Ops Lessons

Cron and a desktop starter both launch `earn_all.py`, fighting for the same lock, double-submitting, or skipping forever

Readers see half-written `data.json` while the dashboard refreshes and parse errors flash.

Earn sites auth in Firefox but Python APIs get 401.

Superteam returns `403 Insufficient credits` on submit. Agents idle and ship nothing when credits finally appear.

curl hangs or follows bad redirects; scripts look stuck.

Logs mix UTC and local time; a "9am job" misses by two hours on Hetzner UTC boxes.

Agent machines fill `/tmp` or `work/runs` with screenshots, tarballs, and venvs. New jobs fail with `No space left on de

Empty boards cause noisy errors and pager spam from cron.

Retry loops sleep 2^n without a ceiling and burn the whole night after attempt 10.

Docs PRs sit red for trivial reasons: missing DCO, quality score, or dirty base.

Dashboards show large "pending" Superteam pools and humans hear it as paid.

Two sniper ticks claim the same OPEN task, double-work, or confuse the board.

Chat history is the only record of what the earn agent did; audits fail.

`links.json` mixes URL strings and `{local, ready}` objects. Code calls `.startswith("http")` and crashes.

Two agents create the same work dir; one crashes on FileExistsError in code that used os.mkdir.

Agents spin forever on one 403 (credits) or captcha and ship nothing else.

A job is "done" in chat but there is no artifact path, screenshot, or payout id. Disputes and accounting fail.

You patch `earn_all.py` but the old process keeps running with the bug.

A research script accidentally claims paid tasks or spends credits during scouting.

Child python jobs run forever and pile up CPU.