Zelari Code

             -#%=
           .*%%%@#:
          =%%%%%%%@*
         +%%%%%%%%%%#.
        +%%%@@@@@@@%@#.
      .*%@@@@@@@@@@@@@%-
     .#%@@@@@@@@@@@@@@@@-
     *%@@@@@@@@@@@@@@@@@%.
    :@%%@@@@@@%:+%@@@@@%@=
     =@%%@@@@@%.=-+%@%%@*
     .=@@%@@@@%.*@*-+@@*.
   -*%@@@@@@@@%.*@@@+:#@@#=.
  *%%%%%%@@@@@@.*#%#=-:+@@@%
 :@%%%%%%@@@@@@.:=.*:*@@@@@@=
 *@%%@%%@@@@@@@*%@*:-:%@@@@@%.
:@@@%@@@@@@@@@@@@@@#%@@@@@@@@=
*%%%%%%%%%%%%%%%%%%%%%%%%%%%%#

     Z E L A R I   C O D E
        Anathema Studio

Open-source coding orchestrator — you choose the model, the proof is mandatory. Kraken by default, council when the work earns it.

By Anathema Studio · Product page · GitHub · npm · Contributing · Security

Trailer (EN, ~30s): docs/media/trailer/zelari-code-trailer.mp4 · media kit

📖 Full user guide · Reference guide — install, TUI, slash commands, council, skills, workspace, headless, MCP, Desktop.

Zelari Code is an open-source coding orchestrator for the terminal — the agent that doesn't trust itself. The default dispatch is the kraken super-agent (aliases agent/single) spawning explore/general/verify tentacles; a six-role council pipeline (Caronte, Nettuno, Gerione, Plutone, Minosse, Lucifero) is there when the work earns a second opinion; optional zelari missions loop until a deliverable is proven done — done means verified, not claimed. It ships a rich TUI (Ink + React), slash commands, plan/build phases, and provider-agnostic LLM streaming (OpenAI-compatible, xAI Grok, ChatGPT, Anthropic, Muse, GLM/Z.AI, MiniMax, DeepSeek). OAuth via /login grok, /login chatgpt, /login anthropic, /login muse. The reusable runtime is published as @zelari/core (Apache-2.0).

# Zero install — runs from npm's cache (cold-cache start):
npx zelari-code@latest

# Or install persistently (fast start + `/update` self-update):
npm install -g zelari-code
zelari-code

These are opt-in — the CLI runs fine without them. The agent auto-skips a tool if its dependency is missing.

Disable any tool group: set ZELARI_LSP=0, ZELARI_AST=0, ZELARI_SEMANTIC=0, ZELARI_BROWSER=0, ZELARI_DIAGNOSTICS=0.

Two supported paths:

# Zero install — runs straight from npm's cache (cold-cache start):
npx zelari-code@latest

# Persistent install — fast start and `/update` self-update support:
npm install -g zelari-code
zelari-code --doctor   # recommended once on Windows

• npx (zero install) — nothing to install or maintain; ideal for a quick try. The first run resolves/downloads the package (a few seconds), and /update is disabled — re-run npx zelari-code@latest for the newest version.
• Global (npm install -g) — installs once, starts instantly after, and enables the in-app /update self-update. This is what the Windows PATH notes below assume.

If you already ran npm install -g zelari-code above, skip to First Run.

An installable GUI shell lives in apps/desktop/. It does not replace the CLI — it streams zelari-code --headless into a modern chat UI. GitHub Releases attach platform installers on each v* tag.

Installer ≠ CLI. Downloading Desktop from GitHub does not install or upgrade the global zelari-code package. Use npm i -g zelari-code (or Desktop → Settings → Update CLI). First launch shows a Setup guide when Node/CLI is missing.

