TechRadar collects feeds throughout the week. A separate, single codex exec run
turns the stored titles/excerpts into a weekly briefing. There is no LiteLLM,
OpenAI API key, provider proxy, or API SDK in this runner. It uses the CLI's saved
ChatGPT login and is subject to your plan's usage limits and model availability.
The source records are sent to OpenAI for summarization; this is not local inference.
This does not replace the per-article HTTP client: disable ai_analysis and
ai_translation, and use filter.method: keyword in the collector configuration
if you want Codex to be the only AI component. Keep RSS collection enabled and
storage.local.retention_days: 0. Collect daily, not just on briefing day.
Official references:
Requires Python 3.12+ (standard library only) and Codex CLI supporting
--ignore-user-config, --ignore-rules, --ephemeral and --output-last-message.
The invocation flags were checked against Codex CLI 0.163.0. No real model request
was made during development.
Create a dedicated authentication directory and sign in interactively:
mkdir -p "$HOME/.techradar-codex"
CODEX_HOME="$HOME/.techradar-codex" codex -c 'cli_auth_credentials_store="file"' login --device-auth
CODEX_HOME="$HOME/.techradar-codex" codex login statusChoose ChatGPT/subscription authentication, not an API key. Keep that directory private and outside the repository; never commit its contents. Use a dedicated OS account/container without unrelated secrets or personal files. The read-only sandbox limits writes; it is not a substitute for filesystem isolation.
From the repository root, first check what will be included without calling Codex:
python3 scripts/weekly_codex.py --data-dir output --dry-runGenerate one edition (yesterday plus the six preceding days, Europe/Zurich):
python3 scripts/weekly_codex.py --data-dir output \
--codex-home "$HOME/.techradar-codex"Use --end-date 2026-10-18 to choose an inclusive collection window ending on a
specific date. --model MODEL_ID is optional; otherwise Codex selects its default.
--interests "..." customizes the briefing. --max-items defaults to 150 newest
observed unique URLs, with 500 characters of title and 1,500 of excerpt per item.
The count of omitted items is recorded. It is not a relevance-ranking algorithm.
Each edition contains:
output/weekly/START_END/
briefing.md # Briefing with source references
sources.json # Exact supplied records, database paths and collection observations
prompt.txt # Exact prompt for this edition
The script opens collection databases read-only, deduplicates identical URLs across daily databases, and preserves each source observation. Related stories with different URLs can be grouped by Codex. Publication dates can be older than the collection window. Feed summaries may be incomplete; no full-text crawling or independent verification is performed.
Existing editions are never overwritten. An empty data window, failed CLI call,
timeout, empty response or missing/unknown citation IDs produces an error and no
published edition. Citation-ID validation does not prove that every claim is
supported. No automatic retries are made against your subscription. To regenerate
an edition deliberately, select another --archive-dir.
The additional image runs only Python and Codex, independently of the collector. Build this checkout on the ZimaBoard (or build amd64 elsewhere and transfer it):
docker compose -f docker/docker-compose-codex.yml build weeklyIf weekly-reports was created by root, grant the container's node user (UID 1000)
write access to that directory only. Ensure the collection output directory is
readable by that user. Adjust bind-mount paths in the Compose file to match your
ZimaOS persistent storage; /data must point at the collector's output folder.
Authenticate into the dedicated named volume:
docker compose -f docker/docker-compose-codex.yml run --rm --entrypoint codex weekly \
-c 'cli_auth_credentials_store="file"' login --device-authFollow the displayed sign-in instructions in your browser. Then run a dry check:
docker compose -f docker/docker-compose-codex.yml run --rm weekly \
--data-dir /data --archive-dir /reports --codex-home /auth --dry-runGenerate the briefing by running the same command without --dry-run:
docker compose -f docker/docker-compose-codex.yml run --rm weeklyThe image uses pinned Codex CLI 0.163.0; check release/security information before upgrading. The image build and actual ChatGPT login/request still need validation on your deployment host. If container sandbox initialization fails, investigate that host's sandbox support rather than bypassing it.
For scheduling, use your host's persistent task scheduler. Run the command once every Monday at 09:00 Europe/Zurich, which covers the previous Monday–Sunday. For a Linux host running cron in that timezone, an example is:
0 9 * * 1 cd /ABSOLUTE/PATH/TO/TechRadar && /usr/bin/docker compose -f docker/docker-compose-codex.yml run --rm -T weekly >> /ABSOLUTE/PATH/TO/techradar-weekly.log 2>&1Use the actual Docker path from command -v docker. Verify whether your ZimaOS
scheduler persists custom jobs across updates before relying on host cron. The
Compose app itself is a one-shot runner, not an always-running scheduler. Keep
collection running separately. Subscription expiry/rate limits cause a failed
job; review the scheduler log and authenticate again if needed.
Prompts travel over stdin, not command-line arguments. The CLI runs in a fresh empty directory with read-only sandboxing, no approvals, no loaded user config or rules, web search disabled and the shell tool disabled. Only basic runtime env variables and the dedicated CODEX_HOME are inherited, not API/cloud key variables. Raw CLI output/errors are not copied into the briefing or logs. Timeouts terminate the CLI process group on POSIX. Isolation on the deployment host remains essential.
Offline tests (no subscription use):
python3 -m unittest discover -s tests -p test_weekly_codex.py -v