herdr-pm

herdr-pm

put a technical pm/cto on every agent tab you've got running.

if you live in herdr you probably have a dozen agents going at once — one per repo, half of them idle, half mid-task, all of them quietly waiting for you to come back and decide the next move. herdr-pm-init is the thing that comes back for you. it walks your live agent tabs, drops a conductor on each one, and that conductor reads the session, figures out where things stand, and asks you what's next as a short scored menu — then drives the agent that's already sitting there to do it.

key idea: the conductor doesn't rewrite your code. it conducts. it reads context, makes the call, hands a tight brief to the agent that already has the full history, and verifies the result like a human would — opening a browser to screenshot a UI change, picking the right model/effort for the task, planning before mission-critical work, running a cleanup pass, and getting a cross-model second opinion (a codex review of claude's work, when you've got codex) before merging. when you pick several independent things, it fans them out as parallel executor lanes in git worktrees and merges them back, instead of queueing everything through one agent. one manager per tab, split in right beside its agent — separate context, better decisions. no single brain juggling ten projects.

the idea

• you have N agents running in herdr, one per project.
• each one needs a human to look at it and say "ok, do X next."
• that human is the bottleneck.
• so: spawn a small pm per tab that does the looking, proposes the move (scored, pick-many), and once you choose, drives the existing agent to ship it.
• you stay in the loop on decisions, not babysitting.

how it works

/herdr-pm-init  ->  discovers your live agent tabs (one herdr call)
                ->  you pick which to manage
                ->  asks exactly two setup qs (backend? how autonomous?) — mode is auto-detected
           ->  installs the operator skill, splits one conductor in beside each agent
                     |
              conductor (side-by-side in the agent's tab — you answer it right there)
                reads the managed session  (scrollback first, sessionr if it needs deep history)
                cross-checks git reality
                asks you 5 scored impact×effort options  ->  you pick "1,3"
                writes a tight mission brief, dispatches it with a freshly-minted done-marker
                independent picks run as parallel worktree lanes; merges come back as missions
                waits, reads the result, verifies *that mission's* marker + git reality
                ends each turn with a mini-changelog + what's next (state = custom status + toast)

everything the live loop does is plain herdr / sessionr / git — no daemons, no glue. the helpers check Herdr's server protocol/capabilities before using newer primitives like custom status, notifications, diagnostics, and native worktrees (verified against herdr 0.6.9 / protocol 13). durable state (assignments, missions) lives under ~/.local/state/herdr-pm/<slug>/, and touch ~/.local/state/herdr-pm/PAUSE is an out-of-band all-stop for every PM. the only scripts are a handful of tiny stdlib-python helpers (setup ×3 + the conductors' id-resolver and ops kit: dispatch/await/keys/lanes/read/status/diagnose/review/spawn-exec + the monitor-mode fleet detector pm_monitor.py).

the two skills

skill	who runs it	what it does
herdr-pm-init	you	the launcher. discover tabs → two setup qs → install the brain → split one conductor in per tab → hand off.
herdr-pm-agent	each spawned conductor	the brain. reads the session, asks you scored decisions in-pane, drives the existing agent, and runs independent work as parallel worktree lanes. mounted automatically by the launcher.

light vs power