Agent Skill

Introduction

The rgctl agent skill is a structured instruction set that teaches AI coding agents (Claude Code, Antigravity, Codex, Cursor) how to use the rgctl CLI to answer structural questions about a codebase. When installed into a repository, the skill gives your agent the ability to automatically map natural-language questions to the right rgctl commands, interpret the results, and report findings -- all without the developer needing to know the CLI syntax.

The agent pack embeds skills and optional chat commands into the rgctl binary. rgctl install copies them into per-product paths (.cursor/skills/rgctl/, .claude/skills/rgctl/, .agents/skills/rgctl/, .agent/skills/rgctl/, OpenCode, Pi, … — see Agent commands). You get a meta skill rgctl, eight workflow skills (rgctl-discover, rgctl-gql, …), and with --with-commands slash/prompt files that map to the same workflows. No external downloads. Once installed, the agent follows a structured loop: parse the question, route to the right workflow or CLI command, run rgctl -f json, summarize results.

The skill turns rgctl from a CLI tool into an always-available architectural advisor inside your editor.

Use Cases

The agent skill unlocks several powerful workflows where the combination of structural graph analysis and AI reasoning produces results that neither could achieve alone.

Refactoring with Confidence

Before renaming, extracting, or restructuring a function, the agent can automatically check the blast radius, trace callers, and verify that no hidden dependencies will break. Instead of manually running commands, you ask the agent a natural-language question and it handles the analysis.

Monolith-to-Microservice Migration

The agent can generate a complete migration roadmap, explain the ordering rationale, identify high-risk extraction targets, and guide you through each step -- all from conversational prompts like "generate a migration plan for this repo" or "what should we extract first?"

Porting a Function from One Language to Another

When rewriting a function from Java to Go (or any language pair), the agent can extract the function's data-flow graph, call neighborhood, and field mutations from the source language, then verify that the target-language implementation preserves the same logic and data-flow structure. The graph provides a language-independent ground truth.

Writing Better Test Cases

The agent can analyze a function's control-flow graph, identify all branch paths, trace data dependencies, and use that structural information to generate test cases that cover the actual code paths rather than guessing at coverage.

Continuous Architecture Review

With the skill installed, every code review conversation has access to architectural context. The agent can check policy compliance, detect coupling drift, and flag high-impact changes before they are merged.

Example Project

This guide uses the CoolStore (example/coolstore). Make sure you have run discover first:

rgctl -r example/coolstore discover --with-cfg

Step-by-Step

1. Install the Skill

Install the rgctl agent skill into your repository:

rgctl -r example/coolstore install --skill --with-commands --tools cursor,claude,codex,antigravity,agents

Text mode lists each created or updated path. Typical layout (Cursor example):

• .cursor/skills/rgctl/SKILL.md — meta router + references/
• .cursor/skills/rgctl-gql/SKILL.md — per-workflow skills (eight workflows)
• .cursor/commands/rgctl-gql.md — slash command stub (with --with-commands)

What happened:

• rgctl unpacked the embedded agent pack (generated at build time from skills/rgctl/ and skills/rgctl/workflows/).
• Claude uses .claude/skills/ and colon-style commands (/rgctl:gql).
• Codex / agents / zed share .agents/skills/ (install dedupes).
• Cursor uses .cursor/skills/ and /rgctl-gql commands.
• Workflow prose lives in skills/rgctl/workflows/*.md; installed references/workflows.md is assembled from those fragments.
• No network — content matches your rgctl binary version.

See Agent commands for the full flag table and registry (install --list-agents).

2. Verify with JSON Output

Check the install status programmatically:

rgctl -r example/coolstore -f json install --skill --with-commands --tools cursor \
  | jq '{schema_version, scope, agents, with_commands, writes: [.writes[] | {agent, workflow, kind, status}]}'

Output (schema version 2, abbreviated):

{
  "schema_version": 2,
  "command": "install",
  "skill": "rgctl",
  "repo": "/path/to/example/coolstore",
  "scope": "local",
  "agents": ["cursor"],
  "with_commands": true,
  "with_policy": false,
  "force": false,
  "writes": [
    {
      "agent": "cursor",
      "workflow": null,
      "kind": "meta",
      "path": "/path/to/example/coolstore/.cursor/skills/rgctl/SKILL.md",
      "status": "unchanged"
    },
    {
      "agent": "cursor",
      "workflow": "gql",
      "kind": "skill",
      "path": "/path/to/example/coolstore/.cursor/skills/rgctl-gql/SKILL.md",
      "status": "created"
    }
  ]
}

The status field for each write is one of:

• created -- new file written
• unchanged -- file already exists with identical content (idempotent)
• overwritten -- existing file replaced (with --force)
• skipped_exists -- file differs but --force was not set (exit code 1)

3. Install for specific agents

Limit adapters with --tools (comma-separated registry ids):

rgctl -r example/coolstore install --skill --with-commands --tools claude
rgctl -r example/coolstore install --skill --with-commands --tools codex,agents
rgctl -r example/coolstore install --skill --with-commands --tools antigravity
rgctl -r example/coolstore install --skill --with-commands --tools cursor
rgctl install --list-agents   # all ids and paths

--host is deprecated; it still maps to a subset of tools but prints a warning.

