codex-cto
/codex-cto
Odwraca orkiestrację: Codex CLI pełni rolę CTO i planuje oraz przegląda, a Claude Code wykonuje pracę natywnymi narzędziami, co daje zewnętrzny nadzór architektoniczny.
- Autor
- tdimino ↗
- Repozytorium
- claude-code-minoan ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 19.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 6
- Skrypty
- 1
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 19.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 1 skrypt w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/cto-invoke.sh
Co robi i kiedy
Odwraca orkiestrację: Codex CLI pełni rolę CTO i planuje oraz przegląda, a Claude Code wykonuje pracę natywnymi narzędziami, co daje zewnętrzny nadzór architektoniczny.
Argumenty
/codex-cto [objective]
Oryginalny opis (EN)
Inverted orchestration—Codex CLI acts as CTO (planning and reviewing), Claude Code executes with native tools. Use when a task benefits from external architectural oversight before and after implementation.
Jak działa
- Protocol
- Step 1: Get the Plan
- Step 2: Execute Tasks
- Step 3: Submit for Review
- Step 4: Iterate (if revise)
- Model Selection
- Options
- Artifacts
- Diagnostics
- Examples
Podgląd SKILL.md
---
name: codex-cto
description: Inverted orchestration—Codex CLI acts as CTO (planning and reviewing), Claude Code executes with native tools. Use when a task benefits from external architectural oversight before and after implementation.
argument-hint: [objective]
disable-model-invocation: true
---
# Codex CTO — Inverted Orchestration
Codex plans and reviews. Claude Code executes. This inverts the `codex-orchestrator` pattern: instead of Claude Code delegating to Codex subagents, Codex directs Claude Code as the engineering team.
## Protocol
```
1. CTO plans → cto-invoke.sh plan "<objective>"
2. You execute → native tools (Read, Edit, Write, Bash, Glob, Grep)
3. CTO reviews → cto-invoke.sh review "<objective + results>"
4. Loop → if "revise", fix and re-submit for review (max 5 iterations)
```
## Step 1: Get the Plan
Run the plan phase with the user's objective:
```bash
~/.claude/skills/codex-cto/scripts/cto-invoke.sh plan "<OBJECTIVE>" [--model MODEL]
```
Parse the returned JSON. It contains:
- `objective` — what to accomplish
- `analysis` — CTO's assessment of the codebase
- `strategy` — high-level approach
- `tasks[]` — ordered task list with IDs, actions, targets, descriptions, acceptance criteria, and dependencies
- `verification` — commands to run after all tasks complete
## Step 2: Execute Tasks
Work through each task in dependency order using native tools. For each task:
1. Read the `description` and `acceptance_criteria`
2. Execute the action (`create_file`, `modify_file`, `delete_file`, `run_command`, `verify`)
3. Verify each acceptance criterion before moving to the next task
4. Track what was done and what the outcome was
Do not deviate from the plan without reason. If a task is blocked or the plan has an error, note it and continue with unblocked tasks. Report blockers in the review submission.
## Step 3: Submit for Review
After executing all tasks, run the review phase. Include the original objective and a summary of what was done:
```bash
~/.claude/skills/codex-cto/scripts/cto-invoke.sh review "<OBJECTIVE>. Results: <summary of what was done, any issues encountered>"
```
Parse the returned JSON:
- `verdict: "approve"` — done. Report success to the user.
- `verdict: "revise"` — check `task_reviews` for feedback and `revised_tasks` for new/updated tasks. Fix issues and re-submit.
- `verdict: "abort"` — fundamental problem. Report the CTO's `reason` to the user and stop.
## Step 4: Iterate (if revise)
On a "revise" verdict:
1. Read each `task_reviews` entry with status `fail` or `partial`
2. Execute the fixes described in `feedback` and any `revised_tasks`
3. Re-submit for review (step 3)
4. Maximum 5 review iterations. If the CTO issues 3 consecutive "revise" verdicts, escalate to the user.
## Model Selection
The `--model` flag passes through to `codex exec`. Available models:
| Model | ID | Notes |
|-------|----|-------|
| GPT-5.4 Pro | `gpt-5.4-pro` | Default for plan phase. Deepest reasoning for architectural decomposition. |
| GPT-5.4 | `gpt-5.4` | Default for review phase. Unified coding + reasoning flagship. |
| GPT-5 Mini | `gpt-5-mini` | Cost-optimized, fast iteration. |
| GPT-5.3-Codex | `gpt-5.3-codex` | Previous generation. Still functional. |
Omit `--model` to use the default from `~/.codex/config.toml`.
## Options
| Option | Description |
|--------|-------------|
| `--model <id>` | Override model (default: config default) |
| `--dry-run` | Print command without executing |
## Artifacts
Plan and review JSON files are saved to `.codex-cto/<PID>/` in the working directory. Each invocation creates sequentially numbered files (`plan-001.json`, `review-001.json`, etc.). PID namespacing prevents conflicts between concurrent sessions.
## Diagnostics
Reuse codex-orchestrator's status script for CLI diagnostics:
```bash
~/.claude/skills/codex-orchestrator/scripts/codex-status.sh
```
## Examples
```bash
# Simple feature
/codex-cto Add a health check endpoint at /health returning {status: "ok"}
# With model override
/codex-cto Refactor the auth module to use JWT --model gpt-5-mini
# Dry run to inspect the command
~/.claude/skills/codex-cto/scripts/cto-invoke.sh plan "Add rate limiting" --dry-run
```
## Sister Skill
This skill is the inverse of `codex-orchestrator`. Use `codex-orchestrator` when Claude Code should delegate one-shot specialist tasks to Codex. Use `codex-cto` when Codex should direct Claude Code through a plan-execute-review loop.
Pliki w paczce (6)
- README.md 5 KB
- SKILL.md 4 KB
- agents/cto.md 3 KB
- scripts/cto-invoke.sh 4 KB
- templates/plan-schema.json 1 KB
- templates/review-schema.json 1 KB
Wersje
W bazie nie ma innych skilli o nazwie codex-cto.
Migawka bazy z 13.11.2026 · wersja paczki 01adcd5ee41f (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100review
Łączy Claude Code z lokalnym Codexem: Claude planuje i recenzuje, a Codex wdraża i daje niezależną drugą opinię. - 90Codex Integration
Deleguje złożone zadania kodowe do Codex CLI i koordynuje pracę między Claude a Codex. - 100orka
Prowadzi projekt programistyczny jako CTO: planuje pracę, pisze karty zadań, rozdziela je do agentów CLI w izolowanych worktree, a potem recenzuje, testuje, ocenia i scala. - 100cto-legends
Kieruje zwykłe zapytania do pasującego modułu Legends, pokazuje dostępne możliwości i aktualizuje moduły przez CLI. - 100yourteam
Uruchamia zespół wielu agentów AI, od CEO po QA, który buduje produkt etapami, z panelem w przeglądarce. - 95codex-orchestrator
Pełni rolę strategicznego orkiestratora: Claude planuje, a agenty Codex wykonują kod i zmiany w plikach, także w badaniach bazy kodu.