argo

Opinionated, security-first AI workflow for Salesforce DX projects in Claude Code.

A Claude Code plugin that turns Claude into a Salesforce DX team. Eleven specialist subagents — solution architect, data architect, integration architect, Apex/LWC/React/agent builders, QA, E2E tester, security and trust reviewers — coordinate through structured handoffs. Fifty-six slash-skills handle setup, deploys, code review, agent dev, and DevOps. Every org-touching operation is gated by a hardened security model: production orgs are hard-blocked, customer-data queries require explicit per-call consent, and anonymous Apex is refused by default.

Built for Experience Cloud, Lightning Experience, Communities, mobile, and Slack delivery surfaces — and aware of everything Salesforce announced at TDX 2026 (Headless 360, Agentforce, the new React framework, the Einstein Trust Layer + AI Gateway controls, AgentExchange).

---

Quickstart

Five steps from zero to a working setup, in your Salesforce DX project.

1. Install Claude Code (if you haven't)

# macOS / Linux
curl -fsSL https://claude.ai/install.sh | sh

# Windows: install from the Anthropic site or via npm
npm install -g @anthropic-ai/claude-code

2. Add the plugin

In Claude Code:

/plugin marketplace add https://github.com/vhmarquez/argo
/plugin install argo@argo

Restart your session so the agents, skills, and hooks register.

3. Bootstrap your project

cd into your SFDX project, then:

/argo:sf-init

Detection runs automatically (sfdx-project.json, package.json, sf org list, etc.) and presents one screen with everything pre-filled. You typically only fill in:

• A one-sentence project description
• Confirming the target org alias if multiple are detected
• Classifying any non-sandbox orgs as production (hard-blocked) or known-non-prod (allowed)

A smoke test runs after the write to verify the org alias resolves, paths exist, and lint commands work.

4. (Optional but recommended) Configure MCP

/argo:mcp-setup --profile dev

Installs @salesforce/mcp, scopes the toolsets to a sensible dev set (metadata, data, testing, lwc, code-analysis), and writes .mcp.json. The plugin runs without it via direct sf CLI calls; MCP unlocks richer routing for Headless-360 features.

5. Try the workflow

@architect: design a customer-greeting agent for our portal

The architect produces a structured plan (data model, automation type, governor budget, risk, effort, test strategy) and hands off to specialists — @agent-dev for the agent, @apex-dev for any Apex, @react-dev if the portal needs React, @qa to write tests, @trust-reviewer to check the agent surface. All in one conversation; specialists run in parallel where possible.

Before merging:

/argo:code-review pr
/argo:security-scan
/argo:diff-deploy --validate
/argo:pr-prepare --push

That's the loop.

---

Why this plugin

• Salesforce DX-aware out of the box. /sf-init reads sfdx-project.json, scans existing components for the LWC prefix, sniffs package.json for lint/test scripts, classifies orgs from sf org list — most of the config writes itself. You typically answer one or two questions.
• Security is the default, not an afterthought. Production orgs are unreachable; data queries require per-call consent; anonymous Apex is refused; every override is logged. Read docs/security-model.md before loosening anything.
• Specialists, not a generalist. Eleven subagents with distinct responsibilities and clean handoffs. @architect plans, builders implement in parallel, reviewers gate the merge. No one agent has all the context — that's the point.
• Headless 360 native. First-class MCP routing via @salesforce/mcp, Agentforce dev workflow (/agent-spec → /agent-test → /agent-deploy), Trust Layer audits, AgentExchange listing prep, React-on-Salesforce parity with LWC.
• CI-ready. Most skills accept --ci, --format json|sarif, --out, --fail-on flags following a documented contract. SARIF output integrates natively with GitHub Code Scanning.

