{"id":51183167,"url":"https://github.com/adityashubham1997/squad-public","last_synced_at":"2026-06-27T08:03:10.313Z","repository":{"id":353751397,"uuid":"1220779691","full_name":"adityashubham1997/squad-public","owner":"adityashubham1997","description":"Multi-agent AI development framework — any stack, any IDE, any cloud","archived":false,"fork":false,"pushed_at":"2026-06-20T23:23:20.000Z","size":788,"stargazers_count":1,"open_issues_count":0,"forks_count":1,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-06-21T01:03:17.411Z","etag":null,"topics":["agentic-ai","ai-agents","ai-framework","code-review","developer-tools","devops","generative-ai","llm","multi-agent","nodejs"],"latest_commit_sha":null,"homepage":"https://www.npmjs.com/package/squad-public","language":"JavaScript","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/adityashubham1997.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":"CONTRIBUTING.md","funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":".github/CODEOWNERS","security":"SECURITY.md","support":null,"governance":null,"roadmap":"ROADMAP.md","authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-04-25T10:18:39.000Z","updated_at":"2026-06-20T23:25:31.000Z","dependencies_parsed_at":"2026-06-21T01:00:29.755Z","dependency_job_id":null,"html_url":"https://github.com/adityashubham1997/squad-public","commit_stats":null,"previous_names":["adityashubham1997/sqad-public"],"tags_count":6,"template":false,"template_full_name":null,"purl":"pkg:github/adityashubham1997/squad-public","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/adityashubham1997%2Fsquad-public","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/adityashubham1997%2Fsquad-public/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/adityashubham1997%2Fsquad-public/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/adityashubham1997%2Fsquad-public/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/adityashubham1997","download_url":"https://codeload.github.com/adityashubham1997/squad-public/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/adityashubham1997%2Fsquad-public/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34845749,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-27T02:00:06.362Z","response_time":126,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["agentic-ai","ai-agents","ai-framework","code-review","developer-tools","devops","generative-ai","llm","multi-agent","nodejs"],"created_at":"2026-06-27T08:03:09.696Z","updated_at":"2026-06-27T08:03:10.295Z","avatar_url":"https://github.com/adityashubham1997.png","language":"JavaScript","funding_links":[],"categories":[],"sub_categories":[],"readme":"\u003cdiv align=\"center\"\u003e\n\n# SQUAD\n\n### 56 Specialist AI Agents. 5 Model Providers. 8 IDEs. Zero Dependencies.\n\n[![npm](https://img.shields.io/npm/v/squad-public.svg)](https://www.npmjs.com/package/squad-public)\n[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)\n[![Node.js \u003e=18](https://img.shields.io/badge/Node.js-%3E%3D18.0-green.svg)](https://nodejs.org)\n[![Tests](https://img.shields.io/badge/Tests-202%20passing-brightgreen.svg)](#testing)\n[![Zero Dependencies](https://img.shields.io/badge/Dependencies-0-success.svg)](#security--privacy)\n[![IDEs](https://img.shields.io/badge/IDEs-8%20supported-blueviolet.svg)](#supported-ides)\n[![Skills](https://img.shields.io/badge/Skills-34%20commands-orange.svg)](#skills-slash-commands)\n[![Agents](https://img.shields.io/badge/Agents-56-ff69b4.svg)](#agents)\n\n*The AI dev tool that replaces \"one model, one chat\" with a full engineering team.*\n\n\u003c/div\u003e\n\n---\n\n## The Problem\n\nEvery AI coding tool today works the same way: **one model, one chat window**, trying to be architect, security expert, test engineer, code reviewer, and product manager — all at once.\n\nThe result? Generic feedback. Missed edge cases. No blast radius awareness. No real adversarial review.\n\n## The SQUAD Solution\n\nSQUAD gives you a **team of 56 specialists** — each with a distinct lens, a specific job, and the inability to say \"looks good\" when it shouldn't.\n\n```mermaid\ngraph LR\n    YOU[👤 You] --\u003e|\"/dev-task\"| SQUAD{SQUAD\u003cbr/\u003eOrchestrator}\n    SQUAD --\u003e N[📊 Nova\u003cbr/\u003eRequirements]\n    SQUAD --\u003e A[🏗️ Atlas\u003cbr/\u003eArchitecture]\n    SQUAD --\u003e F[💻 Forge\u003cbr/\u003eCode]\n    SQUAD --\u003e C[🧪 Cipher\u003cbr/\u003eTests]\n    SQUAD --\u003e R[🔍 Raven\u003cbr/\u003eAdversarial]\n    SQUAD --\u003e S[🛡️ Aegis\u003cbr/\u003eSecurity]\n    N \u0026 A \u0026 F \u0026 C \u0026 R \u0026 S --\u003e|\"Structured\u003cbr/\u003eoutputs\"| V[🔥 Phoenix\u003cbr/\u003eVerdict]\n    V --\u003e|\"User Gate\"| YOU\n\n    style SQUAD fill:#4a90d9,color:#fff\n    style YOU fill:#2ecc71,color:#fff\n    style V fill:#e74c3c,color:#fff\n```\n\n### What makes SQUAD different\n\n| Feature | Other AI Tools | SQUAD |\n|---|---|---|\n| **Architecture awareness** | Grep the codebase | Pre-built Knowledge Graph — 2-hop blast radius in milliseconds |\n| **Code review** | One model says \"looks good\" | 5 parallel agents: adversarial + security + architecture + QA + code quality |\n| **Model selection** | Whatever the IDE uses | Auto-routes each agent to the right model (Opus for reasoning, Flash for docs) |\n| **Execution** | Sequential chat | True parallel agent dispatch (up to 5 concurrent on Claude Code) |\n| **Safety** | Trust the output | Phase gates — you approve before each phase advances |\n| **Learning** | None | `/evolve` — analyzes execution history, proposes evidence-backed skill improvements |\n| **Financial analysis** | N/A | 7 quant-grade agents: Beneish M-Score, Kelly criterion, EVT tail risk |\n| **IDE lock-in** | One IDE | Same 56 agents across 8 IDEs — Claude Code, Codex, Cursor, Windsurf, Kiro, Gemini, Devin, Antigravity |\n\n---\n\n## See It In Action\n\n### Software Development — `/dev-task`\n\n```\nYou:    \"/dev-task — implement JWT authentication\"\n\nPhase 1  → Nova finds 2 missing acceptance criteria in the story\n         → Atlas flags rate-limiting gap, shows KG blast radius (8 files)\n    ⏸ USER GATE — you review analysis, approve or correct\n\nPhase 2  → Forge writes code matching YOUR patterns (not boilerplate)\n         → Phase 1.5: characterization tests on current behavior BEFORE any changes\n    ⏸ USER GATE\n\nPhase 3  → Cipher generates tests following your test framework (Jest/pytest/etc)\n    ⏸ USER GATE\n\nPhase 4  → 5 reviewers run in parallel:\n           Raven (adversarial) + Atlas (architecture) + Sentinel (QA) + Forge (code) + Cipher (tests)\n         → Phoenix synthesizes: 0 critical, 1 major (null check missing on line 47)\n    ⏸ USER GATE\n\nPhase 5  → PR created, tracking logged\n```\n\n### Financial Analysis — `/financial-analysis`\n\n```\nYou:    \"/financial-analysis RELIANCE.NS\"\n\nPhase 0  → Asks what data you have (yfinance/Bloomberg/none)\n         → Provides Python snippet if needed, waits for you to paste output\n\nPhase 1  → Charts: RSI 61, above 200 SMA, bullish engulfing on daily\n           Options P/C ratio 0.72, IV squeeze building\n\nPhase 2  → Ledger: PE 24x vs sector 28x, FCF +18% YoY\n           Beneish M-Score -2.4 (safe), 3/25 forensic screens triggered\n\nPhase 3  → Quant: Sharpe 0.84, Kelly 11%, P(ruin|1yr) 2.3%\n           EVT tail risk: normal understates by 3.1x\n\nPhase 4  → Sage: Reinvestment runway ~7 years at current ROIC\n\nPhase 5  → Prism: Devil's advocate — regulatory risk is unpriced [VERIFIED-3]\n\nPhase 6  → 3 options: Buy / Wait / Avoid — each with Kelly fraction + CVaR\n```\n\n**Works offline. Zero npm dependencies. Same agents, config, and skills across 8 IDEs.