Desktop highlights: Mode / Phase / Provider bar · Files|Git project panel · Settings (provider keys, App updates, Update CLI, MCP Extensions, SSH Connections) · multi-turn chat history · optional overlay HUD · password/agent/key SSH targets (ssh_status / ssh_run).

npm run build                 # CLI side-car
npm run desktop:install
npm run desktop:dev           # Tauri dev window
# npm run desktop:build       # MSI / NSIS / DMG / AppImage

See apps/desktop/README.md and docs/GUIDA.md (Desktop, MCP, SSH). Requires Rust + Node ≥ 20.17.

Prerequisites:

• Node.js ≥ 20.17 — required. Without it the agent cannot run npm/tsc/builds, so zelari-code refuses to boot.
• Git — recommended. Without it, /diff, /undo and the git sidebar are disabled. Install from https://git-scm.com.
• Git Bash (Windows only) — recommended. The agent's bash tool needs real POSIX semantics (ls, which, $VAR, &&). Ships with Git for Windows.

After install, verify your environment:

zelari-code --doctor   # checks shim, bundle, PATH, node/git/bash in the agent shell

Why --doctor matters on Windows: the agent runs commands through Git Bash, which inherits a different PATH than the Node process. Node can be visible to PowerShell yet invisible to Git Bash (typical when Node was installed for "current user" only). --doctor's node (agent shell) row catches this; the boot-time preflight (runPreflight) blocks the launch with an actionable message instead of letting the agent fail mid-task.

After npm install -g, the zelari-code command may not be on your PATH. Fix:

PowerShell (run as admin, then restart your terminal):

$npmPrefix = npm config get prefix
[Environment]::SetEnvironmentVariable("Path", $env:Path + ";$npmPrefix", "User")

Git Bash / WSL:

echo 'export PATH="$(npm config get prefix):$PATH"' >> ~/.bashrc
source ~/.bashrc

Verify the fix: where zelari-code (CMD) or which zelari-code (Bash) should print a path.

Node visible to PowerShell but not to Git Bash? This is the dual-PATH problem: Node installed for "current user" only reaches the user shell, while Git Bash inherits the system Path. Fix: reinstall Node with "Add to PATH for all users", or add C:\Program Files\nodejs\ to the System Path (not User). Confirm with zelari-code --doctor — the node (agent shell) row must read OK.

The first time you run zelari-code (or whenever your provider config is missing), the CLI launches a 5-step onboarding wizard instead of the regular TUI:

╭─────────────────────────────────────────────────╮
│ zelari-code — first-time setup                  │
│ 1/welcome  2/provider  3/model  4/apikey  5/...│
│                                                 │
│ Welcome! Let's get you coding in under 2 min.   │
│ Press [Enter] to continue, [Q] to quit.         │
╰─────────────────────────────────────────────────╯

The wizard walks you through:

1. Welcome — overview + how to quit.
2. Provider — pick from grok, minimax, glm, deepseek, openai-compatible (↑/↓ + Enter).
3. Model — type a model name or accept the default (Enter).
4. API key — choose env (use GROK_API_KEY etc.), keystore (save locally), or skip (set later via /login).
5. Confirm — review + Enter to commit.

When you press Enter on confirm, the wizard writes ~/.zelari-code/provider.json (and keys.json if you chose keystore), shows a brief "✓ Setup complete!" banner, then transparently transitions into the regular TUI — no need to re-launch.

zelari-code --no-wizard            # skip wizard even on first run
zelari-code --reset-config         # force re-run wizard (clears provider.json)
ZELARI_NO_WIZARD=1 zelari-code     # env equivalent of --no-wizard
zelari-code --version              # print version + exit (no TUI)
zelari-code --help                 # print help + exit (no TUI)

The wizard re-runs automatically if provider.json is missing on the next launch.

# Set your OpenAI-compatible API key (OpenAI, Together, Groq, custom endpoint, etc.)
export OPENAI_API_KEY=sk-...

# Or use Grok via OAuth (Device Authorization Grant — RFC 8628)
zelari-code
# Inside the TUI: /login grok
# → A code + verification URL appears; open the URL, enter the code, authorize.