4. Update After Upgrading rgctl

When you upgrade rgctl, the embedded skill may have changed. Update it with --force:

rgctl -r example/coolstore install --skill --with-commands --force

This overwrites any existing skill files, even if they have been modified locally.

5. The Agent Loop

Once installed, the agent follows a 5-step loop for every structural question:

1. USER PROMPT     "What's the impact of changing priceShoppingCart?"
2. TOOL CALL       rgctl -f json blast-radius priceShoppingCart
3. GRAPH FACTS     Parse JSON: score 40.35, 5 callers, impact zone 7
4. LLM REASONING   Summarize: moderate risk, spans service and REST layers
5. ACTION          Report findings, suggest next steps

The skill includes a decision table that maps 19 categories of natural-language questions to the right rgctl command. The agent does not need to be told which command to use -- it matches the user's intent automatically.

Use Case: Refactoring with Confidence

Scenario

You want to refactor priceShoppingCart in the CoolStore application -- perhaps splitting it into separate pricing methods for items and shipping. Before making changes, you need to understand the full impact.

What the Agent Does

When you ask the agent: "What happens if I change priceShoppingCart?"

The agent follows the skill's decision table (row 10: "Impact if I change X") and runs:

rgctl -r example/coolstore -f json blast-radius priceShoppingCart

Output:

{
  "metrics": {
    "score": 40.35,
    "direct_callers_count": 5,
    "impact_zone_size": 7
  },
  "target": {
    "canonical_fqn": "ShoppingCartService::priceShoppingCart",
    "file_path": "example/coolstore/./src/main/java/com/redhat/coolstore/service/ShoppingCartService.java",
    "signature": "public void priceShoppingCart(ShoppingCart sc) {"
  },
  "topology": {
    "direct_callers": [
      {"fqn": "com.redhat.coolstore.service.ShoppingCartService.checkOutShoppingCart"},
      {"fqn": "com.redhat.coolstore.rest.CartEndpoint.add"},
      {"fqn": "com.redhat.coolstore.rest.CartEndpoint.dedupeCartItems"},
      {"fqn": "com.redhat.coolstore.rest.CartEndpoint.delete"},
      {"fqn": "com.redhat.coolstore.rest.CartEndpoint.set"}
    ]
  }
}

The agent then reports:

priceShoppingCart has a blast-radius score of 40.4/100 (moderate risk). It is called by 5 functions: checkOutShoppingCart in the service layer, and 4 REST endpoint methods (add, dedupeCartItems, delete, set). The total impact zone is 7 functions. If you change the signature, all 5 callers will need updating.

Going Deeper: Call Neighborhood

The agent can then trace the call neighborhood to understand what priceShoppingCart itself depends on:

rgctl -r example/coolstore -f json cpg calls priceShoppingCart

Output (truncated):

{
  "edges": [
    {"direction": "out", "name": "initShoppingCartForPricing"},
    {"direction": "out", "name": "getShoppingCartItemList"},
    {"direction": "out", "name": "setCartItemPromoSavings"},
    {"direction": "out", "name": "getCartItemTotal"},
    {"direction": "out", "name": "setCartItemTotal"},
    {"direction": "out", "name": "setShippingTotal"},
    {"direction": "out", "name": "setCartTotal"}
  ]
}

The agent reports: "priceShoppingCart calls 7 functions: it initializes pricing, iterates cart items to compute totals and promo savings, calculates shipping, and sets the final cart total. A refactoring that splits item pricing from shipping pricing should preserve these call relationships."

Verifying Field Mutations

Before splitting the function, check which fields it modifies:

rgctl -r example/coolstore -f json cpg mutations \
  --type ShoppingCart --exclude-ctors

Output:

{
  "mutations": [
    {"function": "setShoppingCartItemList", "member": "shoppingCartItemList", "line": 35},
    {"function": "setCartItemTotal", "member": "cartItemTotal", "line": 71},
    {"function": "setShippingTotal", "member": "shippingTotal", "line": 79},
    {"function": "setCartTotal", "member": "cartTotal", "line": 87},
    {"function": "setCartItemPromoSavings", "member": "cartItemPromoSavings", "line": 95},
    {"function": "setShippingPromoSavings", "member": "shippingPromoSavings", "line": 103}
  ]
}

The agent reports: "ShoppingCart has 6 mutable fields. The priceShoppingCart method writes to cartItemTotal, shippingTotal, cartTotal, cartItemPromoSavings, and shippingPromoSavings through their setters. If you split the method, ensure both halves still set all required fields."