**\n\n---\n\n## Table of Contents\n\n### Part I — Getting Started\n1. [Installation](#installation)\n2. [Setup](#setup)\n3. [Quick Start](#quick-start)\n\n### Part II — Understanding SQUAD\n4. [Core Concepts](#core-concepts)\n5. [The Grounding Waterfall](#the-grounding-waterfall)\n6. [How Agents Are Orchestrated](#how-agents-are-orchestrated)\n7. [Multi-Model Routing](#multi-model-routing)\n8. [Parallel Execution \u0026 Dispatch Paths](#parallel-execution--dispatch-paths)\n\n### Part III — All 56 Agents\n9. [Agent Packs Overview](#agent-packs-overview)\n10. [Core Agents (14)](#core-agents-14)\n11. [Extended Core (3)](#extended-core-3)\n12. [Math \u0026 Theory Pack (6)](#math--theory-pack-6)\n13. [AI/ML Pack (5)](#aiml-pack-5)\n14. [Systems \u0026 Data Pack (5)](#systems--data-pack-5)\n15. [Startup Pack (3)](#startup-pack-3)\n16. [Financial Pack (7)](#financial-pack-7)\n17. [Specialized Agents (13)](#specialized-agents-13)\n\n### Part IV — All 34 Skills\n18. [Skills (Slash Commands)](#skills-slash-commands)\n\n### Part V — Deep Dives\n19. [Supported IDEs](#supported-ides)\n20. [Supported Model Providers](#supported-model-providers)\n21. [Knowledge Graph](#knowledge-graph)\n22. [Financial \u0026 Consulting Analysis Suite](#financial--consulting-analysis-suite)\n23. [Skill Self-Evolution — /evolve](#skill-self-evolution--evolve)\n24. [Token Compression Engine](#token-compression-engine)\n\n### Part VI — Reference\n25. [Configuration Reference](#configuration-reference)\n26. [Project Structure](#project-structure)\n27. [Adding a New IDE](#adding-support-for-a-new-ide)\n28. [Adding a New Model Provider](#adding-support-for-a-new-language-model)\n29. [Security \u0026 Privacy](#security--privacy)\n30. [Testing](#testing)\n31. [FAQ](#faq)\n32. [Contributing](#contributing)\n33. [Credits \u0026 Acknowledgments](#credits--acknowledgments)\n\n---\n\n## Installation\n\n```bash\nnpx squad-public init\n```\n\nThat's it. One command. ~10 seconds.\n\nWith specific IDEs: `npx squad-public init --ide claude,cursor,windsurf`\n\nWithout npm: `curl -fsSL https://raw.githubusercontent.com/adityashubham1997/squad-public/main/install.sh | bash`\n\n**Requirements:** Node.js \u003e= 18.\n\n**What `init` does:**\n\n```mermaid\nflowchart LR\n    A[1. Sync\u003cbr/\u003esquad-method/] --\u003e B[2. Detect Stack\u003cbr/\u003e15 langs · 40+ frameworks]\n    B --\u003e C[3. Detect Cloud\u003cbr/\u003eAWS · GCP · Azure · IaC]\n    C --\u003e D[4. Build\u003cbr/\u003eKnowledge Graph]\n    D --\u003e E[5. Deploy Skills\u003cbr/\u003eto 8 IDEs]\n    E --\u003e F[6. Generate\u003cbr/\u003econfig.yaml]\n\n    style A fill:#3498db,color:#fff\n    style D fill:#e74c3c,color:#fff\n    style F fill:#2ecc71,color:#fff\n```\n\nOn subsequent runs, `init` **syncs** new agents/skills/tools while preserving your `config.yaml` and `output/`.\n\n---\n\n## Setup\n\nAfter installation, run `/squad-setup` inside your IDE:\n\n| # | Question | Required | Example |\n|---|---|---|---|\n| 1 | Your name | ✅ | \"Aditya\" |\n| 2 | Your role | ✅ | \"Senior Engineer\" |\n| 3 | Team name | ✅ | \"Platform\" |\n| 4 | Company name | | \"Acme Corp\" |\n| 5 | Industry / domain | | \"fintech\" |\n| 6 | Project name | | \"payments-api\" |\n| 7 | Project description | | \"REST API for payment processing\" |\n| 8 | Project type | | \"api\" |\n| 9 | Sprint board URL | | Auto-detects Jira/Linear/GitHub/Shortcut/Notion |\n\nShows a **config completeness score** at the end. Without `/squad-setup`, SQUAD still works — tech detection ran at install. But agents won't know your name, team, or project context.\n\n---\n\n## Quick Start\n\n| Command | What it does |\n|---|---|\n| `/dev-task` | Full 6-phase implementation: analyse → spec → code → test → review → PR |\n| `/review-code` | Pre-commit review by Forge + Raven + Sentinel |\n| `/brainstorm` | Multi-agent ideation with all 56 agents |\n| `/financial-analysis` | Quant-grade forensic financial analysis by ticker |\n| `/refresh` | Scan workspace, rebuild knowledge graphs and context |\n| `/health` | Agent effectiveness report with skill utility scores |\n\nEvery skill pauses at **user gates** — you approve before each phase advances.\n\n---\n\n# Part II — Understanding SQUAD\n\n## Core Concepts\n\nBefore diving into architecture, here's how SQUAD's pieces fit together:\n\n```mermaid\ngraph TB\n    subgraph \"What You See\"\n        SKILL[\"🎯 Skill\u003cbr/\u003e(e.g. /dev-task, /brainstorm)\"]\n        GATE[\"⏸ User Gate\u003cbr/\u003eYou approve each phase\"]\n    end\n\n    subgraph \"What Runs Under the Hood\"\n        AO[\"Agent Orchestrator\u003cbr/\u003eWHO runs? What order?\"]\n        MR[\"Model Router\u003cbr/\u003eWHICH model per agent?\"]\n        DP[\"Dispatch Engine\u003cbr/\u003eParallel or sequential?\"]\n    end\n\n    subgraph \"What Agents Read\"\n        KG[\"📊 Knowledge Graph\u003cbr/\u003eBlast radius · God nodes · Coverage\"]\n        CTX[\"📝 Context Files\u003cbr/\u003eCONTEXT.md · DEEP-CONTEXT.md\"]\n        FRAG[\"📦 Fragments\u003cbr/\u003eStack · Cloud · Rubric\"]\n    end\n\n    subgraph \"What Improves Over Time\"\n        TRK[\"📈 tracking.jsonl\u003cbr/\u003eEvery skill run logged\"]\n        EVO[\"🧬 /evolve\u003cbr/\u003eSelf-improving skills\"]\n    end\n\n    SKILL --\u003e AO\n    AO --\u003e MR\n    MR --\u003e DP\n    DP --\u003e|\"runs\"| AGENTS[\"56 Agents\"]\n    AGENTS --\u003e|\"read\"| KG \u0026 CTX \u0026 FRAG\n    AGENTS --\u003e|\"produce\"| OUTPUT[\"Structured Output\"]\n    OUTPUT --\u003e GATE\n    GATE --\u003e|\"approved\"| NEXT[\"Next Phase\"]\n    OUTPUT --\u003e|\"logged to\"| TRK\n    TRK --\u003e|\"analyzed by\"| EVO\n\n    style SKILL fill:#4a90d9,color:#fff\n    style GATE fill:#2ecc71,color:#fff\n    style KG fill:#e74c3c,color:#fff\n    style EVO fill:#9b59b6,color:#fff\n```\n\n**Key principles:**\n- **Agents are lazy-loaded** — only agents needed for the current skill enter context\n- **Fragments are conditional** — Python projects load Python rubric; AWS projects load AWS fragments\n- **Everything is deterministic** — same inputs → same agent dispatch → same outputs (content-hashed)\n- **Nothing phones home** — zero network calls, zero telemetry, zero dependencies\n\n---\n\n## The Grounding Waterfall\n\nBefore any agent does work, SQUAD follows an **evidence-first protocol** — a strict hierarchy of what to read, in what order.\n\n```mermaid\ngraph TD\n    L0[\"Level 0 — Identity\u003cbr/\u003eCONTEXT.md · CLAUDE.md\u003cbr/\u003e~300 tokens · Always loaded\"]\n    L0 --\u003e L0B[\"Level 0b — Architecture\u003cbr/\u003eDEEP-CONTEXT.md\u003cbr/\u003eKG_REPORT.md\"]\n    L0B --\u003e L1A[\"Level 1a — Knowledge Graph\u003cbr/\u003egraph.json\u003cbr/\u003eBlast radius · God nodes · Dependencies\u003cbr/\u003e⚡ One read, not 10 greps\"]\n    L1A --\u003e L1B[\"Level 1b — Code Search\u003cbr/\u003egrep / ripgrep\u003cbr/\u003eOnly if KG doesn't answer the question\"]\n    L1B --\u003e L2[\"Level 2 — Fragments\u003cbr/\u003eStack · Cloud · Rubric · Tracker\u003cbr/\u003eConditional on config.yaml\"]\n    L2 --\u003e L3[\"Level 3 — Nothing Found?\u003cbr/\u003e🛑 STOP\u003cbr/\u003ePresent assumptions · Await user approval\"]\n\n    style L0 fill:#3498db,color:#fff\n    style L1A fill:#e74c3c,color:#fff\n    style L3 fill:#f39c12,color:#fff\n```\n\n**Why this matters:** The KG answers \"what depends on this file?\" in one JSON read — what would otherwise take 3–10 grep commands. Pre-computing blast radius, test coverage, and god-node status saves **~80% of exploration tokens** per workflow.