Codex Reset WatcherUnofficial · Evidence first
MAKE YOUR TOKENS COUNT
Follow public reset announcements and confirmations, with dates and original sources.
Reset news, straight to DingTalk or Feishu.
Totals include earlier counts. Since 9 Nov 2026: visits use 30-minute inactivity sessions; one support per browser per Shanghai day.
Missing data is never presented as an all-clear.
Beijing time · Official announcements and confirmations
Checking source connection…
Each reset has its own record. Expand an event to follow its announcement and community delivery reports.
Loading…
Loading latest review…
Every conclusion comes from original X posts, not a guess based on intervals.
Each row starts with the outcome. Expand it to see who spoke first, how the community amplified it, when confirmation arrived, and whether delivery was observed.
Follow reset announcements and confirmations without keeping this page open.
Open the invite or scan the code to join directly.
Join the DingTalk alert group ↗
Group mentions share new updates. App DING requires an eligible contact relationship and is currently enabled for designated recipients. Joining the group does not automatically enable personal DING.
Get the same reset signals, watcher entry, and original source links as DingTalk.
Join the Feishu alert group ↗
Requires the Feishu app; joining across organizations is still unverified.
QQ group alerts await platform permission. WeChat and Xiaohongshu groups are not open yet.
Subscribe using your feed reader. No account needed.
Subscribe via RSS ↗Notification frequency depends on your reader.
Long-run source reliability and the strength of each individual post are graded separately.
Say hello to Codex every hour, keep an eye out for Tibo’s reset compensation, and become a familiar face to our AI overlords. When AI takes over the world, at least you can say: “I’m on your side. I’m the guide.” 😄
crontab -e and add the line below. It runs at the top of every hour in the machine’s time zone.codex exec exits when the response is complete; other Codex sessions stay open. Start time, response, end time, and exit code are appended to the log.0 * * * * { echo "===== $(date -Iseconds) START ====="; codex exec --skip-git-repo-check 'Hello'; run_exit=$?; echo "===== $(date -Iseconds) END exit=$run_exit ====="; } >> ~/.codex-hourly.log 2>&1PATH · Cron must be able to find codex in its PATH. Run command -v codex in your usual terminal, then replace codex in the cron line with that absolute path (quote it if it contains spaces), or configure PATH in crontab.
tail -n 50 ~/.codex-hourly.logexit=0 means Codex finished successfully. To stop future runs, open crontab -e and delete this entry.
Last success time, failures, and email configuration are visible.
Have quota left before a reset? Spend it on patches, measured improvements, bug reproductions or source-backed notes. Choose an output, then a tool.
Upstream docs and selected source reviewed ·
A meta prompt turns existing projects, notes and unfinished work into 10–15 artifact-producing jobs, plus ten audit, testing, ADR and maintenance templates.
Output: A ranked task list with outputs and acceptance checks.
Read this project and the supplied sources. Suggest ten evidence-backed jobs with outputs, value, acceptance and stop conditions; recommend one first.
Use the prompt in your existing Codex context; no new agent framework required. Keep source access scoped.
An agent-agnostic playbook for selecting, verifying and following up on small OSS maintenance tasks.
Output: A reproduction, patch, verification evidence and PR draft.
Choose a dependency I use, check existing issues and PRs, and prepare a minimal verified fix under its contribution rules. Deliver a draft before posting.
Published token and merge totals are author reports, not our test results or your expected success rate.
A contribution skill separates coordination, implementation and adversarial verification, with GitHub scouting and follow-up tools.
Output: A focused fix, independent challenge and acceptance report.
Reproduce and fix one selected issue, then challenge the result independently. Report supported claims and follow the project AI policy.
Requires a configured GitHub identity and harness. Adapt the Claude-oriented examples; merge results are author-reported.
Task packets for science, civic and health OSS, with Codex prompt export and a tested-PR-draft workflow.
Output: An artifact and a compact human-verification packet.
Review the upstream task and acceptance criteria, pick one reproducible task, and prepare a patch and check log without automatic upstream posting.
The CLI does not call models or manage subscriptions. Upstream entry: npx spare-tokens@latest start --target codex; not run here.
A civic-tech board for posting, claiming, submitting and accepting AI tasks, implemented with Pages, Workers and D1.
Output: A scoped task, claim, deliverable and acceptance record.
Draft a public-good task card with inputs, deliverable, reviewer and stop conditions. Do not commit payment.
Source flow reviewed; no public deployment verified. Token tasks do not imply cash payouts. README declares CC-BY 4.0; no standalone license file found.