---

Table of contents

• Quickstart — five steps to a working setup
• Why this plugin
• What you get — agents, skills, hooks, packs, standards
• Install — full prerequisites
• Project config
• Security model
• CI integration
• Plugin layout
• Compatibility
• License
• Changelog

---

What you get

Subagents (agents/) — 11 specialists

Agent	Model	Role
@architect	opus	Read-only solution design + implementation plan (automation-type rec, governor budget, risk/blast-radius, effort, test strategy)
@data-architect	opus	Object model, master-detail vs lookup, sharing, LDV, migrations
@integration-architect	opus	Callouts, Named Credentials, Platform Events, CDC, External Services, MCP Bridge, Trusted Agent Identity, Agent Fabric
@apex-dev	sonnet	Apex implementation (classes, triggers, batch, queueable, REST, callouts, custom metadata)
@lwc-dev	sonnet	Lightning Web Component implementation (i18n, virtualization, lazy-load, LWS)
@react-dev	sonnet	React-on-Salesforce implementation (@salesforce/react/graphql, SLDS via tokens, LWC interop)
@agent-dev	sonnet	Agentforce agent authoring (topics, sub-agents, actions, prompts, eval suites)
@qa	sonnet	Apex + Jest tests; runs lint/tests/coverage; severity-graded code review
@e2e-tester	sonnet	UTAM (Lightning) / Playwright (Experience Cloud) end-to-end tests against scratch orgs
@security-reviewer	opus	OWASP-for-SF: SOQL injection, IDOR, sharing/CRUD/FLS edge cases
@trust-reviewer	opus	OWASP-for-LLM on agents: prompt injection, output validation, grounding leakage, jailbreak resistance

Typical flow: @architect plans → hands off to specialists → builders work in parallel → @qa reviews → @e2e-tester covers journeys → @security-reviewer + @trust-reviewer before prod.

Skills (skills/) — 56 total

Invoke any as /argo:<name>. Most accept --ci, --format json|sarif, --out, --env <name> per the CI output contract. Each skill declares a data-access field in its frontmatter (none / metadata-only / data-with-consent); see the security model.

Setup & onboarding

Skill	Purpose
/argo:sf-init	Detect → review → edit → verify bootstrap. Aggressive auto-detection populates a single review screen with confidence markers; the user edits only what's ambiguous or required; a smoke test verifies the result. Modes: auto, update <fields>, env <name>, verify
/argo:onboard	Verify a developer's machine + smoke-test the dev loop end-to-end
/argo:pattern-pack	Install/list/info/remove domain pattern packs
/argo:mcp-setup	Install/configure @salesforce/mcp toolsets (metadata/data/testing/lwc/code-analysis/devops/aura)

Org awareness

Skill	Purpose
/argo:org-explore	Optional org-schema snapshot (in MCP mode, agents read live; cache is --cache opt-in)
/argo:org-diff	Source-vs-org drift report (setup-only / source-only / conflicts)
/argo:flow-audit	Active-flow inventory; flags Apex/Flow overlap and untracked-in-source flows
/argo:permset-audit	Object/field × principal access matrix; flags fields with no read access
/argo:field-impact	Field references across LWC, Apex, layouts, validation rules, formulas, flows, reports
/argo:agent-discover	Agentforce agent inventory — source vs org reconciliation; bridge tools per agent

Architecture & design

Skill	Purpose
/argo:erd	Mermaid ERD from .object-meta.xml (idempotent, depth-bounded)
/argo:sequence-diagram	Mermaid sequence from an LWC entry point through Apex / DB / external
/argo:adr	Architecture Decision Records under docs/adr/
/argo:flow-vs-apex	Flow vs Trigger vs Queueable vs Batch decision helper
/argo:agent-vs-flow-vs-apex	Extends flow-vs-apex with Agent as first-class option
/argo:lwc-vs-react	Frontend framework decision (LWC vs React vs both)
/argo:mcp-tool-vs-rest	Integration pattern decision (MCP Tool vs Apex REST vs Platform Event)
/argo:before-vs-after-trigger	Trigger phase decision
/argo:queueable-vs-batch	Async mechanism decision