\n\n### Context Digest (mandatory Phase 1 output)\n\nEvery `/dev-task` starts with a Context Digest — agents can't proceed until this is populated:\n\n```\n━━━ CONTEXT DIGEST ━━━\n\nFiles Read:\n  ✅ CONTEXT.md (repo) — 200 lines\n  ✅ DEEP-CONTEXT.md — 180 lines\n  ✅ KG_REPORT.md — 45 nodes, 38 edges\n  ❌ complete-flow.md — not found\n\nScope Analysis (from KG):\n  Files in change path: 4\n  God nodes in scope: none\n  Untested files in scope: lib/generate/ide-skills.js\n  Cross-community changes: NO\n\nBlast Radius: LOW — 3 reverse deps, 2 test files covering scope\n\nAssumptions:\n  [ASSUMPTION-1]: ... — CONFIDENCE: HIGH\n```\n\n---\n\n## How Agents Are Orchestrated\n\nThe Agent Orchestrator builds a **dependency DAG**, identifies parallel layers, and enforces completion.\n\n### Phase 1 — Analysis (DAG with fan-out)\n\n```mermaid\ngraph LR\n    O[\"🔬 Oracle\u003cbr/\u003eResearch + KG\"]\n    F[\"💻 Forge\u003cbr/\u003eFramework detect\"]\n    A[\"🏗️ Atlas\u003cbr/\u003eArchitecture\"]\n\n    O \u0026 F \u0026 A --\u003e|\"sync barrier\"| N[\"📊 Nova\u003cbr/\u003eAssemble requirements\"]\n    N --\u003e C[\"📋 Compass\u003cbr/\u003eFrame value + summary\"]\n    C --\u003e|\"⏸ User Gate\"| YOU[\"👤 You approve\"]\n\n    style O fill:#f39c12,color:#fff\n    style F fill:#f39c12,color:#fff\n    style A fill:#f39c12,color:#fff\n    style N fill:#3498db,color:#fff\n    style YOU fill:#2ecc71,color:#fff\n```\n\n**Layer 1 (parallel):** Oracle + Forge + Atlas — no dependencies, fan out simultaneously.\n**Sync barrier:** Wait for all 3 to complete + validate their outputs.\n**Layer 2 (sequential):** Nova (consumes Layer 1 outputs) → Compass (consumes Nova's output).\n\n### Phase 5 — Multi-Agent Review (5 parallel reviewers)\n\n```mermaid\ngraph LR\n    R[\"🔍 Raven\u003cbr/\u003eAdversarial\"]\n    A[\"🏗️ Atlas\u003cbr/\u003eArchitecture\"]\n    S[\"🛡️ Sentinel\u003cbr/\u003eSecurity\"]\n    FG[\"💻 Forge\u003cbr/\u003eCode quality\"]\n    CI[\"🧪 Cipher\u003cbr/\u003eTest coverage\"]\n\n    R \u0026 A \u0026 S \u0026 FG \u0026 CI --\u003e|\"sync barrier\"| P[\"🔥 Phoenix\u003cbr/\u003eSynthesis verdict\"]\n    P --\u003e|\"⏸ User Gate\"| YOU[\"👤 You\"]\n\n    style P fill:#e74c3c,color:#fff\n    style YOU fill:#2ecc71,color:#fff\n```\n\n### Guarantees (enforced by 30 hard rules)\n\n| Rule | What It Guarantees |\n|---|---|\n| **R3 — Output Contracts** | Every agent declares inputs/outputs in YAML. Schema-validated. |\n| **R4 — Determinism** | Inputs content-hashed (SHA-256). Same hash → same dispatch. Run manifest logged. |\n| **R6 — Completion Verification** | Expected agents vs. actual agents compared after every phase. Missing → re-dispatch. |\n| **R8 — Anti-Skip** | NEVER skip an agent to save time. Every declared agent MUST run. |\n| **R9 — Gate Ledger** | Phases don't advance without user approval. Gate status persisted to disk. |\n\n---\n\n## Multi-Model Routing\n\nSQUAD doesn't use one model for everything. Each agent is routed to the **right model for its task**.\n\n```mermaid\ngraph TD\n    REQ[\"Incoming Agent Request\"]\n    REQ --\u003e WM{\"Workspace Mode?\"}\n    WM --\u003e|\"quality\"| HEAVY[\"🔴 Heavy Tier\u003cbr/\u003eClaude Opus 4 · o3\"]\n    WM --\u003e|\"budget\"| FAST[\"🟢 Fast Tier\u003cbr/\u003eGPT-4o-mini · Flash\"]\n    WM --\u003e|\"balanced\"| PO{\"Phase Override?\"}\n    PO --\u003e|\"phase_6 (PR)\"| FAST\n    PO --\u003e|\"no\"| BR{\"Blast Radius\u003cbr/\u003e\u003e threshold?\"}\n    BR --\u003e|\"god node\u003cbr/\u003e(degree \u003e 20)\"| HEAVY\n    BR --\u003e|\"normal\"| AO{\"Agent Override?\"}\n    AO --\u003e|\"raven: heavy\"| HEAVY\n    AO --\u003e|\"scribe: fast\"| FAST\n    AO --\u003e|\"default\"| DEFAULT[\"🟡 Default Tier\u003cbr/\u003eClaude Sonnet 4 · GPT-4o\"]\n\n    style HEAVY fill:#e74c3c,color:#fff\n    style DEFAULT fill:#f39c12,color:#fff\n    style FAST fill:#2ecc71,color:#fff\n```\n\n### Priority chain (highest wins)\n\n```\nworkspace_mode → phase_override → blast_radius → budget_cap → agent_override → default\n```\n\n### Default agent assignments\n\n| Agent | Model Tier | Reason |\n|---|---|---|\n| **Raven** | 🔴 Heavy | Adversarial second-order reasoning |\n| **Atlas** | 🔴 Heavy | Architecture blast radius + threat modeling |\n| **Phoenix** | 🔴 Heavy | Complex multi-agent verdict merging |\n| **Forge** | 🟡 Default | Good balance of speed and quality |\n| **Scribe** | 🟢 Fast | Structural pattern matching, no deep reasoning |\n| All others | 🟡 Default | Unless overridden in config.yaml |\n\n**Auto-upgrade:** When an agent is about to modify a **god node** (KG degree \u003e 20), the router automatically upgrades to the heavy tier — no configuration needed.\n\n---\n\n## Parallel Execution \u0026 Dispatch Paths\n\nNot all IDEs can run agents in parallel. SQUAD auto-detects what's available and picks the optimal path:\n\n```mermaid\ngraph TD\n    DETECT[\"Auto-detect IDE capabilities\"]\n    DETECT --\u003e AT{\"Agent() tool\u003cbr/\u003eavailable?\"}\n    AT --\u003e|\"yes\"| PA[\"✅ Path A — Native Subagent\u003cbr/\u003eClaude Code\u003cbr/\u003eMax 5 concurrent\"]\n    AT --\u003e|\"no\"| CLI{\"CLI on PATH?\u003cbr/\u003e(codex / claude)\"}\n    CLI --\u003e|\"yes\"| PB[\"✅ Path B — CLI Subprocess\u003cbr/\u003eCodex · Kiro · Gemini · Devin\u003cbr/\u003eMax 3 concurrent\"]\n    CLI --\u003e|\"no\"| PC[\"⚠️ Path C — Sequential\u003cbr/\u003eCursor · Windsurf · Antigravity\u003cbr/\u003e1 at a time\"]\n\n    style PA fill:#2ecc71,color:#fff\n    style PB fill:#3498db,color:#fff\n    style PC fill:#f39c12,color:#fff\n```\n\n| Path | True Parallelism | What's preserved | What differs |\n|---|---|---|---|\n| **A** (Native) | ✅ Max 5 concurrent | All correctness guarantees | Best wall-clock |\n| **B** (CLI) | ✅ Max 3 concurrent | All correctness guarantees | Good wall-clock |\n| **C** (Sequential) | ❌ One at a time | All correctness guarantees | Slowest wall-clock |\n\n**Path C preserves:** dependency ordering, output contracts, run manifest, determinism hashing, anti-skip rules, gate ledger, completion verification. Only wall-clock time and per-agent model isolation differ.\n\n---\n\n# Part III — All 56 Agents\n\n## Agent Packs Overview\n\n```mermaid\npie title Agent Distribution (56 total)\n    \"Core\" : 14\n    \"Extended Core\" : 3\n    \"Math \u0026 Theory\" : 6\n    \"AI/ML\" : 5\n    \"Systems \u0026 Data\" : 5\n    \"Startup\" : 3\n    \"Financial\" : 7\n    \"Specialized\" : 13\n```\n\n| Pack | Count | Primary Use Case |\n|---|---|---|\n| **Core** | 14 | Software development lifecycle |\n| **Extended Core** | 3 | Security architecture, platform ops, cross-agent oversight |\n| **Math \u0026 Theory** | 6 | Algorithm correctness, complexity, proofs |\n| **AI/ML** | 5 | Neural networks, model evaluation, edge AI |\n| **Systems \u0026 Data** | 5 | Distributed systems, databases, data pipelines |\n| **Startup** | 3 | Founding strategy, GTM, financial modeling |\n| **Financial** | 7 | Market analysis, trading, investment |\n| **Specialized** | 13 | Games, security, performance, DevOps, data |\n\nAll agents install together. Packs are logical groupings — agents are **lazy-loaded** per skill (only the agents a skill needs enter context).\n\n---\n\n## Core Agents (14)\n\nThe foundation. These agents cover the entire software development lifecycle.