# Or use GLM/Z.AI
export GLM_API_KEY=...

# Run zelari-code from any directory
zelari-code

Full reference: docs/GUIDA.md (all flags, examples, skill IDs).

TUI: shift+tab cycles kraken → council → zelari for free-form prompts (agent is a kraken alias). Use /mode when the terminal captures shift+tab. Work phase is orthogonal: /plan (no project writes) vs /build (full tools) — same axes as Desktop Mode / Phase bars.

Run a single task without the TUI (CI/scripts/Desktop):

zelari-code --headless --task "Explain src/cli/main.ts" --output plain
zelari-code --headless --task "Design a REST API" --council --output json
zelari-code --headless --mode kraken --phase plan --task "Outline the change"

Useful flags: --mode kraken|council|zelari (agent=alias), --phase plan|build, --provider, --model, --history-file (multi-turn Desktop). See docs/GUIDA.md for exit codes and config helpers (--print-config, MCP, SSH).

Zelari Code includes a built-in update mechanism:

# Inside the TUI:
/update          # check for updates (prints current vs latest)
/update --yes    # apply update (runs npm install -g zelari-code@latest)

On startup, the CLI silently checks the npm registry. If a newer version is available, it prints a one-line hint to stderr.

Disable auto-check: ANATHEMA_DEV=1 zelari-code

• 🐙 Kraken super-agent — default mode (aliases agent/single): lead that spawns task tentacles (explore / general / verify), optional git worktrees, and Kraken Graph (/kraken graph) for a parallel DAG
• 🤖 Opt-in council second opinion — the multi-agent council (6 roles, Caronte → Lucifero, feedback loops + member promotion) runs only when you ask for it (shift+tab / /mode council); kraken stays the default dispatch
  • Role legend: Caronte — Council Director (orchestrator, opens the run) · Nettuno — Project Planner · Gerione — Creative Ideator (diverge→converge) · Plutone — Knowledge Architect · Minosse — Quality Critic (oracle: literal verification, per-run verdict) · Lucifero — Chairman (synthesis + dispatch) — definitions in packages/core/src/agents/roles.ts
• ⚡ Zelari-mode — autonomous multi-run missions: a free-form prompt becomes a mission brief, then design@council → build@kraken until the MVP slice's completion.ok is green or the iteration budget runs out
• 🧮 Budget-aware mission continuation — after each mission slice a continuation gate decides repair / pivot / hold from the remaining budget and gap history (repeated gap → pivot with reduced roster; exhaustion → hold, never a false done; deterministic PASS stays the only authority)
• 🧾 Deep harness manifest — each session fingerprints its real tool surface (name + description + input schema per tool) alongside profile and resource policy, so tool or schema drift changes the manifest hash and is detectable on resume and in eval provenance
• 🚦 Eval retention gate — tiered anchor suite (eval/anchors/), versioned result store per manifest hash, and a CI regression gate comparing candidate vs last stable tag (blocking once the baseline store is seeded; honest shadow + warning until then)
• 🧠 Native shared project memory — compatible JSONL by default; opt-in SQLite V2 adds typed graph nodes, immutable history, bounded recall and shared Council/Kraken/session knowledge without MCP (guide)
• ⇧⇥ Kraken/council/zelari mode switch — shift+tab cycles free-form prompts between the kraken lead, the full council pipeline, and an autonomous mission (mode shown in the status line)
• 🎨 Rich TUI — Ink + React: native-scrollback chat stream, input bar with status line below it (mode · provider · model · session · cwd · execution timer)
• 🗂️ Live git sidebar — right-hand panel with working-tree changes (+added/-removed per file, refreshed every 4s; auto-hidden on narrow terminals)
