claude-to-grok

Claude → Grok (CCW runtime map)

Runtime invariants

1. Preserve CCW execution — keep every CCW step, [Stop] gate, handoff contract, review loop, commit cycle, and quality gate mandatory.
2. Translate identifiers — replace Claude Code tool/agent names with the Grok tools in the maps; keep prompt what, deliverable paths, and acceptance criteria unchanged.
3. Use the required specialist — spawn exclusively the plugin-qualified CCW agent type named by the recipe (dev-workflows-fullstack:<name> or the backend/frontend prefix).
4. Register progress — fulfill CCW TaskCreate/TaskUpdate duties with todo_write (session todos; duty-level equivalent only — see Progress tracking).
5. Handle runtime failures — when a mapped tool or agent cannot complete, follow the Runtime failure procedure. Map only documented equivalents.

Recipe invocation remains exclusively user-slash initiated (disable-model-invocation on recipe-*). This skill activates for translation during orchestration.

Tool name map (Claude Code → Grok)

Claude Code (in CCW text)	Grok tool	Notes
Agent / Task	spawn_subagent	Fresh call per phase. Then wait — see Waiting for a subagent. Use resume_from only when CCW explicitly continues the same subagent session.
Bash	run_terminal_command	git commit, ls, verification shells
Read	read_file
Write	write	New file or full overwrite
Edit / MultiEdit	search_replace	Prefer exact old→new edits; use write when full-file write is clearer
Grep	grep
Glob / LS / ListDir	list_dir + grep / run_terminal_command as needed	Discover paths via list_dir or shell only when list_dir is insufficient
WebSearch	web_search
AskUserQuestion	ask_user_question	All [Stop] confirmations
TaskCreate / TaskUpdate	todo_write	Duty equivalent — see Progress tracking
SendMessage (resume)	spawn_subagent + resume_from	Prefer fresh spawn_subagent per CCW isolation; resume_from only when CCW continues the same session

Orchestrator write tools: use mapped Edit/Write tools only inside specialists that CCW allows to edit.

Waiting for a subagent

CCW text says “invoke Agent and use the result”; it does not name a Grok wait API.

After spawn_subagent:

1. Read the returned subagent id (or task id from the tool result).
2. Call get_command_or_subagent_output with that id (use a positive timeout_ms when waiting for completion).
3. WHEN the subagent is still running or the wait times out, wait again with the same id.
4. WHEN it completes with a final payload, use that payload as the specialist result and continue the CCW flow (Stop, next specialist, etc.).
5. WHEN it fails, is cancelled, or completes without a final payload, run the Runtime failure procedure.

Treat the spawn result as pending unless it already contains a completed final payload. spawn_subagent often runs in the background by default.

Subagent type map

Canonical form:

dev-workflows-fullstack:<agent-name>

CCW subagent_type text	Grok subagent_type
requirement-analyzer	dev-workflows-fullstack:requirement-analyzer
task-executor	dev-workflows-fullstack:task-executor
task-executor-frontend	dev-workflows-fullstack:task-executor-frontend
quality-fixer	dev-workflows-fullstack:quality-fixer
quality-fixer-frontend	dev-workflows-fullstack:quality-fixer-frontend
document-reviewer	dev-workflows-fullstack:document-reviewer
technical-designer	dev-workflows-fullstack:technical-designer
technical-designer-frontend	dev-workflows-fullstack:technical-designer-frontend
work-planner	dev-workflows-fullstack:work-planner
task-decomposer	dev-workflows-fullstack:task-decomposer
codebase-analyzer	dev-workflows-fullstack:codebase-analyzer
code-verifier	dev-workflows-fullstack:code-verifier
security-reviewer	dev-workflows-fullstack:security-reviewer
acceptance-test-generator	dev-workflows-fullstack:acceptance-test-generator
design-sync	dev-workflows-fullstack:design-sync
integration-test-reviewer	dev-workflows-fullstack:integration-test-reviewer
prd-creator	dev-workflows-fullstack:prd-creator
ui-analyzer	dev-workflows-fullstack:ui-analyzer
ui-spec-designer	dev-workflows-fullstack:ui-spec-designer

WHEN the installed plugin is backend-only or frontend-only, use:

• dev-workflows:<agent-name>
• dev-workflows-frontend:<agent-name>

Use the plugin-qualified subagent type. WHEN it is unavailable, inspect the registered subagent types and follow the Runtime failure procedure. Use a bare name only when inspection shows that exact bare name is registered.

Progress tracking (TaskCreate / TaskUpdate → todo_write)

Fulfill the same registration duty with todo_write:

CCW intent	Grok action
Register all flow steps after scale determination	todo_write (merge: false or first write) listing every step; include “Map preloaded skills to concrete rules” when the agent requires it
Mark current step in progress	todo_write merge: true, status: in_progress
Complete a step	merge: true, status: completed
Specialist internal TaskCreate first/last tasks	Same via todo_write inside that subagent when it has the tool

Duty limits

• todo_write is session todo state, not Claude’s Task board API.
• Use stable string IDs (e.g. flow-req-analyzer, flow-prd).
• MCP scheduled tasks represent scheduled jobs; use todo_write for CCW progress registration.
• IF progress registration is required and todo_write is unavailable → Runtime failure procedure.

Subagent prompt construction

Keep CCW prompt rules. Always include:

1. Concrete deliverable paths with every placeholder resolved
2. Expected action
3. Scope boundary block from the recipe when the recipe requires it
4. Runtime footer WHEN the selected agent instructions or delegated prompt refer to Claude Code tool names:

Runtime (Grok): Agent/Task→spawn_subagent then get_command_or_subagent_output;
Bash→run_terminal_command; Read→read_file; Edit/MultiEdit→search_replace; Write→write;
AskUserQuestion→ask_user_question; TaskCreate/TaskUpdate→todo_write (duty).
subagent_type prefix: dev-workflows-fullstack: (or dev-workflows: / dev-workflows-frontend:).
Preserve every CCW step and gate.

Orchestrator tool surface (mapped)

CCW permitted	Grok
Agent	spawn_subagent + wait via get_command_or_subagent_output
AskUserQuestion	ask_user_question
TaskCreate / TaskUpdate	todo_write
Bash	run_terminal_command
Read	read_file

Implementation Edit/Write/MultiEdit remain specialist-only per CCW.

Runtime failure procedure

WHEN a mapped operation cannot complete (including: no map row, agent not registered, spawn_subagent failed, subagent failed/cancelled, no final payload, mapped tool missing from the current surface):

1. Record the attempted CCW obligation, mapped Grok operation, and observed result.
2. Inspect the currently available Grok tools or subagent types relevant to that operation.
3. Retry only when the result indicates a transient or still-running state.
4. Continue with a documented equivalent when the current environment exposes one that appears in this skill’s maps.
5. WHEN no executable equivalent is available, report using the format below.

Runtime failure output

## Unresolved CCW→Grok runtime failure

- CCW obligation:
- Mapped Grok operation:
- Observed result:
- Available tools or subagent types checked:
- Missing capability:
- User action or information required:

Scope

This skill covers runtime identifier translation, progress-duty translation, subagent waiting, and runtime-failure reporting. CCW defines recipe bodies, workflow gates, and quality policy.