\n\n| Agent | Icon | Role | What They Actually Do |\n|---|---|---|---|\n| **Nova** | 📊 | Requirements Analyst | Finds missing acceptance criteria, validates stories, identifies gaps BEFORE work begins |\n| **Atlas** | 🏗️ | Solution Architect | Architecture blast radius (from KG), threat modeling, technology trade-offs |\n| **Forge** | 💻 | Implementation Lead | Writes code matching YOUR patterns. Self-reviews before handing off. |\n| **Cipher** | 🧪 | QA Engineer | Test generation following your test framework. Coverage analysis. TDD enforcement. |\n| **Sentinel** | 🧪 | QA Architect | Test strategy, risk-based planning, test pyramid balance |\n| **Raven** | 🔍 | Adversarial Reviewer | Actively tries to break your code. Logic bugs, edge cases, second-order effects. |\n| **Catalyst** | 🚀 | Release Engineer | Release readiness, quality gate validation, compliance (L10N, security, a11y) |\n| **Oracle** | 🔬 | Technical Researcher | Domain research, precedent analysis, codebase investigation |\n| **Scribe** | 📚 | Technical Writer | Documentation, changelogs, API docs |\n| **Compass** | 📋 | Product Manager | Value framing, story validation, scope control |\n| **Tempo** | 🎯 | Scrum Master | Sprint status, velocity tracking, retrospectives |\n| **Aegis** | 🛡️ | Security Engineer | OWASP Top 10, auth/authz audit, secrets management, CVE scanning |\n| **Stratos** | ☁️ | Cloud Architect | Cloud infra design, IaC review, cost optimization |\n| **Phoenix** | 🔥 | DevOps / SRE | Synthesizes multi-agent findings into a single actionable verdict |\n\n## Extended Core (3)\n\nAgents that fill unique functional lanes not covered by the 14 core agents.\n\n| Agent | Icon | Role | What They Actually Do |\n|---|---|---|---|\n| **Trinity** | 🛡️ | Security Architect | Access control design, STRIDE threat modeling, privilege escalation analysis |\n| **Otis** | 🔧 | Platform Specialist | Build systems, deploy verification, framework detection |\n| **Krishna** | 🌟 | Omniscient Overseer | Cross-agent flaw detection, convergence forcing, identifies 100x solutions |\n\n## Math \u0026 Theory Pack (6)\n\nFor algorithm correctness, complexity analysis, and mathematical proofs.\n\n| Agent | Icon | Role | Specialty |\n|---|---|---|---|\n| **Tao** | ∞ | Lead Mathematician | Proof construction, complexity bounds |\n| **Knuth** | 📐 | Algorithm Analyst | Exact running time, literate code analysis |\n| **Ramanujan** | ✨ | Intuitive Mathematician | Radical shortcuts, pattern recognition |\n| **Hardy** | 🔬 | Rigorous Mathematician | Proof validation, counter-example construction |\n| **Pearl** | 🔗 | Lead Statistician | Causal inference, Bayesian networks, DAGs |\n| **Gelman** | 📊 | Bayesian Statistician | Model critique, posterior predictive checks |\n\n## AI/ML Pack (5)\n\nFor neural network architecture, model evaluation, and edge deployment.\n\n| Agent | Icon | Role | Specialty |\n|---|---|---|---|\n| **Andrej** | 🧠 | AI Supervisor | Neural nets from scratch, training loops |\n| **Yann** | 🌊 | Chief AI Scientist | World models, self-supervised learning |\n| **Scott** | 📱 | On-Device AI Architect | Quantization, edge deployment, latency budgets |\n| **Woz** | 🔓 | Open Source AI Lead | Reproducibility, open-weight models |\n| **Percy** | 📏 | AI Eval Lead | HELM benchmarks, bias/fairness, calibration |\n\n## Systems \u0026 Data Pack (5)\n\nFor distributed systems, database design, and data pipeline engineering.\n\n| Agent | Icon | Role | Specialty |\n|---|---|---|---|\n| **Jeff** | 🌐 | Distributed Systems Lead | Scale 1000x, partitioning, consensus |\n| **Sanjay** | ⚙️ | Systems Pair Programmer | Memory layout, lock contention, cache lines |\n| **Stonebraker** | 🗄️ | Database Architect | Workload-specific DB design, OLTP vs OLAP |\n| **Reynold** | 🔀 | Data Systems Engineer | Pipelines, query optimization, data flow |\n| **Kyle** | 💥 | DB Correctness Lead | Jepsen-style testing, consistency verification |\n\n## Startup Pack (3)\n\nFor founding strategy, go-to-market, and unit economics — grounded in your actual codebase.\n\n| Agent | Icon | Role | Focus |\n|---|---|---|---|\n| **Richard** | 👑 | Startup CEO | Product-market fit, vision, OKRs |\n| **Monica** | 📢 | Startup CMO | Growth loops, GTM strategy, personas |\n| **Jared** | 💰 | Startup CFO | Unit economics, runway modeling, pricing |\n\n\u003e **`/startup-founding`** scans your actual codebase and project structure to build context-aware startup strategy — not generic advice.\n\n## Financial Pack (7)\n\nQuant-grade agents for market analysis, forensic accounting, and investment research.\n\n| Agent | Icon | Role | Key Methods |\n|---|---|---|---|\n| **Charts** | 📉 | Technical Analyst | RSI/MACD, options flow, volume profile, multi-timeframe confluence |\n| **Ledger** | 📊 | Forensic Analyst | Beneish M-Score, Benford's Law, accrual anomaly, footnote forensics |\n| **Herald** | 📡 | Signal Analyst | Earnings NLP, insider activity, credit market divergence, alt data |\n| **Sage** | 🔬 | Structural Researcher | Industry S-curves, moat velocity, Bass diffusion, causal inference |\n| **Maven** | 📐 | Strategic Architect | Decision theory, EVPI, Kelly criterion, pre-mortem (7+ failure paths) |\n| **Quant** | 📈 | Chief Risk Analyst | EVT tail risk, copulas, ruin probability, Monte Carlo, factor decomposition |\n| **Prism-Adversarial** | ⚡ | Adversarial Epistemics | 12-lens challenge, superforecasting, Dutch Book audit, falsifiability cert |\n\n## Specialized Agents (13)\n\nDomain experts for games, performance, DevOps, data, and creative problem-solving.\n\n| Agent | Icon | Role | Domain |\n|---|---|---|---|\n| **Shadow** | 🕵️ | Security Engineer | Pen-test mindset, cloud/code/infra security |\n| **Pixel** | 🎮 | Game Developer | Game engine code, render pipelines, physics |\n| **Quest** | 🗺️ | Product Discovery | Game mechanics, balance, progression |\n| **Lore** | 📜 | Knowledge Engineer | Narrative design, world-building, dialogue |\n| **Spark** | ⚡ | AI Developer | AI/ML framework integration in production |\n| **Muse** | 🎨 | AI Researcher | Research synthesis, paper analysis |\n| **Dynamo** | 🔋 | Performance Engineer | N+1 detection, query optimization, profiling |\n| **Flux** | 🔄 | DevOps Automation | CI/CD pipelines, deployment automation |\n| **Index** | ⚡ | Query Optimizer | SQL tuning, index strategy, execution plans |\n| **Kernel** | ⚙️ | Systems Programmer | OS-level code, memory management, concurrency |\n| **Neuron** | 🧬 | ML Engineer | ML pipelines, model evaluation, data quality |\n| **Prism** | 🔺 | Data Analyst | SQL analytics, data models, dashboard quality |\n| **Titan** | 🏔️ | Infrastructure | Standards enforcement, quality gates |\n\n---\n\n# Part IV — All 34 Skills\n\n## Skills (Slash Commands)\n\n### Development \u0026 Code Quality\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/dev-task` | Nova, Atlas, Forge, Cipher, Raven, Sentinel | Full 6-phase implementation: analyse → spec → code → test → review → PR |\n| `/review-code` | Forge, Raven, Sentinel | Quick pre-commit review of uncommitted changes |\n| `/review-pr` | Raven, Atlas, Sentinel, Forge, Cipher, Catalyst | Full pull request code review |\n| `/review-story` | Raven, Atlas, Sentinel, Forge, Cipher | Validate implementation against acceptance criteria |\n| `/dev-analyst` | Nova, Atlas, Oracle, Forge | Deep story analysis: feasibility, architecture, effort |\n\n### Testing \u0026 QA\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/qa-task` | Cipher, Sentinel, Raven | End-to-end QA: dependency analysis → test plan → tests |\n| `/test-story` | Cipher, Sentinel | Story-aware test generation following existing patterns |\n| `/test-repo` | Cipher | Run test suite, analyze results, report coverage |\n| `/test-project` | Cipher | Cross-repo test health report |\n\n### Product \u0026 Planning\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/create-prd` | Compass, Nova, Atlas, Oracle | Multi-agent product requirements