Agent skill
grace-setup-subagents
Create GRACE subagent presets for the current agent shell. Use when you want GRACE worker and reviewer agent files scaffolded for Claude Code, OpenCode, Codex, or another shell.
Install this agent skill to your Project
npx add-skill https://github.com/osovv/grace-marketplace/tree/main/plugins/grace/skills/grace/grace-setup-subagents
SKILL.md
Create GRACE subagent files for the current shell by reusing the shell's own agent-file conventions.
Purpose
$grace-multiagent-execute works best when the shell already has GRACE-specific worker and reviewer presets.
This skill scaffolds those presets into the correct local agent directory for the current shell.
The controller remains in the main session. Workers should expect compact execution packets, fresh one-module ownership, scoped reviews, and controller-managed graph updates.
Default Roles
By default, create these subagents:
grace-module-implementergrace-contract-reviewergrace-verification-reviewergrace-fixer
The main session remains the controller. This skill does not create a controller agent.
Process
Step 1: Detect the Current Shell
Use the current environment and project structure to determine where the skill is running.
Prefer, in this order:
- explicit environment hints from the shell
- project-local config directories
- user-level agent directories
Typical examples:
- Claude Code projects often use
.claude/ - OpenCode projects often use
.opencode/ - Codex projects often use
.codex/
If detection is ambiguous, ask the user which shell to target.
Step 2: Find a Real Agent File Example
Do not guess the target file format if a local example exists.
Search for an existing agent file for the current shell:
- first in the current project
- then in the user's global config for that shell
- then in nearby projects if needed
Use a real example to infer:
- file extension
- frontmatter or config structure
- model field names
- tool/permission layout
If no reliable local example exists:
- look for official shell documentation
- if documentation is still unclear, ask the user for a canonical sample or doc link
Step 3: Choose Scope and Target Directory
Default to project-local setup unless the user explicitly asks for global setup.
Create the GRACE presets under the shell's local agent directory in a grace/ subfolder when the shell supports subfolders cleanly.
If the shell does not support nested subfolders for agents, place the generated files directly in the local agent directory with GRACE-prefixed names.
Step 4: Read the Role Prompts
The role prompt bodies live in references/roles/:
references/roles/module-implementer.mdreferences/roles/contract-reviewer.mdreferences/roles/verification-reviewer.mdreferences/roles/fixer.md
These are the shared role bodies. Reuse them. Only the shell-specific wrapper should change.
These shared prompts assume the newer multi-agent workflow:
- workers receive execution packets instead of rereading full XML artifacts whenever possible
- reviewers default to scoped gate review and escalate only when evidence suggests wider drift
- verification is split across module, wave, and phase levels
- the controller owns
docs/verification-plan.xmlin addition to plan and graph artifacts
Step 5: Render Shell-Specific Agent Files
For each role:
- wrap the shared role body in the file format discovered in Step 2
- preserve the shell's conventions for:
- metadata fields
- model names
- permissions or tool declarations
- subagent mode flags
- write the file into the target directory
If the shell has no first-class subagent file concept, create the nearest useful equivalent and explain the limitation.
Step 6: Report What Was Created
After scaffolding, report:
- detected shell
- target directory
- created files
- any assumptions copied from the local example
- any fields the user should adjust manually, such as model aliases
Rules
- Prefer copying the shell's real local format over inventing one
- Keep prompts aligned with
grace-multiagent-execute,grace-reviewer,grace-fix, andgrace-verification - Do not overwrite existing files without user intent
- Do not create architecture-planning agents here; this skill is for execution support
- Do not introduce worker-pool or worker-reuse assumptions into the generated presets
- If the shell supports agents differently, create the nearest working equivalent and explain the difference
Recommended Agent Skills
Expand your agent's capabilities with these related and highly-rated skills.
grace-verification
Design and enforce testing, traces, and log-driven verification for a GRACE project. Use when modules need stronger automated tests, execution-trace checks, or a maintained verification-plan.xml that autonomous and multi-agent workflows can trust.
grace-refactor
Refactor GRACE-governed code safely: rename, move, split, merge, or extract modules while keeping contracts, graph, verification, and semantic markup synchronized.
grace-fix
Debug an issue using GRACE semantic navigation. Use when encountering bugs, errors, or unexpected behavior - navigate through the graph, verification plan, and semantic blocks to analyze the mismatch and apply a targeted fix.
grace-cli
Operate the optional `grace` CLI against a GRACE project. Use when you want to lint GRACE artifacts, resolve modules from names or file paths, inspect shared/public module context, or inspect file-local/private markup through `grace lint`, `grace module find`, `grace module show`, and `grace file show`.
grace-plan
Run the GRACE architectural planning phase. Use when you have requirements and technology decisions defined and need to design the module architecture, create contracts, map data flows, and establish verification references. Produces development-plan.xml, verification-plan.xml, and knowledge-graph.xml.
grace-reviewer
GRACE integrity reviewer. Use for fast scoped gate reviews during execution, or full integrity audits at phase boundaries and after broader code, graph, or verification changes.
Didn't find tool you were looking for?