mirror of
https://github.com/jjbrunton/Sauci.git
synced 2026-10-31 18:19:43 +00:00
jjbrunton
9957fa43f2
Add a repository-owned Codex skill and helper for bounded Claude Code execution, model and effort selection, resumable follow-ups, and independent acceptance checks.
74 lines
3.9 KiB
Markdown
74 lines
3.9 KiB
Markdown
---
|
|
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.
|