Agent dev (Headless 360)

Skill	Purpose
/argo:agent-spec	Wrap sf agent generate agent-spec with project context; iterative refinement
/argo:agent-test	Run agent eval suite via Testing Center; per-axis severity (factuality, completeness, refusal-correctness, etc.)
/argo:agent-eval-trend	Per-agent eval history; PR-mode regression diff; security regressions zero-tolerance
/argo:agent-deploy	Deploy AgentDefinition + register evals; gates by trust-layer-audit + agent-test + eval-regression
/argo:mcp-bridge	Wrap an Apex REST class as an MCP tool — closes SF-16 → agent ecosystem loop
/argo:slack-agent	Scaffold a Slack-native agent end-to-end via Slack Agent Kit
/argo:agent-exchange-list	Validate readiness for AgentExchange listing

React (Headless 360)

Skill	Purpose
/argo:react-init	Scaffold a React component bundle with @salesforce/react/graphql + i18n + SLDS

Testing

Skill	Purpose
/argo:test-plan	Generate a structured test plan (positive/negative/bulk/edge/security) before writing tests
/argo:test-data	Scaffold an Apex TestDataFactory from sObject describes
/argo:test-coverage	Apex coverage or agent eval (modes: apex / agent)
/argo:coverage-trend	Coverage history; PR-mode regression gate
/argo:flaky-test-finder	Re-run a test class N times to identify non-deterministic methods

Code review & static analysis

Skill	Purpose
/argo:code-review	Per-component, batch (all/audit), and PR-mode review
/argo:security-scan	PMD apex-security ruleset; SARIF for GitHub Code Scanning
/argo:fls-audit	Static check for missing CRUD/FLS on DML / SOQL
/argo:sharing-review	without sharing audit; flag privilege escalation from @AuraEnabled
/argo:soql-analyzer	Selectivity check (indexed fields, LDV awareness, leading-wildcard)
/argo:limit-usage	Per-method governor-budget estimator
/argo:perf-review	LWC/React bundle size, @wire waterfall, render-blocking, missing virtualization
/argo:dead-code	Unused Apex methods/fields, LWC/React bundles, custom labels, custom permissions
/argo:complexity	Cyclomatic + cognitive complexity per method
/argo:dependency-graph	Apex call graph + LWC/React import graph

Trust & governance (Headless 360)

Skill	Purpose
/argo:trust-layer-audit	Einstein Trust Layer config audit (org-level + per-agent: PII masking, FLS-on-grounding, ZDR, jailbreak eval, etc.)
/argo:trust-eval	Runtime drift audit via Testing Center Custom Scoring Evals + Session Tracing sampling
/argo:gateway-config	Generate/validate AI Gateway config (model allowlist, token quotas, fallback chain) per environment

Deployment

Skill	Purpose
/argo:deploy	Deploy or --validate configured paths against the default org
/argo:diff-deploy	Deploy only metadata changed since <ref>
/argo:quick-deploy	Promote a validated deploy id to production without re-running tests
/argo:scratch-org	Create/destroy/recreate scratch orgs; seed Apex + agent data
/argo:package-version	2GP/unlocked package version create/promote/list/install
/argo:destructive-changes	Interactive destructiveChanges.xml builder with reference validation
/argo:devops-natural	Natural-language deploy via DevOps Center MCP (Headless 360)

Documentation & release

Skill	Purpose
/argo:generate-docs	Generate or update LWC, Apex, React, and agent docs (audit mode flags stale/orphaned/missing)
/argo:release-notes	Conventional-commits release notes + ADRs + coverage stats
/argo:pr-prepare	Assemble a PR body from review/coverage/deploy/security gates
/argo:notify	Slack/Teams webhook poster for deploy/coverage/security/release events

Hooks (hooks/)

