Fullstack Forge

A production engineering skill suite for AI coding agents.

One audit system. Forty-two specialist skills. Evidence before confidence.

Fullstack Forge gives AI coding agents a repeatable way to build features and to audit, fix, verify, and report on real full-stack applications. It discovers the actual stack, selects only applicable modules, gathers reproducible evidence, separates safe fixes from risky decisions, and refuses to call missing evidence a pass.

It works as an open-format Agent Skill collection and as a dependency-light TypeScript CLI.

Two modes

Build mode — use while starting a project or implementing a feature, so the agent follows a production-quality engineering workflow. Audit mode — use to inspect, harden, and gate work that already exists, including work built in Build mode.

A coherent project journey: forge new frames the product once; each feature then runs forge feature <slug> through frame → plan → implement → check → done; before release, the independent backstop runs regardless of how the code was built — forge all audit (or --scope changed for a diff) → forge all fix --safe → forge all verify → forge ship. Build mode's frame and plan are recorded guidance; check and done are CLI-derived, never agent-asserted; and build evidence under .forge/build/ satisfies zero forge ship or forge all audit gates — both always re-derive their own evidence independently. See docs/BUILD_MODE.md for the complete guide.

npm install --save-dev github:thethunderbolt/fullstack-forge-skill#v0.2.0 && npx forge init all

Codex, Claude Code, Antigravity, Gemini CLI, Cursor, Windsurf, GitHub Copilot, and generic Agent Skills are supported. Commands · Platforms · Contributing · Security · Releases

Why this exists

“Check best practices” is not an audit. Production readiness crosses product logic, interface states, identity, authorization, data integrity, hostile inputs, failure recovery, deployment, operations, and specialized features such as payments or AI tools. Fullstack Forge turns those concerns into concrete procedures with stable findings and an honest completion contract.

flowchart LR
  D["Discover"] --> A["Select applicable audits"]
  A --> I["Inspect"]
  I --> R["Report"]
  R --> F["Fix"]
  F --> V["Verify"]
  V --> S["Ship"]

Install

From a release archive

Download the archive for your agent from the latest release, verify it against SHA256SUMS.txt, and extract it at the project root. Archives contain real copies, never symlinks.

With npm, from the Git tag

Fullstack Forge is not published to the npm registry. The working npm-based installation resolves the package directly from its Git tag:

npm install --save-dev github:thethunderbolt/fullstack-forge-skill#v0.2.0
npx forge init all --dry-run
npx forge init all

After npm registry publication

Registry publication has not happened yet. Once it does — and not before — the command below will work. It does not work today, and this section is retained only to document the intended future form:

# NOT YET AVAILABLE — the package is not on the npm registry.
npm install --save-dev fullstack-forge-skill

The installer writes .fullstack-forge/install-manifest.json. It will not overwrite unowned or modified files, follows no destination symlinks, and uninstalls only unchanged files it owns.

Quick start

Build mode, starting a project or a feature:

/forge-new
/forge-feature auth-login

forge new
forge feature auth-login   # identity triggers escalate this feature to high tier, recorded
forge feature auth-login plan
forge feature auth-login check --allow-run
forge feature auth-login done

Audit mode, inspecting and gating what already exists:

$fullstack-forge audit this application before release
$forge-security audit the changed authentication flow
/forge-ui audit the running application at mobile and desktop widths

forge discover audit
forge ui audit
forge ux audit
forge security audit --json
forge uploads audit
forge queries audit
forge all audit --scope changed --base origin/main
forge all audit --scope full
forge all fix --safe
forge ship --allow-run

Discovery creates ignored local artifacts at .forge/project-profile.json and .forge/architecture-map.md. Audits generate .forge/report.json and .forge/report.md. Build mode creates .forge/build/ (project.json, features/<slug>.json, DECISIONS.md, DESIGN.md) — agent context only, never a ship-gate input. See docs/BUILD_MODE.md.

The 42 skills

Every command skill is self-contained and supports audit, fix, verify, and report. Each one defines when it applies, inputs, an executable and manual procedure, evidence rules, stable IDs, severity, safe/risky fixes, verification, standards, stack guidance, limitations, and the same completion contract. Together they enumerate 957 explicit inspection criteria and 212 discipline-specific inspection steps — forge-database reads schema, types, cascades, and migration history, while forge-realtime traces connect authorization, channel isolation, reconnection, and backpressure — so specialized risks remain visible instead of disappearing behind a generic “best practices” instruction.

The audit module set stays closed at 42. Build mode ships two additional command skills, forge-new and forge-feature, alongside them — see Two modes above.