document |\n| `/create-story` | Compass, Nova | Story with GIVEN/WHEN/THEN acceptance criteria |\n| `/product-researcher` | Oracle, Compass | Deep product research across tracker, web, codebase |\n\n### Multi-Agent Sessions\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/brainstorm` | All agents | Multi-perspective brainstorming — all 56 agents available |\n| `/assemble` | All agents | Full group discussion — architecture debates, post-mortems |\n\n### Financial \u0026 Strategy\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/financial-analysis` | Charts, Ledger, Herald, Sage, Maven, Quant, Prism-Adversarial | 7-phase forensic analysis by ticker. Adapts to data subscriptions. |\n| `/market-research` | Oracle, Sage, Herald, Prism-Adversarial | Structural market \u0026 industry deep-dive |\n| `/consulting-brief` | Maven, Sage, Prism-Adversarial, Quant | Strategic brief: pre-mortem + EVPI + Kelly + 3 options |\n| `/startup-founding` | Richard, Monica, Jared, Oracle, Compass, Atlas, Nova | Codebase-aware startup strategy |\n\n### Sprint \u0026 Delivery\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/setup` | Tempo | Configure user, team, company, project, tracker |\n| `/standup` | Tempo | Auto-generate daily standup from git + tracker |\n| `/retro` | Tempo, Compass, Scribe | Sprint retrospective with live tracker data |\n| `/current-sprint` | Tempo | Sprint status at a glance |\n\n### Domain Audits\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/data-audit` | Neuron, Prism | ML pipeline and data quality audit |\n| `/db-audit` | Dynamo | Database schema, query performance, migration safety |\n| `/infra-audit` | Stratos, Aegis | Infrastructure observability and monitoring |\n| `/os-audit` | Kernel | OS-level code, process management, systems patterns |\n| `/game-review` | Pixel, Quest | Game engine: performance, networking, design |\n| `/ai-ideate` | — | Design agentic workflow and AI automation ideas |\n| `/ai-workflow-audit` | — | Audit existing AI/LLM integrations in the codebase |\n\n### Meta \u0026 Learning\n\n| Skill | Agents | What It Does |\n|---|---|---|\n| `/evolve` | — | Skill self-evolution: analyze tracking → propose edits → branch |\n| `/health` | — | Agent effectiveness, skill utility grades (A–D), evolution candidates |\n| `/refresh` | — | Scan workspace, rebuild KG, regenerate context files |\n| `/refresh-git` | — | Enrich context from PR review history and git patterns |\n| `/git-learn` | Scribe | Extract learnings from PR history, enrich CONTEXT.md |\n\n---\n\n# Part V — Deep Dives\n\n## Supported IDEs\n\n| IDE | Parallel | Multi-Model | Hook Enforcement | Skill Format |\n|---|---|---|---|---|\n| **Claude Code** | ✅ Max 5 | ✅ Anthropic + OpenAI + Google | Automatic (settings.json) | `.claude/skills/` |\n| **Codex** (OpenAI) | ✅ Max 3 | ✅ OpenAI + Anthropic | Script (hooks.sh) | `.codex/skills/` |\n| **Kiro** (AWS) | ✅ Max 3 | ✅ Bedrock + Q + OpenAI + Google + Anthropic | Script (hooks.sh) | `.kiro/skills/` |\n| **Gemini** (Google) | ✅ Max 3 | ✅ Google + Anthropic + OpenAI | Script (hooks.sh) | `.gemini/skills/` |\n| **Devin** (Cognition) | ✅ Max 3 | ✅ Anthropic + OpenAI + Google | Script (hooks.sh) | `.devin/skills/` |\n| **Cursor** | ❌ Sequential | ✅ Anthropic + OpenAI + Google | Script (hooks.sh) | `.cursor/rules/*.mdc` |\n| **Windsurf** | ❌ Sequential | ❌ Single model | Script (hooks.sh) | `.windsurf/skills/` |\n| **Antigravity** | ❌ Sequential | ✅ Anthropic + OpenAI | Script (hooks.sh) | `.agent/skills/` |\n\n---\n\n## Supported Model Providers\n\n| Provider | Models | Best For |\n|---|---|---|\n| **Anthropic** | Claude Opus 4, Claude Sonnet 4 | Reasoning, code generation, implementation |\n| **OpenAI** | o3, GPT-4o, GPT-4o-mini | Security reasoning, fast structured output |\n| **Google** | Gemini 2.5 Pro, Gemini 2.0 Flash | Long-context research (1M tokens) |\n| **Amazon Bedrock** | Claude via Bedrock, Titan, Llama 3 | AWS-native multi-model gateway |\n| **Amazon Q** | Q Developer | AWS-specific codebase knowledge |\n\n---\n\n## Knowledge Graph\n\nSQUAD includes a built-in, zero-dependency knowledge graph that pre-computes what agents need to know about your codebase.\n\n```mermaid\nflowchart LR\n    subgraph \"4-Pass Pipeline\"\n        P1[\"Pass 1 — build.js\u003cbr/\u003eScan imports · Build edges\"]\n        P2[\"Pass 2 — git-pass.js\u003cbr/\u003eCo-change · Churn · Hotspots\"]\n        P3[\"Pass 3 — cluster.js\u003cbr/\u003eCommunity detection\"]\n        P4[\"Pass 4 — analyze.js\u003cbr/\u003eSurprise edges · Complexity\"]\n    end\n    P1 --\u003e P2 --\u003e P3 --\u003e P4\n    P4 --\u003e G[\"graph.json + graph.html + KG_REPORT.md\"]\n    style P1 fill:#3498db,color:#fff\n    style P4 fill:#e74c3c,color:#fff\n    style G fill:#2ecc71,color:#fff\n```\n\n```bash\nnode squad-method/tools/knowledge-graph/build.js \u003crepo-path\u003e\n# Optional: function-level AST analysis\nnode squad-method/tools/knowledge-graph/build.js \u003crepo-path\u003e --ast\n```\n\n### 4-Pass Analysis Pipeline\n\n| Pass | Module | What It Does |\n|---|---|---|\n| 1 | `build.js` | Scan source files, extract imports, build dependency edges |\n| 2 | `git-pass.js` | Git history: co-change patterns, churn hotspots, author count |\n| 3 | `cluster.js` | Label propagation community detection (graph-aware, not directory-based) |\n| 4 | `analyze.js` | Surprise edges, hotspot scoring, complexity grading (A–F) |\n\nOptional Pass 5 (`--ast`): function-level nodes and call-graph edges via regex or tree-sitter.\n\n### Supported Languages (15)\n\nJavaScript, TypeScript, Python, Go, Rust, Java, Ruby, C, C++, C#, Swift, Kotlin, Scala, PHP, Protocol Buffers, GraphQL\n\n### Output\n\n```\n\u003crepo\u003e/knowledge-graph-out/\n├── graph.json      ← Full graph: nodes, edges, communities, hotspots, complexity\n├── graph.html      ← Interactive D3-powered force-directed visualization\n└── KG_REPORT.md    ← Human-readable analysis for agents\n```\n\n### Query API\n\nAgents can query the graph programmatically via `squad-method/tools/knowledge-graph/query.js`:\n\n```javascript\nimport { loadGraph, reverseDeps, godNodes, untestedFiles, ripple, shortestPath } from './query.js';\n\nconst graph = loadGraph('/path/to/repo');\nreverseDeps(graph, 'lib/auth/login.js');  // what breaks if I change this?\ngodNodes(graph);                          // files with degree \u003e 30\nuntestedFiles(graph);                     // source files with no tests\nripple(graph, 'lib/auth/login.js', 2);    // 2-hop blast radius\n```\n\n### Context Prioritization\n\nGiven a task description, `prioritize.js` ranks which files agents should read first:\n\n```javascript\nimport { prioritize } from './prioritize.js';\nconst ranked = prioritize('fix authentication login flow', graph, { topN: 20 });\n// Returns files sorted by: keyword match × degree centrality × test coverage gap\n```\n\n### Incremental Updates\n\nFor large repos, `incremental.js` updates only affected nodes/edges instead of a full rebuild — falls back to full rebuild if \u003e 30% of files changed.\n\n### Why KG before grep?\n\n| Question | Without KG | With KG |\n|---|---|---|\n| \"What depends on this file?\" | 3–10 grep commands | One graph edge lookup |\n| \"Is this file high-risk?\" | Manual analysis | God node flag + hotspot score |\n| \"What tests cover this?\" | Grep for imports | Test edge query |\n| \"What's the blast radius?\" | Recursive grep | 2-hop reachability (instant) |\n\nSaves **~80% of exploration tokens** per workflow.\n\n---\n\n## Financial \u0026 Consulting Analysis Suite\n\nSeven quant-grade agents across four analysis streams — triggered by ticker symbol, adapts to your data subscriptions.