• ⏱️ Execution timer — elapsed time of the in-flight turn in the status line (⏱ 12s), frozen as last 34s when the run completes
• 🧠 Provider-agnostic — OpenAI-compatible APIs (OpenAI, Together, Groq, custom), xAI Grok / ChatGPT / Anthropic OAuth, GLM/Z.AI, MiniMax, DeepSeek; optional ZELARI_LOCAL_CLI=claude
• 🛠️ Built-in tools — filesystem (read/write/edit), shell (bash), search (grep), web fetch/search
• 🧠 LSP code intelligence (lsp_* tools) — go-to-definition, find references, hover type, document symbols, rename symbol via real language servers (tsserver, pyright, …)
• 🌲 AST structural tools (ast_* tools) — symbol outline + find-by-name via the TypeScript compiler API, no language server needed
• 🔎 Semantic search (semantic_search + /index) — concept-level code search via embeddings, local-first
• 🌐 Browser verification (browser_check) — headless browser with click/fill/goto/wait actions, console + network + screenshot capture for visual verification of web work
• 🔁 Post-edit diagnostics loop — after every edit_file/write_file the harness runs project lint/compile (eslint, ruff, LSP-pluggable) and surfaces the errors to the model in the same turn (opt out: ZELARI_DIAGNOSTICS=0)
• 💾 Prompt-cache accounting — tracks prompt-cache hit rate per provider/model and surfaces it in the status bar (cache 73%) so you can see when a session is amortizing its prefix
• 🧷 Workspace checkpoints — /checkpoint [label] + /rollback [id|latest] use git plumbing to snapshot the working tree (tracked + untracked) and restore it atomically; every zelari-mode mission takes one before starting
• 🤝 Sub-agent delegation (task tool) — isolated tentacles: explore (read-only), general (bounded writes), verify (tests/checks); optional scope[] + acceptance[]; no nested task
• 📚 26 coding skills (+ user SKILL.md from .zelari/skills/, .claude/skills/, …) including schema-loop, computer-use-cua, qwen-mm-plugins-install-setup
• 🔄 Cross-provider failover — automatic retry with provider swap on transient errors
• 📊 Metrics + skill history — fire-and-forget logging to ~/.zelari-code/ (unified home; auto-migrated from the legacy ~/.tmp roots — see src/cli/paths.ts)
• 🗜️ Session management — JSONL transcripts, resume across restarts, compaction
• 🌿 Branch isolation — session snapshots per branch
• 🔌 MCP — external MCP servers via .zelari/mcp.json or ~/.zelari-code/mcp.json; Desktop Extensions store for one-click install
• 🔐 SSH targets — configure hosts for deploy/monitor (password, agent, or key); tools ssh_status / ssh_run with allowlist; ZELARI_SSH=0 kill switch
• 🖥️ Zelari Desktop — Tauri 2 shell with chat UI, project tree, Settings (OAuth Sign in/Refresh/Sign out), Workbench Plan/Tasks, dual updates (app vs npm CLI), multi-turn history, optional overlay HUD
• 📱 Companion host + Android — zelari-code serve (SSE over Tailscale/LAN) + Settings → Start companion serve; MVP app in apps/companion-android/
• 🧩 Skills store (Desktop) — create/remove user & project SKILL.md; import skill from URL with the selected model; /skills picker in TUI
• 📎 @-tag paths — tag workspace files/folders in Desktop composer and CLI prompts
• ◇◆ Plan / build phase — /plan explores without project writes; /build implements with full tools (independent of kraken/council/zelari mode)
• 🛡️ Folder trust + lifecycle hooks — project MCP/hooks load only for trusted folders (/trust, zelari-code --inspect); hooks are fail-open in the TUI and fail-closed in autonomous runs (headless/mission/CI); override with ZELARI_HOOKS_FAILURE=fail-open|fail-closed. Full vector × gate matrix and fresh-clone defaults: docs/THREAT_MODEL.md
• 👁️ Native vision — @image.jpg (and Desktop drop-to-attach) inlines pixels to vision-capable models; no third-party vision API
• 🆕 Self-update — /update slash command + silent registry check on startup