Use Case: Migration Planning

Scenario

You need to decompose the CoolStore application into microservices and want a data-driven extraction plan.

What the Agent Does

When you ask: "Generate a migration plan for this codebase"

The agent follows the skill's decision table (row 1: "Generate a migration plan") and runs:

rgctl -r example/coolstore discover \
  --with-cfg --with-harmonic --export-migration-hints \
  --migration-preset hybrid_default --migration-order scheduled

Then reads the plan:

{
  "preset": "hybrid_default",
  "order_mode": "scheduled",
  "steps": [
    {
      "step": 1,
      "label": "main.webapp.bower_components.matches-selector",
      "max_blast": 0.0,
      "avg_pagerank": 0.0000634
    },
    {
      "step": 2,
      "label": "com.redhat.coolstore.model",
      "max_blast": 0.0,
      "avg_pagerank": 0.0000532
    }
  ]
}

The agent reports:

The migration plan contains N steps using the hybrid_default preset (balanced weighting of PageRank, harmonic centrality, and blast radius). The first extraction targets are low-risk leaf modules with zero blast radius. Step 2 extracts the domain model (com.redhat.coolstore.model), which has no blast-radius impact and can be safely moved to a shared library.

Investigating a Migration Step

When the user asks: "Tell me more about the coolstore model community"

The agent queries community members:

rgctl -r example/coolstore -f json gql \
  "MATCH (f:Function) WHERE f.community_id = '13' RETURN f LIMIT 20"

And checks blast radius on key functions:

rgctl -r example/coolstore blast-radius getShoppingCart --depth 3

Use Case: Porting a Function to Another Language

Scenario

You are migrating priceShoppingCart from Java to Go (or TypeScript, Rust, Python, etc.) and need to ensure the new implementation preserves the same data flow and logic.

What the Agent Does

Step 1: Extract the structural blueprint from the source language.

The agent captures the function's data-flow graph, which is language-independent:

rgctl -r example/coolstore -f json cpg flows \
  ./src/main/java/com/redhat/coolstore/service/ShoppingCartService.java \
  --line 68 --variable sc --function priceShoppingCart --direction forward

Output:

{
  "direction": "forward",
  "function": "priceShoppingCart",
  "steps": [
    {"code": "for-each sc.getShoppingCartItemList()", "line": 64},
    {"code": "sc.setCartItemTotal(sc.getCartItemTotal() + sci.getPrice() * sci.getQuantity())", "line": 68},
    {"code": "ps.applyShippingPromotions(sc)", "line": 81},
    {"code": "sc.setCartTotal(sc.getCartItemTotal() + sc.getShippingTotal())", "line": 83}
  ],
  "reduction_percent": 66.67
}

Step 2: Extract the call neighborhood.

rgctl -r example/coolstore -f json cpg calls priceShoppingCart

This gives the agent the complete list of outgoing calls that the new implementation must replicate.

Step 3: Extract field mutations.

rgctl -r example/coolstore -f json cpg mutations \
  --type ShoppingCart --exclude-ctors

This lists every field that priceShoppingCart writes to, which the new implementation must also write.

Step 4: Extract the PDG.

rgctl -r example/coolstore -f json cpg pdg priceShoppingCart

Output (truncated):

{
  "control_deps": 18,
  "data_deps": 5,
  "edges": [
    {"kind": "data", "source": "node_7", "target": "node_8", "variable": "sci"},
    {"kind": "data", "source": "node_7", "target": "node_8", "variable": "sc"},
    {"kind": "data", "source": "node_4", "target": "node_6", "variable": "sc"},
    {"kind": "control", "source": "node_1", "target": "node_5"}
  ]
}

How the agent uses this:

The agent now has four language-independent facts about priceShoppingCart:

1. Data flow path: iteration over cart items, accumulation of totals, shipping calculation, final total.
2. Call contract: the 7 functions it must call in the same order.
3. Mutation contract: the 5 ShoppingCart fields it must write.
4. Dependency graph: 18 control edges and 5 data edges encoding the computation structure.

When writing the Go implementation, the agent verifies each fact against the new code. If a data-flow step is missing, a call is omitted, or a field mutation is skipped, the agent flags it.

The agent can report: "The Java implementation of priceShoppingCart has 4 data-flow steps, calls 7 functions, and mutates 5 fields on ShoppingCart. Your Go implementation should replicate the same iteration pattern (for-each over cart items), accumulate the same totals, and call equivalent functions for shipping calculation and promo application."

Use Case: Writing Better Test Cases

Scenario

You need to write tests for priceShoppingCart and want to ensure you cover all branch paths and edge cases.

