships in dashboard.html, fed by scripts/collect-stats.mjs. Committed stats (data/stats.json) store only opaque per-repo aliases — never repo/owner names, issue titles, branches, or commit messages. See docs/dashboard-spec.md for setup and daily GitHub Actions automation.

Tests

npm test                     # runs installer selftests + all Node & shell test suites
npm run build:plugins        # regenerates .claude-plugin manifests & hooks (--check to verify)

License

MIT — see LICENSE.

Navigation Menu

Repository navigation

About

Repository files navigation

engsys

Mental Model: Model vs. Coding Harness vs. Meta-Harness

The Three-Part File Taxonomy Inside `engsys`

Where to Start: The 4-Stage Graduation Ladder

Stage 1 — Interactive Team (Start Here — 3 Minutes)

Option A — Install into your project (CLI or Clone)

Option B — Install as Claude Code Plugins (Zero Files Copied Into Repo)

What Stage 1 Gives You Immediately

Stage 2 — Cross-Family Review & Compounding Memory

2.1 Cross-Family Model Review & Worker Dispatch (`Codex`, `Grok`, `DeepSeek`)

2.2 The Agentic Memory Engine (`docs/memory-spec.md`)

Stage 3 — Single-Host Autonomous Fleet (Monsters + Health Desk)

Stage 4 — Multi-Repo / Multi-Host Federated Fleet

CLI Reference

Project Installer (`engsys` / `./install`)

Fleet Host CLI (`scripts/fleet` in a fleet instance)

Repository Layout & Documentation Map

Deep-Dive Documentation (`docs/`)

Activity Dashboard (Optional)

Tests

License

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.claude-plugin		.claude-plugin
.github/workflows		.github/workflows
core		core
data		data
docs		docs
lib		lib
optional-agents		optional-agents
scripts		scripts
stacks		stacks
team-images		team-images
tests		tests
.env.example		.env.example
.gitignore		.gitignore
.nojekyll		.nojekyll
DEPLOY.md		DEPLOY.md
LICENSE		LICENSE
README.md		README.md
dashboard.html		dashboard.html
engsys.config.example.yaml		engsys.config.example.yaml
index.html		index.html
install		install
package.json		package.json

Latest commit

History

Folders and files

Repository files navigation

engsys

Mental Model: Model vs. Coding Harness vs. Meta-Harness

The Three-Part File Taxonomy Inside engsys

Where to Start: The 4-Stage Graduation Ladder

Stage 1 — Interactive Team (Start Here — 3 Minutes)

Option A — Install into your project (CLI or Clone)

Option B — Install as Claude Code Plugins (Zero Files Copied Into Repo)

What Stage 1 Gives You Immediately

Stage 2 — Cross-Family Review & Compounding Memory

2.1 Cross-Family Model Review & Worker Dispatch (Codex, Grok, DeepSeek)

2.2 The Agentic Memory Engine (docs/memory-spec.md)

Stage 3 — Single-Host Autonomous Fleet (Monsters + Health Desk)

Stage 4 — Multi-Repo / Multi-Host Federated Fleet

CLI Reference

Project Installer (engsys / ./install)

Fleet Host CLI (scripts/fleet in a fleet instance)

Repository Layout & Documentation Map

Deep-Dive Documentation (docs/)

Activity Dashboard (Optional)

Tests

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

The Three-Part File Taxonomy Inside `engsys`

2.1 Cross-Family Model Review & Worker Dispatch (`Codex`, `Grok`, `DeepSeek`)

2.2 The Agentic Memory Engine (`docs/memory-spec.md`)

Project Installer (`engsys` / `./install`)

Fleet Host CLI (`scripts/fleet` in a fleet instance)

Deep-Dive Documentation (`docs/`)