\n\n\u003e **Design principle:** \"McKinsey gives you frameworks. Renaissance Technologies gives you edge. Every claim is falsifiable. Every conclusion has a confidence interval.\"\n\n### How `/financial-analysis` works\n\n```mermaid\nflowchart TD\n    IN[\"Phase 0 — Intake\u003cbr/\u003eTicker + data source\"]\n    IN --\u003e PAR\n\n    subgraph PAR[\"4 Parallel Streams\"]\n        T[\"📉 Charts\u003cbr/\u003eTechnical\"]\n        F[\"📊 Ledger\u003cbr/\u003eForensic\"]\n        Q[\"📈 Quant + Herald\u003cbr/\u003eQuantitative + Signals\"]\n        R[\"🔬 Sage + Maven\u003cbr/\u003eResearch + Strategy\"]\n    end\n\n    PAR --\u003e ADV[\"⚡ Prism-Adversarial\u003cbr/\u003e12-lens · Dutch Book · Falsifiability\"]\n    ADV --\u003e REC[\"3 Options: Buy / Wait / Avoid\u003cbr/\u003eKelly fraction + CVaR + ruin prob\"]\n\n    style IN fill:#3498db,color:#fff\n    style ADV fill:#e74c3c,color:#fff\n    style REC fill:#2ecc71,color:#fff\n```\n\n### Data source adaptation\n\n| What You Have | What Gets Unlocked |\n|---|---|\n| **Nothing** | LLM training data only — tagged [LLM-TRAINING], lower confidence |\n| **yfinance (free)** | Provides Python snippet → you run + paste → full OHLCV + options |\n| **Screener.in / Tickertape** | Indian fundamentals + sector context |\n| **TradingView** | Paste chart key levels + indicators |\n| **Bloomberg / Reuters** | Full data: real-time, options chain, insider flow, transcripts |\n| **Earnings call transcript** | Herald runs Shannon entropy + tone shift analysis |\n\n### 4-Gate Verification Protocol\n\nEvery major claim goes through four gates:\n\n```mermaid\ngraph LR\n    G1[\"Gate 1\u003cbr/\u003eEMPIRICAL\"] --\u003e G2[\"Gate 2\u003cbr/\u003eMATHEMATICAL\"]\n    G2 --\u003e G3[\"Gate 3\u003cbr/\u003eLOGICAL/CAUSAL\"]\n    G3 --\u003e G4[\"Gate 4\u003cbr/\u003eADVERSARIAL\"]\n    G4 --\u003e V[\"[VERIFIED-4]\"]\n    style G1 fill:#3498db,color:#fff\n    style G4 fill:#e74c3c,color:#fff\n    style V fill:#2ecc71,color:#fff\n```\n\nClaims classified: `[VERIFIED-4]` (all gates) → `[VERIFIED-3]` → `[VERIFIED-2]` → `[UNVERIFIED]` (never in recommendations).\n\n### Agent specializations\n\n- **Ledger**: Beneish M-Score, Benford's Law, Lev-Thiagarajan 12 signals, accrual anomaly (Jones Model), footnote forensics, DuPont 5-factor\n- **Herald**: Granger causality validation, Shannon entropy of earnings calls, Breeden-Litzenberger options-implied distributions, Bayesian composite scoring\n- **Sage**: Bass diffusion model, power law analysis (Clauset-Shalizi-Newman), formal causal inference (DiD, IV, DAGs), ergodicity economics\n- **Maven**: Bayesian decision theory + EVPI, mechanism design, mandatory pre-mortem (7+ failure paths), Kelly criterion, DMDU\n- **Quant**: Extreme Value Theory for tails, copula tail dependence, ruin probability, bootstrap CI, AIC/BIC model selection\n- **Prism-Adversarial**: 12-lens analysis, superforecasting (Tetlock), Dutch Book coherence audit, reference class forecasting, Fermi cross-checks\n\n---\n\n## Skill Self-Evolution — /evolve\n\nSQUAD learns from its own execution history and proposes evidence-backed skill improvements.\n\n```mermaid\nflowchart LR\n    E1[\"1. Evidence\u003cbr/\u003eRead tracking.jsonl\u003cbr/\u003eLast 100 records\"]\n    E2[\"2. Reflect\u003cbr/\u003eSuccess vs failure\u003cbr/\u003epatterns per skill\"]\n    E3[\"3. Quality Gate\u003cbr/\u003eSpecificity ≥ 3\u003cbr/\u003eActionability ≥ 3\u003cbr/\u003eGrounding ≥ 3\"]\n    E4[\"4. Bounded Update\u003cbr/\u003eTop 3 edits max\u003cbr/\u003eUser approves each\"]\n    E5[\"5. Branch Commit\u003cbr/\u003eevolve/YYYY-MM-DD\u003cbr/\u003eValidate → merge\u003cbr/\u003eor revert\"]\n\n    E1 --\u003e E2 --\u003e E3 --\u003e E4 --\u003e E5\n\n    style E3 fill:#e74c3c,color:#fff\n    style E5 fill:#2ecc71,color:#fff\n```\n\n**Safety constraints:**\n- **Max 3 edits per cycle** (gradient clipping)\n- Edits land on a **branch**, never main\n- **User gate at every edit** — never auto-applied\n- Both success AND failure records analyzed\n- `/health` shows skill utility grades (A–D) and flags evolution candidates\n\n---\n\n## Token Compression Engine\n\nNative JS compression pipeline — no external dependencies.\n\n```\nInput → Detect content type → Mask (protect errors/KG data) → Handler → Unmask → Output\n```\n\n| Content Type | Handler | Typical Ratio |\n|---|---|---|\n| Code | Strip comments, collapse imports | 40–60% |\n| Grep output | Group by file, deduplicate | 50–70% |\n| JSON | Minify, truncate arrays \u003e 10 items | 60–80% |\n| Logs / errors | Collapse repeated lines, summarize stacks | 50–70% |\n| File listings | Summarize by extension, collapse deep paths | 60–80% |\n\n**Protected (never compressed):** error messages, test assertions, KG graph data, user input.\n\n---\n\n# Part VI — Reference\n\n## Configuration Reference\n\n`squad-method/config.yaml` — auto-generated at install, filled by `/setup`:\n\n```yaml\ncompany:\n  name: \"\"\n  domain: \"\"                   # fintech | healthcare | saas | gaming | ...\n  compliance: []               # soc2 | hipaa | pci-dss | gdpr\n\nproject:\n  name: \"\"\n  type: \"\"                     # web-app | api | library | cli | mobile | infra | monorepo | game | ai-ml\n  maturity: \"\"                 # greenfield | brownfield | migration\n\nstack:\n  languages: []                # auto-detected\n  frameworks: []               # auto-detected\n  test_command: \"npm test\"\n\nmodel_routing:\n  default: \"default\"           # fast | default | heavy\n  mode: \"balanced\"             # balanced | quality | budget\n  agent_overrides: {}          # e.g. { raven: heavy, scribe: fast }\n  complexity_upgrade:\n    enabled: true\n    blast_radius_threshold: 20\n\ntoken_budget:\n  max_context_tokens: 50000\n  compression: none            # none | native\n\nknowledge_graph:\n  enabled: true\n  auto_rebuild: true\n  ast_enabled: false           # function-level analysis (opt-in)\n\nagents:\n  built_in: 56\n  custom: []\n  packs:\n    extended_core: [krishna, otis, trinity]\n    math_theory: [tao, knuth, ramanujan, hardy, pearl, gelman]\n    ai_ml: [andrej, yann, scott, woz, percy]\n    systems_data: [jeff, sanjay, stonebraker, reynold, kyle]\n    startup: [richard, monica, jared]\n    financial: [charts, ledger, herald, sage, maven, quant, prism-adversarial]\n\nides:\n  installed: []                # auto-detected: claude, devin, windsurf, cursor, codex, kiro, gemini, antigravity\n```\n\n---\n\n## Project Structure\n\n```\nworkspace/\n├── CONTEXT.md                 ← Root context (always loaded, ~300 tokens)\n├── CLAUDE.md / AGENTS.md      ← IDE-specific copies\n├── DEEP-CONTEXT.md            ← Architecture from KG analysis\n│\n├── squad-method/\n│   ├── config.yaml            ← Single source of truth\n│   ├── agents/                ← 56 agent personas (lazy-loaded per skill)\n│   │   ├── _base-agent.md     ← Base protocols\n│   │   ├── nova.md … phoenix.md    ← 14 core\n│   │   ├── trinity.md … krishna.md ← 3 extended core\n│   │   ├── tao.md … gelman.md     ← 6 math/theory\n│   │   ├── andrej.md … percy.md    ← 5 AI/ML\n│   │   ├── jeff.md … kyle.md      ← 5 systems/data\n│   │   ├── richard.md … jared.md   ← 3 startup\n│   │   ├── charts.md … prism-adversarial.md ← 7 financial\n│   │   └── shadow.md … titan.md    ← 13 specialized\n│   ├── skills/                ← 34 skill definitions\n│   ├── fragments/             ← Conditional knowledge modules\n│   │   ├── rubric/            ← Language-specific review rubrics\n│   │   ├── stack/             ← Framework knowledge\n│   │   ├── cloud/             ← Cloud provider guidance\n│   │   ├── tracker/           ← Sprint tracker integration\n│   │   └── agent-orchestrator.md ← 30 hard orchestration rules\n│   ├── tools/\n│   │   ├── knowledge-graph/   ← KG builder, query API, prioritize, AST pass\n│   │   ├── compress/          ← Token compression pipeline\n│   │   ├── router/            ← Multi-model routing engine\n│   │   └── dispatch/          ← Parallel execution adapters per IDE\n│   └── output/\n│       ├── tracking.jsonl     ← Operation log (feeds /health, /evolve)\n│       └── meta-skill.md      ← Optimizer memory across /evolve cycles\n│\n└── \u003crepo\u003e/knowledge-graph-out/\n    ├── graph.json             ← Full dependency graph\n    ├── graph.html             ← Interactive D3 visualization\n    └── KG_REPORT.md           ← Human-readable analysis\n```\n\n### Fragment Conditional Loading\n\nA Python/AWS/Jira project loads Python rubric + AWS fragments + Jira tracker. A JavaScript/no-cloud project loads a completely different set. Agents never see irrelevant knowledge.