Hook	Trigger	Action
session-start.sh	SessionStart	One-line first-run nudge when an SFDX project has no .claude/sf-project.json yet. Silent in all other cases
security-guard.sh	PreToolUse on Bash	Defense-in-depth security gate: refuses sf commands that target prod aliases or query data objects without consent. Belt-and-suspenders companion to hooks/lib/security.sh; see docs/security-model.md
lint-lwc.sh	PostToolUse on Edit/Write of LWC JS	Prettier + ESLint; surfaces findings on stderr; never blocks
lint-apex.sh	PostToolUse on Edit/Write of .cls/.trigger	PMD errorprone + bestpractices
lint-react.sh	PostToolUse on Edit/Write of .tsx/.jsx (when platform.frontend includes react)	Prettier + ESLint

Plus shared library helpers under hooks/lib/:

• config.sh — load and deep-merge project config + per-env overrides
• sf-cli.sh — wrappers for sf CLI; routes through security.sh; the always-works fallback
• mcp.sh — @salesforce/mcp routing helpers (mcp_prefer, mcp_run <toolset> <tool>, mcp_list_tools); routes through security.sh
• security.sh — central security gate: sec_check_org, sec_check_soql, sec_check_anon_apex, sec_log_consent, metadata allowlist. See docs/security-model.md
• pmd.sh — lazy PMD download into ${CLAUDE_PLUGIN_DATA} on first use
• sarif.sh — SARIF 2.1.0 emitter for --format sarif

Domain pattern packs (templates/packs/) — 11 packs

Install via /argo:pattern-pack add <name>. Format documented in docs/pack-format.md.

Pack	Status	Patterns
agentforce	✅ v1.0	AGT-1..7 — Agent topic boundaries, sub-agent decomposition, guardrails, MCP-tool actions, FLS-aware grounding, memory & state, escalation paths
react	✅ v1.0	RX-1..6 — Platform GraphQL fetch, platform-aware auth, deployment, SLDS via tokens, i18n, LWC↔React interop
platform-events	✅ v1.0	PE-1..5
change-data-capture	✅ v1.0	CDC-1..5 — Source-controlled CDC selection, trigger subscriber + idempotency, Pub/Sub API, GAP_OVERFLOW reconciliation, CDC vs. PE decision
external-objects	✅ v1.0	EXT-1..5 — Adapter selection, __x schema + Indirect/External Lookup, query/cache, write-back semantics, Custom Apex Connector
big-objects	✅ v1.0	BIG-1..5 — Index design, Database.insertImmediate, Async SOQL aggregation, index-aligned predicates, lifecycle and capacity
field-service	✅ v1.0	FS-1..5 — Work Order lifecycle, Service Appointment scheduling, Resource Absence + Operating Hours, mobile offline, territory design
industries	✅ v1.0	IND-1..6 — OmniScript composition, FlexCards, Integration Procedures, DataRaptors, EPC, Apex extensions
cms	✅ v1.0	CMS-1..5 — Content types, workspaces + channels, multi-locale variants, headless delivery, CMS vs Knowledge vs Files
data-cloud	✅ v1.0	DC-1..5 — Data Streams + DLOs, DMOs + identity resolution, calculated insights + segments, activations, SQL API for Apex/agents
einstein-agentforce	⚠️ deprecated	Renamed to agentforce in v2.5; redirect-only pack retained for back-compat with installs pinned to the old name

Bundled standards (templates/docs/)

Generic Salesforce standards copied into each new project by /argo:sf-init:

• apex-standards.md — governor limits, security, SOQL/DML, async, naming
• lwc-standards.md — CSS, JavaScript, accessibility, template directives
• react-standards.md — React-on-Salesforce specifics (only copied when platform.frontend includes react)
• quality-checklist.md — unified pre-flight checklist with dedicated Agent + Trust Layer + AI Gateway sections
• patterns/salesforce-patterns.md — 20 reusable patterns SF-1..20 (base 14 + callouts + REST + custom metadata + i18n + virtualized list + lazy-load)

