The Lead Stack: Agent-Agnostic Workflows

intent-brief → spec → plan → diff → review-report → release

A high-performance repository of "Skills" and RTK-powered tools designed for Tech Leads. These workflows are Agent-Agnostic, allowing any LLM agent (Gemini, Claude, GPT) to assist with implementation planning, code review, and automated testing.

Live Web App: https://ai-tech-lead-stack.vercel.app

[!NOTE] Related project — SML Gate (small-language-model-gate, CLI slm-gate) — a local AI routing and pre-processing layer that uses a small, free local model via Ollama to intercept, compress, and answer easy or repetitive prompts before they reach your paid subscription or API cloud model, cutting token spend and protecting your monthly quota. Its mcp-gate layer can sit in front of this stack's MCP server (TLS_ADAPTER=on + DOWNSTREAM_MCP pointing at dist/mcp-server.mjs) to condense tool and skill payloads before they hit your editor's context window.

<a href="https://github.com/zenithfoundry/sml-gate" target="_blank" rel="noopener noreferrer">Explore SML Gate on GitHub →</a>

Table of Contents

• Commands Quick Reference
• Which tier am I on?
  • Tier Decision Guide
  • Architecture: Harness Independence vs. Model Separation
• 🚀 Quick Start
• Install, Link & Uninstall
• Supported Editors & Agents
• Antigravity Setup
• Cursor Setup
• Continue Setup
• Claude Code Setup
• Cline Setup
• Gemini CLI & Gemini Desktop Setup
• Workflow Catalogue
• The Web App
• Docs
• Available Skills
  • Orchestrators
  • Discover & Define
  • Plan & Harden
  • Build & Fix
  • Review & Verify
  • Design & UI
  • Ship & Communicate
  • Internal Skills
• 🧠 The Methodology: Four Pillars
• 🛠 Technical Architecture: RTK & MCP Synergy
• 🛠 Technical Overview: Skill Discovery & Priority
• How to use in any project
• Branching Strategy
• Requirements
• 🧹 Resetting a Project
• 🧪 CI/CD
• Resources 📚

Commands Quick Reference

| What you're doing | Call this | Key principle | | :-------------------------------- | :---------------------- | :------------------------------------------- | | Leading a multi-agent team | /dev-team | Orchestrates sub-agents safely in parallel. | | Deep architecture planning | /plan | Full codebase audit, solid vertical slices. | | Fast lean tasks | /plan-quick | High velocity for smaller changes. | | Breaking down tickets | /vertical-slice | Creates ClickUp-ready tasks (<= 2d). | | Local pre-commit check | /code-review | 4 gates (Spec, SOLID, A11y, Evidence). | | Visual testing | /verify-changes | Playwright-powered before/after screenshots. | | Fixing QA/Regression feedback | /regression-bug-fix | Maps impact and remediates safely. | | Merging to main | /pr-automator | Synthesized diffs with visual proof. | | Full feature loop (Sandbox) | /feature-orchestrator | End-to-end implementation from idea. | | Asking codebase questions | /ask | High-density technical advice. |

Which tier am I on?

| Your plan | Loop to call | Dev-team to call | Capabilities & Isolation | | :------------------------------ | :----------------------- | :---------------------- | :-------------------------------------------------------------------------------------------- | | API keys (Gemini+Anthropic) | reflexion-loop | dev-team-orchestrator | Dual-model SDK enforcement (validateDistinctModels), 3+ parallel lanes, uncapped. | | $100-a-month subscription | reflexion-loop-sub-max | dev-team-sub-max | Max 2 parallel lanes, git worktrees, L0–L3 cross-vendor verify, 60 turn budget. | | $20-a-month subscription | reflexion-loop-sub-pro | dev-team-sub-pro | Single-lane pair (no worktrees), L0–L3 cross-vendor verify, 20 turn budget, capped at M size. |

Tier Decision Guide

• no keys + $20/mo -> reflexion-loop-sub-pro + dev-team-sub-pro; ceiling M; Risk-2 refused at intake and escalated if discovered mid-flight
• no keys + $100/mo -> reflexion-loop-sub-max + dev-team-sub-max; ceiling XL with a Tech-Lead confirmation gate
• API keys -> reflexion-loop + dev-team-orchestrator

[!NOTE] Platform facts (as of August 2026) — verify current pricing and quotas with the vendor.

