Ten fetched records are not ten pieces of news
An update assistant can produce a busy-looking brief simply by listing everything it just found. That mixes older releases, preview builds and items already waiting for review.
In an experiment on November 14, 2026, we collected five release records from each of the public Claude Code and Codex repositories. Four were prereleases. Of the six stable records, only one was published on the current date in Shanghai. A second live collection found no new metadata, while keeping that one review item in the queue.
The useful result was a small, sourced draft and a record of what had already been seen. We did not install an update, send a notification or enable a new schedule.
Define a small inbox before asking for a daily brief
This exercise monitors anthropics/claude-code and openai/codex. The audience is someone using coding agents for recurring work who wants a short list to review, not every preview build.
The rules are explicit: use Asia/Shanghai for the reporting day; keep stable releases in the main list; preserve the source URL and publication time; distinguish initial discovery from a subsequent change. The sample reads the first five records by creation time from each repository. Both responses say there are more pages, so this is a bounded inbox, not a complete daily news scan.
The work ran in this Codex conversation with Python and an existing GitHub CLI login. It is not a Claude Code or Onevium runtime test. Download the actual snapshots, scripts and drafts. Saved snapshots can be replayed without a network connection.
The first revision was to ask for less data
The initial REST request for twenty Codex releases timed out after receiving part of its response. A five-record request also timed out. We did not parse those fragments into an empty list or claim there were no updates.
Release responses can include installation-asset information that this brief does not need. We changed the collection to a GraphQL query selecting just the release ID, tag, prerelease and draft flags, publication time, URL and pagination flag. GitHub documents that GraphQL queries return the selected fields; the successful projected response in this task was 2,121 bytes. This is an observed response size, not a general speed benchmark.
The collector uses an existing authenticated gh session and only the two named public repositories. It refuses GraphQL errors, missing repositories and invalid records before saving a snapshot. If authentication or networking fails, fix that access problem; do not replace the previous state with an empty result.
Publication time and discovery time answer different questions
The v2.1.325 record has publication time 2026-11-13T20:40:01Z. In Shanghai, that is November 14 at 04:40:01. Filtering on the first ten characters of the UTC timestamp would put it on the wrong local day.
The first successful collection was at 10:06:10 Shanghai time on November 14. That is when this task observed the record, not when the vendor released it. The Codex stable record in the same sample, rust-v0.169.0, has a November 13 local publication date; discovering it today does not make it today’s release.
| Field or rule | What it decides |
|---|---|
| publishedAt converted to Asia/Shanghai | Which local publication day a record belongs to |
| observed_at | When this collector saw the response |
| Repository plus release ID | Whether this is the same tracked record |
| isPrerelease | Whether it belongs in this stable-only inbox |
| hasNextPage | Whether there is more history outside the sample |
The list is ordered by creation time, while the brief filters publication time. That difference is another reason not to promise full-day coverage from five records. The GitHub release schema supplies the fields; the reporting policy is our own.
Keep seen records separate from pending work
The first run establishes a baseline. It records all ten identities and creates one pending review item. It does not announce ten newly released versions.
We then fetched the same two sources again at 10:07:04 and compared against the saved state. There were zero new or changed selected metadata records, zero newly queued items and one retained pending item. The earlier draft did not disappear merely because it had been seen.
This sample fingerprints only the selected metadata. It does not detect edits to the release-note body. It also does not implement delivery acknowledgements or automatic removal after approval. A record falling out of the five-item window is not evidence that it was deleted. These are specific limits to address before turning the exercise into an unattended service.
Read the candidate and write a useful next step
The metadata brief initially contains only a candidate and a link. We then read the full Claude Code v2.1.325 release note and wrote a separate short draft. Its practical focus is confirmation behavior around scheduled Skills and sandboxed commands, plus a reported session-resume fix.
The proposed next step is to check whether an already-authorized recurring task now waits for confirmation, and whether resumed work retains its recent context. That is a task to verify, not a claim that we upgraded and tested the product. It is also not a recommendation to relax all permissions.
The working pack includes this revised draft in both languages. No notification was sent. For your own brief, ask the assistant for four things: what the source says changed, who should care, one concrete check to run, and what remains unverified. Keep vendor claims and your test results visibly separate.
Replay the result, then collect another snapshot
Unzip the pack and work inside release-brief/. Python 3.10 or later is sufficient for the offline replay; it uses the system’s IANA timezone data. Use new output directory names:
python3 brief.py snapshot-first.json --out my-first
python3 brief.py snapshot-second.json --previous my-first/state.json --out my-second
Compare the six generated files per run with first-run/ and second-run/. summary.json contains the counts; state.json keeps seen identities and the pending queue; today.json lists the local-day candidates; delta.json records newly queued items; the two Markdown files are the metadata briefs.
To fetch fresh metadata, first confirm that your existing GitHub CLI login can read the two public repositories:
python3 collect.py --out fresh-snapshot.json
python3 brief.py fresh-snapshot.json --previous my-second/state.json --out my-next
Live results will change. The checked-in snapshots preserve this particular run; repeating the Python export does not repeat the AI’s reading or editorial judgment.
Check the quiet run before adding a schedule
We also exercised three explicitly injected invalid inputs: a missing source, a duplicate release identity and a timestamp without a timezone. All were rejected. These are labelled validation controls, separate from the real network timeouts. A failed check must leave the earlier state available.
Before scheduling a workflow like this, decide how far to paginate, how to catch up after missed runs, how to track edited notes, and when an item moves from pending to approved or delivered. Start with a local draft you can inspect. The scheduled-runs guide covers the corresponding Onevium entry point; this exercise has not installed a schedule.
A quiet second run can be a correct result. It should mean that the checked window has no newly observed candidates, while still showing its coverage and outstanding review work.