Telar

The agentic engineering framework — plan, build, review, and ship, with agents. By Zeki Yugnak · v0.17.0 — 51 agents, 126 skills, 24 commands, 4 hooks, 7 rules, 30 scripts

Telar is a multi-agent engineering framework for Claude Code and Codex that takes a feature from idea to production — orchestrated planning, adversarial review gates, a persistent knowledge base, cross-model verification, and a new OKF domain-knowledge layer. It spans cross-platform apps end to end: mobile (React Native & Flutter) with deep native integration, web (Astro, Next.js/Tailwind/shadcn, Vite/TanStack admin panels), Rust service layers, and desktop — with stack-aware orchestration, reviewers, and rubrics across every domain.

Config-driven hybrid model roster (v0.13): assign the best model to each pipeline role — architect / moderator / developer / reviewer / tester — in external-tools.yaml, and mix Claude (Sonnet 5 / Opus 4.8 / Fable 5), GPT-5.6-sol (via the Codex adapter) and Kimi K3 (via a native kimi CLI adapter or a generic Anthropic-compatible compat adapter) freely. Onboarding a new model is a YAML edit, not a code change; pricing is fail-closed and reviews stay cross-model. Benchmark-tuned defaults: Sonnet 5 (high) implements, Opus 4.8 (high) escalates on critical work, Kimi K3 is an opt-in $0-marginal cost mode, and review runs Opus + GPT-5.6-sol in parallel.

Dual-host — Claude or Codex (v0.14): the same project can be driven natively by either agentic host. runtime.host + symmetric routing.profiles give each host its own roster — the active host plays architect/moderator/developer, the other vendor reviews cross-model — so neither is permanently secondary. Codex can now be the moderator/implementer (gpt-5.6-terra for implementation, gpt-5.6-sol for architecture/moderation at xhigh), with Claude reached via a first-party claude.sh adapter. The anti-nest rule (a model whose home host is the active host runs in-harness; everything else via its adapter) keeps a host from re-invoking its own CLI, and a .tl-telar/context/orchestration-lock.json mutex stops two hosts from orchestrating the same plan. Configs without runtime/profiles keep the previous Claude-primary behavior.

📖 Full documentation

zekiyugnak.github.io/telar-framework →

The full reference — every agent, skill, command, and rule, plus the orchestration deep-dive and end-to-end examples — lives on the documentation site. This README is just the entry point.

Getting Started	Install and run your first command
Agents	46 specialized subagents
Skills	116 reference modules
Commands	23 slash-command workflows
Rules	7 always-on standards
Orchestration	The opt-in pipeline, in depth
Examples	End-to-end command flows
Configuration	All config keys — .tl-telar-thresholds.json and external-tools.yaml (incl. cc_features)

Installation

Claude Code

# Add the marketplace
claude plugin marketplace add zekiyugnak/telar-framework

# Install the plugin
claude plugin install tl-telar@telar

After installation, restart Claude Code to load the plugin. All commands are namespaced under /tl-telar:.

Codex

Telar also ships a generated Codex plugin from the same repository. Install it from the GitHub marketplace source, not from a local checkout:

# Latest development build
codex plugin marketplace add zekiyugnak/telar-framework --ref develop
codex plugin add tl-telar@telar
codex plugin list --marketplace telar

For the stable main branch, use:

codex plugin marketplace add zekiyugnak/telar-framework --ref main
codex plugin add tl-telar@telar

If you previously added the marketplace with another ref, refresh it before reinstalling:

codex plugin remove tl-telar@telar
codex plugin marketplace remove telar
codex plugin marketplace add zekiyugnak/telar-framework --ref develop
codex plugin add tl-telar@telar

Start a new Codex thread and invoke Telar with @tl-telar or specific bundled skills such as $create-app, $review-plan, or $orchestrate.

If you already have a user-level Codex skill with the same name, prefer @tl-telar so Codex resolves the Telar plugin context.