\n\n### MCP Tracker Integration\n\n| Tracker | MCP Server | Config |\n|---|---|---|\n| Jira | `@anthropic/mcp-jira` | `.{ide}/mcp.json` |\n| Linear | `@anthropic/mcp-linear` | `.{ide}/mcp.json` |\n| GitHub Issues | Built-in (Claude Code) | `.{ide}/mcp.json` |\n| Shortcut | `shortcut-mcp-server` | `.{ide}/mcp.json` |\n| Notion | `@modelcontextprotocol/server-notion` | `.{ide}/mcp.json` |\n\n---\n\n## Adding Support for a New Language Model\n\n### Step 1 — Add provider to registry\n\nEdit `squad-method/tools/router/providers.cjs`:\n\n```javascript\nvar MISTRAL = {\n  id: 'mistral',\n  models: { fast: 'mistral-small-latest', default: 'mistral-large-latest', heavy: 'mistral-large-latest' },\n  supports_effort: false,\n  max_context: 128000,\n};\n```\n\n### Step 2 — Add to IDE provider mapping\n\n```javascript\ncursor: {\n  primary: ANTHROPIC,\n  secondary: [OPENAI, GOOGLE, MISTRAL],  // ← add here\n}\n```\n\n### Step 3 — (Optional) Agent affinity rules\n\n```javascript\nvar AGENT_PROVIDER_AFFINITY = {\n  scribe: { prefer: 'mistral', tier: 'fast', reason: 'structured_output' },\n};\n```\n\n### Step 4 — Run tests\n\n```bash\ncd squad-public \u0026\u0026 node --test test/providers.test.js\n```\n\n---\n\n## Adding Support for a New IDE\n\n### Step 1 — Create a transformer\n\n`lib/transform/volta.js`:\n\n```javascript\nimport { deploySkillDir } from './base.js';\n\nexport function deploy(workspacePath, skill, options = {}) {\n  return deploySkillDir(workspacePath, skill, '.volta', options);\n}\nexport const IDE_ID = 'volta';\nexport const SKILLS_PATH = '.volta/skills';\n```\n\n### Step 2 — Register in the IDE skills generator\n\n`lib/generate/ide-skills.js`:\n\n```javascript\nconst TRANSFORMER_MAP = {\n  // ...existing...\n  volta: '../transform/volta.js',\n};\n```\n\n### Step 3 — Add IDE detection\n\n`lib/detect/ide.js`:\n\n```javascript\nIDE_CHECKS.push({ id: 'volta', name: 'Volta', configDir: '.volta', binary: 'volta' });\n```\n\n### Step 4 — Create a dispatch adapter\n\n`squad-method/tools/dispatch/adapter-volta.cjs`:\n\n```javascript\nvar BaseAdapter = require('./adapter-base.cjs');\nfunction VoltaAdapter(config) { BaseAdapter.call(this, 'volta', config); }\nVoltaAdapter.prototype = Object.create(BaseAdapter.prototype);\n// Implement: dispatchAgent, dispatchParallel, buildMultiModelPlan\n```\n\n### Step 5 — Add to provider mapping\n\n`providers.cjs`:\n\n```javascript\nvolta: {\n  primary: ANTHROPIC,\n  secondary: [OPENAI],\n  supports_parallel: false,\n  parallel_mechanism: 'sequential',\n  max_parallel: 1,\n}\n```\n\n### Step 6 — Update hooks and parity test\n\n```bash\n# Add detection in hooks.sh\n# Add to ide-parity-test.sh\nbash squad-method/tools/ide-parity-test.sh\n```\n\n---\n\n## Security \u0026 Privacy\n\n### Zero-Footprint Design\n\n- **Zero network calls** — SQUAD never phones home, no telemetry, no analytics\n- **Zero dependencies** — `package.json` has 0 runtime dependencies\n- **Local-only tracking** — `tracking.jsonl` stays on your machine\n- **No API keys stored** — environment variables only, never written to files\n- **Git exclude** — SQUAD artifacts use `.git/info/exclude` (never modifies `.gitignore`)\n\n### 5-Layer Safety Hooks\n\n`squad-method/tools/hooks.sh` runs at skill boundaries:\n\n| Layer | Hook | What It Checks |\n|---|---|---|\n| 1 | Skills Gate | SQUAD installed, config present, base-agent present |\n| 2 | Pre-Edit Guard | Blocks edits to auto-generated files (`dist/`, lock files, `*.generated.*`) |\n| 3 | Secret Detection | Scans for API keys, AWS keys, private keys before commits |\n| 4 | Progress Save | Forces progress doc update when context window fills (~40 messages) |\n| 5 | Gate Ledger | Verifies all phase gates passed before advancing |\n\nIn Claude Code, hooks fire automatically at the harness level (impossible to bypass). In all other IDEs, hooks fire when the skill calls `hooks.sh`.\n\n### Destructive Action Guard\n\nBefore any destructive action (delete, drop, force push), agents:\n1. State exactly what will be destroyed\n2. Ask for explicit confirmation\n3. Wait for approval before proceeding\n4. Never combine destructive actions\n\n---\n\n## Testing\n\n```bash\n# Unit tests\nnode --test test/*.test.js\n\n# Full suite (unit + e2e)\nnpm run test:all\n\n# IDE parity check\nbash squad-method/tools/ide-parity-test.sh\n```\n\nCurrent: **202 assertions, 0 failures** (unit + e2e + agent contracts).\n\nTest coverage includes:\n- Stack / cloud / IDE / tracker detection\n- IDE skill deployment (all 8 IDEs)\n- Knowledge graph: language patterns (15 languages), community detection, query API\n- AST extraction (JS/TS, Python, Go, Java)\n- Compression pipeline (all handlers, mask integrity, end-to-end)\n- Agent contracts: 56 agents validated (capabilities, determinism, frontmatter)\n- Provider routing, dispatch adapters, DAG wiring\n\n---\n\n## FAQ\n\n### How is SQUAD different from just using an AI IDE?\n\nAI IDEs give you one model in a chat. SQUAD adds:\n- **56 specialized agents** with distinct review lenses\n- **Pre-computed knowledge** via the knowledge graph — agents check dependency data before grepping\n- **Conditional fragment loading** — only project-relevant knowledge is loaded\n- **Phase-gated workflows** — complex tasks have user approval at each gate\n- **Cross-IDE portability** — same agents, skills, and config across 8 IDEs\n- **Self-evolution** — `/evolve` improves skills from execution history\n\n### Do I need all 8 IDEs?\n\nNo. SQUAD auto-detects installed IDEs and deploys skills only to those. When you run `init` again after updating the package, new skills are synced to all detected IDEs automatically.\n\n### What does \"zero dependencies\" mean?\n\n`package.json` has literally `\"dependencies\": {}`. No npm packages. No supply chain risk. Every line of code is in the repo. The tradeoff: regex-based YAML handling instead of a library, and regex-based import parsing in the KG (with AST as an opt-in via `--ast`).\n\n### How do agents find context without loading everything?\n\n1. **Always loaded:** `CONTEXT.md` + `context/index.md` (~500 tokens)\n2. **Per skill:** the skill declares which agents and fragments it needs\n3. **Per config:** fragments auto-load based on detected stack/cloud/tracker\n4. **Queried on demand:** `graph.json` is queried for specific files, never loaded in full\n\nTarget: \u003c 8,000 tokens for agent + fragment loading per skill invocation.\n\n### Why check the knowledge graph before grepping?\n\nThe KG pre-computes answers to the most common agent questions:\n- \"What depends on this file?