What the Agent Does

Step 1: Examine the control-flow graph.

rgctl -r example/coolstore -f json inspect priceShoppingCart cfg

Output (truncated):

{
  "edges": [
    {"kind": "next", "source": "block_5", "target": "block_7"},
    {"kind": "iftrue", "source": "block_7", "target": "block_8"},
    {"kind": "iffalse", "source": "block_7", "target": "block_1"},
    {"kind": "iftrue", "source": "block_9", "target": "block_10"},
    {"kind": "iffalse", "source": "block_9", "target": "block_3"},
    {"kind": "iftrue", "source": "block_12", "target": "block_13"},
    {"kind": "iffalse", "source": "block_12", "target": "block_14"},
    {"kind": "jump", "source": "block_13", "target": "block_12"},
    {"kind": "iftrue", "source": "block_15", "target": "block_16"},
    {"kind": "iffalse", "source": "block_15", "target": "block_0"}
  ]
}

The agent identifies all branch points:

priceShoppingCart has 4 branch conditions:

1. Null check: sc != null (block_7) -- test with null and non-null input
2. List check: sc.getShoppingCartItemList() != null && size > 0 (block_9/10) -- test with empty and populated cart
3. Loop: for-each sci : sc.getShoppingCartItemList() (block_12/13) -- test with 0, 1, and many items
4. Price threshold: sc.getCartItemTotal() >= 25 (block_15) -- test below and above the threshold

Step 2: Examine data dependencies.

rgctl -r example/coolstore -f json inspect priceShoppingCart pdg --edge-layer data

The agent identifies the key data flows:

Variable sc flows through 3 data-dependency edges. Variable sci flows through the loop body. Test cases should verify that each accumulation step (cartItemTotal, cartItemPromoSavings, shippingTotal) produces correct values.

Step 3: Check field mutations to verify test assertions.

rgctl -r example/coolstore -f json cpg mutations \
  --type ShoppingCart --exclude-ctors

The agent reports: "Your tests should assert on 5 ShoppingCart fields after calling priceShoppingCart: cartItemTotal, shippingTotal, cartTotal, cartItemPromoSavings, and shippingPromoSavings. Here are the test cases derived from the CFG:"

The agent can then generate concrete test cases:

1. Null input: call priceShoppingCart(null) -- no fields should be mutated.
2. Empty cart: cart with no items -- totals should be zero.
3. Single item below threshold: one item priced below 25 -- no shipping insurance.
4. Single item above threshold: one item priced at or above 25 -- shipping insurance applied.
5. Multiple items with promotions: verify promo savings accumulate correctly.
6. Boundary case: cart total exactly 25 -- verify threshold behavior.

The NL-to-Command Decision Table

The skill embeds a decision table that maps natural-language patterns to CLI commands. Here are the key mappings:

The agent handles disambiguation (e.g., adding --class or --file when a symbol name is ambiguous) and error recovery (e.g., running discover --with-cfg if slicing fails because CFG data is missing).

Install options reference

See Agent commands for the full table. Summary:

How the pack is distributed

At rgctl build time, rgctl-agent-pack-codegen generates the pack from agent-pack/manifest.yaml, agent-pack/agents/registry.toml, and skills/rgctl/workflows/, then embeds it as a zip in the binary. This means:

• No network access needed to install.
• Pack version matches the CLI version.
• Upgrading rgctl and running install --skill --with-commands --force refreshes skills and commands.
• Target repos do not need a checkout of the rgctl source tree.

Benefits

• Zero-configuration AI integration. One command installs everything the agent needs to understand your codebase structurally.
• Natural-language interface. Developers ask questions in English; the agent translates to CLI commands.
• Language-independent analysis. Data-flow graphs, call neighborhoods, and mutation tracking work across all 9 Tier 1 languages.
• Refactoring safety net. Blast radius, call tracing, and mutation analysis catch breaking changes before they ship.
• Test case generation. CFG branch analysis and data dependencies produce higher-coverage tests.
• Migration guidance. The agent can generate, explain, and walk through a complete migration roadmap.
• Cross-language porting. Data-flow and dependency graphs provide a structural blueprint that transcends language syntax.
• Always in sync. The skill is embedded in the binary, so it always matches the CLI version.

Related Guides

• Agent commands — install flags, adapters, workflow ↔ CLI table
• Discovering and Indexing a Codebase -- the discover step that all agent queries depend on
• Blast Radius Analysis -- the most common agent query for refactoring safety
• Hybrid CPG -- mutations, flows, and call neighborhoods used in porting and testing
• Program Slicing -- statement-level analysis for data-flow tracing
• Inspecting CFG, PDG, and Dominance -- the CFG data that drives test case generation
• Migration Planning -- the migration roadmap the agent can generate and explain
• CI Policy Checks -- policy validation the agent runs for continuous architecture review