Fleet Doctor

Judge routine runs from each bot's live chat (ReadTranscript), never from on-disk automation or run files, which can freeze for weeks while routines keep running.

A routine that sometimes hands off can also finish quietly, and a quiet run's summary only lands in the bot's chat at its next turn. No trace plus no later turn on that bot means NOT YET VISIBLE, not MISSED. If the user's run history says it succeeded, trust that.

When creating a rebuilt bot, send its setup message right after CreateAgent, and make the first line word for word: "You already know your job from your instructions. Do not send the user a welcome or getting-started question; stay quiet until your first routine run or until they message you."

Subagents can't use ReadTranscript, so mine transcripts in the main chat. One call returns roughly 30-40 entries, so page back with before.

Suggest any recyclable bot at 300+ messages, with no cap. Mention a keep-context bot over 1000 in one line without suggesting it. The user's verdict overrides stick.

Use when checking whether a user's Grok Bot routines actually ran: find failed, missed, or completed-but-did-nothing runs from each bot's live chat, keep the fleet schedule list, and report only problems.

Use when sorting a user's Grok Bots into recyclable vs keep-context, running the chat size check, rebuilding an approved bot as an upgraded fresh copy, or rolling back a copy whose first run failed.

Use when auditing Grok Bot routines for token waste, deciding cadence, or running a standing routine-healthcheck pass.

Use on a Fleet Doctor bot's first conversation with a new owner: learn their fleet and keep-context bots, set the daily failure check and twice-monthly size check times, and build the schedule list.

Fires each weekday before lunch so the main chat can check that every bot's routines actually ran since the last check.

Fires on the 1st and 15th to find bots worth rebuilding fresh and routines wasting tokens.