1. Google confirmed a $100/month AI Ultra tier at I/O 2026 at roughly 5x Pro quotas, and cut the top tier from $250 to $200.
2. On Antigravity, all paid tiers ($20 Pro, $100 Ultra, $200 Ultra Max) run THE SAME MODEL LINEUP with the same context limits. The extra cost buys rate limits and weekly-cap headroom, not better model access.
3. Google has not published what a single AI credit buys in tokens, requests or compute time. Budget by observation rather than arithmetic: individual frontier-model sessions have been reported consuming a large share of a monthly allowance, and multi-day lockouts occur when a quota is exhausted.
4. The Antigravity CLI routes through the SAME credit pool as the IDE. Switching surfaces does not restore quota.
5. Gemini CLI stopped serving Google AI Pro, Ultra and free Gemini Code Assist individual users on 18 June 2026. The consumer replacement is Antigravity CLI (agy). Enterprise Gemini Code Assist licences are the exception.

Architecture: Harness Independence vs. Model Separation

To choose the right tier for your environment, distinguish between these two independent axes:

| Axis | Description | | :----------------------- | :-------------------------------------------------------------------------------------------------------------------------------------------------------- | | HARNESS INDEPENDENCE | Does the skill run under any agent? reflexion-loop does NOT, because scripts/reflexion-loop.ts calls model endpoints directly, bypassing the harness. | | MODEL SEPARATION | Do the writer and the auditor differ? reflexion-loop guarantees it in code; the subscription tiers get it from the harness instead. |

Subscription tiers obtain model separation FROM THE HARNESS rather than from direct API calls. Where the harness offers models from more than one vendor, L0 separation matches the API loop's guarantee. Where it does not, the tiers fall back through L1 to L3 and DISCLOSE the level achieved. The difference is enforcement location, not assurance level.

The real tradeoff is throughput: the subscription tiers are throughput-limited (quota, lanes, crew ceiling, critique passes), and they cannot enforce distinct models in code the way validateDistinctModels does — which is why disclosure is mandatory.

When using subscription tiers that rely on the harness for model separation, the orchestrator targets specific isolation levels:

| Level | Description | | :----------------------- | :------------------------------------------------------------------------ | | L0 (Cross-Vendor) | Writer and reviewer run on models from different vendors. | | L1 (Cross-Family) | Writer and reviewer run on different model families from the same vendor. | | L2 (Fresh Sub-Agent) | Same model, fresh sub-agent context. | | L3 (Degraded) | Same model, same context. |

[!NOTE] IDE & CLI Model Selection (as of June 2026): Consumer Google AI Pro/Ultra access via legacy standalone gemini CLI stopped on 18 June 2026. The active consumer CLI is Antigravity CLI (agy). When configuring cross-vendor model pairing (L0), verify available models via your agent harness model picker (e.g. Antigravity Agent Manager, Cursor Composer model dropdown, or Claude Code sub-agent configuration).

The subscription tiers resolve their critic with ./.ai/rtk-run run resolve-critic --writer <anthropic|google|openai>, which walks enterprise gemini (needs GOOGLE_CLOUD_PROJECT) -> agy -> another harness model -> same model. The last rung forces PROVISIONAL and writes a STRONG criticAdvisory into state.json. Set TLS_CRITIC_MODEL to pin the agy model.

🚀 Quick Start

Three Ways to Run the Tech-Lead-Stack MCP

The MCP server is built as a standalone artifact (dist/mcp-server.mjs), which is why it can be reached in more than one way. There are three paths. Two of them — Direct and install.sh — reach the same MCP server (install.sh just automates the setup); the SLM Gate path puts a gateway in front of it.

Prerequisite for every path: build the artifact once.

pnpm run mcp:build   # bundles src/mcp-server + src/lib/ai into dist/mcp-server.mjs

(install.sh runs this for you — see Path C.)

---

Path A — Direct MCP Point your IDE's MCP config straight at the stack's mcp:start:

{
  "mcpServers": {
    "tech-lead-stack": {
      "command": "npm",
      "args": [
        "--prefix",
        "/path/to/tech-lead-stack",
        "--silent",
        "run",
        "mcp:start"
      ]
    }
  }
}

Best when: you use one IDE you configure by hand, working against the stack's own repo. The server automatically falls back to the stack's own skills, so nothing else is required.

---

Path B — Through the SLM Gate (mcp-gate) Instead of pointing your IDE at the stack directly, run the SLM Gate's mcp-gate and set its DOWNSTREAM_MCP env var to the stack's MCP artifact. The gate becomes the front door and forwards tool calls downstream to the tech-lead-stack MCP.

Best when: you want the gate's layer in front of the stack — small-language-model / model routing, request filtering, or aggregating several MCP servers behind a single endpoint — rather than talking to the stack in isolation. The mcp-gate configuration (flags beyond DOWNSTREAM_MCP) lives in the @zenithfoundry/slm-gate repo's own docs. (The stack is consumable by other tools the same way — e.g. voice-relay via STACK_REPO — because it is just a standalone artifact.)

slm-gate is the worked example, not a requirement. Any proxy that spawns or forwards to dist/mcp-server.mjs produces this same topology; only the registered name differs.

**One consequence worth kno