claude

Claude Code Advisor

Use this skill to route Codex work to the local Claude Code CLI through the bundled companion runtime. Codex remains the orchestrator. Claude Code is the advisor, checker, or reviewer.

Invocation Contract

Canonical forms:

$claude setup
$claude advise [--max-turns <n>] <question>
$claude do [--background] [--write] [--model sonnet|opus] [--max-turns <n>] <prepared task>
$claude rescue [--background] [--write] [--resume] [--model sonnet|opus] [--max-turns <n>] <task>
$claude review [--base <ref>]
$claude adversarial-review [--base <ref>] [focus]
$claude monitor [job-id]
$claude status [job-id]
$claude result [job-id]
$claude cancel [job-id]
$claude resume-candidate

If Codex passes slash-style text through to this skill, normalize it before routing:

/claude <subcommand> <args> -> $claude <subcommand> <args>
/claude:setup -> $claude setup
/claude:advise <question> -> $claude advise <question>
/claude:do <prepared task> -> $claude do <prepared task>
/claude:rescue <task> -> $claude rescue <task>
/claude:review [--base <ref>] -> $claude review [--base <ref>]
/claude:adversarial-review [focus] -> $claude adversarial-review [focus]
/claude:monitor [job-id] -> $claude monitor [job-id]
/claude:status [job-id] -> $claude status [job-id]
/claude:result [job-id] -> $claude result [job-id]
/claude:cancel [job-id] -> $claude cancel [job-id]

Do not add or depend on undocumented plugin manifest fields for custom slash commands. The guaranteed Codex surface is the $claude skill mention.

Hard Rules

• Always call the bundled companion at <plugin root>/scripts/claude-companion.mjs using an absolute path. The plugin root is the claude-code-advisor directory that contains .codex-plugin/, skills/, and scripts/.
• Do not call claude directly from the skill instructions.
• Return a concise human summary unless the user asks for raw JSON.
• Review and adversarial-review are read-only.
• Write-capable Claude work requires explicit --write.
• Every default companion launch must be non-interactive. Keep the runtime's MCP isolation flags: --mcp-config '{"mcpServers":{}}' --strict-mcp-config --no-chrome. These are essential for unattended runs; without them Claude Code can stop at the MCP permission picker instead of producing an answer.
• Do not use project MCP servers unless the user explicitly asks for MCP. Background mode must refuse the current directory and ancestor directories with .mcp.json unless --allow-mcp is explicit. Use --allow-mcp only after that explicit user approval.
• Keep read-only prepared local tasks (do and rescue) on local tools by default: Read,Glob,Grep. advise may use web tools. For do and rescue, enable WebFetch or WebSearch only when the task needs web access and --allow-web is explicit.
• Do not pass --model sonnet for advice, review, adversarial-review, rescue, or monitor work unless the user explicitly asks for Sonnet. Let Claude Code use the user's configured default model.
• Pass --effort xhigh for advice, review, adversarial-review, rescue, and background jobs unless the user explicitly asks for another effort.
• Sonnet is reserved for explicit junior-agent delegation governed by the tasks-for-sonnet skill. Never send Sonnet a vague task.
• Before $claude do --model sonnet, load tasks-for-sonnet and prepare the task from that skill. The prompt must include the role, absolute paths or a pinned commit, a word cap, "What Must Be True", "Known Constraints", and "Mechanical Verification".
• Use $claude do --model opus for complex/high-judgment Claude tasks when the user asks for or agrees to a stronger Claude worker. Treat Opus as the Claude-side choice for tasks you would not hand to a junior agent.
• Foreground advise, do, and rescue use a larger default max-turn budget for prepared work. review and adversarial-review stay tight and structured with one default turn. If a foreground prepared task hits the max-turn limit, rerun with --max-turns <higher> or narrow the task.
• Do not auto-resume a Claude job when the companion says explicit selection is required.