Codex plugin installation is separate from the optional adapters.codex entry in .tl-telar/external-tools.yaml. The plugin loads Telar into Codex; the adapter lets Telar delegate selected orchestration work to the Codex CLI when explicitly enabled.

Quick start

Goal	Command
New app from scratch	/tl-telar:create-app [description]
Add a feature	/tl-telar:add-feature [feature]
Run tests	/tl-telar:test-app [scope]
Code review	/tl-telar:review-code [path]
End-to-end orchestrated build	/tl-telar:orchestrate <task>
Release to stores	/tl-telar:release-app

See the command reference for all 23 workflows.

Highlights

• 47 specialized agents — mobile & native platform experts, native bridges, architecture, security, testing, release, orchestration; a full web stack (Astro/Next.js/Vite-TanStack plus a framework-agnostic React expert and web security/performance/accessibility specialists); Rust services; and desktop (Electron/Tauri).
• 126 skills — reusable reference modules, decision frameworks, and ready-to-use feature blueprints for React Native, Flutter, the web stack, and Rust services.
• Two-stage review gates — requirement compliance and code quality, with adversarial and collaborative reviewers.
• 🆕 Risk-tiered review (new) — rigor is front-loaded into the plan (mandatory data_contracts/edge_cases/test_plan per Work Unit, enforced by the plan gate), so implementation review stays fast. The per-WU adversarial roster scales with each WU's risk_tier — trivial→Code-only, standard→Code+Maintainability (UI a11y/perf handled by CI lenses), critical→full roster + up-front design gate + human checkpoint. A sensitive-path Security floor (auth/token/payment/migration/…) is never droppable by tier, retries re-review incrementally (sticky-pass), and cross-model review stays always-on when configured.
• Orchestrated mode (opt-in) — design + plan review gates, a 4-phase IMPLEMENT/VALIDATE/REVIEW/COMMIT loop, and blocking quality gates. Independent Work Units run in parallel — a pure readiness scheduler dispatches concurrent WUs whose dependencies are met and file scopes are disjoint (bounded by execution.max_parallel_wus). The orchestrator honors your git policy and never auto-commits.
• Persistent knowledge base — typed JSONL facts captured via /tl-telar:self-reflect and re-primed into context each session.
• 🆕 OKF knowledge layer (new) — optional support for the Open Knowledge Format: a docs/knowledge/ domain-knowledge bundle agents consult (orientation) before touching a table or schema, while ADRs and the schema spec stay the source of truth. Produced and kept healthy by the okf-knowledge-curator agent and the okf-knowledge-authoring skill; a no-op in projects without a bundle.
• Optional external AI — Codex/Gemini adapters with a budget circuit breaker and cross-model review (disabled by default).

The workflow

1. Explore — read the existing codebase first (codebase-first rule)
2. Requirements — establish what to build in REQUIREMENTS.md
3. Research — decide how to build it in RESEARCH.md
4. Plan — break work into atomic tasks in PLAN.md + PROGRESS.md
5. Build — iterate with simulator verification
6. Debug — find the root cause before fixing
7. Verify — require fresh evidence before claiming done
8. Review — two-stage gates: requirement compliance + code quality

Repository layout

Directory	Contents
agents/	47 agent definitions
skills/	117 skill modules (incl. blueprints/ and orchestration/)
commands/	23 slash commands
rules/	7 always-on rules
hooks/	Session and pre-build hooks
scripts/	Automation and validation utilities
resources/	Decision trees, platform guides, checklists, rubrics
docs/	The documentation site (this is also published via GitHub Pages)

Contributing

To extend the plugin: add agents in agents/, skills in skills/, and commands in commands/. Validate with the scripts in scripts/ (validate-agents.js, validate-skills.js, validate-blueprints.js). See AGENTS.md for the auto-generated agent directory.

License

MIT © 2026 Zeki Yugnak. See LICENSE and THIRD_PARTY_NOTICES.md.