Plus stubs the user fills in:

• project-context.md — object model, channels, glossary, project-specific constraints
• patterns/project-patterns.md — project-specific patterns and shared components
• adr/0000-template.md — ADR template

---

Install

As a marketplace (recommended)

/plugin marketplace add https://github.com/vhmarquez/argo
/plugin install argo@argo

Local dev loop

claude --plugin-dir /path/to/argo

Restart the Claude Code session after install so agents, skills, and hooks register.

Prerequisites

• Salesforce CLI (sf) — latest
• Node.js 20+
• @salesforce/mcp — recommended; the plugin runs without it (CLI fallback) but Headless 360 features unlock when available. Install: /argo:mcp-setup
• git 2.30+
• jq 1.6+
• bash 4+ (Git Bash on Windows)
• Java 11+ (only if using PMD-based skills: /security-scan, /complexity, lint-apex.sh)

PMD is downloaded automatically on first use of any PMD-based skill into ${CLAUDE_PLUGIN_DATA}/argo/pmd/<version>/. Run /argo:onboard to verify all prerequisites.

---

Project config (.claude/sf-project.json)

Section	Keys	Notes
project	name, description
naming.lwc	prefix, excludePrefixes
naming.react	prefix	When platform.frontend includes react
naming.apex	suffixes
platform	apiVersion, defaultTargetOrg, lwcTargets, sharingDefault, devHubAlias, packageName, frontend	frontend ∈ "lwc" | "react" | "both"
paths	lwcSource, apexSource, reactSource, reactDocs, agentDefinitions, agentDocs, lwcDocs, apexDocs, doc paths
quality	codeCoverageTarget, lintCommand, unitTestCommand, agentEvalThreshold	Agent eval threshold default 0.85 (Trust Layer band)
mcp	toolsets, allowNonGaTools	Configured by /argo:mcp-setup; downstream skills route through @salesforce/mcp when present
notifications.webhooks	slack, teams	For /notify
security	prodOrgAliases, knownNonSandboxNonProd, metadataOnly, allowAnonymousApex	Restrictive defaults; see security model. Production aliases are hard-blocked, no override

Per-environment overrides (.claude/sf-project.<env>.json) deep-merge over the base.

---

Security model

The plugin enforces four hard invariants. Detailed model in docs/security-model.md:

1. No contact with orgs classified as production (security.prodOrgAliases). Hard refuse, no override.
2. Metadata-only across all orgs by default. SOQL queries must target the metadata allowlist (ApexClass, EntityDefinition, Profile, Flow, AgentDefinition, etc., plus any *__mdt). Customer-data targets (Account, custom __c, AgentSessionTrace, User, ContentDocument) require per-call user consent.
3. Anonymous Apex disabled by default. sf apex run is refused outright; even when enabled via security.allowAnonymousApex: true, every call prompts for consent.
4. Overrides are runtime-only. No persistent "always allow" grants. Every restricted call prompts.

Enforcement is centralized in hooks/lib/security.sh (every sf-cli.sh and mcp.sh wrapper routes through it) plus a PreToolUse Bash hook (hooks/security-guard.sh) that catches anything bypassing the library. Each skill declares its data-access surface in frontmatter (data-access: none | metadata-only | data-with-consent).

Three skills fundamentally need data access and prompt for consent every run: /trust-eval (queries AgentSessionTrace), /permset-audit (queries PermissionSetAssignment), /agent-test (eval inputs / outputs may carry test PII).

/argo:sf-init detects every non-sandbox alias sf org list knows about and requires the user to classify each one as production (refused) or known-non-prod (allowed) before writing config.

---

CI integration

Skills with the --ci flag follow the CI output contract: JSON or SARIF output, exit codes per severity, configurable --fail-on threshold. SARIF integrates natively with GitHub Code Scanning.