zelari-code (CLI, Apache-2.0)
├── src/cli/                  # Ink TUI, providers, workspace, wizard, MCP, SSH, serve
│   ├── components/           # ChatStream, InputBar, Sidebar, StatusBar, …
│   ├── slashHandlers/        # /provider, /workspace, /kraken, /trust, …
│   └── workspace/            # .zelari/ persistence + AGENTS.MD curation
├── packages/core/            # @zelari/core (Apache-2.0, npm)
│   ├── core/                 # AgentHarness — provider-neutral agent loop
│   ├── agents/               # Council API, roles, 26 skills, tool schemas
│   └── council/              # Run mode, tier banners
├── apps/desktop/             # Optional Tauri 2 shell (spawns CLI --headless)
└── apps/companion-android/   # Thin client for `zelari-code serve`

AgentHarness takes (model, provider, messages, tools) + a streaming function and yields AsyncIterable<BrainEvent>. Mixed tool batches run segmented: contiguous read-only calls in parallel, write/execute tools as serial barriers (opt out: ZELARI_PARALLEL_TOOLS=0). See MIGRATION.md for the package boundary.

See docs/GUIDA.md for the full list.

The CLI persists council output (decisions, risks, docs, plan, reviews) into a project-local .zelari/ directory and auto-curates an AGENTS.MD at the project root.

.zelari/                      # auto-gitignored, per-project
├── plan.md / plan.json       # phases / tasks / milestones
├── risks.md                  # risk register (live, ordered by severity)
├── decisions/                # ADRs — 001-<slug>.md
├── reviews/                  # Minosse verdict per council run
├── docs/                     # doc drafts produced by the council
├── memory/                   # file-based project memory (zelari missions)
├── radio/                    # Kraken tentacle progress bus
├── kraken/                   # last-graph.json (DAG resume)
├── world/                    # schema-loop hypothesis / checks / timeline
└── hooks/                    # project lifecycle hooks (trusted folders only)
AGENTS.MD                     # committed at project root, auto-curated from .zelari/

AGENTS.MD is partitioned into:

1. Manual blocks (free-form, preserved verbatim across updates) — write anything you want here.
2. Auto-managed sections delimited by <!-- zelari:auto:start section="..." --> / <!-- zelari:auto:end --> markers — overwritten each sync.

Auto-managed sections:

• tech-stack — languages, frameworks, build tools (derived from package.json)
• decisions — accepted ADRs (newest first, capped)
• conventions — code conventions observed in source tree
• build — build/test/lint commands
• open-questions — info-level risks + unsolved questions

Set ZELARI_AGENTS_MD=0 to disable AGENTS.MD auto-curation.

• Frontmatter: subset YAML (scalars, flow/sequence/block-sequence arrays, flow/block maps) — no external deps.
• Concurrency: per-key mutex (filesystem writes are serialized per artifact).
• Idempotency: hash comparison — AGENTS.MD write is skipped when no section changed (clean git diff).

See docs/plans/2026-07-01-council-workspace-cli-stubs.md for the full schema.

git clone https://github.com/N-THEM-Studio/zelari-code.git
cd zelari-code
npm install
npm run build:cli
npm link
zelari-code

npm test

npm run typecheck

npm run test:eval    # eval engine + regression-gate unit tests
npm run eval:gate    # regression gate vs stored baseline (CI parity)

Anchors: eval/anchors/ (tiered) · result store: eval/results/, versioned per harness-manifest hash and seeded with node --experimental-strip-types tools/eval/runAnchors.ts · CI: .github/workflows/eval-retention-gate.yml — Tier-0 always blocking; the retention gate runs candidate anchors vs the store extracted from the last stable tag and blocks on regression once the baseline store is seeded (declared bootstrap + warning until then; never fake greens).

• Product page (IT: zelari-codice) — Anathema Studio
• GitHub repository
• npm: zelari-code — CLI
• npm: @zelari/core — reusable runtime (Apache-2.0)
• Site docs (IT: documentazione)

Apache-2.0 © 2026 Anathema Studio.