feat(agents): add Claude delegation skill

Add a repository-owned Codex skill and helper for bounded Claude Code execution, model and effort selection, resumable follow-ups, and independent acceptance checks.

.codex/skills/delegate-to-claude/SKILL.md

@@ -0,0 +1,73 @@
+---
+name: delegate-to-claude
+description: Delegate a bounded coding, investigation, or review task from Codex to Claude Code, then inspect and independently assess Claude's result. Use when the user asks Codex to delegate to Claude or when a second-model execution or review pass is explicitly desired. Do not use for ordinary Codex subagents or as a substitute for Codex's final review.
+---
+
+# Delegate to Claude Code
+
+Use Claude Code as an executor or independent reviewer while Codex remains the coordinator responsible for scope, evidence, and acceptance.
+
+## Preconditions
+
+1. Confirm `claude` is installed and authenticated by running `scripts/claude_delegate.sh check` from this skill directory.
+2. Read the repository's applicable `AGENTS.md`, `CLAUDE.md`, and task-specific instructions before writing the work order.
+3. Do not delegate secrets, provider writes, production mutations, releases, or destructive operations unless the user explicitly authorized that exact action. Claude receives no broader authority than Codex has.
+
+## Work order
+
+Write the task into a temporary Markdown file outside the repository unless the user requested a durable task artifact. Make it self-contained and include:
+
+- role: executor or read-only reviewer;
+- exact objective and boundaries;
+- working directory and relevant files;
+- repository instructions Claude must read;
+- allowed side effects and forbidden actions;
+- required checks and expected evidence;
+- requested return format: findings or changed files, commands run, results, and blockers.
+
+The helper adds an executor marker and tells Claude not to spawn further agents. Do not pass the user's whole conversation when a concise work order is sufficient.
+
+## Run
+
+From the target repository directory:
+
+```bash
+/absolute/path/to/this-skill/scripts/claude_delegate.sh run /absolute/path/to/work-order.md /absolute/path/to/result.json
+```
+
+The default permission mode is `auto`, which lets Claude decide whether tools are safe to use. Override it only when the task requires a stricter mode:
+
+```bash
+CLAUDE_DELEGATE_PERMISSION_MODE=plan scripts/claude_delegate.sh run work-order.md result.json
+```
+
+Optional environment variables:
+
+- `CLAUDE_DELEGATE_MODEL`: Claude model alias or full model name; default `sonnet`.
+- `CLAUDE_DELEGATE_EFFORT`: `low`, `medium`, `high`, `xhigh`, or `max`; default `medium`.
+- `CLAUDE_DELEGATE_MAX_BUDGET_USD`: API budget ceiling when supported by the current authentication mode.
+- `CLAUDE_DELEGATE_PERMISSION_MODE`: default `auto`; prefer `plan` for read-only planning/review when tools are unnecessary.
+
+The helper loads project and local Claude settings while excluding user-level integrations by default, and disables browser/MCP integrations. This keeps delegation focused on the checked-out repository. If a task genuinely requires an external connector, handle that action in Codex under the original authorization rather than silently widening Claude's access.
+
+Never choose `bypassPermissions` merely to avoid a permission failure. Use it only when the user explicitly requests it and the execution is inside an appropriately isolated sandbox.
+
+For a follow-up in the same Claude session, obtain `session_id` from the JSON result and run:
+
+```bash
+scripts/claude_delegate.sh resume SESSION_ID follow-up.md follow-up-result.json
+```
+
+If a call outlives the command yield, continue waiting on that same process/session rather than launching a duplicate.
+
+## Accept or iterate
+
+Claude's summary is evidence to inspect, not proof. Codex must:
+
+1. inspect the actual diff or artifacts;
+2. run proportionate independent checks;
+3. compare the result with the original scope and repository rules;
+4. send a focused follow-up with concrete failures when iteration is needed;
+5. report Claude's contribution separately from Codex's verification.
+
+Do not claim success from Claude's self-report alone. Codex retains the decision to accept, commit, publish, deploy, or make any external change.

.codex/skills/delegate-to-claude/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Delegate to Claude Code"
+  short_description: "Delegate scoped work to Claude Code"
+  default_prompt: "Use $delegate-to-claude to delegate this scoped task to Claude Code and review the result."

.codex/skills/delegate-to-claude/scripts/claude_delegate.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+die() {
+  printf 'claude_delegate: %s\n' "$*" >&2
+  exit 1
+}
+
+usage() {
+  cat >&2 <<'EOF'
+usage:
+  claude_delegate.sh check
+  claude_delegate.sh run <prompt-file> [result-json]
+  claude_delegate.sh resume <session-id> <prompt-file> [result-json]
+EOF
+  exit 2
+}
+
+command -v claude >/dev/null 2>&1 || die "claude CLI is not installed or not on PATH"
+
+if [[ "${1:-}" == "check" ]]; then
+  claude --version
+  auth_status="$(claude auth status 2>/dev/null)" || die "unable to read Claude authentication status"
+  [[ "$auth_status" == *'"loggedIn": true'* ]] || die "Claude CLI is not authenticated"
+  printf 'authentication: logged in\n'
+  exit 0
+fi
+
+mode="${1:-}"
+case "$mode" in
+  run)
+    [[ $# -ge 2 && $# -le 3 ]] || usage
+    prompt_file="$2"
+    result_file="${3:-}"
+    resume_id=""
+    ;;
+  resume)
+    [[ $# -ge 3 && $# -le 4 ]] || usage
+    resume_id="$2"
+    prompt_file="$3"
+    result_file="${4:-}"
+    [[ "$resume_id" =~ ^[0-9a-fA-F-]{36}$ ]] || die "session id must be a UUID"
+    ;;
+  *) usage ;;
+esac
+
+[[ -f "$prompt_file" ]] || die "prompt file does not exist: $prompt_file"
+[[ -r "$prompt_file" ]] || die "prompt file is not readable: $prompt_file"
+
+prompt="ROLE: EXECUTOR — do the work yourself; do not spawn subagents. Stay within the work order's authorization and scope. At the end, report changed files or findings, commands and checks run with results, and any blockers. WORK ORDER: $(<"$prompt_file")"
+
+effort="${CLAUDE_DELEGATE_EFFORT:-medium}"
+permission_mode="${CLAUDE_DELEGATE_PERMISSION_MODE:-auto}"
+case "$effort" in low|medium|high|xhigh|max) ;; *) die "invalid effort: $effort" ;; esac
+case "$permission_mode" in acceptEdits|auto|manual|dontAsk|plan|bypassPermissions) ;;
+  *) die "invalid permission mode: $permission_mode" ;;
+esac
+
+model="${CLAUDE_DELEGATE_MODEL:-sonnet}"
+args=(--print --output-format json --model "$model" --effort "$effort" --permission-mode "$permission_mode" --setting-sources project,local --strict-mcp-config --no-chrome)
+[[ -n "${CLAUDE_DELEGATE_MAX_BUDGET_USD:-}" ]] && args+=(--max-budget-usd "$CLAUDE_DELEGATE_MAX_BUDGET_USD")
+[[ -n "$resume_id" ]] && args+=(--resume "$resume_id")
+
+if [[ -n "$result_file" ]]; then
+  result_dir="$(dirname "$result_file")"
+  [[ -d "$result_dir" ]] || die "result directory does not exist: $result_dir"
+  claude "${args[@]}" "$prompt" | tee "$result_file"
+else
+  claude "${args[@]}" "$prompt"
+fi