X9 codex delegation
Claude Code only; never use in a Codex session. Use when substantial, well-scoped work should actually be delegated to Codex — «отдай в Codex», «делегируй Codex», «передай эту часть работы в Codex, чтобы распределить лимиты», "use Codex", implementation, migration, codebase analysis, logged-in browser work, UI verification, or a second-opinion review. Do not use for checking or planning subscription quotas without a concrete task, trivial work, choosing between Claude models, or merely drafting a prompt/brief that is not explicitly being sent to Codex.From its SKILL.md
npx -y skills add xonika9/agent-skills --skill x9-codex-delegationAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.4 KB, 763 tokens by cl100k_base, as published. Nobody here has run it
Delegate from Claude Code to Codex
Use x9-browser-session for browser routing. Use the plugin route when it is available; the raw CLI remains a fallback for an explicitly requested delegation.
When Claude and OpenAI/Codex are backed by separate subscriptions or usage pools, delegation can distribute substantial work instead of exhausting only the Claude allowance. Treat that as a user-specific routing benefit, not a promise of lower cost: API billing, plan limits, and available runtimes vary.
Choose the route
- Implementation or fixes: use the installed
codex:codex-rescuesubagent when available. It is write-capable; the active project's local-change authority applies. - Read-only review or one-shot analysis: use the plugin route when callable, otherwise raw
codex exec. - Logged-in browsing: delegate the whole bounded browse-and-return task and follow the browser adapter. Shell read-only mode does not change the separately approved browser authority.
- Plugin commands marked user-only: report the exact command for the user; do not pretend it was invoked programmatically.
Raw fallback
Read references/codex-cli.md and use its one-shot or resume recipe after checking the current command help. That reference owns exact flags and known differences between exec and resume; this skill owns the delegation contract, not the volatile command surface. Prefer the model configured in ~/.codex/config.toml unless the user explicitly chooses another.
Brief contract
Pass a self-contained outcome, relevant files/sources, constraints, authority, deliverable, and completion bar. Add these failure rules when tools or current facts are load-bearing:
- Report failed or cancelled calls exactly; do not substitute memory.
- Surface an unresolved fork only when it materially changes the result.
- Attach evidence appropriate to the task: paths/diff/tests for code; URLs and exact facts for research.
Verify delegated work
For write-capable work, the delegate's report is advisory:
- Inspect
git statusand the full relevant diff; preserve unrelated changes. - Run focused verification in the orchestrator's environment.
- Confirm the result matches the brief rather than only the delegate's narration.
Stop retrying when the same blocking cause repeats, the budget is exhausted, or a fresh run cannot add evidence. Report the blocker instead of using a universal round count.
Record the delegation
Logging is off by default. Append to ~/.local/share/x9/codex-delegation-log.md only when the user explicitly opts in for the current task/session or an existing user-owned policy records that choice. Delegation must work normally without a log; do not ask again when the current scope already contains the answer.
When logging is enabled, read the current table header before writing and keep each row to minimal non-sensitive telemetry: route, actual model/effort when observable, wall-clock time, Codex tokens, whether rework was needed, and a generic success/failure outcome. Never record task content, secrets, repository names, or personal paths. Use н/д for unavailable data and ? when rework is not known yet; never invent telemetry. Do not replace an existing file or follow a symlink; report that logging was skipped instead.
The log is optional permanent operational history, not a completion gate. A missing or unwritable log must not fail the delegated task.
Done
- The actual Codex model/runtime is inherited or explicitly confirmed when load-bearing.
- Tool failures and degraded results are visible.
- Write results have orchestrator-owned diff review and verification evidence.
- When logging was enabled, the minimal telemetry row is recorded or the logging failure is visible.
What ships with it: 1 file
2.4 KB alongside SKILL.md
references/
- codex-cli.md2.4 KB
Gives 0 of the 12 instructions most plan spec skills give in 763 tokens
Counted across 1,360 of the 2,617 authors here whose files we hold, read 2026-09-06
- Ask one question at a timein 73 of 1360
- Write the spec using the templatein 22 of 1360
- Ask clarifying questions if neededin 19 of 1360, across 18 files
- Wait for user confirmation before proceedingin 19 of 1360
- Save plans to the plans directoryin 17 of 1360, across 13 files
- Check for product marketing context firstin 16 of 1360, across 5 files
- Read the plan file completelyin 16 of 1360
- Order tasks by dependencyin 16 of 1360
- Gather context from the conversationin 15 of 1360, across 9 files
- Explore the codebase instead of askingin 15 of 1360, across 13 files
- Wait for explicit user approvalin 14 of 1360, across 13 files
- Quiz the user on the breakdownin 13 of 1360, across 7 files
Said here and by no other author read
- Use x9-browser-session for browser routing
- Use codex-rescue for implementation or fixes
- Use the plugin route for read-only review
- Delegate the whole browse-and-return task
- Report failed or cancelled calls exactly
- Inspect git status and the full relevant diff
Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.