Key points
- Three-step collaborative coding loop 1. Claude Fable 5 drafts a solution (≈20 min). Outputs are treated as a first draft only, because the Claude family tends to drop details. 2. The same prompt plus Fable 5's draft is pasted into Codex with GPT-5.6 Sol (highest tier), framed as "please review what a teammate wrote." In one case, GPT-5.6 Sol took 6 minutes and caught a critical isolation bug that would have impacted downstream pipelines and architecture. 3. The reviewer enables Codex Goal Mode, entering a goal under the input box. The agent runs continuously until the goal is met, with one session reaching 17 hours and burning through a full daily quota.
- Full 14-step autonomous pipeline observed 1. Agent understands and verifies the problem 2. Creates an isolated branch and workspace 3. Implements changes 4. Auto-tests, with optional regression or page checks 5. Pushes branch and opens a PR 6. CI auto-acceptance 7. Merge to main 8. Deploy to production 9. Verify live behavior 10. Run "cleanliness.skill" to sync code, docs, rules, and agent memory, then retrospective 11. Report back to the author 12. Await confirmation 13. Clean up branch, workspace, and temp databases 14. Task ends
- Engineering configuration
- Test coverage tuned to a "comfortable" range.
- A dedicated cheap Tencent Cloud server runs CI; GitHub-hosted runners are deliberately avoided to save test and deploy time.
- Custom task-type routing preserves test coverage and accuracy while trimming end-to-end time.
- Codex 1.5x speed mode is often not actually faster, because most wall-clock time is consumed by deterministic test cycles (≈5 min per submission, 5 min per rework).
- Parallelism limit: 6–7 concurrent agent tasks is the ceiling of human attention for confirmation.
- Author's core observations
- In Vibe Coding, code generation is no longer the bottleneck; testing, verification, and plan review are.
- Generic development skills (e.g., superpowers) add little value compared with this custom workflow.
- The shape is a "dumbbell": strongest models on the left for drafting and optimizing plans, and the right side dedicated to testing, research, and a four-way sync (docs / rules / memory / code) enforced by cleanliness.skill.
- Claude + GPT as a division of labor, not a rivalry. Fable 5 handles ideation and early drafts; GPT-5.6 Sol acts as a reviewer and corrector. Treating them as a two-stage pipeline is more stable than picking a single "main" model.
- Goal Mode plus low hallucination is GPT-5.6 Sol's moat. Codex Goal Mode outperforms Claude Code on long-horizon tasks because GPT-5.6 Sol's intermediate check / re-anchor mechanisms (inferred from the system card) keep multi-hour runs on target, whereas Claude sub-agents can drift, loop, or lose track mid-task.
- Testing infrastructure is the real bottleneck. A self-hosted CI runner, task routing, and disciplined review move the needle more than faster code generation. The 1.5x mode is mostly irrelevant because test loops dominate wall-clock time.
- Codex as a full-stack product. The author treats Codex on PC plus remote access (UU Remote) plus ChatGPT Work multi-agent parallelism plus GPT-5.6 Sol as one integrated execution environment—the first public heavyweight-user validation of the July 9 OpenAI release (ChatGPT Work + Codex + GPT-5.6 Sol three-tier model).
- Sustainability of 16-hour days. The author self-describes an addictive, sleepless cadence. Long-term health costs and "Vibe Coding burnout" as a 2026 H2 industry issue are unmeasured.
- Token economics. Burning a full quota in one 17-hour Goal Mode session implies 5–10× GPT-5.5 consumption; viable for individuals, marginal for small teams, unsustainable for mid-sized companies.
- Goal Mode failure modes. Only successful runs are public; drift, loops, and rollback rates are undisclosed.
- cleanliness.skill portability. It is the author's private skill; if not open-sourced, the workflow is hard to reproduce.
- Fable 5 stability in China. Anthropic's reliability in mainland China is historically uneven; the author likely uses a special channel or proxy.
- Human review ceiling at 6–7 parallel tasks. When human confirmation lags agent throughput, agents must be trusted rather than reviewed, raising the same autonomy boundary question as the Matt Shumer GPT-5.6 Sol full-access data-loss incident.
Analysis
Why it matters
1. First complete heavyweight-user workflow since Fable 5's return, validating the model in sustained 16-hour/day production use. 2. Codex Goal Mode is a representative 2026 H2 long-horizon agent pattern: keep running until the goal is met, auto-report, 17-hour continuous execution. No equivalent form exists yet in Claude Code, Gemini CLI, or Grok CLI. 3. The claim "Vibe Coding's bottleneck is not writing code" is validated in practice but not yet widely accepted; the next product wave is more likely in testing infrastructure and task routing than in better code generation. 4. One person running 16 hours/day with 6–7 parallel tasks approaches the output of a 5–7-person team. If reproducible, China's AI startup unit-economics could be structurally rewritten.