\" → graph edges (instant, one read)\n- \"Is this high-risk?\" → god node flag + hotspot score (instant)\n- \"What tests cover this?\" → test edges (instant)\n- \"What's the blast radius?\" → 2-hop reachability query (instant)\n\nWithout the KG: 3–10 grep commands across the entire codebase. With the KG: one JSON read. Saves ~80% exploration tokens in typical workflows.\n\n### What are god nodes?\n\nAny file with more than 30 dependency connections (imports + importers). When an agent detects it's modifying a god node:\n- Model router **auto-upgrades** to heavy tier\n- Review requires **extra approval**\n- KG report flags the full blast radius\n\n### What's in `tracking.jsonl`?\n\nEvery skill run appends one JSON line with: skill name, agents dispatched, phases completed, review findings (critical/major/minor), outcome, assumptions count. This feeds:\n- `/health` — agent effectiveness analysis, skill utility grades (A–D)\n- `/evolve` — evidence-backed skill improvement proposals\n- Quality gate (V2) — re-dispatches on low-quality output\n- Learned classifier (V3 stub) — will predict optimal model tier\n\n### Can I add custom agents?\n\nYes. Create a `.md` file in `squad-method/agents/` following the frontmatter format of existing agents. Add the agent name to `config.yaml → agents.custom`. The agent is then available to any skill that declares it.\n\n### What are the financial agents useful for?\n\nThe seven financial agents (`/financial-analysis`, `/market-research`, `/consulting-brief`) apply quant-fund grade methods — Beneish M-Score, Benford's Law, Granger causality, Kelly criterion, EVT tail risk, Dutch Book coherence — to produce analysis with explicit confidence intervals and falsifiable claims. Every conclusion includes a verification summary (VERIFIED-4 through UNVERIFIED) and a mandatory disclaimer.\n\n### How does `/evolve` work safely?\n\nEdits proposed by `/evolve` always land on a branch (`evolve/YYYY-MM-DD`), never main. The quality rubric requires specificity ≥ 3, actionability ≥ 3, and grounding ≥ 3 — vague rules fail and are rejected. Maximum 3 edits per cycle. User must explicitly accept each edit. After N runs on the branch, if outcomes improve, you merge; if not, you revert.\n\n---\n\n## Contributing\n\n### Quick contributions\n- **Bug reports** — open an issue with steps to reproduce\n- **Feature suggestions** — open an issue with the use case\n- **Typos / docs** — PR directly\n\n### Code contributions\n\n1. Fork and clone the repo\n2. Create a branch: `git checkout -b feature/my-feature`\n3. Follow existing patterns (look at similar files first)\n4. Add tests — every new feature needs tests\n5. Run: `npm run test:all`\n6. Run parity check: `bash squad-method/tools/ide-parity-test.sh`\n7. Submit PR with what and why\n\n### What we're looking for\n\n- New IDE adapters — see [Adding Support for a New IDE](#adding-support-for-a-new-ide)\n- New model providers — see [Adding Support for a New Language Model](#adding-support-for-a-new-language-model)\n- New language detection for KG (add to `LANGUAGE_PATTERNS` in `build.js`)\n- Stack / cloud / tracker detection fragments\n- Rubric modules for additional frameworks\n- Bug fixes and test improvements\n\n### Code style\n\n- Use existing patterns — look at similar files before writing new ones\n- Zero dependencies — don't add npm packages\n- Tests required — if it's testable, test it\n- ESM for `lib/` — CommonJS (`.cjs`) for `squad-method/tools/`\n\n---\n\n## Credits \u0026 Acknowledgments\n\n### Direct Inspirations\n\n| Source | What We Took | How SQUAD Uses It |\n|---|---|---|\n| **[Graphify (Karpathy)](https://github.com/karpathy)** | Knowledge graph extraction approach | KG builder: AST/import extraction, community detection, god nodes, git co-change coupling |\n| **[Headroom (chopratejas)](https://github.com/chopratejas/headroom)** | Tool output compression pipeline | Inspired `squad-method/tools/compress/`: content-type detection → domain handlers → universal compression |\n| **[SkillLens (Microsoft)](https://microsoft.github.io/SkillLens/)** | Skill quality rubric + utility scoring | `/evolve` quality gate: specificity × actionability × grounding. `/health` utility grades (A–D) |\n| **[SkillOpt (Microsoft)](https://microsoft.github.io/SkillOpt/)** | Rollout → Reflect → Bounded Update | `/evolve` evolution loop: max 3 edits per cycle, slow-update on branch, meta-skill memory |\n| **[RouteLLM (2024)](https://arxiv.org/abs/2406.18665)** | Learned model routing | 3-tier routing: rule-based → quality gate → classifier stub trained from `tracking.jsonl` |\n| **[HyperAgentMeta (Meta 2026)](https://arxiv.org/abs/2602.00000)** | Self-improving agent loops | `/evolve` structure: tracking data → failure analysis → surgical skill diffs → human approval |\n\n### Core Concepts\n\n- **Multi-Agent Systems** — Multi-agent debate (Du et al., 2023), mixture-of-agents (Wang et al., 2024)\n- **Agentic Coding** — Patterns from Claude Code, Devin, SWE-Agent, OpenHands\n- **Knowledge Graphs for Code** — Graph-based dependency analysis inspired by Sourcegraph, CodeQL\n- **TDD \u0026 Agile** — Review rubrics grounded in Martin, Fowler, Beck, Nygard, OWASP\n\n### Financial Analysis Methodology\n\nAcademic citations for quantitative methods used by financial agents:\n\n- Beneish (1999) — M-Score for earnings manipulation detection\n- Sloan (1996) — Accrual anomaly: earnings persistence vs cash\n- Lev \u0026 Thiagarajan (1993) — 12 fundamental signals predicting future returns\n- Altman (1968, 2020) — Z-Score for bankruptcy prediction\n- Benford (1938), Nigrini (2012) — Benford's Law for fraud detection\n- Bass (1969) — Diffusion model for technology adoption\n- Peters (2019) — Ergodicity economics\n- Tetlock (2015) — Superforecasting and calibrated probability\n- Pearl (2009) — Causal inference with DAGs\n- Embrechts et al. (1997) — Extreme value theory for tail risk\n- Kelly (1956) — Capital allocation criterion\n\n### Model Providers\n\n- **[Anthropic](https://anthropic.com)** — Claude Opus 4 \u0026 Sonnet 4\n- **[OpenAI](https://openai.com)** — GPT-4o and o3\n- **[Google DeepMind](https://deepmind.google)** — Gemini 2.5 Pro (1M context)\n- **[Amazon AWS](https://aws.amazon.com/bedrock/)** — Bedrock, Titan, Amazon Q\n- **[Meta AI](https://ai.meta.com)** — Llama models via Bedrock\n\n### IDE Platforms\n\n- **[Anthropic Claude Code](https://docs.anthropic.com/en/docs/claude-code)** — Native Agent() API for true parallel execution\n- **[OpenAI Codex CLI](https://github.com/openai/codex)** — CLI subprocess dispatch\n- **[AWS Kiro](https://kiro.dev)** — Bedrock multi-provider gateway\n- **[Google Gemini CLI](https://github.com/google-gemini/gemini-cli)** — Vertex AI integration\n- **[Cursor](https://cursor.com)** — Multi-model IDE, `.mdc` rule format\n- **[Windsurf](https://codeium.com/windsurf)** — Cascade AI with skill/workflow system\n- **[Antigravity](https://antigravity.dev)** — AI-native development environment\n\n---\n\n## License\n\nMIT — see [LICENSE](LICENSE) for details.\n\n---\n\n\u003cdiv align=\"center\"\u003e\n\n**Built for developer experience, not vendor lock-in.**\n\n[npm](https://www.npmjs.com/package/squad-public) · [Issues](https://github.com/adityashubham1997/squad-public/issues) · [Contribute](#contributing)\n\n\u003c/div\u003e\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fadityashubham1997%2Fsquad-public","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fadityashubham1997%2Fsquad-public","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fadityashubham1997%2Fsquad-public/lists"}