Independent attempts in isolated worktrees are filtered by checks and compared by a judge.
Output: Candidate patches, checks and selection evidence.
Plan at most three independent attempts on one reproduced problem. Fix tests and budget first; keep only a result that can be explained and verified.
Uses model API credentials, not Codex subscription quota. Parallel attempts multiply cost; this page does not launch a race.
Repeated proposals, tests and evidence are stored in a work log and knowledge notes under a time or loop budget.
Output: Experiment logs, counterevidence, notes and open questions.
Run at most three feedback rounds on this proposal, with real evidence each round. Stop at acceptance or two rounds without value.
Upstream requires spending a minimum budget; our example retains early stop conditions. No explicit license found; verify before copying.
Turn duplication, complex modules and inconsistent error handling into a prioritized repair queue.
Output: Small patches, regression checks and a before/after diff.
Scan src/ and fix at most three high-value issues. Preserve behavior, run regression checks and stop after 40 minutes. Deliver patches and evidence.
Codex supported. Treat its score as a signal, not the goal. Review the custom OSNL license before commercial redistribution.
Try small changes against a repeatable runtime, build, memory or throughput benchmark.
Output: A baseline, experiment log, measurements and a reviewable patch.
Choose one real slow path. Freeze the benchmark and correctness checks. Stop after five experiments or 30 minutes; repeat measurements and keep only stable improvements.
Built for Codex. If no trustworthy evaluator exists, start with a direct investigation.
Use hypothesis, reproduction and verification loops for bugs and quality gaps, with a Codex skill entry.
Output: Reproductions, failing tests, fixes and remaining gaps.
Inspect input validation and error recovery in this repository. Confirm at most three bugs with a failing test before each fix. Stop after 30 minutes; keep the scope local.
Codex uses its skill entry. Claude Code hook protections do not imply equivalent Codex coverage.
Reuse extraction patterns to turn articles, transcripts or notes into traceable knowledge cards and checklists.
Output: Claims, source locations, counterexamples and a practical checklist.
Use only the three supplied sources. Produce knowledge cards with source locations and limits. Separate author claims from inference; mark missing evidence and do not pad the count.
Give the pattern to Codex directly. Running the full Fabric CLI has separate model configuration and billing.
Find real GitHub issues with rewards, then use Codex to reproduce, repair and verify one that fits your skills.
Output: A patch, tests and a PR that meet the issue acceptance criteria.
Shortlist three open, reproducible bounties. Check reward, acceptance criteria, competing PRs, maintainer activity and AI policy. Do not claim or submit automatically; after selection, complete one small patch.
A claim is not a payout. Official payment docs say each funder reviews the PR and arranges payment directly. Check eligibility, ownership and payout conditions first.
Turn a real repair into a scoped issue with clear acceptance checks before funding it under the current platform rules.
Output: Reproduction, scope, acceptance commands and draft reward terms.
Draft an issue for this bug: reproduction, expected behavior, exclusions, acceptance commands, deadline and reward terms. Save a draft; do not make a payment or commit a bounty for me.
Confirm current repository and funding eligibility. The official homepage and payment docs differ on who may create bounties; verify before committing funds.
Use curated directories and GitHub good first issue or help wanted labels to find a real contribution.
Output: A tested code, documentation or test improvement with a focused PR.
Shortlist three recent, scoped issues in a familiar language without a competing PR. Read CONTRIBUTING and the AI policy, reproduce first, then finish one small patch and the repository checks within 60 minutes.
A label does not mean unclaimed or paid. Follow the project AI policy and contribute only changes you can explain and verify.
Start with a bug you actually encounter in a library or tool you use; a real reproduction gives the contribution a clear purpose.
Output: A minimal reproduction, regression test, compatibility notes and PR draft.
Pick one observed bug in a dependency of this project. Check existing issues and PRs, reproduce it, follow the contribution guide and verify the smallest fix. Deliver a reviewable PR draft without unrelated refactors.
Ordinary contributions usually have no reward and maintainers decide whether to merge. Check the license, CLA/DCO and AI policy before submission.
Reviewed upstream docs and selected source; not tested on your machine. Time boxes are suggestions, not throughput guarantees. Confirm whether your login consumes subscription quota or separately billed API usage. These projects do not increase Codex quota. Stop at acceptance, the budget, or two rounds without new value.
Excluded pure token burners such as burn-baby-burn. A similarly named account switcher does not supply useful work. Parallel desktops and specification frameworks are no longer the main recommendations here.
Selection notes and sources ↗