One Codex task ran for more than 31 hours, according to author and operator kfinds, and continued through application and computer restarts. Its useful achievement was not simply lasting that long: it produced a linked teaching package whose state and checks could be inspected in files. The account is a sanitized, author-documented case study—not a benchmark, independent replication, or claim of error-free work.
What the 31-hour Codex task produced
Kfinds describes a Windows workflow labeled as using GPT-5.6 Sol. A single parent task turned an original mathematical framework into a versioned teaching package, with Codex doing long-horizon construction, implementation, verification, auditing, and documentation. The author retained authority over the definitions, permitted transformations, research direction, and acceptance criteria.
As an Amazon Associate I earn from qualifying purchases.
The deliverable was a connected set of artifacts rather than one textbook file:
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →- Source or master chapters, student chapters, and teacher chapters.
- Conventional hand-work exercises and a separate AI-assisted exercise track.
- Worked answers organized by difficulty.
- Five deterministic exercise and data generators, alongside reproducibility scripts and datasets.
- Manifests, hashes, audit reports, execution records, and restart checkpoints.
The AI-assisted exercises were curricular material: learners use an AI system to generate, inspect, test, or audit mathematical objects under stated rules. In the project map described by kfinds, theory feeds the student and teacher texts; manual and AI-assisted exercises connect to worked answers; generators produce datasets and summaries that can be rerun and hash-checked; and execution records support manifests, audits, and checkpoints.
How continuity was documented
The central distinction in the account is between remembering a conversation and retaining inspectable project state. Kfinds says the task resumed after restarts from verified artifacts and checkpoints, rather than depending only on a reconstructed narrative of what had happened. That is a reported explanation from this one workflow, not a controlled test showing that artifacts outperform conversation history in every project.
The author identifies seven practices associated with continuity and auditability:
- Version project state as artifacts. Keep the evolving work in files that can be inspected and compared.
- Set definitions and authority boundaries before execution. Specify what the agent may change and which decisions remain with the human.
- Write acceptance criteria and invariants explicitly. State what must be true for an output to count as complete.
- Turn qualitative requirements into executable checks. Have generators and scripts test requirements where possible.
- Check student and teacher outputs against one another. Treat consistency between related deliverables as something to verify, not assume.
- Preserve negative results and rejected paths. Keep evidence of what did not work rather than allowing it to vanish from project history.
- Resume from verified artifacts. Re-establish progress from checked files and records, not a guess based on a long chat narrative.
OpenAI’s Goals guide describes a durable, thread-scoped objective in terms of an outcome, verification surface, constraints, boundaries, iteration policy, and a condition for stopping when blocked. That guidance aligns with the case’s emphasis on clear finish criteria; it does not establish how the particular Windows task recovered after restarts.
What the reported records show—and do not show
The figures below are attributed to kfinds’ recovered local execution records and artifact audits in 2026. They are author-reported details, not independent benchmark measurements.
| Reported item | What the account says |
|---|---|
| Task duration | More than 31 hours for one parent task. |
| Recovered local session log | Approximately 338.55 MiB. |
| Context compression | 47 events. |
| Associated records | 21,991. |
| Recorded UI checkpoint | 512 changed files, with +84,401 and -720 lines. This is a checkpoint snapshot, not a final project total. |
| Experiment checks | 67/67 reported checks passed. |
| Student–teacher cross-checks | 36/36 reported checks passed. |
| Reruns and validation | Deterministic reruns at 100,000 and 250,000 events, plus a separate five-million-row validation. |
These counts show the kinds of checks the author says were run, but the public account does not establish independent reproduction of the metrics. Nor does a set of passing checks prove that a large mathematical teaching package contains no errors: kfinds explicitly disclaims “zero errors.”
Why this is not proof that long-running agents are reliable
The evidence concerns one personally operated workflow. It does not compare competing methods, establish that another project would survive restarts the same way, or support a hardware recommendation. The computer hardware and Windows build are not specified. The task is labeled GPT-5.6 Sol, but that alone does not establish general performance or reliability for a model or product version.
The public deposit is sanitized and limited to timestamped execution evidence and Chapter 1 artifacts. The foundational corpus and full internal research protocol are withheld, and private chain-of-thought is not part of the deposit. The article and its technical discussion therefore offer a useful account of an auditable workflow, but not enough material for a complete independent reproduction.
OpenAI’s Codex Cloud help page describes a separate cloud workflow: tasks run on OpenAI-managed computers and can continue while a user’s own computer is asleep. It gives a default saved-VM recovery window of up to seven days after the last turn start or task resume, and distinguishes saved VM recovery from conversation-history retention. Those cloud details should not be conflated with kfinds’ local Windows account or its artifact-based recovery.
OpenAI’s account of its internal Codex controls says sandbox boundaries govern where Codex can write and whether it can reach the network, approval policy determines when it must ask, and agent-native telemetry helps people understand and audit actions. This provides general context for governance; it is not evidence about the specific run.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.The reported weak point: observability
Kfinds calls observability the weakest product-level part of the experience. Very long histories were difficult to navigate and export as one official report. After a restart, much of the client-visible history had disappeared, although local session data and produced artifacts remained recoverable.
The author argues that native exports of checkpoints, token composition, file manifests, and audit events would make long-running work easier to inspect. That limitation matters to the case’s central lesson: artifact recovery helped preserve project state, but it did not make the full run easy to review from the application’s visible history.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