Finding contract

Every finding includes:

id · instance_id · section · title · severity · confidence · status · location · evidence
impact · recommendation · safe_fix · verification · standards · analyzer_id
trace · evidence_snapshot · verification_plan · fix_attempts

Statuses are PASS, FAIL, WARNING, NOT_APPLICABLE, NOT_VERIFIED, and BLOCKED. Severities are CRITICAL, HIGH, MEDIUM, LOW, and INFO. A pass requires code/line evidence, a successful check, running-app inspection, a behavior-demonstrating test, or verified configuration output. Silence is not a pass.

PASS            direct evidence satisfied the stated check
FAIL            reproducible evidence shows a defect
WARNING         risk exists without a proven defect
NOT_APPLICABLE  discovery shows the module is outside scope
NOT_VERIFIED    required behavior or environment evidence is missing
BLOCKED         approval, access, or a required tool is unavailable

The CLI has a typed safe-fix registry. It can replace actual-looking values in environment templates with explicit placeholders, add noopener noreferrer to proven JSX target="_blank" links, and add X-Content-Type-Options: nosniff to an existing global Vercel header rule. Every write is bound to a confirmed finding, exact post-audit hash, structural parser, repository-contained path, and finding-specific verification. Identity, tenant, upload policy, data, migration, secret rotation, financial, legal, architecture, production, and destructive changes remain approval-bound.

CLI

forge <section> <audit|fix|verify|report> [options]
forge all audit --scope changed [--base <ref>]
forge new [--tier <light|standard|high>] [--stack <name>] [--non-goal <item:reason>]
forge feature <slug> [frame|plan|check|done|accept-risk|abandon|status]
forge resume
forge init <platform|all> [--global] [--dry-run]
forge update [platform] [--dry-run]
forge uninstall [platform] [--dry-run]
forge doctor | validate | package | list
forge tool <name>

forge new and forge feature are Build mode; every other verb above is Audit mode. See docs/BUILD_MODE.md for the full build flag surface.

The CLI includes bounded TypeScript-compiler and structured-config analyzers for supported JavaScript/TypeScript security, auth, authorization, tenancy, upload, query, cache, accessibility, AI, payment, and integration shapes. Keyword scanners remain secondary discovery signals and never establish a pass. Unsupported languages or frameworks are reported as NOT_VERIFIED with the missing adapter named. Project commands execute only after their local definitions are shown and --allow-run is supplied.

When the audited project has Playwright installed, forge tool inspect-rendered-ui <url> captures desktop, tablet, and mobile screenshots plus browser console output into .forge/evidence/ui/ as direct running-application evidence for UI, UX, and accessibility audits. Without Playwright or a reachable URL the tool reports BLOCKED and rendered-state criteria stay NOT_VERIFIED — visual evidence is captured, never fabricated.

Audit reports also include typed, revision-bound ship-gate evidence and structured analyzer coverage for each detected language/framework. See commands, analyzer support, and coverage policy.

Platform support

These paths were verified against current primary platform documentation on 2026-07-18. Some platforms also scan .agents/skills/. See platform support for global paths, aliases, caveats, and primary sources.

Canonical and generated architecture

src/fullstack-forge/ is the canonical source. npm run generate renders command skills from the ordered module catalog and synchronizes six platform roots with per-file SHA-256 ownership manifests. Synchronization refuses modified or unowned managed paths. CI fails if a generated copy drifts.

See architecture, development, release process, the traceability matrix that maps every requirement to evidence, and the v0.1.0 historical verification record. The v0.1.1 verification record separates local, CI, publication, asset-download, and clean-room evidence.

Safety and limitations

Fullstack Forge is an engineering audit aid, not a compliance certificate, penetration test, legal opinion, accessibility conformance claim, financial audit, or substitute for production access. Static analyzers are bounded, can produce false positives, and do not imply complete coverage of an unsupported stack. Runtime, provider, database, browser, assistive-technology, and operator checks stay NOT_VERIFIED until actually performed.

Build mode cannot force analysis quality during frame or plan — the CLI records what an agent decides, it does not grade it — and most discipline criteria resolve NOT_VERIFIED or NOT_APPLICABLE without runtime evidence. Build state is agent context only; the independent backstop for correctness remains forge all audit and forge ship.

Read the security model and report vulnerabilities through SECURITY.md.

Research and attribution

The implementation adapts concepts—not third-party code or substantial prose—from public standards, official platform documentation, and open-source skill repositories. Exact revisions, access dates, licenses, and handling decisions are in research sources, the license matrix, and third-party notices.

Contributing and license

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. Fullstack Forge is licensed under Apache-2.0.