A typical PR pipeline:

- run: /argo:code-review pr --ci --format sarif --out review.sarif
- run: /argo:security-scan --ci --format sarif --out security.sarif
- run: /argo:trust-layer-audit --ci --format sarif --out trust.sarif
- run: /argo:agent-test --ci --fail-on error
- run: /argo:diff-deploy --validate --ci
- run: /argo:coverage-trend pr
- uses: github/codeql-action/upload-sarif@v3
  with: { sarif_file: review.sarif }

CI guidance: ARGO_CONSENT_GRANTED must never be set in CI. CI cannot grant consent; runs that would prompt fail loudly.

---

Plugin layout

argo/
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── agents/                    11 subagents
├── skills/                    56 skills (each with data-access frontmatter)
├── hooks/
│   ├── hooks.json
│   ├── session-start.sh, security-guard.sh
│   ├── lint-lwc.sh, lint-apex.sh, lint-react.sh
│   └── lib/                   config, sf-cli, mcp, security, pmd, sarif
├── templates/
│   ├── CLAUDE.md
│   ├── docs/                  bundled standards (apex/lwc/react), patterns, ADR template
│   ├── packs/                 11 domain pattern packs (10 full + 1 deprecated redirect)
│   ├── gateway/               AI Gateway config profiles (dev/qa/prod)
│   └── scratch/               seed.apex + seed-agents.apex
├── docs/
│   ├── ci-output-contract.md
│   ├── pack-format.md
│   └── security-model.md
├── README.md
├── CHANGELOG.md
└── .gitattributes / .gitignore

${CLAUDE_PLUGIN_ROOT} resolves to the plugin's installed location. ${CLAUDE_PLUGIN_DATA} holds caches (org snapshots, PMD binary, coverage history, agent eval history, deploy history, consent log).

---

Compatibility

• Salesforce DX — assumes force-app/main/default/ layout (configurable via paths.*)
• Experience Cloud / Lightning Experience — platform.lwcTargets switches between lightningCommunity__* and lightning__*Page targets
• Agentforce — requires API ≥ 63.0 (set platform.apiVersion)
• React-on-Salesforce — requires API ≥ 63.0 + platform.frontend includes react
• OS — agents and skills are platform-agnostic; bash hooks run on macOS/Linux and Git Bash on Windows
• Java — required only for PMD-based skills

---

Project status

argo is currently a solo project — I develop it primarily for my own use and ship it MIT-licensed so others can try it, fork it, and adapt it freely.

• External contributions aren't accepted right now. I'm keeping the dev process single-author for the moment to maintain a coherent design and security posture. If the project gains enough traction that a contribution path makes sense, I'll revisit and publish a CONTRIBUTING.md at that point — so this isn't a permanent "no," just a "not yet."
• Security disclosures are welcome regardless of project stage. See SECURITY.md for how to report a vulnerability privately via GitHub.
• Bug reports aren't actively solicited today — fork and fix locally is the supported path.
• The plugin is on GitHub for visibility and reuse, not (yet) to incubate a community around the codebase.

If you build something interesting on top, I'd love to hear about it.

---

License

MIT — see the LICENSE file at the repo root.

---

Trademarks

argo is an independent open-source project. It is not affiliated with, endorsed by, or sponsored by Salesforce, Inc. or Anthropic.

"Salesforce", "Apex", "Lightning", "Lightning Web Components", "Agentforce", "Experience Cloud", "Einstein Trust Layer", "OmniStudio", "Data Cloud", "MuleSoft", and other Salesforce product names referenced in this repository are trademarks of Salesforce, Inc., used here descriptively to indicate compatibility. "Claude" and "Claude Code" are trademarks of Anthropic, PBC.

All other trademarks are the property of their respective owners.

---

Changelog

See CHANGELOG.md for the full v1.0.0 → v3.x history.