Cognia

A collection of specialised AI agents and skills for comprehensive project analysis — covering architecture, backend, frontend, iOS, Android, UX, technical quality, product ownership, performance, security, reverse engineering, and test engineering.

Agents

Agents auto-detect the platform(s) present in a project (backend, frontend, iOS, Android, or mixed) and apply only the relevant analysis playbook(s). Every agent writes its findings to a mandatory output file in cognia/.

| Agent | Domain | Description | |-------|--------|-------------| | cognia-android | Android mobile | Full Android codebase audit: screens, components, networking, navigation, state management | | cognia-arch | System architecture | Component map, data flow, service boundaries, scalability, and architectural risk assessment | | cognia-backend | Backend / API | Endpoint inventory, service catalogue, integrations, database schema, auth, and background jobs | | cognia-frontend | Frontend / Web | Page and route inventory, component catalogue, state management, API integration, and build config | | cognia-ios | iOS mobile | Full iOS codebase audit: screens, components, networking, navigation, state management | | cognia-perf | Performance analysis | Auto-detects platform(s) and audits for bottlenecks: slow queries, large bundles, blocking threads, memory issues — with a prioritised improvement roadmap | | cognia-po | Product ownership | Feature inventory, user stories, requirements gaps, business value mapping, and backlog recommendations | | cognia-review | Code review (changeset) | Deep PR / branch / changeset review across structural, semantic, architectural, regression, performance, security, test-alignment and communication-quality dimensions — with file:line findings, suggested patches, and a Change Risk Score | | cognia-reverse | Reverse engineering | Extracts business domain, user roles, workflows, business rules, and integrations from code — output written for business analysts and product owners | | cognia-sec | Security analysis | Auto-detects platform(s) and audits for vulnerabilities: injection, broken auth, insecure storage, exposed secrets, dependency CVEs — with a CVSS-aligned remediation roadmap | | cognia-tech | Technical quality | Code quality, tech debt, dependency audit, security signals, and test coverage gaps | | cognia-test | Test engineering | Audits existing test coverage for correctness and quality, identifies missing unit/integration/e2e tests, and produces a prioritised test backlog with acceptance criteria | | cognia-ux | UI/UX audit | Page inventory, user flow mapping, design consistency, accessibility audit (WCAG), and UX improvement recommendations | | cognia-ux-design | UI/UX design | Produces wireframes, information architecture, design system, design tokens, component API contracts, and developer handoff artifacts for new or redesigned UIs |

Skills

All 13 agents ship with runtime wrappers for Claude Code (.claude/skills/), Codex CLI (.codex/skills/), and Cursor (.cursor/rules/). Three agents additionally include deep procedure documents and standards under .github/skills/:

| Skill folder | Used by | Contents | |---|---|---| | .github/skills/cognia-arch | cognia-arch | Mermaid diagram procedure, HTML template, STANDARDS.md | | .github/skills/cognia-tech | cognia-tech | Multi-step analysis procedure, scale decomposition, output format | | .github/skills/cognia-ux | cognia-ux-design | Wireframe procedure, design system, WCAG compliance, handoff artifacts |

Installation

Run the interactive installer via npx — no local install required:

npx cognia

The installer will ask:

• Scope: Global (recommended — available in all projects) or Local (current project only)
• Runtime: All (Claude Code + Codex CLI + Cursor), Claude Code only, Codex CLI only, or Cursor only

Non-interactive flags

npx cognia --global           # global install, all runtimes
npx cognia --local            # local install (current project only)
npx cognia --global --claude  # global, Claude Code only
npx cognia --global --codex   # global, Codex CLI only
npx cognia --global --cursor  # global, Cursor only
npx cognia --uninstall        # remove all installed files

Script shortcuts (if installed locally)

npm run install:global   # global install, all runtimes
npm run install:local    # local install
npm run uninstall        # remove

What gets installed

| Location | Contents | |----------|----------| | ~/.copilot/agents/ or .github/agents/ | 13 .agent.md canonical agent definitions | | ~/.copilot/skills/ or .github/skills/ | Deep skill procedures for cognia-arch, cognia-tech, and cognia-ux-design | | ~/.copilot/standards/ or .github/standards/ | Shared standards: core.md (evidence rules, output contracts, handoff format) and preflight.md (setup and context policy) | | ~/.claude/agents/ | Claude Code agent wrappers (13 agents) | | ~/.claude/skills/ | Claude Code skill wrappers (13 skills) | | ~/.codex/skills/ | Codex CLI skill wrappers (13 skills) | | ~/.cursor/rules/ | Cursor rule wrappers (13 .mdc files) |

For a local install, all paths above are relative to the project root (.claude/, .codex/, .cursor/, .github/) instead of ~/.

Usage

GitHub Copilot

All agent definitions live in .github/agents/*.agent.md. Reference them in VS Code's agent mode or any Copilot-compatible tool:

@cognia-arch analyse the architecture of this service
@cognia-backend map all API endpoints

Claude Code

Invoke agents with /agent-name or @agent-name:

/cognia-arch
@cognia-tech
/cognia-sec

Cursor

After installing, Cursor's AI picks up the relevant agent rule automatically based on what you ask. You can also invoke agents explicitly:

Use cognia-arch to analyse the architecture of this project.
Run cognia-backend on the codebase and write the analysis.

Each agent rule is stored in .cursor/rules/ (local) or ~/.cursor/rules/ (global) as a .mdc file. Cursor loads the rule into context when the request matches the rule's description.

Codex CLI

$cognia-arch
$cognia-backend

Cross-Agent Rules

1. Evidence first — every finding cites at least one concrete file path.
2. Tag confidence — claims are marked Confirmed (directly evidenced) or Inferred (best-fit interpretation).
3. Missing evidence — states Not found in scanned files rather than guessing.
4. Output files are mandatory — analysis is not complete until the designated output file is written.
5. No domain creep — each agent respects its scope; cross-domain observations belong in a handoff note.

License

MIT