Agent Manage Bot

use this as the first conversation after someone imports the Agent Manage Bot template — ask setup questions one at a time, then wire rooms, routines, and rules

use this when authoring or running Mac CLI launches or research collect → deep-brief → publish pipelines — write failure branches (stall, dirty exit, quota, Mac-down, miss detection, orphan output, unacked handoff) before the happy path, and apply run limits, demotion, and the retry/cost ceiling

use this when launching any long Mac CLI job (Claude Code, Codex, Kimi, Agy, Muse, DSH) — tmux launcher, stream heartbeats, stale-log watchdog, no-exit-while-waiting, Jev shadow pre-launch log, exception paths, and run limits

use this when running a Mac CLI implementer (Opus 5.5 via Claude Code by default) with two blind dual-LGTM reviewer CLIs on GitHub PRs

use this when acting as engineering manager of an agent coding team and you need the Plan–Design–Build–Test–Deploy–Maintain loop, not just a single PR

use this when filing notes to the shared Obsidian agent-context vault on the Mac, or wiring any agent to read context before a task and post a learning note after

use this when deciding whether to fan out, open a browser, retry, run a high-risk mutation, verify research citations, check completion before reporting done, or whether an expensive run is worth its tokens — call Jev via OpenRouter through the jev-shadow wrapper

Weekly check that the public playbook, the vault agent card, and this bot's profile agree, plus weekly receipts and a kill review of every routine.

Daily kick of one rotating research CLI on the Mac to gather fresh agent-harness design tips into a dated pack.

Daily curation of the morning's Lab pack into KEEP-only cards on the public Harness Lab page, never inventing content.

Nightly health audit of the Mac CLIs, tmux sessions, watchdogs, and vault so problems surface before the morning.

Twice-daily sweep of open PRs for red CI, conflicts, and stalls, routed to the coding EM bot; quiet when the board is clean.

Checks the handoff ledger for coding kicks the coding EM bot hasn't picked up, re-wakes it once, then tells the owner if still stuck.

Weekday tally of Jev shadow-gate verdicts and the would-be token savings, to inform whether any gate is ready for promotion.