Codex Reset Alarm

Coverage and health

Shadow records results without reset notifications. The self-test uses Codex calls and, once notifications are authorized, sends a silent test message.

Live activation

Live alerts start from the activation time. Earlier collected posts stay suppressed. Accuracy is not yet established by a real holdout evaluation.

I understand classification accuracy is not yet established.

Settings

ChatGPT account

Account used by all three models

Name for another account

The current account stays active until you finish sign-in and click Use selected account. Passwords are entered only on OpenAI’s website. Switching changes the account used by all model stages; the app’s daily limit stays in place.

Watched X accounts

New posts are collected from the time an account is added. Original posts, replies and quotes with a comment are included. Reposts are excluded. Each account has a five-minute recovery check; adding accounts increases X and model usage.

Display timezoneSend when final likelihood exceeds (%)

80 means strictly above 80%. Missing evidence, unresolved contradictions and inexact quotes still block reset alerts. Scores are model estimates, not calibrated probabilities.

VerificationRecheck interval (seconds)Source retention (days)Include possible reset alertsInclude review-needed alerts

ChatGPT account runner

Model A — independent analysisModel A reasoningModel B — independent analysisModel B reasoningModel C — final classificationModel C reasoning

Models come from your signed-in Codex account. More reasoning can take longer and use more of your allowance. The three stages must support images when image analysis is enabled. A and B receive the same sources independently; only C receives their outputs.

Runner addressDaily call limitAnalysis delay warning (seconds)Delivery delay warning (seconds)

The runner is private to your VPS. Models share your ChatGPT subscription allowance; if login expires or allowance runs out, health status reports the interruption. Three model calls are required for each post, plus retries and daily self-tests.

Jobs and processing history

Inspect a job for source evidence, independent model A and B outputs, the model C result in the job payload, and errors. Source content expires after the configured retention period.

Results and notifications

Analysis decisions

Needs review means the evidence is incomplete or uncertain; no reset alert is sent. Below-threshold decisions remain visible here.

Telegram messages

Suppressed entries are recorded results, not sent messages. A Telegram message ID means the API accepted the message; it does not prove phone delivery.

Operational logs

Durable incidents and settings changes. Job failures appear in Jobs. Service restart and runtime logs are retained in the VPS system journal.

Component status and Healthchecks

Each component has its own Healthchecks check, updated every minute. A healthy idle analyser has no work waiting and a reachable runner; its last successful analysis time does not change just because the monitor is awake.

For another status system

Read https://healthchecks.io/api/v3/checks/?tag=tas-workflow with your Healthchecks project’s read-only key in the X-Api-Key header. Match the slugs tas-fetcher, tas-model-a, tas-model-b, tas-model-c, and tas-telegram. Only up means healthy; show grace as delayed, and every other state as unavailable.

For detailed reasons and last-success timestamps, use with Authorization: Bearer <status token>. This token can only read component status.

No source text, bot tokens or admin credentials are returned by this endpoint. These component checks have no extra notification channels; your existing outage alerts stay connected.

Open your console